Generalized Ricci Flow

Generalized Ricci Flow
arXiv:2008.07004v2 [math.DG] 28 Aug 2020
Mario Garcia-Fernandez and Jeffrey Streets

Universidad Autónoma de Madrid, and Instituto de Ciencias Matemáticas,
Cantoblanco, 28049 Madrid, Spain
E-mail address: mario.garcia@icmat.es
University of California, Irvine, CA 92617

E-mail address: jstreets@uci.edu
2020 Mathematics Subject Classification. 53C55, 53D18, 53E20, 53E30
Contents
Chapter 1. Introduction 1
1.1. Outline 2
1.2. On pedagogy 4
1.3. Acknowledgements 5
Chapter 2. Generalized Riemannian Geometry 7

2.1. Courant algebroids 7
2.2. Symmetries of the Dorfman bracket 12
2.3. Generalized metrics 20
2.4. Divergence operators 28
Chapter 3. Generalized Connections and Curvature 33

3.1. Generalized connections 33
3.2. Metric compatible connections 37
3.3. The classical Bismut connection 42
3.4. Curvature and the First Bianchi Identity 46
3.5. Generalized Ricci curvature 49
3.6. Generalized scalar curvature 53
3.7. Generalized Einstein-Hilbert functional 58
Chapter 4. Fundamentals of Generalized Ricci Flow 67

4.1. The equation and its motivation 67
4.2. Examples 71
4.3. Maximum principles 75
4.4. Invariance group and solitons 78
4.5. Low dimensional structure 84
Chapter 5. Local Existence and Regularity 87

5.1. Variational Formulas 87
5.2. Short time existence 91
5.3. Curvature evolution equations 94
5.4. Smoothing estimates 101
5.5. Results on maximal existence time 104
5.6. Compactness results for generalized metrics 108
Chapter 6. Energy and Entropy Functionals 111

6.1. Generalized Ricci flow as a gradient flow 111
6.2. Expander entropy and Harnack estimate 118
6.3. Shrinking Entropy and local collapsing 122
6.4. Corollaries on nonsingular solutions 125
iii
iv CONTENTS
Chapter 7. Generalized Complex Geometry 129

7.1. Linear generalized complex structures 129
7.2. Generalized complex structures on manifolds 136
7.3. Courant algebroids and pluriclosed metrics 144
7.4. Generalized Kähler geometry 150
Chapter 8. Canonical Metrics in Generalized Complex Geometry 167
8.1. Connections, Torsion, and Curvature 167
8.2. Canonical metrics in complex geometry 172
8.3. Examples and rigidity results 178
Chapter 9. Generalized Ricci Flow in Complex Geometry 185
9.1. Kähler-Ricci flow 185
9.2. Pluriclosed flow 187
9.3. Generalized Kähler-Ricci flow 192
9.4. Reduced flows 199
9.5. Torsion potential evolution equations 202
9.6. Higher regularity from uniform parabolicity 204
9.7. Metric evolution equations 209
9.8. Sharp existence and convergence results 213
Chapter 10. T-duality 221
10.1. Topological T-duality 221
10.2. T-duality and Courant algebroids 224
10.3. Geometric T-duality 227
10.4. Buscher rules and the dilaton shift 231
10.5. Einstein-Hilbert action 235
10.6. Examples 237
Bibliography 243
CHAPTER 1
Introduction
The Ricci flow is a nonlinear parabolic partial differential equation for a Rie-
mannian metric that has yielded new results and insights in topology, Riemann-
ian geometry, complex geometry, partial differential equations, and mathematical
physics. The equation was introduced by Hamilton [92], who used it to classify
compact Riemannian 3-manifolds with positive Ricci curvature. In the ensuing
decades Hamilton went on to prove various results and formulate precise conjec-
tures on the Ricci flow in low dimensions. The importance of Ricci flow was then
permanently enshrined by Perelman, in his spectacular resolution of the Poincarè
Conjecture and Thurston Geometrization Conjecture [140, 141, 142].
Parallel to the story of Ricci flow in Riemannian geometry is that of Kähler-
Ricci flow in complex geometry. The fundamental result of Cao [35] gave a proof
of the classical celebrated theorems of Aubin-Yau [9, 186] and Yau [186] using
Ricci flow. Since then the Kähler-Ricci flow has yielded many further results in
complex geometry. In recent years the “Analytic minimal model program” of Song-
Tian [154] has attracted much attention, aiming at fundamental breakthroughs in
complex and algebraic geometry. A third parallel thread informs the study of
Ricci flow, coming from mathematical physics. In the thesis of Friedan [68], it
was observed that the Ricci flow equation arises in the context of renormalization
group flow for nonlinear sigma models, and this connection was strengthened by
Perelman, who noted [140] that this physical point of view loosely suggests some
of his fundamental monotone quantities.
More recently a natural extension of the Ricci flow equation has appeared in
the context of generalized geometry, a relatively new subject drawing inspiration
from Poisson geometry, complex geometry, and mathematical physics. An early
appearance of this subject was the discovery of generalized Kähler structure in
the seminal work of Gates-Hull-Roček [74] in mathematical physics. Twenty years
later these structures were rederived by Gualtieri [88] within the general frame-
work of Hitchin’s generalized geometry program. Like generalized geometry itself,
the generalized Ricci flow equation also has several different origins coming from
considerations in classical geometry, complex geometry, and mathematical physics.
The relationships between these points of view, and the realization that a priori
different extensions of the Ricci flow equation are unified by a single equation, the
generalized Ricci flow, have only been achieved recently. Thus,
The primary purpose of this book is to provide an intro-

duction to the fundamental geometric, algebraic, topo-
logical, and analytic aspects of the generalized Ricci flow
equation.
1
2 1. INTRODUCTION
Our focus is mostly on foundational results, although we will discuss some more
technical global existence and convergence results in the penultimate chapter. As
the mathematical study of the generalized Ricci flow equation is more nascent,
the results described inevitably fall short of the depth of results attained for the
Ricci flow. However it seems apparent that such depth of results are waiting to be
attained in the context of generalized geometry, beyond tautologically recapturing
what is known for Ricci flow. Thus,
The secondary purpose of this book is to formulate ques-

tions and conjectures about the generalized Ricci flow as
an invitation to the reader.
These questions and conjectures will be interspersed througout the text.
1.1. Outline
We begin in Chapter 2 by introducing fundamental concepts of generalized
Riemannian geometry. The starting point is to extend the smooth structure of a
manifold, in the form of the Lie bracket on the tangent bundle, to new structures
on the direct sum of the tangent and cotangent bundle, specifically the Dorfman
and Courant brackets. These brackets come equipped with a symmetry group that
includes the natural action of diffeomorphisms, but is enlarged by the action of B-
fields. We then introduce generalized Riemannian metrics from the point of view
of reducing the structure group of T ⊕ T ∗ to a maximal compact subgroup. We
show the equivalence between a generalized metric and a pair (g, b) consisting of a
classical Riemannian metric and skew-symmetric two-form b.
Having introduced the fundamental properties of generalized geometry, in Chap-
ter 3 we introduce and analyze special generalized metrics. We begin with the
theory of generalized connections, leading eventually to a derivation of natural
curvature quantities associated to a generalized metric that naturally involve the
curvature of the classical Bismut connection. Given this, we seek to define a class
of canonical geometries through a variational principle generalizing the derivation
of the classical Einstein equation. Specifically we generalize the Einstein-Hilbert
action using a scalar curvature quantity natural to generalized geometry. Critical
points of this functional satisfy a natural coupling of the classical Einstein equation
with the equations for a harmonic three-form, again expressed naturally in terms
of the Ricci curvature of the Bismut connection. With this in mind, we give the
classification of Bismut-flat connections. This is a special case of a classical result
of Cartan-Schouten, and we show that all examples are covered by semisimple Lie
groups with bi-invariant metrics and torsion tensor determined by the Lie bracket.
In Chapter 4 we introduce the generalized Ricci flow equation as a tool for con-
structing canonical generalized geometric structures, specifically generalized Ein-
stein structures. We first introduce the equation, as expressed in terms of both
classical and generalized objects. We explain the invariance properties of the equa-
tion, and introduce natural classes of solutions, illustrated with basic examples.
We discuss self-similar solutions of generalized Ricci flow, called solitons, and de-
rive some of their basic properties. We end by discussing special properties satisfied
by generalized Ricci flow in low dimensions, and special classes of solutions that
illustrate connections to further extensions of the Ricci flow equation.
1.1. OUTLINE 3
We prove the fundamental analytic properties of generalized Ricci flow in Chap-

ter 5. We first show that the equation is well-posed for arbitrary initial data on
compact manifolds. We then exhibit evolution equations for the Riemannian and
Bismut curvature tensors, and use these to establish estimates on derivatives of
curvature in the presence of bounds on the Riemannian curvature. This leads to
the basic fact that the Riemannian curvature tensor must blow up at any finite
time singularity of the flow. We end by recording compactness results for solutions
to generalized Ricci flow. In Chapter 6 we show that generalized Ricci flow is the
gradient flow for a natural energy functional. The construction leads furthermore
to natural entropy functionals which are monotone along the generalized Ricci flow.
As consequences of this we establish a noncollapsing result for broad classes of solu-
tions to generalized Ricci flow, as well as results on the limiting behavior of certain
nonsingular solutions of the flow.
The remainder of the text focuses on the role of generalized Ricci flow in gen-
eralized complex geometry. We start in Chapter 7, using generalized geometry
to motivate and introduce generalized complex structures, natural objects which
unify and extend classical complex structures as well as symplectic structures. Af-
ter recording their most basic properties we describe other constructions in complex
geometry and show their relationship to generalized geometry. In particular, we de-
scribe pluriclosed structures (also known as strong Kähler with torsion structures)
using generalized complex geometry. Lastly we define generalized Kähler geometry
and discuss its basic properties and local structure.
Having established the fundamental geometric properties of generalized com-
plex structures, in Chapter 8 we discuss canonical metrics in the setting of gener-
alized complex geometry, extending the discussion of Chapter 3. We recall basic
aspects of different Hermitian connections relevant in complex, non-Kähler geome-
try, and express the generalized Einstein equations in this setting in terms of their
curvatures. In the setting of generalized Kähler geometry these equations can be
reduced to certain scalar PDE using an extension of the classical transgression for-
mula for the first Chern class. We end the chapter with a discussion of examples
of and rigidity results for canonical metrics in this setting.
In Chapter 9 we establish fundamental properties of the relationship between
generalized Ricci flow and generalized complex geometry. We begin with the classi-
cal point of view, introducing the Kähler-Ricci flow, then introduce the pluriclosed
flow, an extension of the Kähler-Ricci flow to complex, non-Kähler geometry. We
show that the pluriclosed flow is gauge-equivalent to generalized Ricci flow, and fur-
thermore preserves generalized Kähler structure, yielding the generalized Kähler-
Ricci flow system. We describe all of these flows directly in the language of gen-
eralized geometry, and describe natural deformation classes of generalized Kähler
structure which are preserved along the flow. We prove the short-time existence of
solutions to all these equations and formulate conjectures on the sharp existence
time.
In the second half of this chapter we prove global existence and convergence
results for the generalized Ricci flow in complex geometry. To begin we exhibit
natural reductions of the generalized Ricci flow in the setting of complex geome-
try, generalizing the reduction of Kähler-Ricci flow to a scalar equation modeled
on the parabolic complex Monge-Ampère equation. Next we establish higher reg-
ularity of the flow in the presence of uniform parabolicity estimates, generalizing
4 1. INTRODUCTION
the classic estimates of Evans-Krylov and Calabi/Yau for the complex parabolic
Monge-Ampère equation to the system of equations determined by the pluriclosed
flow. We then give geometric settings where these uniform parabolicity estimates
can be verified, leading to sharp global existence and convergence results for the
flow. First we prove the theorem of Tian-Zhang on the sharp existence time for
Kähler-Ricci flow. In the case of pluriclosed flow we prove global existence on com-
plex manifolds admitting Hermitian metrics of nonpositive bisectional curvature,
and prove exponential convergence to a flat, Kähler, metric on complex tori. This
result includes a complete description of the generalized Kähler-Ricci flow, and
yields a classification result for generalized Kähler structures on tori.
We finish in Chapter 10 by describing the relationship between the generalized
Ricci flow and the physical phenomenon of T-duality. While T-duality first arose
in physics, we recall the relatively recent, purely mathematical, formulation of this
concept, focusing on the most fundamental topological and geometric aspects. We
then show that T-duality naturally transforms solutions to the generalized Ricci
flow. As corollaries we immediately obtain some global existence and convergence
results for the generalized Ricci flow by applying T-duality to Ricci flow solutions
whose behavior is known.
1.2. On pedagogy
This book assumes a working familiarity with fundamental concepts of Rie-
mannian geometry, i.e. smooth manifolds, tangent/cotangent/tensor bundles, dif-
ferential forms, Lie groups, Riemannian metrics, connections, and curvature. We
direct the reader to e.g. [121, 122, 143, 184] for background in these areas. We
also assume some mild familiarity with the language and basic results of ordinary
and partial differential equations, and direct the reader to e.g. [61] for general back-
ground in the area. We also assume familiarity with complex geometry, although
we briefly review some of this material at a more basic level within. We refer to
[47, 87, 136] for further background.
Notably, we will not assume familiarity with the Ricci flow equation itself, and
instead will introduce the fundamental concepts of generalized Ricci flow whole
cloth, without an independent discussion of the classical Ricci flow concepts in-
evitably carried within. This choice is made in part to save space, but also since by
now there are many excellent introductory texts on the Ricci flow e.g. [5, 26, 50,
52, 135]. Having said all of this, we will at times refer back to and discuss results
specific to the Ricci flow when it aids in exposition. Moreover, prior familiarity with
the Ricci flow equation would of course help in reading this text, as would reading a
Ricci flow text in parallel. Ultimately we aim to make the text self-contained given
the stated prerequisites, which will of course entail some overlap with existing texts.
Generalized geometry employs many structures (e.g. Dirac structures, Courant
algebroids) which have a rich history and outlook of their own, independent of the
role they play in generalized geometry. Moreover, generalized geometry as a subject
by itself appears related to a variety of questions in algebraic geometry, Poisson
geometry, mathematical physics, etc. We will make no attempt at an exhaustive
exploration of these various avenues, rather focusing on the aspects which are most
relevant to understanding the analysis of generalized Ricci flow. One imagines that
there are likely fruitful further generalizations of Ricci flow incorporating yet more
1.3. ACKNOWLEDGEMENTS 5
structure from generalized geometry. We point out some possibilities for this as
well as some preliminary results in the literature.
1.3. Acknowledgements
The first named author is grateful to his collaborators Luis Álvarez Cónsul,
Raúl González Molina, Roberto Rubio, Carl Tipler, and Carlos Shahbazi for useful
conversations and insight on different aspects of this text. He would like to specially
thank Nigel Hitchin and Marco Gualtieri for introducing him into the subject of
generalized geometry. He would like also to thank Gil Cavalcanti, Pedram Hek-
mati, Pavol Ševera, Rafael Torres, Fridrich Valach, and Dan Waldram for helpful
and inspiring conversations. He acknowledges support from the Spanish MICINN
via the ICMAT Severo Ochoa project No. SEV-2015-0554, and under grants No.
PID2019-109339GA-C32 and No. MTM2016-81048-P.
The second named author is grateful to Vestislav Apostolov, Matthew Gibson,
Joshua Jordan, Man-Chun Lee, Yury Ustinovskiy, and Micah Warren, for collab-
orations underpinning many results in this text. Furthermore, his understanding
of this subject has benefitted greatly from conversations with Jess Boling, Ryushi
Goto, Paul Gauduchon, Marco Gualtieri, Nigel Hitchin, Chris Hull, and Martin
Roc̆ek. He would like to thank David Streets for a careful proofreading. Finally, he
would like to offer greatest thanks to Mark Stern and Gang Tian for their support,
encouragement, guidance, and collaboration regarding this subject. He gratefully
acknowledges support from the National Science Foundation via DMS-1454854,
DMS-1301864, DMS-1006505, and from the Alfred P. Sloan Foundation.
CHAPTER 2
Generalized Riemannian Geometry
On the road to understanding Riemannian geometry from the modern point of

view, one begins with the notion of a smooth manifold, and derives various objects
canonically associated to the smooth structure, such as the tangent and cotangent
bundles, and the Lie bracket. Geometric structure enters by assigning a Riemannian
metric, which measures the lengths of vectors in the tangent bundle. This way, one
can associate an energy to a smooth curve, thought of as the trajectory of a point
particle moving along the manifold, which leads to geodesics and Jacobi fields.
Generalized geometry changes this story in a fundamental way. Whereas vec-
tors and the tangent bundle have primacy in Riemannian geometry, the starting
point of generalized geometry is to put vectors and covectors on equal footing, and
to treat sections of T ⊕ T ∗ as the fundamental object of geometry. This leads to
a new bracket structure on such sections, the Dorfman bracket, which involves a
delicate combination of the classical operators of differential geometry. This point
of view also inspires the definition of a generalized Riemannian metric, which mea-
sures sections of T ⊕ T ∗ . Such generalized metrics are equivalent to a choice of a
classical Riemannian metric g together with a skew-symmetric two-form b. Similar
to Riemannian metrics, a generalized metric can be used to define an energy for a
surface mapping into the manifold, thought of as the trajectory of a ‘string’ moving
along the target space.
2.1. Courant algebroids

The fundamental object underlying generalized Riemannian geometry is a Courant
algebroid. These first arose in the work of Courant [55] and Dorfman [58], partly
in an effort to describe the geometry of equivariant moment map level sets, and
was later axiomatized by Liu, Weinstein, and Xu [130]. Before giving the general
definition of Courant algebroid we begin by describing the main example, with fur-
ther examples to follow. The linear algebra of T ⊕ T ∗ plays a central role in the
subject, and we begin with some basic definitions.
Definition 2.1. Given V a vector space of dimension n, define an inner prod-
uct on V ⊕ V ∗ via
1
(2.1) hX + ξ, Y + ηi := 2 (ξ(Y ) + η(X)) .
The inner product h, i is obviously symmetric and has signature (n, n), and thus
determines a copy of O(n, n) consisting of endomorphisms preserving this product.
Note that for an n-dimensional vector space V there is a natural decomposition
Λ2n (V ⊕ V ∗ ) = Λn V ⊗ Λn V ∗ .
7
8 2. GENERALIZED RIEMANNIAN GEOMETRY
As there is the natural determinant pairing between Λn V and Λn V ∗ , this induces a

canonical identification of Λ2n (V ⊕V ∗ ) with R, and hence the positive real numbers
correspond to a canonical orientation. The Lie group preserving both the symmetric
product and this orientation will be isomorphic to SO(n, n).
Definition 2.2. Given M a smooth manifold, we denote T ⊕T ∗ := T M ⊕T ∗M .
The Dorfman bracket on sections of T ⊕ T ∗ is defined by
(2.2) [X + ξ, Y + η] = [X, Y ] + LX η − iY dξ.
Note that this bracket restricts to the Lie bracket when acting on tangent vectors.
However, some of the fundamental properties of the Lie bracket do not extend to
the Dorfman bracket. We first observe that the familiar Leibniz rule still holds for
the first-order differential operator [X + ξ, ·] acting on sections of T ⊕ T ∗ . Before
stating this we define the natural projection
π : T ⊕ T ∗ → T, X + ξ 7→ X,
which will be called the anchor map.
Lemma 2.3. Given a, b ∈ Γ(T ⊕ T ∗ ) and f ∈ C ∞ (M ), the Dorfman bracket
satisfies
[a, f b] = f [a, b] + π(a)f b.
Proof. Setting a = X + ξ, b = Y + η, we directly compute using the properties

of Lie derivatives,
[X + ξ, f (Y + η)] = [X, f Y ] + LX f η − if Y dξ
= f [X + ξ, Y + η] + (Xf )Y + (Xf )η
= f [X + ξ, Y + η] + (Xf )(Y + η),
as required.
The differential operator [X + ξ, ·] is furthermore compatible with the inner

product h, i, in the following sense:
Lemma 2.4. Given a, b, c ∈ Γ(T ⊕ T ∗ ), the Dorfman bracket satisfies
π(a) hb, ci = h[a, b], ci + hb, [a, c]i .
Proof. Setting a = X + ξ, b = Y + η, b = Z + ζ and using the identity

i[X,Y ] = [LX , iY ] we compute
h[a, b], ci + hb, [a, c]i = 12 (i[X,Y ] ζ + iZ (LX η − iY dξ) + i[X,Z] η + iY (LX ζ − iZ dξ))
= 21 (LX iY ζ + LX iZ η)
= 12 iX d(ζ(Y ) + η(Z)),
as required.
The main structural property of the Dorfman bracket is the Jacobi identity:
Lemma 2.5. Given a, b, c ∈ Γ(T ⊕ T ∗ ) one has
[a, [b, c]] = [[a, b], c] + [b, [a, c]].
2.1. COURANT ALGEBROIDS 9
Proof. Setting a = X + ξ, b = Y + η, c = Z + ζ, one has, using the identities

i[X,Y ] = [LX , iY ] and L[X,Y ] = LX LY − LY LX ,
[[a, b], c] + [b, [a, c]]
= [[X, Y ] + LX η − iY dξ, Z + ζ] + [Y + η, [X, Z] + LX ζ − iZ dξ]
= [[X, Y ], Z] + L[X,Y ] ζ − iZ d(LX η − iY dξ)
+ [Y, [X, Z]] + LY (LX ζ − iZ dξ) − i[X,Z] dη
= [X, [Y, Z]] + LX LY ζ − LX iZ dη − LY iZ dξ + iZ diY dξ
= [X, [Y, Z]] + LX (LY ζ − iZ dη) − i[Y,Z]dξ
= [a, [b, c]],
as claimed.
Despite these favorable properties, unlike the Lie bracket of vector fields, the
Dorfman bracket is not skew-symmetric. In fact, one can easily show that
(2.3) [a, b] + [b, a] = 2d ha, bi .
The skew-symmetrization of the Dorfman bracket, called the Courant bracket, will
be useful for some calculations.
Definition 2.6. The Courant bracket on sections of T ⊕ T ∗ is defined by
[a, b]c = 12 ([a, b] − [b, a]).
Notice that
(2.4) [a, b]c = [a, b] − d ha, bi .
Furthermore, a straightforward calculation leads to the following explicit formula:
[X + ξ, Y + η]c = [X, Y ] + LX η − LY ξ + 12 d (ξ(Y ) − η(X)) .
We summarize the main properties of the Courant bracket in the following lemma.
Lemma 2.7. The following hold:
(1) Given a, b ∈ Γ(T ⊕ T ∗ ) and f ∈ C ∞ (M ), the Courant bracket satisfies
[a, f b]c = f [a, b]c + π(a)f b − ha, bi df.
(2) Given a, b, c ∈ Γ(T ⊕ T ∗ ), one has
[a, [b, c]c ]c + [c, [a, b]c ]c + [b, [c, a]c ]c
= 13 d (h[a, b]c , ci + h[b, c]c , ai + h[c, a]c , bi)
Proof. We set a = X + ξ, b = Y + η, c = Z + ζ. To prove (1), we directly

compute as in Lemma 2.3
1
[X + ξ, f (Y + η)]c = [X, f Y ] + LX f η − Lf Y ξ + d (f ξ(Y ) − f η(X))
2
1
= f [X + ξ, Y + η]c + (Xf )Y + (Xf )η − ξ(Y )df + 2 (ξ(Y ) − η(X)) df
1
= f [X + ξ, Y + η]c + (Xf )(Y + η) − 2 (ξ(Y ) + η(X)) df
= f [X + ξ, Y + η]c + (Xf )(Y + η) − hX + ξ, Y + ηi df,
as required.
As for (2), using Lemma 2.3 and the fact that [ξ, c] = 0 for any closed one-form
ξ, we obtain
[[a, b]c , c]c = [[a, b]c , c] − d h[a, b]c , ci
= [([a, b] − d ha, bi), c] − d h[a, b]c , ci
= [[a, b], c] − d h[a, b]c , ci .
Using these properties we compute
[[a, b]c , c]c + [[c, a]c , b]c + [[b, c]c , a]c
1
= 4 ([[a, b], c] − [c, [a, b]] − [[b, a], c] + [c, [b, a]] + cyclic)
1
= 4 ([a, [b, c]] − [b, [a, c]] − [c, [a, b]]
−[b, [a, c]] + [a, [b, c]] + [c, [b, a]] + cyclic)
1
= 4 ([a, [b, c]] − [b, [a, c]] + cyclic)
1
= 4 ([[a, b], c] + cyclic)
1
= 4 ([[a, b]]c c]c + d h[a, b]c , ci + cyclic)
1
= 4 ([[a, b]c , c]c + [[c, a]c , b]c + [[b, c]c , a]c
+d (h[a, b]c , ci + h[b, c]c , ai + h[c, a]c , bi)) ,
from which the claim follows.
∗
The structure (T ⊕ T , h, i , [, ], π) was formalized into a general definition, that
of a Courant algebroid, first given in [130].
Definition 2.8. A Courant algebroid 1 is a vector bundle E → M together
with a nondegenerate symmetric bilinear form h, i a bracket [, ] on Γ(E), and a
bundle map π : E → T M such that, given a, b, c ∈ Γ(E) and f ∈ C ∞ (M ), one has
(1) [a, [b, c]] = [[a, b], c] + [b, [a, c]]
(2) π[a, b] = [πa, πb]
(3) [a, f b] = f [a, b] + π(a)f b
(4) π(a) hb, ci = h[a, b], ci + hb, [a, c]i
(5) [a, b] + [b, a] = D ha, bi,
where D : C ∞ (M ) → Γ(E) denotes the composition of three maps: the exterior dif-
ferential d acting on functions, the natural map π ∗ : T ∗ → E ∗ , and the isomorphism
E ∗ → E provided by the symmetric product.
In the sequel, we will abuse notation and denote by π ∗ the composition of
π ∗ : T ∗ → E ∗ with the natural isomorphism E ∗ → E provided by the symmetric
product. The fact that our explicit structure (T ⊕ T ∗ , h, i , [, ], π) fulfills the con-
ditions of a Courant algebroid follows from Lemma 2.3, Lemma 2.4, Lemma 2.5,
equation (2.3), and the definition of π. Note that, in this explicit situation, we have
that
Df = 2df.
We leave as an exercise to check from the previous axioms that the image of π ∗
is isotropic. To finish this section, we unravel Definition 2.8 for the particular class
of Courant algebroids which is of primary interest for the present book.
1We retain the term Courant algebroid despite the fact that all brackets appearing in this
definition are Dorfman brackets, not Courant brackets.
2.1. COURANT ALGEBROIDS 11
Definition 2.9. A Courant algebroid is exact if it fits into an exact sequence

of vector bundles
π∗ π
0 −→ T ∗ −→ E −→ T −→ 0.
For exact Courant algebroids notice that, by definition, the image of π ∗ coin-
cides with the kernel of π. Further consider an isotropic splitting σ : T → E of π.
Such a splitting can be constructed by choosing an arbitrary splitting σ0 : T → E,
and setting
(2.5) σ = σ0 − 12 π ∗ τ,
where τ ∈ Sym2 T ∗ is defined by τ (X, Y ) = hσ0 X, σ0 Y i. Using these structures we
can give a classification result for exact Courant algebroids, incorporating the first
appearance of a closed three-form H, which features prominently in generalized
geometry.
Proposition 2.10. Given an exact Courant algebroid E with isotropic splitting
σ, the map F : T ⊕ T ∗ → E defined by
F (X + ξ) = σX + 12 π ∗ ξ
is an isomorphism of orthogonal bundles, for the symmetric product (2.1). Via this
isomorphism the anchor map is given by π(X + ξ) = X and the Dorfman bracket
is
(2.6) [X + ξ, Y + η]H := [X + ξ, Y + η] + iY iX H,
where H ∈ Λ3 T ∗ is a closed three-form defined by
(2.7) H(X, Y, Z) = 2 h[σX, σY ], σZi .
Proof. We first note that
σX + 21 π ∗ ξ, σY + 12 π ∗ η = 12 (ξ(πσY ) + η(πσX)) = 21 (ξ(Y ) + η(X)),
noting that hσX, σY i = 0, which proves that F is an isometry. Next we transport
the Courant algebroid structure from E to T ⊕ T ∗ using F . It follows from the
definition of F that the induced anchor map on T ⊕T ∗ is πE (X +ξ) := πF (X +η) =
X and that DE = π ∗ d = 2d. The induced bracket, denoted [, ]E , is defined via
F [X + ξ, Y + η]E := [F (X + ξ), F (Y + η)]
(2.8)
= [σX, σY ] + 21 [σX, π ∗ η] + 12 [π ∗ ξ, σY ] + 41 [π ∗ ξ, π ∗ η].
It is an exercise using the axioms of a Courant algebroid to show that [π ∗ ξ, π ∗ η] = 0
for any ξ, η. For the second term on the right hand side we first observe that
π[σX, π ∗ η] = [πσX, ππ ∗ η] = [πσX, 0] = 0.
Thus [X, η]E ∈ T ∗ , and we can compute, using axiom (5),
h[σX, π ∗ η], σZi = hσX, D hπ ∗ η, σZii − hπ ∗ η, [σX, σZ]i
= X hπ ∗ η, Zi − η(π[σX, σZ])
= Xη(Z) − η([X, Z])
= dη(X, Z) + Zη(X).
We conclude that
[X, η]E = LX η.
The third term of (2.8) is similar, where using axiom (5) we conclude
[ξ, Y ]E = −[Y, ξ]E + 2d hξ, Y i = −LY ξ + dξ(Y ) = −iY dξ.
Finally, we decompose the first term on the right hand side of (2.8) into tangent and
cotangent pieces with respect to F . From axiom (2) it follows that π[σX, σY ] =
[X, Y ]. We define a tensor H ∈ (T ∗ )⊗3 via
H(X, Y, Z) = 2 h[σX, σY ], σZi ,
and then it follows that
[σX, σY ] = σ[X, Y ] + 21 π ∗ H(X, Y )
It is an exercise to show that H is indeed a tensor, and moreover is totally skew-
symmetric. Putting the above observations back into (2.8) shows that
[X + ξ, Y + η]E = [X, Y ] + LX η − iY dξ + iY iX H,
as claimed. The Jacobi identity of axiom (1) for [X + ξ, Y + η]E ensures that
dH = 0.
2.2. Symmetries of the Dorfman bracket

A classic fact in smooth manifold theory is that the only automorphisms of the
tangent bundle preserving the Lie bracket structure are given by the tangent maps
associated to diffeomorphisms of the underlying manifold. In this section we prove
an analogous result for T ⊕ T ∗ equipped with the symmetric product (2.1) and the
Dorfman bracket, and use this to obtain Ševera’s classification of exact Courant
algebroids [150]. In particular the symmetry group of the Dorfman bracket is
enlarged to include so-called B-field transformations. These extra symmetries play
a central role throughout generalized geometry. With the group of generalized
diffeomorphisms at hand, we provide a different interpretation of this group by
means of a two-dimensional variational problem. This leads to a first-principles
derivation of the Dorfman bracket and a conceptual explanation for some of its
most basic features collected in Definition 2.8.
Recall that a bundle automorphism of T ⊕ T ∗ is given by a pair (f, F ), where
f ∈ Diff(M ) is diffeomorphism of M and F : T ⊕ T ∗ → T ⊕ T ∗ is a vector bundle
automorphism, which fit into a commutative diagram
F
T ⊕ T∗ / T ⊕ T∗
f
M /M
Here, the vertical arrows denote the bundle projections.
Definition 2.11. Given M a smooth manifold, a bundle automorphism (f, F )
of T ⊕ T ∗ is said to be a Courant automorphism if F is orthogonal and the natural
action of (f, F ) on sections of T ⊕ T ∗ preserves the Dorfman bracket.
Given a diffeomorphism f of M , we can define an orthogonal bundle automor-
phism via (f, f ), where

f∗ 0
(2.9) f := .
0 (f ∗ )−1
2.2. SYMMETRIES OF THE DORFMAN BRACKET 13
It follows from naturality of the various operations in formula (2.2) that f preserves
the Dorfman bracket. We next discuss a natural class of Courant automorphisms,
given by (closed) B-field transformations, which are not associated to diffeomor-
phisms of the manifold.
Definition 2.12. Given M a smooth manifold and B ∈ Λ2 T ∗ , define the
associated B-field transformation via

1 0 X
eB (X + ξ) = = X + ξ + iX B.
B 1 ξ
Lemma 2.13. Given M a smooth manifold and B ∈ Λ2 T ∗ , the map eB is

orthogonal with respect to h, i.
Proof. Given X + ξ, Y + η ∈ T ⊕ T ∗ we directly compute

eB (X + ξ), eB (Y + η) = hX + ξ + iX B, Y + η + iY Bi
1
= 2 (η(X) + iX iY B + ξ(Y ) + iY iX B)
1
= 2 (η(X) + ξ(Y ))
= hX + ξ, Y + ηi .

Proposition 2.14. Given M a smooth manifold and B ∈ Λ2 T ∗ , we have

(2.10) [eB (X + ξ), eB (Y + η)] = eB [X + ξ, Y + η] + iY iX dB,
for any X + ξ, Y + η ∈ Γ(T ⊕ T ∗ ). Consequently, the map eB is an automorphism

of the Dorfman bracket if and only if dB = 0.
Proof. Fix X + ξ, Y + η ∈ Γ(T ⊕ T ∗ ) and B ∈ Γ(Λ2 T ∗ ), and then compute

using the Cartan formula
[eB (X + ξ), eB (Y + η)] = [X + ξ + iX B, Y + η + iY B]
= [X + ξ, Y + η] + [X, iY B] + [iX B, Y ]
= [X + ξ, Y + η] + LX iY B − iY d(iX B)
= [X + ξ, Y + η] + LX iY B − iY LX B + iY iX dB
= [X + ξ, Y + η] + i[X,Y ] B + iY iX dB
= eB [X + ξ, Y + η] + iY iX dB.
The result follows.
Having exhibited that closed B-field transformations are Courant automor-

phisms of the form (Id, eB ) in Lemma 2.13 and Proposition 2.14, we now show that
these, together with the differentials of diffeomorphisms (f, f ), give all possible
Courant automorphisms.
Proposition 2.15. Let M be a smooth manifold and suppose (f, F ) is a Courant
automorphism. Then F = f ◦ eB , that is, F can be expressed as a composition of
a diffeomorphism and a closed B-field transformation.
Proof. Since the given f is a diffeomorphism, we can define a bundle auto-

−1
morphism via (f, f ), as in (2.9). Setting Φ = f ◦ F , we obtain that (Id, Φ) is also
a bundle automorphism. Fixing sections a, b ∈ Γ(T ⊕ T ∗ ) and a smooth function
p, we obtain, using Lemma 2.3, that
Φ([pa, b]) = Φ (p[a, b] − ((πb)p) a)
= pΦ([a, b]) − ((πb)p) Φ(a).
On the other hand, again using Lemma 2.3, one has
[Φ(pa), Φ(b)] = p[Φ(a), Φ(b)] − (πΦ(b)p)Φ(a).
As Φ preserves the Courant bracket, the two quantities above are equal, and hence
π ◦ Φ = π since a, b and p are arbitrary. On the other hand, the equality
2 b, Φ−1 dp = 2 hΦ(b), dpi = πΦ(b)p = πbp = 2 hb, dpi
implies that Φ−1 (and hence Φ) acts as the identity on T ∗ . Together these facts
imply that

Id 0
Φ= ,
B Id
where B ∈ (T ∗ )⊗2 . Since Φ is also orthogonal, it follows that B ∈ Λ2 T ∗ . Lastly,
since Φ preserves the Dorfman bracket, Proposition 2.14 implies that B is closed.
Thus F = f ◦ eB , as claimed.
Notice that the proof of Proposition 2.15 implies that a Courant automorphism
(f, F ) is automatically compatible with the anchor map, in the sense that
π ◦ F = f ◦ π.
Proposition 2.10 and Proposition 2.14 make clear an elementary way to construct
new Courant algebroids, by ‘twisting’ the standard Dorfman bracket by a three-form
H. This is key to Ševera’s classification of exact Courant algebroids in Theorem
2.21 below.
Definition 2.16. Given M a smooth manifold and H ∈ Λ3 T ∗ , we define the
H-twisted Dorfman bracket on sections of T ⊕ T ∗ via
[X + ξ, Y + η]H := [X + ξ, Y + η] + iY iX H.
Proposition 2.17. Given M a smooth manifold and H ∈ Λ3 T ∗ such that

dH = 0, the triple (T ⊕ T ∗ , h, i , [, ]H , π) defines a Courant algebroid.
Proof. Most of the axioms in the definition of Courant algebroid are routine
to check. The main difficulty is to verify axiom (1). Since H is closed and the
computation is local, we can assume H = dB. We note that we can rephrase
equation (2.10) as
e−B [eB (X + ξ), eB (Y + η)] = [X + ξ, Y + η] + e−B iY iX dB
= [X + ξ, Y + η] + iY iX dB
= [X + ξ, Y + η]H .
Using this, Lemma 2.5, and Lemma 2.13 we compute for any a, b, c ∈ Γ(T ⊕ T ∗ )
[[a, b]H , c]H + [[c, a]H , b]H + [[b, c]H , a]H
= e−B [eB [a, b]H , eB c] + e−B [eB [c, a]H , eB b] + e−B [eB [b, c]H , eB a]
= e−B [[eB a, eB b], eB c] + e−B [[eB c, eB a], eB b] + e−B [[eB b, eB c], eB a] = 0.
The proposition follows.
Our next goal is to provide a classification of exact Courant algebroids due to
Ševera [150], which follows from Proposition 2.10 and Proposition 2.17. We first
give the following abstract definition.
Definition 2.18. Given M a smooth manifold and E and E ′ Courant alge-
broids over M , an isomorphism between E and E ′ is a pair (f, F ), where f is a
diffeomorphism of M and F : E → E ′ is an orthogonal Courant automorphism,
which fit into a commutative diagram
F / E′
E
f
M /M
We say that (f, F ) is small if f lies in the identity component of Diff(M ). We say
that E and E ′ are in the same small isomorphism class if E and E ′ are isomorphic
via a small isomorphism.
Theorem 2.19. Given M a smooth manifold, the small isomorphism classes
of exact Courant algebroids on M are in one-to-one correspondence with the coho-
mology group H 3 (M, R).
Proof. Let E be an exact Courant algebroid over M . By Proposition 2.10
we can choose an isotropic splitting σ of E, such that E is isomorphic via a small
isomorphism to (T ⊕ T ∗ , h, i , [, ]Hσ , π), for Hσ the closed three-form on M given by
(2.7). To the small isomorphism class of E we want to associate the Ševera class
[Hσ ] ∈ H 3 (M, R).
Let us check that this is well-defined. First, if σ ′ is a different choice of isotropic
splitting, via the orthogonal automorphism E ∼ = T ⊕ T ∗ induced by σ we can
identify
σ ′ (X) = eB (X)
for some B ∈ Λ2 T ∗ . Therefore, from (2.10)
Hσ′ = Hσ + dB,
and the class [Hσ ] is independent of the choice of isotropic splitting.
Let E ′ be another exact Courant algebroid over M , that we can identify with
(T ⊕ T ∗ , h, i , [, ]H ′ , π) for a closed three-form H ′ on M . If E ′ is isomorphic to E
there exists (f, F ) : E → E ′ which exchanges the brackets on E and E ′ . Without
loss of generality, we identify E with (T ⊕ T ∗ , h, i , [, ]H , π), where we set H = Hσ .
−1
We denote Φ = f ◦ F , and arguing as in the proof of Proposition 2.15 we have
that π ◦ Φ = π and that Φ acts as the identity on T ∗ . Thus, Φ = (Id, eB ) for some
B ∈ Λ2 T ∗ . Since (f, F ) exchanges the brackets we obtain
−1
eB [a, b]H = f [f eB a, f eB b]H ′ .
Combined with
−1
f [f a, f b]H ′ = [a, b]f ∗ H ′
for any a, b ∈ Γ(T ⊕ T ∗ ), we are led to
(2.11) f ∗ H ′ = H − dB.
If E ′ is in the same small isomorphism class as E, then f is in the identity component
of Diff(M ) and by definition there exists a one-parameter group of diffeomorphisms
ft generated by X ∈ Γ(T ) such that f = f1 . Thus, using that H is closed
Z 1 Z 1
d ∗
f ∗H − H = ft Hdt = ft∗ (diX H)dt = db,
0 dt 0
R1 d ∗ 2 ∗
where b = 0 dt ft (iX H)dt ∈ Λ T , and we conclude that
H ′ = H − dB ′ ,
for B ′ = f∗ (B + b), which implies [H ′ ] = [H].
By Proposition 2.17 the map that associates to any small isomorphism class of
exact Courant algebroids its Ševera class is surjective. Finally, if E and E ′ yield the
same cohomology class, taking isotropic splittings as before we obtain H = H ′ + dB
and therefore by Proposition 2.14 we have that E and E ′ are isomorphic via the B-
field transformation eB , and consequently their small isomorphism classes coincide.

As an immediate consequence of the proof of the previous theorem we obtain the

following complete classification of exact Courant algebroids on a smooth manifold.
Corollary 2.20. Given M a smooth manifold, denote by Diff 0 (M ) the compo-
nent of the identity in Diff(M ) and set ΓM = Diff(M )/ Diff 0 (M ) the corresponding
mapping class group. Then, the isomorphism classes of exact Courant algebroids
on M are in one-to-one correspondence with the quotient H 3 (M, R)/ΓM .
Proof. By the proof of Theorem 2.19, the isomorphism classes of exact Courant
algebroids on M are in one-to-one correspondence with H 3 (M, R)/ Diff(M ). Since
the action of Diff 0 (M ) on H 3 (M, R) is trivial, the proof follows.
Another direct consequence of the proof of Theorem 2.19 is an explicit char-

acterization of the group of automorphisms (f, F ) : E → E of an exact Courant
algebroid E over a smooth manifold M . By a choice of an isotropic splitting of
E, we can identify E with the twisted Courant algebroid (T ⊕ T ∗ , h, i , [, ]H , π) in
Proposition 2.17. To state the result, we abuse notation and denote a pair (f, F )
simply by F .
Proposition 2.21. Given M a smooth manifold and H ∈ Λ3 T ∗ a closed three-
form, the group of automorphisms F : E → E of the twisted Courant algebroid
(T ⊕ T ∗ , h, i , [, ]H , π) is given by
Aut(E) = {f ◦ eB : f ∈ Diff(M ), B ∈ Λ2 T ∗ such that f ∗ H = H − dB},
′
together with the product given, for F = f ◦ eB and F ′ = f ′ ◦ eB , by
′ ′∗
F ◦ F ′ = f ◦ f ′ ◦ eB +f B
.
Proof. The characterization of the group follows from (2.11) and the proof of
Theorem 2.19. Following the notation in the statement, the composition law can
be obtained from
H = f∗ (f∗′ (H − dB ′ ) − dB).

Using the previous result, we can also give an explicit description of the Lie
algebra of infinitesimal automorphisms of an exact Courant algebroid E.
Corollary 2.22. Given M a smooth manifold and H ∈ Λ3 T ∗ a closed three-
form, the Lie algebra of the group of automorphisms of the twisted Courant algebroid
(T ⊕ T ∗ , h, i , [, ]H , π) is given by
Lie Aut(E) = {X + B | X ∈ T, B ∈ Λ2 T ∗ , LX H = −dB},
together with the Lie bracket given by
(2.12) [X + B, X ′ + B ′ ] = [X, X ′ ] + LX B ′ − LX ′ B.
Proof. Let Ft = f t ◦ eBt be a one-parameter family in Aut(E) with F0 = IdE .
Taking derivatives in ft∗ H = H − dBt at t = 0, it follows that
! !
∂ ∂
X := ft , B= Bt
∂t ∂t
|t=0 |t=0
2 ∗
satisfies LX H = −dB. Conversely, given X+B ∈ T ⊕Λ T satisfying LX H = −dB,
we define
Ft = f t ◦ eB t ∈ Aut(E)
where ft is the one-parameter family of diffeomorphisms generated by X and
Z t
Bt = fs∗ Bs ds.
0
Notice that
Z t Z t
dB t = fs∗ dBs ds =− fs∗ LXs Hds = H − ft∗ H
0 0
and hence Ft is well-defined.
To calculate the Lie bracket, by Proposition 2.21 we have
−1
ft f )∗ B
F −1 ◦ Ft ◦ F = f −1 ◦ ft ◦ f ◦ eB+f∗ Bt −(f ,
and hence the adjoint action is
F −1 (X ′ + B ′ ) = f ∗ X ′ + f ∗ B ′ − Lf ∗ X ′ B.
Setting F = Ft and taking derivatives in this expression we obtain (2.12). Observe
that we use the convention of right invariant vector fields for the Lie bracket.
We next provide a different view on the group Aut(E) in Proposition 2.21. As
we will see, this is an infinite-dimensional Lie group which arises naturally from
a 2-dimensional variational problem [150]. Here, we will avoid entering into any
details about the theory of infinite-dimensional manifolds and Lie groups, but rather
focus on formal aspects of the construction. Let M be a smooth manifold. Let Σ
be a smooth compact surface (without boundary, say). To be consistent with our
notation, we will denote the tangent and cotangent bundles of M by T and T ∗ ,
while those of Σ will be denoted T Σ and T ∗ Σ. Consider the infinite-dimensional
space of smooth maps from Σ to M , denoted C ∞ (Σ, M ). Given a closed three-form

H ∈ Λ3 T ∗ on M representing an integral cohomology class [H] ∈ H 3 (M, Z), we
can define a functional
Z
∞
(2.13) SH : C (Σ, M ) → R/Z, ϕ 7→ ϕ∗ H,
Y
where Y is a three-manifold with boundary Σ and ϕ : Y → M is any smooth
extension of ϕ to Y . Since H is assumed to have integral periods over the integer
homology H3 (M, Z) of M , this functional is well-defined. Notice also that when H
is exact, by Stokes’ Theorem the functional reduces to
Z
SdB (ϕ) = ϕ∗ B.
Σ
This functional is known in the mathematical physics literature as the Wess-Zumino
term in the action of a two-dimensional Σ-model (see e.g. [80]), and lies at the core
of the relation between generalized geometry and string theory.
Using that H is closed and applying Stokes’ Theorem again, the variation of
SH is given by Z Z
δSH (X) = ϕ∗ LX H = ϕ∗ iX H,
Y Σ
where X ∈ Γ(ϕ∗ T ) is identified with a vector in the tangent space of C ∞ (Σ, M ) at
ϕ, and X ∈ Γ(ϕ∗ T ) denotes any extension of X along ϕ. Observe that the formula
for δSH makes sense even if H does not have integral periods. Solutions of the
variational problem, that is, critical points of SH , are given by maps ϕ such that
for all X ∈ Γ(ϕ∗ T ) there exists a one-form ξ ∈ Γ(T ∗ Σ) such that
ϕ∗ (iX H) = dξ.
In particular, if X + ξ ∈ Γ(T ⊕ T ∗ ) is a global section satisfying
(2.14) iX H = dξ,
by pulling-back to Σ via ϕ we obtain tangent directions on the space of smooth
maps along which SH is constant.
A natural question is whether these ‘flat directions’ of SH arise from some
symmetries of the functional. As is customary in variational problems, we will try
to provide an answer by looking at the space of Lagrangian densities, given in this
case by the space of closed three-forms on M
Ω3cl (M ) := {H ∈ Λ3 T ∗ : dH = 0}.
Note the group of diffeomorphisms Diff(M ) acts on C ∞ (Σ, M ) via ϕ 7→ f ◦ ϕ, for
f ∈ Diff(M ) and we have
SH (f ◦ ϕ) = Sf ∗ H (ϕ).
This suggests a left action of Diff(M ) on Ω3cl (M ) given by push-forward. Consider
now the semi-direct product of Diff(M ) by the space of two forms on M
G = Diff(M ) ⋉ Γ(Λ2 T ∗ ),
with group structure
(f, B) · (f ′ , B ′ ) = (f ◦ f ′ , B ′ + f ′∗ B).
Then, the left action of Diff(M ) on Ω3cl (M ) extends to a G-action defined by
(f, B) · H = f∗ (H − dB).
In terms of the corresponding functionals, setting H ′ = f∗ (H − dB) we have

Z
−1
SH (ϕ) = SH (f ◦ ϕ) −
′ ϕ∗ (f∗ B).
Σ
Notice that our formula for the G-action reproduces the change of a closed three-
form under an isomorphism of Courant algebroids in (2.11) and, in fact, we have
the following.
Proposition 2.23. Given M a smooth manifold and a closed three-form H,
the isotropy group GH ≤ G of H is isomorphic to the group of automorphisms of
the twisted Courant algebroid (T ⊕ T ∗ , h, i , [, ]H , π).
Proof. The proof is straightforward from Proposition 2.21 and the definition
of the G-action on Ω3cl .
By the previous result, a Lagrangian density for our variational problem, given
by a closed three-form H, determines a geometric gadget, the exact Courant alge-
broid with twisted bracket, such that symmetries of SH are realized as automor-
phisms of (T ⊕ T ∗ , h, i , [, ]H , π). In order to interpret condition (2.14) in the present
setup, we need to regard elements in Γ(T ⊕ T ∗) as symmetries. For this, it is better
to think in terms of the Lie algebra
Lie G = Γ(T ⊕ Λ2 T ∗ )
with Lie bracket (see Corollary 2.22)
[X1 + B1 , X2 + B2 ] = [X1 , X2 ] + LX1 B2 − LX2 B1 .
Given H ∈ Ω3cl , the Lie algebra of its isotropy group is
Lie GH = {X + B : d(iX H + B) = 0}.
Proposition 2.24. Given M a smooth manifold and a closed three-form H,
the following defines a normal Lie subalgebra of Lie GH :
0
Lie GH = {X + dξ − iX H : ξ ∈ T ∗ }.
Proof. Using the identity i[X,Y ] = [LX , iY ] we calculate
[X + dξ − iX H, Y + B] = [X, Y ] + LX B − LY (dξ − iX H)
= [X, Y ] + diX B + iX d(B + iY H) − i[X,Y ] H − d(LY ξ)
= [X, Y ] + d(iX B − LY ξ) − i[X,Y ] H
0
for any Y + B ∈ Lie GH and X + dξ − iX H ∈ Lie GH , as required.
0
To finish, we note that the normal Lie subalgebra Lie GH ⊳ Lie GH is parametrized
∗
by elements in Γ(T ⊕ T ). This vector space provides a natural representation for
G, via the action
(f, B) · (X + ξ) = f∗ (X + ξ + iX B)
and there is a (right) GH -equivariant map
(2.15) Ψ : Γ(T ⊕ T ∗ ) → Lie GH
0
: X + ξ 7→ X + dξ − iX H.
We have now that the twisted Dorfman bracket is recovered via the infinitesimal
action
Ψ(X + ξ) · (Y + η) = [X, Y ] + LX η − iX dξ + iY iX H,
and the Jacobi identity in Lemma 2.5 is a direct consequence of the GH -equivariance
of the map Ψ. The map Ψ determines special elements X + ξ ∈ Γ(T ⊕ T ∗ ) such
that Ψ(X + ξ) = X, that is, iX H = dξ. The geometric content of this condition
is that the classical infinitesimal action of the vector field X on T ⊕ T ∗ via the
Lie derivative is realized through the twisted Dorfman bracket via the differential
operator [X + ξ, ·]H . Going back to our original variational problem, elements
X + ξ ∈ Γ(T ⊕ T ∗ ) such that iX H = dξ provide tangent directions in the space
of smooth maps C ∞ (Σ, M ) along which SH is constant. The condition iX H = dξ
appears also naturally in the context of reduction of Courant algebroids by the
action of a Lie group in [31].
2.3. Generalized metrics

Recall the definition of a classical Riemannian metric, which is a smooth choice
of positive definite inner product on the tangent spaces of a smooth manifold. The
geometric kernel of this fundamental object is clear, as it allows one to assign the
size of a vector tangent to a manifold, and hence by integration along a curve one
obtains concepts of length, and eventually notions of differentiation, curvature, etc.
The forthcoming definition of a generalized metric plays a similar role, when we
substitute curves–regarded as the trajectories of point particles–by maps from a
surface into the manifold–regarded as the trajectories of strings.
To start, we introduce generalized metrics seen through the lens of structure
groups. In particular, recall that a classical Riemannian metric is equivalent to a
reduction of the structure group of the tangent bundle to O(n). If we now take the
bundle T ⊕ T ∗ with its neutral inner product instead of the tangent bundle, then
reductions of its structure group should render interesting geometries. As discussed
in §2.1, the neutral inner product already reduces the structure group to O(n, n),
and as it turns out a further reduction to O(n) × O(n) will yield the right notion
of “generalized metric”. Unwinding this reduction in terms of tensor fields (see
Lemma 2.37) finally yields the fundamental notion.
2.3.1. Linear algebra. We shall first consider the linear algebra of gener-
alized metrics. Let V and W be real vector spaces which fit into a short exact
sequence
/ V∗ π
(2.16) 0 /W /V / 0.
We assume that W is endowed with a metric h, i of split signature (n, n) such that
the image of π ∗ : V ∗ → W ∗ is isotropic. Furthermore, we assume that the first
arrow is given by the composition of 12 π ∗ : V ∗ → W ∗ with the isomorphism W ∼
= W∗
given by the metric. We denote by O(n, n) the group of orthogonal transformations
of W . Of course, (2.16) corresponds to the exact sequence in Definition 2.9 for a
fixed point on the base manifold.
Definition 2.25. A generalized metric on W is an endomorphism G of W
satisfying
(1) hGa, Gbi = ha, bi,
(2) hGa, bi = ha, Gbi,
(3) The bilinear pairing hGa, bi is symmetric and positive definite.
As a basic consequence of the definitions, it follows that G 2 = Id.
Lemma 2.26. A generalized metric on W is equivalent to either
2.3. GENERALIZED METRICS 21
(1) a choice of maximal compact subgroup O(n) × O(n) ≤ O(n, n),

(2) an orthogonal decomposition W = V+ ⊕ V− such that the restriction of h, i
to V+ is positive definite.
Proof. A choice of maximal compact subgroup O(n) × O(n) inside O(n, n)
requires a choice of plane on which the neutral inner product is positive definite.
Call such a choice V+ , and let V− denote the neutral orthogonal complement of
V+ , and note that the neutral inner product is negative definite on V− . Thus (1)
implies (2). Given the orthogonal decomposition W = V+ ⊕ V− as in (2), we can
define a positive definite inner product on W via
G(a, b) = ha+ , b+ iV+ − ha− , b− iV− ,
where a = a+ + a− ∈ V+ ⊕ V− and similarly for b. Using the neutral inner product
we can identify W with its dual, and so identify G with an endomorphism G of W ,
which is easily seen to satisfy the axioms of a generalized metric.
Conversely, given a generalized metric, since G is real and G 2 = Id it follows
that the only possible eigenvalues for G are ±1, with eigenspaces V± . It follows
that V+ is a maximal plane on which the neutral inner product is positive definite,
yielding a maximal compact O(n)×O(n) given by orthogonal transformations which
commute with G. Note that of course the two constructions are inverses, with the
spaces V± playing the same role in each.
A generalized metric determines a series of useful isomorphisms and projections
which we will employ in the sequel, and which clarify the relationship between the
generalized metric and the classical data to be determined below.
Definition 2.27. Given G a generalized metric on W , define projection maps
π± : W → V± via
1
π± a := 2 (a ± Ga) .
Since the pairing h, i vanishes on π V it follows that V± ∩V ∗ = {0}, and hence
∗ ∗
the map π : V± → V has no kernel. Since both are vector spaces of dimension n it
follows that π : V± → V is an isomorphism.
Definition 2.28. Given G a generalized metric on W , define isomorphisms
−1
(2.17) σ± := π|V±
: V → V± .
Using now either σ+ or σ− , we can define an isotropic splitting of (2.16) (see
(2.5))
(2.18) σ = σ± − 12 π ∗ τ± ,
where τ± ∈ Sym2 V ∗ is defined by τ± (X, Y ) = hσ± X, σ± Y i. Recall that, similar to
Proposition 2.10, the isotropic splitting σ determines an isometry F : V ⊕ V ∗ → W
defined by
F (X + ξ) = σX + 21 π ∗ ξ
for the symmetric product (2.1).
Lemma 2.29. A generalized metric G on W is equivalent to pair (g, σ), where g
is a positive definite metric on V and σ : V → W is an isotropic splitting of (2.16),
such that

−1 0 g −1
(2.19) F GF = , F −1 (V± ) = {X ± g(X) : X ∈ V }.
g 0
We refer to the generalized metric determined by (g, σ) as G(g, σ).

Proof. Given a generalized metric G, we can construct σ as in (2.18). Notice
that
σ+ (X) − 21 π ∗ τ+ (X) − σ− + 21 π ∗ τ− , a = 0
for all a ∈ W , and therefore σ is independent of the choice of σ± . We also have
0 < hGσ+ X, σ+ Xi = hσ+ X, σ+ Xi = τ+ (X, X)
for X 6= 0, and therefore g = τ+ is a positive definite metric on V . Thus,
F (X + g(X)) = σX + 12 π ∗ g(X) = σ+ X
and therefore F −1 (V+ ) = {X + g(X) : X ∈ V }. Finally, since F is an isometry and
V− is orthogonal to V+ , (2.19) follows. Conversely, given a pair (g, σ), where g is a
positive definite metric on V and σ : V → W is an isotropic splitting, it is easy to
verify that G defined by (2.19) is a generalized metric on W .
Our next goal is to make our previous discussion more explicit, writing a gener-
alized metric in terms of tensors on the vector space V in (2.16). This will help us to
determine the classical data which is equivalent to the specification of a generalized
metric, in the geometric situation below. We fix an isotropic splitting σ0 : V → W ,
inducing an identification
W = V ⊕ V ∗,
so that σ0 (X) = X and the neutral metric on W is given by (2.1). Given another
isotropic splitting σ : V → V ⊕ V ∗ , it follows that b = σ − σ0 : V → V ∗ , which can
be identified with an element b ∈ V ∗ ⊗ V ∗ . The isotropic condition implies further
that b ∈ Λ2 V ∗ and therefore
σ(X) = X + b(X) = eb (X).
Thus, we can identify the isometry F : V ⊕ V ∗ → V ⊕ V ∗ with eb , that is,

1 0 X
F (X + ξ) = eb (X + ξ) = = X + ξ + iX b.
b 1 ξ
The next proposition follows now as an immediate consequence of Lemma 2.29.
Proposition 2.30. A generalized metric G on V ⊕ V ∗ is equivalent to a pair
(g, b) consisting of a positive definite metric g on V and b ∈ Λ2 T ∗ , such that
−1

b 0 g −b −g −1 b g −1
(2.20) G=e e = .
g 0 g − bg −1 b bg −1
We refer to the generalized metric determined by (g, b) as G(g, b). Furthermore,
given k ∈ Λ2 (V ∗ ), we obtain a new generalized metric via Gk := ek Ge−k , with
associated pair (g, b + k).
In terms of the previous explicit description of a generalized metric G = G(g, b)
on V ⊕ V ∗ , we have that
V± = {X + (b ± g)X, X ∈ V },
that the projections π± : V ⊕ V ∗ → V± in Definition 2.27 are given by
π± (X + ξ) = 21 (X + (b ± g)X) + 21 (ξ ± bg −1 ξ ± g −1 ξ),
and that the linear sections σ± : V → V± in Definition 2.28 are
σ± (X) = X + (b ± g)X.
To conclude our study of the linear theory, we next characterize the tangent
space to the space of generalized metrics. It will convenient to go back to the
abstract setup at the beginning of this section (see (2.16)). We denote
G(W ) ⊂ End(W )
the space of generalized metrics on W .
Lemma 2.31. The tangent space at G to the space of generalized metrics G(W )
is given by

GΦ ∈ End(W ) | Φ ∈ Λ2 W, ΦG + GΦ = 0 .
Proof. Let Gt be a one-parameter family of generalized metrics on W , with

G0 = G. Denote
d
Ġ = Gt ∈ End(W ).
dt t=0
Setting Φ = G Ġ, by Definition 2.25 it follows that
(2.21) Φ ∈ Λ2 W, ΦG + GΦ = 0.
It remains to show that an element Φ ∈ End(W ) satisfying (2.21) can be realized
t
through a curve of generalized metrics. For this, consider Φt = e− 2 Φ ∈ O(W ) and
define
Gt = Φt GΦ−t .
It is easy to see that Gt defines a path of generalized metrics and
d
Ġ = Gt = − 21 ΦG + 21 GΦ = GΦ,
dt t=0
as required.
By the second condition in (2.21), a tangent vector Φ at G induces well-defined

maps
Φ|V− : V− → V+ , Φ|V+ : V+ → V− .
Furthermore, the first condition in (2.21) implies that Φ|V+ is uniquely determined
by Φ|V− , thus leading us to the following result.
Lemma 2.32. The map Φ → Φ|V− defines a linear isomorphism between the
tangent space at G and Hom(V− , V+ ).
Proof. By Lemma 2.31, the linear map Φ → Φ|V− is well-defined and injective.
Given R+ ∈ Hom(V− , V+ ), we can define R− ∈ Hom(V+ , V− ) by
hR− a+ , b− i := − ha+ , R+ b− i
for any a+ ∈ V+ and b− ∈ V− . Then it follows that Φ = R+ + R− ∈ End(W ) is
tangent to G, and furthermore Φ|V− = R+ .
Using Lemma 2.29 and Proposition 2.30, we unravel our previous discussion in
terms of the vector space V in (2.16). By Lemma 2.29, the space of generalized
metrics G(W ) is in one-to-one correspondence with
M(V ) × S(W ),
where M(V ) denotes the space of positive definite metrics on V and S(W ) denotes
the space of isotropic splittings σ : V → W of the sequence (2.16). Upon a choice
of an element (g, σ) ∈ M(V ) × S(W ), the space G(W ) is bijective to
GL(V )/O(V ) × Λ2 V ∗ ,
where O(V ) denotes the space of orthogonal transformations of (V, g). Thus, the
tangent space at G = G(g, σ) can be canonically identified with
Sym2 V ∗ ⊕ Λ2 V ∗ ∼
= V ∗ ⊗ V ∗.
Lemma 2.33. Let G = G(g, σ) be a generalized metric, and consider the induced
identification W = V ⊕ V ∗ as in Lemma 2.29. Given a one-parameter family of
generalized metrics Gt = G(gt , σt ) with G0 = G,
−1
d 0 −g −1 ġg −1 −g ḃ −g −1 ġg −1
(2.22) Ġ = Gt = [1 − eḃ , G0 ] + = ,
dt t=0 ġ 0 ġ ḃg −1
where
d d
gt = ġ ∈ Sym2 V ∗ , σt = ḃ ∈ Λ2 V ∗ .
dt t=0 dt t=0
In particular, the associated tangent vector to the space of generalized metrics in

Lemma 2.31 is
−1
g ġ g −1 ḃg −1
(2.23) Φ = G Ġ = .
−ḃ −ġg −1
Proof. Using the identification W = V ⊕ V ∗ induced by σ, following the

notation in Lemma 2.30 we can write Gt = G(gt , bt ) for
(gt , bt ) ∈ M(V ) × Λ2 V ∗
and b0 = 0. Then, (2.22) follows easily taking derivatives in expression (2.20). We
leave (2.23) to the reader.
Remark 2.34. Given Φ ∈ End(W ) satisfying (2.21) it is an exercise to verify

that there exists (ġ, ḃ) ∈ Sym2 V ∗ ⊕ Λ2 V ∗ such that (2.23) holds.
Given a generalized metric G and a tangent vector Φ, using the isomorphism
σ− : V → V− in (2.17) combined with the map π : V+ → V we can regard the
homomorphism Φ|V− in Lemma 2.32 as an element
π ◦ Φ|V− ◦ σ− ∈ End(V ).
Then, using (2.23) we have the following.
Lemma 2.35. Let G = G(g, σ) be a generalized metric. Given a tangent vector
Φ as in (2.23), we have
(2.24) π ◦ Φ|V− ◦ σ− (X) = g −1 (ġ(X) − ḃ(X)).
Proof. Using the identification W = V ⊕ V ∗ induced by σ we have σ− (X) =

X − g(X), and the proof follows from (2.23).
2.3.2. Generalized metrics on manifolds. We next turn to the geometry,

by introducing the notion of generalized metric on a smooth manifold M . Consider
the exact Courant algebroid (T ⊕ T ∗ , h, i , [, ], π) studied in §2.1.
Definition 2.36. Given M a smooth manifold, a generalized metric on T ⊕ T ∗
is an endomorphism G ∈ Γ(End T ⊕ T ∗ ) satisfying
(1) hG(X + ξ), G(Y + η)i = hX + ξ, Y + ηi,
(2) hG(X + ξ), Y + ηi = hX + ξ, G(Y + η)i,
(3) The bilinear pairing hG(X + ξ), Y + ηi is symmetric and positive definite.
Relying on Lemma 2.26, a generalized metric provides a reduction of the
O(n, n)-bundle of frames of T ⊕ T ∗ to a maximal compact O(n) × O(n), thus show-
ing that generalized metrics are natural analogues of ordinary metrics (regarded as
reductions of the frame bundle of M to the orthogonal group O(n)). The details
of the following lemma are left to the reader.
Lemma 2.37. A generalized metric on T ⊕ T ∗ is equivalent to either
(1) a reduction of the bundle of frames of T ⊕ T ∗ to a maximal compact
subgroup O(n) × O(n) ≤ O(n, n),
(2) an orthogonal bundle decomposition T ⊕ T ∗ = V+ ⊕ V− such that the
restriction of h, i to V+ is positive definite.
We will use the same notation for the bundle maps π± , σ± , σ as in the linear
theory. The next proposition is fundamental to the study of generalized geometry,
indicating the classical data which is equivalent to the specification of a generalized
metric. In particular, a generalized metric is seen equivalent to a classical Riemann-
ian metric g, together with a b-field. We have already seen closed b-fields playing
a role as symmetries of the Courant bracket in §2.2, whereas here the associated
b-field is not closed, and plays a nontrivial role in determining relevant associated
geometric quantities. The proof of the following is straightforward from Proposition
2.30.
Proposition 2.38. Given M a smooth manifold, a generalized metric G on
T ⊕ T ∗ uniquely determines a pair (g, b) consisting of a Riemannian metric on M
and two-form b ∈ Γ(Λ2 T ∗ ). Conversely, such a pair (g, b) determines G via

0 g −1 −b −g −1 b g −1
G = eb e = .
g 0 g − bg −1 b bg −1
We refer to the generalized metric determined by (g, b) as G(g, b). Furthermore,
given k ∈ Γ(Λ2 T ∗ ), we obtain a new generalized metric via Gk := ek Ge−k , with
associated pair (g, b + k).
By Proposition 2.29, a generalized metric G determines an isotropic splitting
on T ⊕ T ∗ given by
σ(X) = eb (X) = X + b(X),
and therefore a preferred presentation of the Dorfman bracket on Γ(T ⊕ T ∗ ):
e−b [eb (X + ξ), eb (Y + η)] = [X, Y ] + LX η − iY dξ + iY iX db.
We also obtain a preferred representative db of the (trivial) Ševera class of the
exact Courant algebroid (T ⊕ T ∗ , h, i , [, ], π). This situation extends naturally to
the case of generalized metrics on general exact Courant algebroids, as considered
in Definition 2.9.
Definition 2.39. Given M a smooth manifold and an exact Courant algebroid

E over M , a generalized metric on E is a bundle endomorphism G ∈ Γ(End E)
satisfying
(1) hGa, Gbi = ha, bi,
(2) hGa, bi = ha, Gbi,
(3) The bilinear pairing hGa, bi is symmetric and positive definite.
The analogue of Proposition 2.38 for the case of a general exact Courant al-
gebroid is provided by the following proposition. The proof follows directly from
Lemma 2.29. Recall that an isotropic splitting σ : T → E determines an isometry
F : T ⊕ T ∗ → E defined by
F (X + ξ) = σX + 12 π ∗ ξ
for the symmetric product (2.1).
Proposition 2.40. Given M a smooth manifold and an exact Courant alge-
broid E over M , a generalized metric G is equivalent to pair (g, σ), where g is a
positive definite metric on M and σ : T → E is an isotropic splitting of (2.16),
such that

0 g −1
F −1 GF = , F −1 (V± ) = {X ± g(X) : X ∈ T }.
g 0
Consequently, a generalized metric determines a preferred presentation of the Dorf-
man bracket on Γ(E),
F −1 [F (X + ξ), F (Y + η)] = [X, Y ] + LX η − iY dξ + iY iX H,
where H is a preferred representative of the Ševera class of E, given by
H(X, Y, Z) = 2 h[σX, σY ], σZi .
Given a generalized metric G on an exact Courant algebroid E over M , our
next goal is to characterize the group of generalized isometries, that is, the group
of Courant automorphisms F : E → E which preserve the generalized metric. Us-
ing the isotropic splitting induced by G, E is isomorphic to the twisted Courant
algebroid (T ⊕ T ∗ , h, i , [, ]H , π), whose automorphism group has been characterized
in Proposition 2.21.
Proposition 2.41. Let G = G(g, σ) be a generalized metric on an exact Courant
algebroid E over M . The group of generalized isometries of G, defined as the group
of Courant automorphisms F : E → E which preserve the generalized metric, is
given by
{f : f ∈ Diff(M ), such that f ∗ H = H, f ∗ g = g} ⊂ Aut(E)
where H ∈ Ω3cl (M ) is the preferred representative of the Ševera class of E deter-
mined by G.
Proof. Consider the isomorphism E ∼ = (T ⊕ T ∗ , h, i , [, ]H , π) given by the
generalized metric G. With this identification, by Lemma 2.29) we have

0 g −1
G= .
g 0
By Proposition 2.21, a general element f ◦ eB ∈ Aut(E) preserves G, that is,
f ◦ eB ◦ G = G ◦ f ◦ eB ,
if and only if
(f ∗ g)−1 (B(X) + ξ) = g −1 ξ, (f ∗ g)(X) = B(g −1 ξ) + g(X)
for all X + ξ ∈ Γ(T ⊕ T ∗). Setting e.g. ξ = 0 this implies f ∗ g = g, and consequently
B = 0, as claimed. Returning again to Proposition 2.21 it follows that f ∗ H =
H.
Remark 2.42. With the notation in Proposition 2.38, the action of F = f ◦ eB

on G is
F ◦ G ◦ F −1 = G(f∗ g, f∗ B).
In particular, by Lemma 2.33 the infinitesimal action of X + B ∈ Lie Aut(E) is
given by
−1
−g B g −1 (LX g)g −1
(X + B) · G = .
−LX g Bg −1
To finish this section, we show that a generalized metric G determines an energy
functional SG on the space of smooth maps C ∞ (Σ, M ) from a surface Σ into our
target manifold M . This functional can be thought of as an analogue of the energy
functional which associates the length squared to a parametrized curve on M ,
leading to the notion of distance and geodesics in Riemannian geometry. Morally,
SG can be similarly used to define a pseudo-metric on the space of loops C ∞ (S 1 , M )
of M , by extremizing the energy with prescribed boundary components. We will
follow our discussion at the end of §2.2.
Let M be an oriented smooth manifold. Let Σ be a smooth compact surface
(oriented and without boundary, say). Consider the infinite-dimensional space of
smooth maps C ∞ (Σ, M ). Let (E, G) be an exact Courant algebroid over M endowed
with a generalized metric G. By Lemma 2.40, G determines a Riemannian metric g
and a closed three-form H ∈ Λ3 T ∗ on M . We will assume that [H] ∈ H 3 (M, Z). We
fix a Riemannian metric η on Σ with volume dVη . Then, we can define a functional
Z Z
1
(2.25) SG : C ∞ (Σ, M ) → R/Z, ϕ 7→ |dϕ|2 dVη + ϕ∗ H.
2 Σ Y
Here, |dϕ|2 denotes the norm squared of dϕ ∈ Γ(Σ, T ∗ Σ ⊗ ϕ∗ T ) with respect to

the metric η ∗ ⊗ ϕ∗ g and Y is a three-manifold with boundary Σ and ϕ : Y → M
is any smooth extension of ϕ to Y . This functional is known in the mathematical
physics literature as the Polyakov action with Wess-Zumino term in the action of
a two-dimensional Σ-model (see e.g. [80]). Note that the second summand in the
definition of SG corresponds to the functional SH in §2.2. Our next result shows
that the critical points of SG provide an interesting generalization of the harmonicity
condition for maps ϕ ∈ C ∞ (Σ, M ).
Proposition 2.43. The critical points ϕ ∈ C ∞ (Σ, M ) of the functional SG are
the solutions of the equation
∇∗ ∇ϕ + ϕ∗ g −1 H, dVη η
= 0,
where ∇∗ ∇ is the rough Laplacian for the product of Levi-Civita connections of η

and g, and ϕ∗ g −1 H, dVη η ∈ Γ(ϕ∗ T ) denotes the product of ϕ∗ g −1 H ∈ Γ(ϕ∗ T ⊗
Λ2 T ∗ Σ) with dVη with respect to η.
Proof. A calculation in coordinates shows that the functional SG can be al-

ternatively written as
Z Z
1 2
SG (ϕ) = |∇ϕ| dVη + ϕ∗ H.
2 Σ Y
This follows simply from the fact that action of the Levi-Civita connection on
functions coincides with the exterior differential. Then, the variation of SG along
X ∈ Γ(ϕ∗ T ), identified with a vector in the tangent space of C ∞ (Σ, M ) at ϕ, is
given by
Z Z
δSG (X) = h∇ϕ, ∇Xi dVη + ϕ∗ iX H
Σ Σ
Z Z
∗
= h∇ ∇ϕ, Xi dVη + ϕ∗ g −1 H, X ⊗ dVη dVη ,
Σ Σ
as claimed.
2.4. Divergence operators

As we have seen in the previous section, natural notions in generalized geometry
are described in terms of the O(n, n)-bundle of frames of T ⊕ T ∗ . Starting with
its definition as a reduction to a maximal compact subgroup, a generalized metric
G is a posteriori seen to be equivalent to a classical Riemannian metric g together
with a b-field. Since the neutral metric on T ⊕ T ∗ is fixed in this geometry, how
does generalized geometry keep track of conformal rescalings of the classical metric
associated to G? In this section we answer this question, by means of the following
device introduced in [70].
Definition 2.44. Let E be an exact Courant algebroid over a smooth manifold
M . A divergence operator on E is a differential operator div : Γ(E) → C ∞ (M )
satisfying
(2.26) div(f e) = π(e)(f ) + f div(e),
for all e ∈ Γ(E) and f ∈ C ∞ (M ).
As a consequence of the Leibniz rule (2.26), the divergence operators on E
form an affine space modeled on Γ(E), since if we fix a divergence operator div,
any other divergence operator div′ is given by
div′ = div + he, ·i
for some e ∈ Γ(E). The following result shows that this space is always non-empty.
Lemma 2.45. Let E be an exact Courant algebroid over M . Given a connection
∇ on T , one obtains a divergence operator on E via
div∇ (e) = tr ∇π(e),
where π : E → T is the anchor map.
Proof. This follows from a simple calculation:

div∇ (f e) = f div∇ (e) + tr df ⊗ π(e) = f div∇ (e) + π(e)(f ).

2.4. DIVERGENCE OPERATORS 29
More explicitly, consider an isotropic splitting σ : T → E and the induced

identification E = T ⊕ T ∗ , with anchor map
π(X + ξ) = X.
Then, it follows that, for e = X + ξ,
div∇ (e) = tr ∇X.
Assuming now that ∇ is torsion-free, so that tr ∇X = tr (∇X − LX ), and that
there exists a density µ ∈ Γ(| det T ∗ |) preserved by ∇, we obtain
div∇ (e)µ = LX µ,
which recovers the standard notion of divergence for the vector field π(e) = X. In
particular, for any choice of metric g on M we can choose its Levi-Civita connection
∇ and the associated Riemannian density µg , obtaining a divergence operator
divg (e)µg = LX µg .
Definition 2.46. Given a generalized metric G on an exact Courant algebroid
E, define its Riemannian divergence as
Lπ(e) µg
divG (e) = ,
µg
where g is the Riemannian metric associated to G in Lemma 2.40, and µg is the
associated Riemannian density.
The relation between divergence operators and conformal geometry appears via
the notion of Weyl structure [67]. Recall that a conformal structure C on a smooth
manifold M is a set of Riemannian metrics, such that if g, g ′ ∈ C there exists a
smooth function f ∈ C ∞ (M ) such that
g ′ = ef g.
Definition 2.47. Let C be a conformal structure on a smooth manifold M . A
Weyl structure is a map
W : C → Γ(T ∗ )
such that W (ef g) = W (g) − df for all g ∈ C and f ∈ C ∞ (M ).
Let us fix an exact Courant algebroid E over M , and denote by G(E) the space
of generalized metrics on E. By Lemma 2.40, there is a bijection
G(E) ∼ = M(T ) × S(E),
where M(T ) and S(E) denote the space of Riemannian metrics on M and the
space of isotropic splittings of E, respectively. Given (g, σ) ∈ M(T ) × S(E), we
denote G(g, σ) the associated generalized metric. Our next result shows that a
divergence operator provides a natural analogue of a Weyl structure in the context
of generalized geometry.
Lemma 2.48. Let div be a divergence operator on E. Define
W : G(E) → Γ(E) : G 7→ div − divG ,
where divG is the Riemannian divergence of G (see Definition 2.46). Then,
W (G(ef g, σ)) = W (G(g, σ)) − n hdf, i
for any (g, σ) ∈ M(T ) × S(E) and any f ∈ C ∞ (M ).
Proof. Consider the identification E = T ⊕ T ∗ given by G = G(g, σ). Then,

f n
setting G ′ = G(ef g, σ) we have µe g = e 2 f µg and therefore
n
′ LX (e 2 f µg )
divG (X + ξ) = n = divG (X + ξ) + n2 df (X) = divG (X + ξ) + n hdf, Xi .
e 2 f µg

Our next goal is to understand a natural compatibility condition between a
generalized metric and a divergence operator, which plays an important role in the
definition of the generalized Ricci flow. Given a generalized metric G on E and
e ∈ Γ(E) we define an endomorphism
[e, G] ∈ End(E),
by
[e, G](e′ ) := [e, Ge′ ] − G[e, e′ ].
It is an easy exercise, using axiom (3) of the Dorfman bracket in Definition 2.8,
to check that [e, G] indeed defines a tensor.
Definition 2.49. Given a generalized metric G on E, we say that e ∈ Γ(E) is
an infinitesimal isometry if [e, G] = 0.
Lemma 2.50. Let G be a generalized metric on E. Then, e ∈ Γ(E) is an
infinitesimal isometry if and only if
LX g = 0, dξ = iX H,
where e = X + ξ via the identification E = T ⊕ T ∗ given by G.
Proof. Consider the identification E = T ⊕ T ∗ given by G = G(g, σ), so that
e = X + ξ. Using the notation in §2.2, we have (see (2.15) and Remark 2.42)
−1
−g (dξ − iX H) g −1 (LX g)g −1
[e, G] = Ψ(X + ξ) · G = ,
−LX g (dξ − iX H)g −1
and the result follows.
Remark 2.51. Following the discussion in §2.2, an infinitesimal isometry pro-
vides a flat direction for the Wess-Zumino functional SH in (2.13).
We introduce the following compatibility condition.
Definition 2.52. Let E be an exact Courant algebroid over M . A pair (G, div)
is compatible if W (G) is an infinitesimal isometry of G, that is,
[W (G), G] = 0.
As a straightforward consequence of Lemma 2.40 and Lemma 2.50 we obtain
the following characterization of compatible pairs (G, div).
Proposition 2.53. Let E be an exact Courant algebroid over M . Then,
(G, div) is a compatible pair if and only if
div = divG + σ(X) + 12 π ∗ ξ, ·
where G = G(g, σ) and
LX g = 0, dξ = iX H.
2.4. DIVERGENCE OPERATORS 31
An interesting class of infinitesimal isometries, and hence of compatible pairs,

is provided by closed 1-forms, that is, given by X = 0 and dξ = 0. More invariantly,
this condition can be expressed in terms of the anchor map as
[W (G), G] = 0, π(W (G)) = 0.
This class of compatible pairs, associated to closed one-forms on the manifold, will
be formally introduced in Definition 3.40 (see also Definition 3.48), and further
studied in §3.6 and §3.7.
CHAPTER 3
Generalized Connections and Curvature
Having established the fundamental properties of generalized metrics, we turn

to the question of connections and curvature. The fundamental theorem of Rie-
mannian geometry states that for each Riemannian metric there is a unique linear
connection which is compatible with the metric and torsion-free. Here the story is
more complicated, once one observes that the condition of having zero torsion does
not determine a metric-compatible generalized connection uniquely. The relevant
connections in generalized Riemannian geometry will be related to certain classi-
cal connections on the tangent bundle, which in most cases will have nontrivial
torsion. We analyze the curvature of these connections from the generalized and
classical points of view, eventually leading to the definition of the generalized Ricci
and scalar curvatures. In particular, we will study in detail the classical Bismut
connection, a metric compatible connection with totally skew-symmetric torsion,
deriving formulas for the curvature of this connection as well as Bianchi identities.
With these fundamental concepts at hand, we turn to the task of defining
canonical generalized geometries. For this, we seek to extend the classical Ein-
stein equation to the setting of generalized geometry. By coupling the classical
Einstein-Hilbert action to the torsion we motivate an Einstein-like equation from a
variational point of view, and give nontrivial examples of generalized Einstein struc-
tures. In particular, we will prove that all Bismut-flat connections are associated
to bi-invariant metrics on semisimple Lie groups.
3.1. Generalized connections

We start with the study of connections in the setting of generalized geometry.
In the spirit of treating a Courant algebroid E as the generalized tangent bundle, we
want to not only differentiate sections of E, but we want to differentiate them in the
tangent as well as cotangent ‘directions’. Parallel to the notion of orthogonal affine
connection in Riemannian geometry, we introduce next the concept of generalized
connection, which allows us to differentiate sections of E along directions in E, in
a way compatible with the inner product h, i. The study of generalized connections
was initiated in [3], and largely expanded in [69, 70, 89, 109].
Definition 3.1. A generalized connection on an exact Courant algebroid E is
a first-order differential operator
D : Γ(E) → Γ(E ∗ ⊗ E)
which satisfies the Leibniz rule and is compatible with h, i, i.e.
Da (f b) = π(a)(f )b + f Da b,
(3.1)
π(a)hb, ci = hDa b, ci + hb, Da ci.
33
34 3. GENERALIZED CONNECTIONS AND CURVATURE
We will use the notation Da b = ha, Dbi for the natural pairing between elements in
E ∗ and elements in E.
Lemma 3.2. Given E an exact Courant algebroid, the space of generalized con-
nections is a nonempty affine space modeled on the vector space
Γ(E ∗ ⊗ o(E)),
where
o(E) = {Φ ∈ End(E) : hΦ, i = −h, Φi}.
Proof. Fix a standard orthogonal connection ∇E on E, and then one can
check that the formula
(3.2) Da b = ∇E
π(a) b
defines a generalized connection on E. The proof that the difference of two gen-
eralized connections lies in Γ(E ∗ ⊗ o(E)) is a straightforward exercise using the
axioms (3.1).
Example 3.3. Using Proposition 2.10, choose an isotropic splitting σ : T → E,
with corresponding isomorphism F : E → T ⊕ T ∗ . Fix an affine connection ∇ on
T , with connection ∇∗ naturally induced on T ∗ by enforcing
X(α(Y )) = α(∇X Y ) + (∇∗X α)(Y ).
Then it is straightforward to check that the connection
∇E = ∇ ⊕ ∇∗
on T ⊕ T ∗ is orthogonal with respect to h, i, and thus will define a generalized con-
nection via (3.2). Finally, pulling-back D via F , we obtain the desired generalized
connection on E.
Definition 3.4. Given E an exact Courant algebroid and D a generalized
connection on E, the torsion TD of D is defined by
(3.3) TD (a, b, c) = hDa b − Db a − [a, b], ci + hDc a, bi,
for any a, b, c ∈ Γ(E).
Using the axiomatic definition of Courant algebroid and the Leibniz rule for
a generalized connection (see Definition 2.8 and Definition 3.1), one can easily
prove that formula (3.3) defines a three-tensor TD ∈ Γ(E ⊗3 ). Another elementary
calculation using that D is compatible with h, i shows that, indeed,
(3.4) TD ∈ Γ(Λ3 E ∗ ).
Further computations show that this definition is equivalent to other formulations
which have appeared (eg. [3, 89]). Rather than giving an abstract proof of the sym-
metries of the generalized torsion, we derive these by means of an explicit formula
in the next result. We will use σ(a, b, c) to denote sum over cyclic pertumutations
on a, b, c.
Lemma 3.5. Given M a smooth manifold let E = T ⊕ T ∗ with H-twisted
Dorfman bracket (2.6). Let ∇ be an affine connection on M , and let ∇E = ∇⊕ ∇∗ .
Given χ ∈ Γ(E ∗ ⊗ o(E)), the torsion of D = ∇E π(·) + χ is given by
X
(3.5) TD (a, b, c) = hT∇ (πa, πb), ci + hχa b, ci − 16 H(πa, πb, πc) .
σ(a,b,c)
3.1. GENERALIZED CONNECTIONS 35
Proof. Denoting D0 = ∇E
π(·) , we have that
TD (a, b, c) = TD0 (a, b, c) + hχa b − χb a, ci + hχc a, bi

X
= TD0 (a, b, c) + hχa b, ci,
σ(a,b,c)
where we have used that hχb a, ci = −hχb c, ai. Setting now a = X + ξ, b = Y + η,

c = Z + ζ, we define
N := Da0 b − Db0 a − [a, b]
= T∇ (X, Y ) + ∇∗X η − ∇∗Y ξ − LX η + iY dξ − iY iX H.
It follows that
TD0 (a, b, c) = hN, ci + hDc a, bi
= hN, ci + h∇E
Z a, bi
= hN, ci + 12 η(∇Z X) + 12 iY (∇∗Z ξ)

= hN, ci + 12 η(∇Z X) − 12 ξ(∇Z Y ) + 21 iZ d(ξ(Y ))
= − 12 H(X, Y, Z) + hT∇ (πa, πb), ci + h∇∗X η − ∇∗Y ξ − LX η + iY dξ, ci
+ 21 η(∇Z X) − 12 ξ(∇Z Y ) + 21 iZ d(ξ(Y ))
= − 12 H(X, Y, Z) + hT∇ (πa, πb), ci + h∇Z X − ∇X Z, bi + h∇Y Z − ∇Z Y, ai
+ 21 iX d(η(Z)) + 12 iZ d(ξ(Y )) − 12 iY d(ξ(Z)) − 12 iZ LX η + 21 iZ iY dξ
= − 12 H(X, Y, Z) + hT∇ (πa, πb), ci + h∇Z X − ∇X Z, bi + h∇Y Z − ∇Z Y, ai
+ 21 (LX iZ − iZ LX )η + 12 (iZ LY − LY iZ )ξ
= − 12 H(X, Y, Z) + hT∇ (πa, πb), ci + hT∇ (πb, πc), ai + hT∇ (πc, πa), bi,
as claimed. The last part of the statement follows from the standard symmetry
T∇ (X, Y ) = −T∇ (Y, X) for the torsion of an affine connection ∇ on T .
With the previous result at hand, we can provide some Lie theoretic motiva-
tion for Definition 3.4. Suppose for a moment that we have an abstract Courant
algebroid E and assume that M is just one point. Then, the axioms in Definition
2.8 tell us that E is a quadratic Lie algebra, that is, a Lie algebra endowed with
an invariant bilinear form h, i. In this setup, a generalized connection is simply an
element χ ∈ E ∗ ⊗ o(E) and the torsion
X
Tχ (a, b, c) = hχa b, ci − 13 h[a, b], ci
σ(a,b,c)
3 ∗
measures the difference in Λ E between the total skew-symmetrization of χ and
the canonical Cartan three-form on the quadratic Lie algebra E.
In order to study torsion-free generalized connection compatible with a general-
ized metric in the next section, we finish this section with a detailed analysis of the
space of generalized connections with fixed torsion, and its relation with divergence
operators in Definition 2.44. Let T be an element in Γ(Λ3 E ∗ ), and consider the set
DT of generalized connections on E with fixed torsion T
DT ⊂ D.
Lemma 3.6. Given E a Courant algebroid and D a connection with torsion T .

The space DT is an affine space modeled on
 
 X 
(3.6) Σ = χ ∈ Γ(E ⊗3 ) : χ(a, b, c) = −χ(a, c, b), χ(a, b, c) = 0 .
 
σ(a,b,c)
Proof. Exercise.
Remark 3.7. The space Σ admits a canonical splitting

(3.7) Σ = Σ0 ⊕ Γ(E),
where e ∈ Γ(E) corresponds to the mixed symmetric tensor χe defined by
χe (a, b, c) = ha, bihe, ci − he, biha, ci
and the orthogonal complement of Γ(E) is given by
n 2n
X o
(3.8) Σ0 = χ ∈ Σ : χ(ei , eei , ·) = 0 .
i=1
Here, 2n is the rank of E (twice the dimension n of the smooth manifold M ), {ei }
is an orthogonal local frame for E and the {e ei } are sections of E defined so that
hei , eej i = δij .
More explicitly, given an arbitrary element χ ∈ Γ(E ∗ ⊗ o(E)) we have a unique
decomposition
(3.9) χ = χ0 + χe ,
where
χea b = ha, bie − he, bia
with
2n
1 X
e= χe eei .
2n − 1 i=1 i
It follows by construction that χ − χe ∈ Σ0 .
The E ∗ -valued skew-symmetric endomorphism χe in the decomposition (3.9)
is reminiscent of the ‘1-form valued Weyl endomorphisms’ in conformal geometry,
which appear in the variation of a metric connection with fixed torsion upon a
conformal change of the metric. Similarly, our E ∗ -valued Weyl endomorphisms χe
enable us to deform a generalized connection D with fixed torsion T inside the space
DT . For later applications in this work, it is important to study the interaction
between the space DT and the notion of divergence operator. This will allow us
to ‘gauge-fix’ the Weyl degrees of freedom in the space of generalized connections
DT corresponding to Γ(E) in (3.7). The basic observation is that any generalized
connection determines a divergence operator as in §2.4. Given this, one can show
that the divergence constrains the Weyl degrees of freedom in the space generalized
connections with fixed torsion, as required.
Definition 3.8. The divergence operator of a generalized connection D on E
is
divD (a) = tr Da ∈ C ∞ (M ).
3.2. METRIC COMPATIBLE CONNECTIONS 37
Lemma 3.9. Given E an exact Courant algebroid fix T ∈ Γ(Λ3 E ∗ ) and div : Γ(E) →
∞
C (M ) a divergence operator on E. Then
DT (div) := {D ∈ D | TD = T, divD = div} ⊂ DT
is an affine space modeled on Σ0 , as defined in (3.8).
Proof. Given D′ , D ∈ DT , we denote D′ = D + χ with χ ∈ Σ (see (3.6)).
Then, we have
2n
X
divD′ (a) = divD (a) − hχeei ei , ai.
i=1
Decomposing χ = χ0 + χe as in (3.9), for e ∈ Γ(E), we obtain
2n
X
χeei ei = (2n − 1)e.
i=1
Imposing now divD′ = divD = div, implies e = 0, which concludes the proof.
3.2. Metric compatible connections

In this section we prove an analogue in generalized geometry of the Fundamen-
tal Lemma of Riemannian geometry, that is, the existence of a unique torsion-free
connection compatible with a given metric. In generalized geometry there is always
the freedom to choose a compatible torsion-free generalized connection, that we
will constrain via the Weyl gauge fixing mechanism introduced in Lemma 3.9. This
will lead us to the definition of canonical operators in the next sections, providing
a weak analogue of the Fundamental Lemma. Throughout this section we fix an
exact Courant algebroid E over a smooth manifold M and a generalized metric G.
Ideas in this direction appeared originally in [89, 100]. Our discussion here follows
closely [70].
Definition 3.10. A generalized connection D on E is compatible with G if
DG ≡ 0.
Equivalently, this condition will hold if
D(Γ(V± )) ⊂ Γ(E ∗ ⊗ V± ).
Thus, the space of G-compatible generalized connections on E, which we denote
D(G), is an affine space modeled on
Γ(E ∗ ⊗ o(V+ )) ⊕ Γ(E ∗ ⊗ o(V− )).
Such a connection is thus determined by four first-order differential operators
+ −
D− : Γ(V+ ) → Γ(V−∗ ⊗ V+ ), D+ : Γ(V− ) → Γ(V+∗ ⊗ V− ),
+ −
D+ : Γ(V+ ) → Γ(V+∗ ⊗ V+ ), D− : Γ(V− ) → Γ(V−∗ ⊗ V− ),
satisfying the Leibniz rule (formulated in terms of the projections π± ). The opera-
± ±
tors D± are called pure type, whereas the operators D∓ are called mixed type. We
furthermore say that the torsion of such a connection D is of pure type if
TD ∈ Γ(Λ3 V+∗ ⊕ Λ3 V−∗ ).
±
The next lemma shows that the mixed-type operators D∓ are fixed, if we vary
a generalized connection D inside D(G) while preserving the torsion TD . And
±
furthermore when the torsion is of pure-type then D∓ are uniquely determined by
the generalized metric and the bracket on E.
Lemma 3.11. Let D ∈ D(G) with torsion T ∈ Γ(Λ3 E ∗ ).
(1) If D′ ∈ D(G) and TD′ = T , then (D′ )± ±
∓ = D∓ .
±
(2) Furthermore, T is of pure-type if and only if the mixed-type operators D∓
are
(3.10) Da− b+ = [a− , b+ ]+ , Da+ b− = [a+ , b− ]− .
Proof. Given a, b, c ∈ Γ(E), we have
(3.11) T (a− , b+ , c+ ) = (Da− b+ − [a− , b+ ], c+ ),
and therefore (2) follows. To prove (1), we subtract from (3.11) the analogous
expression for TD′ (a− , b+ , c+ ).
The second part of the previous lemma provides a weak analogue of the Koszul
formula in Riemannian geometry. The canonical mixed-type operators (3.10) were
first considered in [100] and further studied in [89], and relate to the following
classical connections with skew-symmetric torsion.
Definition 3.12. Let (M n , g, H) be a Riemannian manifold and H ∈ Λ3 T ∗ .
The Bismut connections associated to (g, H) are defined via
(3.12) ∇± 1
X Y, Z = h∇X Y, Zi ± 2 H(X, Y, Z),
where ∇ denotes the Levi-Civita connection of g. These are the unique metric-
compatible connections with torsion ±H.
Our next goal is to show that the canonical mixed type operators in Lemma 3.11
are determined by the classical Bismut connections associated to the generalized
metric via Proposition 2.40, first observed in [100, Theorem 2]. Thus, the Bismut
connection plays a fundamental role in generalized Riemannian geometry. To see
this, we will work with the explicit presentation of the Courant algebroid E given
by G. Recall that G determines an isometry F : T ⊕ T ∗ → E (Proposition 2.40),
and for any X ∈ T we have isomorphisms σ± : T → V± given by (see Definition
2.28)
F (X ± gX) = σ± X.
We need the following basic identity.
Lemma 3.13. Given a generalized metric G on E, consider the associated closed
three-form H ∈ Λ3 T ∗ and isometry F : T ⊕ T ∗ → E in Proposition 2.40. Then,
one has the equality
(3.13) F −1 [σ− X, σ+ Y ] = [X, Y ] + iY iX H + LX iY g + LY iX g − diX iY g.
Proof. We directly compute, using (2.17) and the definition of the twisted
Courant bracket,
F −1 [σ− X, σ+ Y ] = [X − gX, Y + gY ]H
= [X, Y ] + [X, iY g] − [iX g, Y ] + iY iX H
= [X, Y ] + LX iY g − 21 diX iY g + LY iX g − 12 diY iX g + iY iX H
= [X, Y ] + (LX iY g + LY iX g − diX iY g) + iY iX H.

Proposition 3.14. Given a generalized metric G on E, consider the associated
closed three-form H ∈ Λ3 T ∗ with associated Bismut connections ∇± . Then one has
the equality
∇±
X Y = π ◦ [σ∓ X, σ± Y ]± .
Proof. Let N (X, Y ) = π ◦ [σ− X, σ+ Y ]+ . We first compute

N (f X, Y ) = π ◦ π+ [f σ− X, σ+ Y ]
= π ◦ π+ (f [σ− X, σ+ Y ] + Y f σ− X − hσ− X, σ+ Y i Df )
= f π ◦ π+ [σ− X, σ+ Y ]
= f N (X, Y ).
Next we obtain
N (X, f Y ) = π ◦ π+ [σ− X, f σ+ Y ]
= π ◦ π+ (f [σ− X, σ+ Y ] + ((πσ− X)f )σ+ Y − hσ− X, σ+ Y i Df )
= f π ◦ π+ [σ− X, σ+ Y ] + (Xf )π ◦ π+ σ+ Y
= f N (X, Y ) + (Xf )Y.
This implies that N does indeed define a linear connection on T . We next show
that this connection is metric compatible. First observe using Proposition 2.40 that
(3.14) g(N (X, Y ), Z) = hσ+ N (X, Y ), σ+ Zi = hπ+ [σ− X, σ+ Y ], σ+ Zi .
Then we can compute, using Proposition 2.40 and also Property (5) of Definition
2.8,
hπ+ [σ− X, σ+ Y ], σ+ Zi + hπ+ [σ− X, σ+ Z], σ+ Y i
= h[σ− X, σ+ Y ], σ+ Zi + h[σ− X, σ+ Z], σ+ Y i
= π(σ− X) hσ+ Y, σ+ Zi − hD hσ− X, σ+ Zi , σ+ Y i
− hσ+ Z, D hσ− X, σ+ Y ii
= X hσ+ Y, σ+ Zi
= Xg(Y, Z),
which yields the claim comparing against (3.14). It remains to compute the torsion
tensor. First, arguing as in (3.14) we have
T (X, Y, Z) = g(N (X, Y ) − N (Y, X) − [X, Y ], Z)
(3.15) = hπ+ ([σ− X, σ+ Y ] − [σ− Y, σ+ X]) , σ+ Zi − g([X, Y ], Z)
= h[σ− X, σ+ Y ] − [σ− Y, σ+ X], σ+ Zi − g([X, Y ], Z).
Now, using Lemma 3.13, and observing that the final three terms in (3.13) are
symmetric, we obtain
h[σ− X, σ+ Y ] − [σ− Y, σ+ X], σ+ Zi = 2 h[X, Y ] + iY iX H, σ+ Zi
(3.16) = 2 σ− [X, Y ] + i[X,Y ] g + iY iX H, σ+ Z
= g([X, Y ], Z) + H(X, Y, Z).
Using (3.16) in (3.15) yields the claim T = H, and so the result follows.
To proceed further, we must analyze the space of torsion-free G-compatible

generalized connections, which we denote
D0 (G) = D(G) ∩ D0 .
Proposition 3.15. Given a generalized metric G on a Courant algebroid E,
there exists a torsion-free generalized connection on E compatible with G.
Proof. We construct first a G-compatible generalized connection D on E
±
which has torsion of pure-type. For this, we define the mixed-type operators D∓ by
(3.10). To define the pure-type operators, we choose arbitrary metric connections
e + on V+ and ∇
∇ e − on V− , and set
e±
Da± b± := ∇π(a± ) b± .
Finally, we define our torsion-free generalized connections D0 by explicitly removing

the torsion, that is
1
D 0 = D − TD ,
3
where we use the metric h·, ·i to regard
TD ∈ Γ(V+∗ ⊗ o(V+ )) ⊕ Γ(V−∗ ⊗ o(V− )).
Note that the pure-type condition on TD implies that D0 is still G-compatible.
Given a generalized metric, the torsion-free condition does not determine a

compatible generalized connection uniquely. Unlike in Riemannian geometry, the
generalized Levi-Civita connections form an affine space, modeled on the pure-type
mixed symmetric 3-tensors
Σ+ ⊕ Σ− ,
where
Σ± = Γ(V±⊗3 ) ∩ Σ,
and Σ is as in (3.6). There are canonical splittings
(3.17) Σ± = Σ±
0 ⊕ Γ(V± ),
where the first summand corresponds to ‘trace-free’ elements, in analogy with (3.7).
To see this, fix {e+ + ∗
i } an orthonormal local frame for V+ . Given χ ∈ Γ(V+ ⊗o(V+ )),
set
n
1 X
e+ := χ + e+ .
n − 1 i=1 ei i
Then, on V+ the splitting (3.17) corresponds to the following direct sum decompo-
sition
e
(3.18) χ+ = χ+ +
0 + χ+ ,
where χ+
0 is such that
n
X
(χ+ +
0 )e+ ei = 0,
i
i=1
and
e
(3.19) (χ++ )a+ b+ = ha+ , b+ ie+ − he+ , b+ ia+ .
Example 3.16. Using the above discussion we can produce finally an explicit
example of a torsion-free generalized connection compatible with a generalized met-
ric. Let G be a generalized metric on an exact Courant algebroid E over a manifold
M of dimension n. The generalized metric induces an isomorphism E ∼ = T ⊕ T ∗,
∗
where the twisted bracket on T ⊕ T is given by a closed 3-form H in the Ševera
class of E in H 3 (M, R). Via this isomorphism, the generalized metric takes the
simple form:
V± = {X ± g(X) : X ∈ T },
where g is the Riemannian metric induced by the isomorphism π|V+ : V+ → T .
Using Lemma 3.11 and Proposition 3.14, we see that the initial step of the
construction is forced upon us, where we must use the Bismut connections ∇± to
define the mixed type operators. As in the proof of Proposition 3.15, we must
choose metric compatible connections to define the pure-type operators, and it is
natural to still use the Bismut connections. Doing so we obtain the Gualtieri-Bismut
connection DB , introduced in [89], with pure-type operators
πDσB+ X σ+ Y = ∇+
X Y,
πDσB− X σ− Y = ∇−
X Y.
By construction this connection is compatible with G, and has torsion TDB =

∗ ∗
π|V +
H + π|V−
H. Following the proof of Proposition 3.15, we construct a torsion-
free generalized connection D0 ∈ D0 (V+ ) from DB by
1
D 0 = D B − TD B .
3
A computation shows that this has pure-type operators
+1/3
πDσ0 + X σ+ Y = ∇X Y,
−1/3
πDσ0 − X σ− Y = ∇X Y,
where ∇±1/3 denotes the metric connection with skew-symmetric torsion (cf. (3.12))
D E
±1/3
∇X Y, Z = h∇X Y, Zi ± 16 H(X, Y, Z).
To finish this section, we introduce the space of torsion-free metric connections

compatible with a fixed divergence operator div, that is,
(3.20) D0 (G, div) = {D ∈ D(G) | TD = 0, divD = div}.
This space of connections plays an important role in the definition of the generalized
Ricci tensor. By Lemma 3.9, D0 (G, div) is an affine space modeled on
Σ+ −
0 ⊕ Σ0 .
In the next result we show that D0 (G, div) is non-empty. Recall from Definition
2.46 that a generalized metric has an associated Riemannian divergence divG .
Lemma 3.17. Let (G, div) be a pair given a generalized metric G and a di-
vergence operator div on E. Denote divG − div = he+ , ·i + he− , ·i, for e± ∈ V± .
Then,
(1) D0 ∈ D0 (G, divG ), where D0 is defined as in Example 3.16.
(2) D ∈ D0 (G, div), where

1 e+ e

(3.21) D = D0 + χ+ + χ−− ,
n−1
e
and χ±± are defined as in (3.19).
Proof. To prove (1), consider the Levi-Civita connection ∇ of the Riemannian

metric g induced by G and the generalized connection D′ = ∇ ⊕ ∇∗ as in Example
3.3. Then, a simple calculation shows that
L X µg
divD′ (e) = tr ∇X = .
µg
Applying now [69, Proposition 3.6] it follows that D′ and D0 differ by a totally
skew-symmetric element in Γ(E ⊗ o(E)) and therefore divD′ = divD0 .
As for (2), we choose an orthogonal frame {ei } for E with dual frame {e
ei } and
calculate
X
divD (e′ ) = ei , Dei i
he
i
1 X e 1 X e
= divD0 (e′ ) + he ei , (χ++ )ei e′ i + ei , (χ−− )ei e′ i
he
n−1 i n−1 i
1 X
= divG (e′ ) + ei , e+ ihπ+ ei , e′ i − he+ , e′ ihe
he ei , π+ ei i
n−1 i
1 X
+ ei , e− ihπ− ei , e′ i − he− , e′ ihe
he ei , π− ei i
n−1 i
= divG (e′ ) − he+ , e′ i − he− , e′ i
= div(e′ ).
The proof follows from the decomposition (3.17), which shows that D is torsion-free
and metric compatible.
3.3. The classical Bismut connection

Proposition 3.14 shows that the Bismut connection, as presented in Definition
3.12, plays a fundamental role in generalized Riemannian geometry. In this sec-
tion we pause for a moment in the development of the general theory, in order to
present some basic features of this classical object. It is interesting to observe that
Bismut naturally came across this connection in the context of index theory prob-
lems in complex non-Kähler geometry. We note here that historically this family
of connections was considered by Yano [185] in his study of manifolds with an
almost product structure. Remarkably, Cartan-Schouten [37, 36] also encountered
these connections in their study of Riemannian manifolds admitting an “absolute
parallelism”, i.e. a flat metric compatible connection. In particular they showed
that, beyond classical flat Riemannian metrics, the only examples are semisimple
Lie groups with bi-invariant metric and torsion determined canonically by the Lie
algebra structure, as well as an exceptional example on S 7 . In the Lie group case
the torsion is skew-symmetric and closed, thus fitting naturally into our setting,
whereas the exceptional S 7 example has nonclosed torsion. We will go back to this
interesting class of examples in §3.7.
3.3. THE CLASSICAL BISMUT CONNECTION 43
Our first result gives a formula for the curvatures of the Bismut connection
in terms of the Riemann curvature tensor. Throughout this section ∇ will denote
the Levi-Civita connection of a fixed Riemannian metric g. Given H ∈ Λ3 T ∗ , we
will denote by ∇+ the associated Bismut connection as in Definition 3.12. We will
denote by Rm+ , Rc+ , and R+ , the curvature tensor, the Ricci tensor, and the scalar
curvature of ∇+ , respectively. Similarly, we will use the notation Rm, Rc, and R
for the curvatures associated to the metric via ∇. We denote by d∗ the adjoint of
the exterior differential acting on differential forms with respect to the metric g.
Proposition 3.18. Let (M n , g, H) be a Riemannian manifold with H ∈ Λ3 T ∗ ,
dH = 0. Then
Rm+ (X, Y, Z, W ) = Rm(X, Y, Z, W )
+ 21 ∇X H(Y, Z, W ) − 12 ∇Y H(X, Z, W )
1 1
(3.22) − 4 hH(X, W ), H(Y, Z)i + 4 hH(Y, W ), H(X, Z)i ,
+
Rc = Rc − 41 H 2 − 21 d∗ H,
2
R+ = R − 14 |H| ,
where
X
(3.23) H 2 (X, Y ) := hXyH, Y yHi = H(X, ei , ej )H(Y, ei , ej ).
i,j
Proof. It suffices to establish the first formula for local normal coordinate
vectors, i.e. ∇X Y (p) = 0 and [X, Y ] = 0 etc. Using the definition of the Bismut
connection we thus directly compute
Rm+ (X, Y, Z, W )
D E
= ∇+ ∇
X Y
+
Z − ∇+ +
∇
Y X Z − ∇+
[X,Y ] Z, W
= ∇X (∇+ 1 +
Y Z), W + 2 H(X, ∇Y Z, W )
− ∇Y (∇+ 1 +
X Z), W − 2 H(Y, ∇X Z, W )
= ∇X (∇Y Z + 21 g −1 H(Y, Z, ·)), W + 12 H(X, ∇Y Z + 21 g −1 H(Y, Z, ·), W )
− ∇Y (∇X Z + 21 g −1 H(X, Z, ·))W − 21 H(Y, ∇X Z + 12 g −1 H(X, Z, ·), W )
= Rm(X, Y, Z, W ) + 21 ∇X H(Y, Z, W ) − 12 ∇Y H(X, Z, W )
1 1
− 4 hH(X, W ), H(Y, Z)i + 4 hH(Y, W ), H(X, Z)i ,
as required. Taking the trace of this with respect to the first and fourth indices,
we obtain
Rc+ (Y, Z) = Rm+ (ei , Y, Z, ei )
= Rm(ei , Y, Z, ei ) + 21 ∇ei H(Y, Z, ei ) − 12 ∇Y H(ei , Z, ei )
1 1
− 4 hH(ei , ei ), H(Y, Z)i + 4 hH(Y, ei ), H(ei , Z)i
= Rc(Y, Z) − 14 H 2 (Y, Z) − 1 ∗
2 d H(Y, Z),
∗
as required. A final trace, using that the trace of d H is zero, yields the formula
for the scalar curvature.
The classical Bianchi identities in Riemannian geometry reflect the diffeomor-
phism invariance of the curvature tensor. More general forms of these identities
hold for connections with torsion as well (cf. [114, Theorem 5.3]), and play a fun-
damental role in the study of the generalized Ricci flow. We begin with an identity
on the divergence operator acting on the tensor H 2 , and then prove the Bianchi
identity for connections with torsion.
Lemma 3.19. Given (M n , g) a Riemannian manifold and H ∈ Λ3 T ∗ , dH = 0,

one has

div H 2 i = 61 ∇i |H|2 − (d∗ H)mn Himn .
Proof. We choose local normal coordinates for g at some point and compute,
suppressing some appearances of the metric tensor and using that H is closed
2
div H 2 j = ∇i Hij
= ∇i (Himn Hjmn )
= − (d∗ H)mn Hjmn + Himn ∇i Hjmn
= − (d∗ H)mn Hjmn + Himn (∇j Himn − ∇m Hijn + ∇n Hijm )
2
= − (d∗ H)mn Hjmn + 21 ∇j |H| − 2Himn ∇i Hjmn ,
where in the last line we rearranged indices. Comparing the third and fifth lines
2
above yields that Himn ∇i Hjmn = 61 ∇j |H| , and so the lemma follows.
Proposition 3.20. Given (M n , g) a Riemannian manifold and H ∈ Λ3 T ∗ ,

dH = 0, one has
X X
(3.24) R+ (X, Y )Z = H(H(X, Y ), Z) + (∇+
X H)(Y, Z) ,
σ(X,Y,Z) σ(X,Y,Z)
and
X
(3.25) (∇+ +
X R)(Y, Z) + R (H(X, Y ), Z) = 0,
σ(X,Y,Z)
and
X
0= (∇+
X H)(Y, Z, U ) + 2g(H(X, Y ), H(Z, U ))
(3.26) σ(X,Y,Z)
− (∇+
U H)(X, Y, Z),
and
X
(3.27) R+ (X, Y, Z, U ) − R+ (Z, U, X, Y ) = − 1
2 ∇+
U H(X, Y, Z).
σ(X,Y,Z,U)
Proof. The first two formulas are exercises. The third formula follows from
X
dH(X, Y, Z, U ) = (∇+ +
X H)(Y, Z, U ) + 2g(H(X, Y ), H(Z, U )) − (∇U H)(X, Y, Z),
σ(X,Y,Z)
3.3. THE CLASSICAL BISMUT CONNECTION 45
and dH = 0. To show the fourth we first rewrite (3.24) using (3.26)

X
0= R+ (X, Y, Z, U ) − g(H(H(X, Y ), Z), U ) − ∇+X H(Y, Z, U )
σ(X,Y,Z)
X +
= R (X, Y, Z, U ) − g(H(X, Y ), H(Z, U )) − ∇+
X H(Y, Z, U )
σ(X,Y,Z)
X +
= R (X, Y, Z, U ) + g(H(X, Y ), H(Z, U )) − ∇+
U H(X, Y, Z)
σ(X,Y,Z)
Summing this over cyclic permutation of the entries yields

0 = R+ (X, Y, Z, U ) + R+ (Z, X, Y, U ) + R+ (Y, Z, X, U ) − ∇+
U H(X, Y, Z)
+ R+ (U, X, Y, Z) + R+ (Y, U, X, Z) + R+ (X, Y, U, Z) − ∇+
Z H(U, X, Y )
+ R+ (Z, U, X, Y ) + R+ (X, Z, U, Y ) + R+ (U, X, Z, Y ) − ∇+
Y H(Z, U, X)
+ R+ (Y, Z, U, X) + R+ (U, Y, Z, X) + R+ (Z, U, Y, X) − ∇+
X H(Y, Z, U )
X X
+ g(H(X, Y ), H(Z, U )) + g(H(U, X), H(Y, Z))
σ(X,Y,Z) σ(U,X,Y )
X X
+ g(H(Z, U ), H(X, Y )) + g(H(Y, Z), H(U, X)).
σ(Z,U,X) σ(Y,Z,U)
One observes that the first and third columns of curvature terms collectively cancel,
as do the terms of type H 2 , yielding
0 = 2R+ (X, Z, U, Y ) − 2R+ (U, Y, X, Z)
− ∇+ + + +
U H(X, Y, Z) − ∇Z H(U, X, Y ) − ∇Y H(Z, U, X) − ∇X H(Y, Z, U ).
The result follows.
Our next result (cf. [18] Theorem 1.6) gives an analogue of the first Bianchi
identity for Levi-Civita connection, which relates the curvature tensors of the Bis-
mut connections ∇+ and ∇− .
Proposition 3.21. Let (M n , g) be a Riemannian manifold and fix H ∈ Λ3 T ∗ ,
dH = 0. Then
R+ (X, Y, Z, W ) = R− (Z, W, X, Y ).
Proof. Using the formula for the Bismut curvature in Proposition 3.18 sepa-
rately for ∇± and the symmetry of the Riemann curvature tensor we see
R+ (X, Y, Z, W ) − R− (Z, W, X, Y )
1
= 2 (∇X H(Y, Z, W ) − ∇Y H(X, Z, W ) + ∇Z H(W, X, Y ) − ∇W H(Z, X, Y ))
1
+ 4 (− hH(X, W ), H(Y, Z)i + hH(Y, W ), H(X, Z)i
+ hH(Z, Y ), H(W, X)i − hH(W, Y ), H(Z, X)i)
1
= 2 dH(X, Y, Z, W )
= 0,
as required.
3.4. Curvature and the First Bianchi Identity

We turn next to the study of curvature quantities in generalized geometry. We
fix an exact Courant algebroid E over a smooth manifold M .
Definition 3.22. Let G be a generalized metric on E, and D a metric compati-
ble generalized connection. Associated to D there are the two generalized curvature
operators
R± ∗ ∗
D ∈ Γ(V± ⊗ V∓ ⊗ o(V± )),
defined by
R+
D (a+ , b− )c+ = Da+ Db− c+ − Db− Da+ c+ − D[a+ ,b− ] c+ ,
R−
D (a− , b+ )c− = Da− Db+ c− − Db+ Da− c− − D[a− ,b+ ] c− .
In general, the curvatures of a torsion-free generalized connection are not an

invariant of the generalized metric G, but rather depend on D via the choice of
pure-type operators. We illustrate this with the following example.
Example 3.23. Consider the torsion-free metric compatible connection D0
constructed in Example 3.16. We choose a one-form ϕ on M and use this to deform
D0 to a different torsion-free metric connection, via deformation of the associated
divergence operator (see Definition 3.8). We set ϕ± = π± ϕ using the orthogonal
projections π± : E → V± and define
1 ϕ+ ϕ

D = D0 + χ+ + χ−− ,
n−1
ϕ±
where χ± are defined as in (3.19). By Lemma 3.17, D is torsion-free and metric
compatible, and the associated divergence is
L X µg
divD (e) = divD0 (e) − hϕ, ei = − 12 ϕ(X)
µg
for e = X + ξ ∈ T ⊕ T ∗ . Being torsion-free, by Lemma 3.11 the generalized
connection D is completely determined by the corresponding pure-type operators,
given in this case by
+1/3 1
πDσ+ X σ+ Y = ∇X Y + (g(X, Y )g −1 ϕ − ϕ(Y )X),
2(n − 1)
(3.28)
−1/3 1
πDσ+ X σ− Y = ∇X Y + (g(X, Y )g −1 ϕ − ϕ(Y )X).
2(n − 1)
We are now ready to calculate the curvature of D following [69] (e.g. for R+D ).
First, we have
1 1
1/3
Dσ+ X Dσ− Y σ+ Z = σ+ ∇X ∇+ YZ + g(X, ∇+
Y Z)g
−1
ϕ− ϕ(∇+Y Z)X ,
2(n − 1) 2(n − 1)
1/3 1 1
Dσ− Y Dσ+ X σ+ Z = σ+ ∇+
Y ∇X Z − ϕ(Z)∇+ YX + g(X, Z)g −1 ∇+
Yϕ
2(n − 1) 2(n − 1)
!
1 1
− iY d(ϕ(Z))X + iY d(g(X, Z))g −1 ϕ .
2(n − 1) 2(n − 1)
Using the equality
[σ+ X, σ− Y ] = [X, Y ] − g(∇X Y + ∇Y X, ·) + H(X, Y, ·)
3.4. CURVATURE AND THE FIRST BIANCHI IDENTITY 47
we also obtain
1
D[σ+ X,σ− Y ] σ+ Z = σ+ ∇[X,Y ] Z + g −1 H(2[X, Y ] + ∇X Y + ∇Y X, Z, ·)
6
1
− g −1 H(g −1 H(X, Y, ·), Z, ·)
6
1
− ϕ(Z)([X, Y ] − ∇gY X − ∇X Y + g −1 H(X, Y, ·))
4(n − 1)
!
1
+ (g(Z, [X, Y ] − ∇Y X − ∇X Y ) + H(X, Y, Z))ϕ .
4(n − 1)
Finally, the identity
(∇X H)(Y, Z, ·) = ∇X (H(Y, Z, ·)) − H(∇X Y, Z, ·) − H(Y, ∇Z, ·)
leads us to the following expression for the curvature
+
πRD (σ+ X, σ− Y )σ+ Z
(3.29) 1
e Y )Z − (g(X, Z)g −1 ∇+ +
= R(X, Y ϕ − iZ (∇Y ϕ)X)
2(n − 1)
where
e 1 1
R(X, Y )Z = R(X, Y )Z + g −1 (∇X H)(Y, Z, ·) − (∇Y H)(X, Z, ·)
2 6
1 1
+ H(X, g −1 H(Y, Z, ·), ·) − H(Y, g −1 H(X, Z, ·), ·)
12 ! 12
1
− H(Z, g −1 H(X, Y, ·), ·) .
6
The tensor Re is a hybrid of the second covariant derivatives for ∇+ , ∇1/3 , and ∇− ,
and can be identified with the generalized curvature of D0 . More invariantly, it can
be written as
e 1/3 1/3 1/3
R(X, Y )Z = ∇ ∇+ Z − ∇+ ∇ Z + ∇ + Z − ∇+ − Z.
X Y Y X ∇Y X ∇X Y
From the previous formulae, it is transparent that the generalized curvature oper-
ators R±
D depend on the choice of torsion-free compatible connection D.
Our next goal is to define curvature quantities which only depend on the gen-
eralized metric G. Before we address this question in §3.5, we prove an algebraic
property of the curvatures of a torsion-free generalized connection, that we will use
later in Lemma 3.44.
Proposition 3.24 (First Bianchi identity). Let D be a torsion-free generalized
connection compatible with V+ . Then, for any e± , a± , b± , c± ∈ Γ(V± ) we have
X
(3.30) hR±
D (a± , e∓ )b± , c± i = 0.
σ(a,b,c)
Proof. We argue for R+

D , as the other case is symmetric. We decompose
X
hR+D (a+ , e− )b+ , c+ i = I + J + K,
σ(a,b,c)
where
X
I= hDa+ ([e− , b+ ]+ ), c+ i,
σ(a,b,c)
X
J =− h[e− , Da+ b+ ], c+ i,
σ(a,b,c)
X
K =− hD[a+ ,e− ] b+ , c+ i.
σ(a,b,c)
Using that TD = 0 and formula (3.3) we have

X
K= −hDb+ [a+ , e− ], c+ i − h[[a+ , e− ], b+ ], c+ i + hDc+ [a+ , e− ], b+ i
σ(a,b,c)
X
= h[a+ , e− ], Db+ c+ − Dc+ b+ i − h[[a+ , e− ], b+ ], c+ i
σ(a,b,c)
− π(b+ )(h[a+ , e− ], c+ i) + π(c+ )(h[a+ , e− ], b+ i),

where in the last equality we used the compatibility between D and h·, ·i. Using
again this last property, we also have
X
I= π(a+ )(h[e− , b+ ], c+ i) − h[e− , b+ ], Da+ c+ i ,
σ(a,b,c)
X
J= −π(e− )(h[a+ , b+ ], c+ i) + h[e− , c+ ], Da+ b+ i ,
σ(a,b,c)
where in the last equality we used property (C4) of the bracket (see Definition 2.8)).
From this, using property (C5) of the bracket, which implies [e− , a+ ] = −[a+ , e− ],
we obtain
I + J + K = −h[[a+ , e− ], b+ ], c+ i + h[[b+ , e− ], a+ ], c+ i − h[[c+ , e− ], a+ ], b+ i
+ π(b+ )(h[c+ , e− ], a+ i) − π(a+ )(h[c+ , e− ], b+ i) − π(c+ )(h[b+ , e− ], a+ i)
− π(e− )(TD (a+ , b+ , c+ )) − π(e− )(h[a+ , b+ ], c+ i).
Finally, using again (C5) we have an equality
π(c+ )(h[e− , b+ ], a+ i) = hc+ , [a+ , [e− , b+ ]] + [[e− , b+ ], a+ ]i,
and from TD = 0 we conclude
I + J + K = h[[e− , a+ ], b+ ] + [a+ , [e− , b+ ]] − [e− , [a+ , b+ ]], c+ i,
which vanishes by the Jacobi identity (property (C1)).
The algebraic Bianchi identity (3.30) for torsion-free generalized connections is
a remarkable property, as their construction typically involves standard connections
with skew-torsion in the tangent bundle of M . Recall from Proposition 3.20 that a
torsionful connection on a manifold satisfies complicated Bianchi identities which
involve first-order derivatives of its torsion, as well as quadratic terms in H. To
illustrate this, we compare now the generalized curvature operator in Example 3.23
with the curvature of the torsionful generalized connection DB in Example 3.16.
Proposition 3.25. Given G a generalized metric on E, let DB denote the
associated Gualtieri-Bismut connection in Example 3.16. One has
πR± B (σ± X, σ∓ Y )σ± Z ∼
D = Rm± (X, Y )Z,
3.5. GENERALIZED RICCI CURVATURE 49
where Rm± denotes the curvature of the standard Bismut connection in (3.22).
Proof. The proof follows as an exercise from Example 3.16.
3.5. Generalized Ricci curvature

Different approaches to the definition of the generalized Ricci tensor and scalar
curvature can be found in [23, 53, 70, 86, 109, 151, 152] (see also references
therein). In [53, 69, 110, 152] the authors gave a treatment of generalized con-
nections and curvature and their relationship to supergravity theories. To proceed
with our approach to the Ricci tensor in generalized geometry following [70], we
introduce next the definition of Ricci tensor for a metric-compatible generalized
connection.
Definition 3.26. Let G be a generalized metric on E. The Ricci tensors

Rc± ∗ ∗
D ∈ Γ(V∓ ⊗ V± ) of a generalized metric-compatible connection D ∈ D(G) are
defined by
Rc± ±
D (a∓ , b± ) = tr(e± → RD (e± , a∓ )b± ),
for e± ∈ Γ(V± ).
Our next result investigates the variation of the Ricci tensors when we de-
form a torsion-free generalized connection in D(G), while preserving the pure-type
condition of the torsion and the divergence (see Definition 2.44).
Proposition 3.27. Let (G, div) be a pair given by a generalized metric and a
divergence operator on E. Let D, D′ ∈ D(G) with divergence divD = div = divD′ .
If TD = 0 and TD′ is of pure-type then Rc± ± ± ±
D = RcD′ . In particular, RcD = RcD′
for any pair D, D′ ∈ D0 (G, div) (see (3.20)).
Proof. Without loss of generality we will argue for Rc+ −

D , since the case RD
′
is analogous. Setting χ = D − D, the condition divD = divD′ combined with TD′
being of pure-type imply that
χ = χ+ + χ− ,
where (see (3.17))
χ± = χ± ± ± 3
0 + σ ∈ Σ0 ⊕ Λ (V± ).
Let {e+ n
i }i=1 an orthonormal frame for V+ . Given e± ∈ Γ(V± ) we calculate
(3.31)
n
X
Rc+
D′ (a− , b+ ) = he+ + +
i , RD′ (ei , a− )b+ i
i=1
n
X
= Rc+
D (a− , b+ ) + he+ + + +
i , χe+ (Da− b+ ) − Da− (χe+ b+ ) − χ[e+ ,a b+ i
i i i − ]+
i=1
Xn
= Rc+
D (a− , b+ ) + − hχ+ e+ , Da− b+ i + hχ+
e+ i
b , D a− e +
e+ + i i
i i
i=1

+ ιπ(a− ) dhχ+ e+ , b+ i − hχ+
e+ i [a + e+
i , b + i
i − ,ei ]+
n
X
= Rc+
D (a− , b+ ) + hχ+ b , [a− , e+
e+ +
+
i ]i − hχ[a + e+
i , b+ i ,
i − ,ei ]+
i=1
Xn
= Rc+
D (a− , b+ ) + h(χ+ )
0 e
+
i
+
+ b+ , [a− , e ]+ i − h(χ ) + e
0 [a− ,e ]+ i
+
, b + i
i i
i=1
Pn
where have used i=1 χ+e+
e+ ± 3
i = 0, Lemma 3.11, and that σ ∈ Λ V± . To conclude,
i
given x ∈ M and a− ∈ V−|x , using the previous formula we shall prove that
Rc+ +
D′ (a− , b+ ) = RcD (a− , b+ ),
and hence the statement will follow since Rc+ +

D′ and RcD are bilinear. For this, we
construct a special orthonormal frame in order to evaluate formula (3.31). Choose
an orthonormal frame {e+ +
i } around x for V+ which satisfies Da− ei = 0 at x.
This frame can constructed by smooth extension of the parallel transport for the
+
connection D− along a curve starting at x with initial velocity π(a− ). The proof
now follows from the previous formula, using that [a− , e+ +
i ]+ = Da− ei . The last
part of the statement follows from Lemma 3.9.
Definition 3.28. Let (G, div) be a generalized metric G and a divergence

operator div on the exact Courant algebroid E. The Ricci tensors
Rc± ∈ Γ(V∓∗ ⊗ V±∗ )
of the pair (G, div) are defined by Rc± = Rc± D for any choice of generalized con-
nection D ∈ D(G, div). This is well-defined by Proposition 3.27.
In the next result we calculate the variation of the Ricci tensors Rc± when
we fix the generalized metric G and vary the divergence operator div. Recall from
Definition 2.46 that a generalized metric has an associated Riemannian divergence
divG .
Lemma 3.29. Let (G, div) be a pair as in Definition 3.28. Denote divG − div =
he, ·i, for e ∈ Γ(E). Then, setting e± = π± e, we have
Rc+ (G, div)(a− , b+ ) = Rc+ (G, divG )(a− , b+ ) − h[e+ , a− ], b+ i,

Rc− (G, div)(a+ , b− ) = Rc− (G, divG )(a+ , b− ) − h[e− , a+ ], b− i.
3.5. GENERALIZED RICCI CURVATURE 51
Proof. We argue for Rc+ , as the other case is symmetric. Let D0 ∈ D(G, divG )
and D ∈ D(G, div) as in Lemma 3.17. Let {e+ n
i }i=1 be an orthonormal frame for
V+ . Then,
Rc+ (G, div)(a− , b+ )
= Rc+
D (a− , b+ )
= Rc+
D0 (a− , b+ )
n
1 X + e e e
+ he , (χ++ )e+ (Da0− b+ ) − Da0− ((χ++ )e+ b+ ) − (χ++ )[e+ ,a− ]+ b+ i
n − 1 i=1 i i i i
1 X
n
= Rc+ (G, divG )(a− , b+ ) + − hχ+
e+
e+
i , [a− , b+ ]+ i
n − 1 i=1 i

e e+ e+
+ h(χ++ )e+ b+ , [a− , e+
i ] + i + ιπ(a − ) dh(χ + )e
+
+ e , b+ i + h(χ
i + ) + e
[a− ,e ]+ i
+
, b + i
i i i
+ G
= Rc (G, div )(a− , b+ ) − he+ , [a− , b+ ]i + ιπ(a− ) dhe+ , b+ i + C
= Rc+ (G, divG )(a− , b+ ) + h[a− , e+ ], b+ i + C,
where
1 X e+
n
e
C= h(χ+ )e+ b+ , [a− , e+
i ]+ i + h(χ++ )[a− ,e+ ]+ e+
i , b+ i .
n − 1 i=1 i i
Arguing as in the proof of Proposition 3.27, we can now choose a point x ∈ M and
{e+ +
i } which satisfies [a− , ei ]+ = 0 at x. Thus, C = 0 and the statement follows
from axiom (5) in Definition 2.8.
Our next result provides an explicit formula for the Ricci tensors Rc± associated
to a pair (G, div). We follow the notation in Definition 2.46 and Lemma 3.29.
Proposition 3.30. Let (G, div) be a pair as in Definition 3.28. Denote divG − div =
he, ·i, and set ϕ± = g(πe± , ·) ∈ T ∗ . Then, one has
1 1
Rc+ (σ− X, σ+ Y ) = Rc − H 2 − d∗ H + ∇+ ϕ+ (X, Y ),
(3.32) 4 2
1 2 1 ∗
Rc (σ+ X, σ− Y ) = Rc − H + d H − ∇− ϕ− (X, Y ).
−
4 2
Proof. First, observe that Rc± (G, divG ) = Rc± DB by Proposition 3.27. Now,
by Proposition 3.25 we have that Rc± DB
can be identified with the Ricci tensor for
the classical Bismut connection in (3.22). More precisely, we have
1 2 1 ∗
Rc±
D B (σ∓ X, σ± Y ) = Rc − H ∓ d H (X, Y ).
4 2
Applying now Lemma 3.29, we obtain
Rc± (σ∓ X, σ± Y ) := Rc± (G, div)(σ∓ X, σ± Y )
1 1
= Rc − H 2 ∓ d∗ H (X, Y ) − h[e± , σ∓ X], σ± Y i.
4 2
The statement follows from
h[e± , σ∓ X], σ± Y i = − hσ± ∇±
X (g
−1
ϕ± ), σ± Y i
= ∓ g(∇±
X (g
−1
ϕ± ), Y ) = ∓(∇± ϕ± )(X, Y ).
We leave as an exercise to recover formula (3.32) by taking traces in (3.29), when

e ∈ T ∗.
The Ricci tensors Rc± of a pair (G, div) can be assembled naturally into an
endomorphism of E, which we will call the generalized Ricci tensor.
Definition 3.31. Given a pair (G, div) as in Definition 3.28, define the gener-
alized Ricci tensor via
Rc(G, div) = Rc+ − Rc− ∈ End(E),
where Rc± = Rc± (G, div) are regarded as maps Rc± : V∓ → V±∗ ∼
= V± using the
induced metric on V± .
Recall the compatibility condition for pairs (G, div) introduced in Definition
2.52, which states that e ∈ Γ(E), defined by he, i = divG − div, is an infinitesimal
isometry of G, that is,
[e, G] = 0.
Our next result proves that the generalized Ricci tensor is an element of so(E) ∼ =
Λ2 E ⊂ End(E) if and only if (G, div) is a compatible pair. This structural property
of the Ricci tensor in generalized geometry is strongly reminiscent of the symmetric
property of the Ricci tensor in Riemannian geometry. In particular, it implies that
the two Ricci tensors Rc± contain the same information and, furthermore, that
we can regard Rc(G, div) as a tangent vector to the space of generalized metrics
(see Lemma 2.31). This is the basic principle which will allow us to define the
generalized Ricci flow in the next chapter.
Lemma 3.32. Let (G, div) be as in Definition 3.28. Then, we have
G ◦ Rc(G, div) = −Rc(G, div) ◦ G.
Furthermore,
Rc(G, div) ∈ so(E)
if and only if (G, div) is a compatible pair in the sense of Definition 2.52.
Proof. The first part of the statement follows directly by Definition 3.31 and
Lemma 2.32. As for the second part, identify V± ∼
= T via the anchor map π. Regard
Rc± ∈ T ∗ ⊗ T ∗ and consider the decomposition Rc± = h± + k ± into symmetric
and skew-symmetric parts. Then, the property Rc(G, div) ∈ so(E) is equivalent to
(3.33) h+ = h− , k + = −k − .
Using formula (3.32) it follows immediately that Rc(G, divG ) ∈ so(E). Finally,
setting Rc := Rc(G, div), Lemma 3.29 implies that
hRc(a), bi+ha, Rc(b)i
= Rc+ (a− , b+ ) + Rc+ (b− , a+ ) − Rc− (a+ , b− ) − Rc− (b+ , a− )
= − h[e+ , a− ], b+ i − h[e+ , b− ], a+ i + h[e− , a+ ], b− i + h[e− , b+ ], a− i
= − h[e+ , a− ], b+ i − h[e+ , b− ], a+ i + h[e, a+ ], b− i + h[e, b+ ], a− i
− h[e+ , a+ ], b− i − h[e+ , b+ ], a− i
= − h[e+ , a− ], b+ i − h[e+ , b− ], a+ i + h[e, a+ ], b− i + h[e, b+ ], a− i
+ ha+ , [e+ , b− ]i + hb+ , [e+ , a− ]i
= h[e, a+ ], b− i + h[e, b+ ], a− i.
3.6. GENERALIZED SCALAR CURVATURE 53
and therefore Rc ∈ so(E) if and only if [e, G] = 0.
Remark 3.33. By Lemma 3.32, the generalized Ricci tensor Rc(G, divG ) as-
sociated naturally to a generalized metric G is an element in so(E). From the
classical point of view, this tensor incorporates the Ricci curvatures of both Bismut
connections (associated to ±H). The more general quantity Rc(G, div) considered
here, with varying divergence operator div, plays an important role in the interplay
between the generalized Ricci flow and pluriclosed flow in Chapter 9, as well as
T-duality in Chapter 10.
Remark 3.34. When the divergence div in Proposition 3.30 is such that e =
4df , for some smooth function f (so that ϕ+ = −ϕ− = 2df ), the pair (G, div)
is automatically compatible (see Proposition 2.53). In this case, the two natural
quantities h and k which arise from the generalized Ricci tensor agree exactly with
the leading order term in α′ expansion of the so-called β-functions for the graviton
and antisymmetric tensor fields in the sigma model approach to type II string theory
[34]. This establishes an identification between the 1-loop renormalization group
flow of this physical theory and the generalized Ricci flow in the next chapter.
3.6. Generalized scalar curvature

In this section we define a notion of scalar curvature in generalized geometry,
following the approach from [53, 70], using Dirac operators and their associated
Lichnerowicz formulae. Throughout this section we will assume that the smooth
manifold M is endowed with a spin structure. This assumption can be removed
either working locally, or considering the approach to the generalized scalar curva-
ture in [152] via the Schrödinger-type operator introduced in [138] (see (6.1)). Let
G be a generalized metric on an exact Courant algebroid E.
Let CL(V+ ) and CL(V− ) denote the bundles of complex Clifford algebras of V+
and V− , respectively, defined by the relation
(3.34) v · v = hv, vi
for v ∈ V± , and h, i given by the restriction of the neutral pairing to V± . Via the
isometries π : V± → (T, ±g), where g is the Riemannian metric associated to G, we
can fix complex spinor bundles S± for the orthogonal bundles V± . We assume that
S+ (resp. S− ) is associated to the fixed spin principal bundle given by the spin
structure on M , for a choice of irreducible representation of the complex Clifford
algebra on n-dimensional Euclidean space (Rn , g0 ) (resp. (Rn , −g0 )) (see [120],
and note the different convention for the relation (3.34) in the Clifford algebra).
In order to define the generalized scalar curvature, we need to further study the
Koszul-type formula in Lemma 3.11. The freedom in the construction of a torsion-
free generalized connection compatible with the generalized metric G corresponds
to the choice of pure-type operators
+ −
D+ : Γ(V+ ) → Γ(V+∗ ⊗ V+ ), D− : Γ(V− ) → Γ(V−∗ ⊗ V− ).
We define next a pair of Dirac-type operators which are independent of these
choices, once we fix a divergence operator div on E (see Lemma 3.9). For any
±
choice of such connection, and via the isometries π : V± → (T, g), the operators D±
can be identified with standard metric connections on (T, g) and induce canonically
spin connections
S S
(3.35) D++ : Γ(S+ ) → Γ(V+∗ ⊗ S+ ), D−− : Γ(S− ) → Γ(V−∗ ⊗ S− ).
Recall that we denote by D0 (G) the space of torsion-free generalized connections
compatible with G, while D0 (G, div) ⊂ D0 (G) denotes the subspace compatible with
a given divergence operator div.
Definition 3.35. With the assumptions above, given a generalized connection
D ∈ D0 (G), define a pair of Dirac-type operators
/ + : Γ(S+ ) → Γ(S+ ),
D / − : Γ(S− ) → Γ(S− ),
D
given explicitly by
n
X
± S
/ α=
D e± ±
i · De± α,
i
i=1
for α ∈ Γ(S± ), and a choice of orthonormal basis {e± n
i }i=1 of V± , i.e. such that
± ±
ei , ej = ±δij .
Lemma 3.36. The Dirac operators D / + and D

/ − are independent of the choice
of generalized connection D ∈ D0 (G, div).
Proof. Let D ∈ D(G, div) and consider a torsion-free G-compatible general-
ized connection D′ = D + χ ∈ D0 (G) with
χ = χ + + χ − ∈ Σ+ ⊕ Σ− .
Using the decomposition in (3.18) there exists e± ∈ Γ(V± ) such that
e
χ± = χ± ±
0 + χ± .
Then, we obtain
divD′ = div −(n − 1)he+ , ·i − (n − 1)he− , ·i
and therefore divD′ = div implies e+ = e− = 0. Define σ ∈ Γ(Λ3 E ∗ ) by
σ(e1 , e2 , e3 ) = hχe1 e2 , e3 i,
and note that the total skew-symmetrization c.p.(σ) vanishes, since D′ and D are
torsion-free. Decomposing σ = σ + + σ − in pure-types, the result follows from
′± ± ±
/
D / − 21 c.p.(σ ± ) = D
=D / .

±
Remark 3.37. In [70], the canonical Dirac operators D / are defined using
∗
twisted spinor bundles (for a choice of root of det S± ). This twist is required to
S
ensure a canonical choice of the spin connections D±± on more general Courant
algebroids, but it is not necessary in our setting.
Consider the metric connections ∇± and ∇±1/3 on (T, ±g) associated to the
generalized metric G, as in Example 3.16. We will abuse notation and let ∇±
and ∇±1/3 also denote the spin connections induced on a given spinor bundle for
±1/3
±g. Furthermore, by ∇ / we mean the cubic Dirac operator, that is, the Dirac
operator associated to ∇±1/3 . We will use the following Lichnerowicz-type formula
for the cubic Dirac operator due to Bismut [18] (see also [1, Theorem 6.2]). Note
the different sign convention for the Clifford algebra relations (3.34).
Lemma 3.38. The spinor Laplacian (∇± )∗ ∇± and the square of the cubic Dirac

±1/3 2
operator ∇/ are related by
2 1 1
/ ±1/3
∇ − (∇± )∗ ∇± = ∓ R ± |H|2 .
4 48
The following lemma provides the structural property for the definition of the
generalized scalar curvatures. Observe that the canonical mixed-type operators
associated to G in Lemma 3.11 induce canonical spin connections as in (3.35)
S S
D−+ : Γ(S+ ) → Γ(V−∗ ⊗ S+ ), D+− : Γ(S− ) → Γ(V+∗ ⊗ S− ).
We follow the notation in Definition 2.46, Lemma 3.29, and Lemma 3.36.
Proposition 3.39. Let (G, div) be a pair given by a generalized metric G and
a divergence operator div on the exact Courant algebroid E. Denote divG − div =
he, ·i, and set ϕ± = g(πe± , ·) ∈ T ∗ . Then, for any spinor η ∈ Γ(S± ) one has
2 ∗
/ S±± η − D∓
D
S± S
D∓± η
!
(3.36)
1 1 ±
= ∓ R − |H|2 ∓ 2d∗ ϕ± − |ϕ± |2 + 2dϕ± ± 4∇g−1 ϕ± η.
4 12
Proof. We derive a formula for

2 ∗
S S+ S
Se+ := D/ ++ − D− D−+ ,
using Lemma 3.38. A similar formula for Se− , defined as S + replacing + by −,

follows from the analogue of Lemma 3.38 for the opposite Clifford algebra relations
(3.34). Via the isommetry π|V+ : V+ → (T, g), we have
2 ∗
Se+ = ∇e
/ − ∇+ ∇+
e is as in (3.28):
acting on sections of the fixed spinor bundle for g, where ∇
e X Y = ∇+1/3 Y + 1
∇ X (g(X, Y )g −1 ϕ+ − ϕ+ (Y )X).
n−1
Let {e1 , . . . , en } be a local orthonormal frame for (T, g). An endomorphism A ∈
End(T ) satisfies
Xn
A= g(Aei , ej )ei ⊗ ej ,
i,j=1
for {e } the dual frame of {ej }. Since ei ⊗ ej − ej ⊗ ei ∈ so(T ) embeds as 21 ej ei in

j
the Clifford bundle Cl(T ), an endomorphism A ∈ so(T ) corresponds to

1X
A= g(Aei , ej )ej ei ∈ Cl(T ).
4 i,j
Then, given a local spinor η, we have

1 X
e = ∇+1/3 η +
∇η ϕ+ (ej )ej ei η ⊗ ei − ϕ+ (ei )ej ei η ⊗ ej
4(n − 1) i,j
and hence,
1 X
e
/ =∇
∇η /
+1/3
η+ ϕ+ (ej )ei ej ei η − ϕ+ (ei )ej ej ei η
4(n − 1) i,j
1 X
/ +1/3 η +
=∇ ϕ+ (ej )(2δij − ej ei )ei η − ϕ+ (ei )ei η
4(n − 1) i,j
+1/3 1
/
=∇ η − ϕ+ η.
2
From this, we calculate
2 1X 1X 1
e +1/3 2
/ η= ∇
∇ / η− ej ∇e+1/3 (ϕ+ η) − ϕ+ ej ∇e+1/3 η + |ϕ+ |2 η
2 j j
2 j j
4
1X 1X
+1/3 2 1
= ∇ / η− ej (∇e+1/3 ϕ+ )η − (ej ϕ+ + ϕ+ ej )∇e+1/3 η + |ϕ+ |2 η
2 j j
2 j j
4
2 1X 1
= ∇ / +1/3 η − (iek ∇e+1/3
+1/3
ϕ+ )ej ek η − ∇g−1 ϕ+ η + |ϕ+ |2 η,
2 j
4
j,k
2 1 1 1
= ∇ / +1/3 η + (d∗ ϕ+ )η − (dϕ+ )η + |ϕ+ |2 η
2 2 4
1 X −1
− H(ej , g ϕ+ , ek )ej ek η − ∇g−1 ϕ+ η
12
j,k
1 X
− H(g −1 ϕ+ , ej , ek )ek ej η
24
j,k
2 1 1 1
= ∇/ +1/3 η + (d∗ ϕ+ )η − (dϕ+ )η + |ϕ+ |2 η − ∇+
g−1 ϕ+ η,
2 2 4
where we used that ej α+ αej = 2α(ej ) for any α ∈ T ∗ , combined with the standard
formulae
X
d∗ α = − iej ∇ej α,
j
∇α(ej , ek ) = 21 dα(ej , ek ) + 12 (iek ∇ej α + iej ∇ek α).

The proof follows now from Lemma 3.38.
Observe that the right hand side of (3.36) defines a differential operator of order
1 acting on spinors. The degree zero part of this operator has a scalar component
and a component of degree two, the latter given by ∓dϕ± . If we insist in defining
a scalar out of (3.36), there are two natural conditions forced upon us. The first
condition is that ∓∇± g−1 ϕ± is a natural operator, depending only on (G, div), so
that it can be absorbed in the left hand side of (3.36). This is achieved provided
that πe = 0, which implies ϕ+ = −ϕ− and (see Lemma 3.11 and Proposition 3.14)
∇± ± S±
g−1 ϕ± = −∇g−1 ϕ∓ ≡ −De∓ .
Given this, the second condition is that e = ϕ ∈ T ∗ satisfies dϕ = 0. As observed

in [70], this pair of conditions relate naturally to the theory of Dirac generating
operators on Courant algebroids introduced in [3, 150].
Definition 3.40. Let (G, div) be a pair given by a generalized metric G and a
divergence operator div on the exact Courant algebroid E. We say that (G, div) is
closed if e ∈ Γ(E), defined by divG − div = he, ·i, satisfies πe = 0 and de = 0.
By Proposition 2.53, it follows immediately that if (G, div) is closed, then it is
a compatible pair in the sense of Definition 2.52. We are ready to introduce our
definition of the scalar curvatures for a closed pair (G, div).
Definition 3.41. Let (G, div) be a closed pair on the exact Courant algebroid
E. The generalized scalar curvature
S = S(G, div) ∈ C ∞ (M )
of the pair (G, div) is defined by
S := S + − S −
where S± ∈ C ∞ (M ) are defined by the following Lichnerowicz type formulae
1 2 ∗
− S + := D / S++ − D− S+ S
D−+ − DeS−+ ,
(3.37) 2
1 2 ∗
S S− S
− S − := D / −− − D+ D+− − DeS+− .
2
This is well-defined by Lemma 3.11, Lemma 3.36, and Proposition 3.39
Remark 3.42. More explicitly, if (G, div) is a closed pair and we denote
divG − div = hϕ, i, for a closed one-form ϕ ∈ Γ(T ∗ ), one has
1 1
(3.38) S = R − |H|2 − d∗ ϕ − |ϕ|2 .
12 4
An important fact that we will use later in §3.7 is that S(G, div) does not coincide
with the trace of the Ricci tensors Rc± in (3.32). Note also that we can regard
(3.37) as a local formula on M , so that there is no obstruction for the existence
of the spinor bundles. Therefore we can define the generalized scalar curvature for
exact Courant algebroids over any smooth manifold.
Remark 3.43. When the divergence div in Proposition 3.30 is such that e =
4df , for some smooth function f (so that ϕ+ = −ϕ− = 2df ), the generalized scalar
curvature is given by
1
(3.39) S = R − |H|2 + 4∆f − 4|∇f |2 ,
12
where ∆f = −d∗ df = trg ∇2 f . Formula (3.39) agrees exactly with the leading
order term in α′ expansion of the so-called β-function for the dilaton field in the
sigma model approach to type II string theory [34] (cf. Remark 3.34).
We finish this section with an alternative point of view on the Ricci tensor, using
the canonical Dirac operators as above (formula (3.40) below corrects a factor of two
in the statement of [70, Lemma 4.7], taking into account the different normalization
for the Dirac operators).
Lemma 3.44. Assume that M is positive-dimensional. Let V+ ⊂ E be a gener-
alized metric and let div be a divergence operator on E. Then the generalized Ricci
tensors in Definition 3.28 can be calculated by
r±
1 ± S± ± X S±
(3.40) ±
± ιa∓ Rc · η = D S
/ D a∓ − D a∓ D
± / − e±
i · D[e± ,a∓ ]∓ η,
2 i=1
i
where η ∈ Γ(S± ) and a∓ ∈ Γ(V∓ ).

Proof. We give a complete proof for Rc+ , as the other case is symmetric.
It is easy to see that the right hand side of (3.40) is tensorial in a∓ ∈ Γ(V∓ ).
To evaluate (3.40), we choose an orthogonal frame {e+ i } for V+ around x ∈ M
which satisfies Da− e+ i = 0 at the point x. This frame can constructed as in the
proof of Proposition 3.27, by smooth extension of the parallel transport for the
+
V− -connection D− along a curve starting at x with initial velocity π(a− ). With the
conventions above we have
r±
X
/ ± , Da− ]η −
[D e±
i · D
S±
± η
[e ,a ]
i ∓ ∓ |x
i=1
Xn
= e+
i · (De+ Da− − Da− De+ − D[e+ ,a− ] )η
i i i
i=1
Xn
= e+ + +
i · RD (ei , a− ) · η
i=1
= 1
2 tr(d
+
→ R+ +
D (d , a− )) · η
X
− 1
2 (c.p.ijk hR+ + + + + + +
D (ei , a− )ej , ek i)ei · ej · ek · η,
i<j<k
where in the first equality we have used that [a− , e+ +

i ]+ = Da− ei by Lemma 3.11,
+
and (Da− ei )|x = 0 by the choice of frame. The proof follows from the algebraic
Bianchi identity in Proposition 3.24.
Remark 3.45. The formula (3.40) is an interpretation of a formula for the
Ricci tensor of a generalized metric in the physics literature [53], claimed without
proof and later proved in [70]. The present proof relies on the Bianchi identity
(3.24) for the generalized curvature of a torsion-free generalized connection.
+
/ and
Since the right hand side of (3.40) is tensorial in e∓ and the operators D
+
D− only depend on (G, div) (see Lemma 3.11 and Lemma 3.36), formula (3.40)
can be taken as an alternative definition of the Ricci tensors, without relying on
Proposition 3.27. For this, we can regard (3.40) as a local formula, and therefore
there is no obstruction for the existence of the spinor bundles.
3.7. Generalized Einstein-Hilbert functional

Having constructed the generalized Ricci tensor and scalar curvature in the
previous sections, it is natural to seek a notion of canonical metric on generalized
geometry which can be expressed in terms of these curvature quantities. Taking
the point of view of the calculus of variations, we will consider canonical metrics
given as critical points of a functional defined on a natural class of compatible
pairs (G, div) (see Definition 2.52). In Riemannian and Lorentzian geometry the
Einstein-Hilbert functional, defined as
Z
EH(g) = Rg dVg ,
M
is a second order diffeomorphism invariant functional, whose critical points in the
Riemannian setting, n ≥ 3, satisfy Rc ≡ 0, and are dubbed Einstein metrics.
This functional and the attendant Einstein metrics were discovered in the early
3.7. GENERALIZED EINSTEIN-HILBERT FUNCTIONAL 59
20th century and have played a central role in the development of Riemannian
geometry and general relativity. By means of our analysis in generalized geometry
we will introduce a more general, flexible class of metrics which can yield useful
geometric structure on a wider class of manifolds. To begin we consider a certain
generalization of the Einstein-Hilbert functional for generalized metrics.
Definition 3.46. Let E be an exact Courant algebroid on an oriented smooth
manifold M . Define the generalized Einstein-Hilbert functional for a generalized
metric G via
Z Z

EH(G) := S(G)dVg = Rg − 121
|H|2g dVg .
M M
Here we abuse of the notation and denote S(G) := S(G, div G ) (see Definition
3.41). As we show in the next proposition, the critical point equation is naturally
expressed in terms of the generalized scalar curvature S(G) and the generalized
Ricci tensor Rc(G) := Rc(G, divG ) (see Definition 3.31).
Proposition 3.47. Let E be an exact Courant algebroid over an oriented
smooth manifold M n , with n ≥ 3. A generalized metric G on E is a critical point
for the generalized Einstein-Hilbert functional if and only if
GRc(G) = 21 S(G)G.
Equivalently, in terms of the associated pair (g, H), the critical points of the func-
tional are given by solutions of the equation
2
Rc+ − 21 Rg − 1
12 |H| g = 0.
Proof. The proof involves computing variation formulas for the different pieces
of the integrand, and these are carried out in §5.1 below. Specifically, using Lemmas
5.2, 5.3 and 5.4, and applying the divergence theorem one derives
Z
d
EH(gs , bs ) = {(−∆ trg h + div div h − hh, Rci)
ds s=0 M
2
o
1
− 12 2 hH, dKi − 3 h, H 2 + R − 12 1
|H| 1
2 trg h dVg
Z nD E o
2
= h, − Rc + 14 H 2 + 21 R − 12
1
|H| g − 23 hd∗ H, Ki dVg .
M
Thus this variation vanishes for all choices of (h, K) if and only if

2
Rc − 41 H 2 − 21 R − 12
1
|H| g = 0, d∗ H = 0.
The first part of the statement follows combining Proposition 3.30, equation (3.38),
and Lemma 2.33. As for the second part, by Proposition 3.18, the vanishing of d∗ H
is equivalent to vanishing of the skew-symmetric part of Rc+ . We reinterpret the
scalar piece of the first equation in terms of the Bismut scalar curvature to obtain
the equivalent equation
2
Rc+ − 21 Rg − 1
12 |H| g = 0,
as required.
Observe the strong analogy between the critical point equation for this func-
tional, formulated in the generalized geometry language, and the classical Einstein
equation. While these generalized Einstein metrics are of interest, it turns out to be
more useful to adopt a more flexible functional, where we allow for the divergence
operator to vary independent of the generalized metric. Following the definition of
the scalar curvature in Definition 3.41, we introduce the class of compatible pairs
(G, div) of main interest.
Definition 3.48. Let (G, div) be a pair given by a generalized metric G and a
divergence operator div on an exact Courant algebroid E. We say that (G, div) is
exact if e ∈ Γ(E), defined by divG − div = he, ·i, satisfies πe = 0 and e = 2df for
some smooth function f ∈ C ∞ (M ).
Being a special case of the closed pairs in Definition 3.40, an exact pair (G, div)
is automatically compatible in the sense of Definition 2.52. Exact pairs are closely
related to the notion of density in the smooth manifold M . For our purposes it will
be sufficient to assume that the manifold M is orientable and deal with the case
of volume forms. Given a volume form µ on M , there is an associated divergence
operator on the exact Courant algebroid E defined by (cf. 2.46)
Lπe µ
divµ (e) = .
µ
Given now a generalized metric G on E we can write
µ = e−f dVg
for some smooth function f ∈ C ∞ (M ), and therefore we obtain
divG (e′ ) − divµ (e′ ) = π(e′ )f = h2df, e′ i
for any e′ ∈ Γ(E). Therefore (G, divµ ) is exact. Conversely, any exact pair (G, div)
is of the form div = divµ , for some volume form µ uniquely determined up to a
positive multiplicative constant.
Definition 3.49. Let E be an exact Courant algebroid on an oriented smooth
manifold M . Define the generalized Einstein-Hilbert functional for exact pairs
(G, divµ ) via
Z
EH(G, divµ ) = S(G, divµ )µ
(3.41) ZM
1
= R − |H|2g + 2∆g f − |∇f |2g e−f dVg .
M 12
The last explicit formula in (3.41) follows directly from (3.38) taking ϕ = 2df .
Using the identity
∆g (e−f ) = e−f |∇f |2g − e−f ∆g f
and integration by parts, we obtain the more amenable formula
Z
µ 1
(3.42) EH(G, div ) = R − |H| + |∇f |g e−f dVg .
2 2
M 12
This functional is known in the mathematical physics literature as the low energy
string effective action in the sigma model approach to type II string theory [34]. It
will play a key role as an energy functional for the generalized Ricci flow in Chapter
6.
Theorem 3.50. Given E an exact Courant algebroid over a smooth manifold

M of dimension n ≥ 3. Then, an exact pair (G, divµ ) is a critical point for the
generalized Einstein-Hilbert functional if and only if
(3.43) Rc(G, divµ ) = 0, S(G, divµ ) = 0.
Proof. We postpone the proof of the variation until Lemma 6.7, and apply
formula (6.6) directly. Let (Gs , divµs ) be a curve of exact pairs. In terms of the
classical data (gs , bs , µs ) given by (Gs , divµs ) (see Lemma 2.33 and Proposition 2.38)
we have
∂ ∂ ∂
gs = h, bs = k, µs = ( 21 trg h − φ)µ.
∂s s=0 ∂s s=0 ∂s s=0
Thus, the tangent vector to (Gs , divµs ) at s = 0 is given by (see Lemma 2.31)
∂
(Gs , divµs ) = (GΦ, −d(2φ − trg h)).
∂s s=0
Applying formula (6.6) and formula (3.32),
Z
d µs

EH(Gs , div ) = − Rc + 14 H 2 − ∇2 f, h + 21 (−d∗ H − ∇f yH) , k
ds s=0 M
i
+ R − 12 1
|H|2 + 2∆f − |∇f |2 12 trg h − φ e−f dVg
Z Z

= hRc, GΦi e−f dVg + S 12 trg h − φ e−f dVg ,
M M
and hence the statement follows.
It is not difficult to see that any critical point of the generalized Einstein-Hilbert
functional on a compact manifold has necessarily vanishing three-form H = 0 and
constant f . This follows taking the trace in the Ricci tensor Rc+ and comparing
with the formula for the generalized scalar curvature (see the proof of [70, Propo-
sition 5.8]). Thus, Ricci flat and scalar flat pairs (G, divµ ) reduce to the classical
Einstein metrics on a compact manifold. The situation is more interesting if we
consider solutions of (3.43) for closed pairs (G, div), as in Definition 3.40, rather
than exact, as illustrated in the next example.
Example 3.51. Consider the compact four-dimensional manifold M = S 3 ×S 1 .
We use the Lie group structure given by identifying
M∼ = SU (2) × U (1).
Consider generators for the Lie algebra
su(2) ⊕ R = he1 , e2 , e3 , e4 i,
such that
de1 = e23 , de2 = e31 , de3 = e12 , de4 = 0,
for {ej } the dual basis, satisfying ej (ek ) = δjk , and the notation eij = ei ∧ ej and
similarly. We take k ∈ R and define a left-invariant three-form
H = −ke123 .
This corresponds to a constant multiple of the pull-back of the Cartan three-form
on the SU (2) factor, and hence it is bi-invariant and closed. Thus, it defines an
equivariant exact Courant algebroid on E = T ⊕ T ∗ , with Dorfman bracket twisted
by H. To define a closed pair (G, div), we fix x ∈ R a non-zero constant, and define
the bi-invariant metric
g = k(e1 ⊗ e1 + e2 ⊗ e2 + e3 ⊗ e3 + x2 e4 ⊗ e4 ),
and the closed form
ϕ = −xe4 .
Then we define our pair (G, div) taking the generalized metric G determined by g
and the standard isotropic splitting T ⊂ E, and the divergence
div = divG −hϕ, i.
We leave as an exercise to check that this pair satisfies the equations (3.43). As a
hint, observe that g is the product of invariant metrics on the simple group SU (2)
and the Abelian group U (1). Then, arguing as in Proposition 3.53 below it follows
that ∇± are flat. Thus, (3.43) reduces to check that (see (3.32) and (3.38))
1 1
∇± ϕ = 0, |H|2 − d∗ ϕ − |ϕ|2 = 0.
6 4
As we will see in Chapter 6, by considering variations of EH which fix a back-
ground volume element, the critical point equation no longer includes the scalar
curvature vanishing, and we obtain an Euler equation determined only by the
vanishing of the generalized Ricci tensor. We thus adopt the following general
definition:
Definition 3.52. Given an exact Courant algebroid E over a smooth manifold
M n , we say that a generalized metric G, with compatible divergence operator div,
is generalized Einstein if
Rc(G, div) = 0.
It is elementary to see that a special case of such metrics occurs for (G, divG ), where
the associated pair (g, H) has vanishing classical Bismut Ricci tensor.
To finish this section we give examples of generalized Einstein metrics. A basic
source of Einstein metrics in Riemannian geometry arises from the famous uni-
formization theorem. This classical result classifies complete metrics with constant
sectional curvature, showing they are all quotients of the three classical model spaces
of the round sphere (S n , gS n ), the flat Euclidean plane (Rn , gEucl ) and hyperbolic
space (Hn , gHyp ). In generalized geometry, the presence of torsion gives a wider
class of examples. In the examples below we will always have div = divG , so that
the generalized Einstein equation reduces to asking for the classical Bismut Ricci
tensor to vanish, which will in particular be satisfied when the Bismut connection
is flat. We begin with the explicit construction of flat generalized metrics on simple
Lie groups.
Proposition 3.53. Let G be a Lie group with bi-invariant metric g. Define
connections ∇± as the unique metric connections with torsion T (X, Y ) = ±[X, Y ],
acting on left-invariant vector fields. Then ∇± are flat.
Proof. Recall that left-invariant vector fields on a Lie group with bi-invariant
metric g satisfy
g([X, Y ], Z) = g(X, [Y, Z]).
Using this, a direct calculation shows that the Levi-Civita connection takes the
form
∇X Y = 12 [X, Y ].
Furthermore, it follows that the tensor H(X, Y, Z) := g(T (X, Y ), Z) is skew-symmetric,
since it is obviously skew-symmetric in X and Y , while
H(X, Z, Y ) = g(T (X, Z), Y ) = ±g([X, Z], Y )
= ∓ g([Z, X], Y ) = ∓g(Z, [X, Y ]) = −H(X, Y, Z).
Thus we can express the connection with torsion T as
∇± 1
X Y, Z = h∇X Y, Zi + 2 H(X, Y, Z) =
1
2 h[X, Y ], Zi ± 1
2 h[X, Y ], Zi .
It is clear that ∇− +
X Y, Z vanishes identically, and is thus flat. For ∇ , this formula
easily yields
R(X, Y )Z = ∇X (∇Y Z) − ∇Y (∇X Z) − ∇[X,Y ] Z
= ∇X [Y, Z] − ∇Y [X, Z] − [[X, Y ], Z]
= [X, [Y, Z]] − [Y, [X, Z]] − [[X, Y ], Z]
= 0,
by the Jacobi identity.
We next give the classification of generalized Riemannian metrics with vanish-

ing Bismut curvature. As discussed above, this result was originally established
in the work of Cartan-Schouten [37, 36], who proved a more general statement
classifying Riemannian manifolds admitting an absolute parallelism. More recently
a short proof of this theorem which does not rely on the classification of symmet-
ric spaces was given by Agricola-Friedrich [2]. We provide an adaptation of their
proof in our setting which is significantly simplified due to the fact that we already
assume the torsion is skew-symmetric and closed.
Theorem 3.54. Let (M n , g, H) be a complete simply-connected Riemannian
manifold with dH = 0 such that the Bismut connection associated to (g, H) is
flat. Then (M n , g) is isometric to a product of simple Lie groups with bi-invariant
metrics g, and g −1 H(X, Y ) = ±[X, Y ] on left-invariant vector fields.
Proof. For convenience we set

X
(σH )(X, Y, Z, U ) = hH(X, Y ), H(Z, U )i .
σ(X,Y,Z)
First note that since the connection is flat, combining (3.24) and (3.26) and using
that dH = 0 we obtain
(3.44) 0 = dH(X, Y, Z, V ) = σH (X, Y, Z, U ) − (∇+
U H)(X, Y, Z).
Inserting this identity back into (3.26) yields σH = 0 (Exercise). Comparing

against (3.44) we see that ∇+ H = 0. A further elementary computation shows
that ∇+ H − ∇H = 21 σH = 0, thus ∇H = 0. Having shown ∇+ H = 0, Proposition
3.18 yields
1
Rm(X, Y, Z, W ) = 4 (hH(X, W ), H(Y, Z)i − hH(Y, W ), H(X, Z)i)
In particular, the sectional curvatures satisfy

2
1 |H(X, Y )|
K(X, Y ) = ≥ 0.
4 |X| |Y |2 − hX, Y i2
2
This implies a splitting of the torsion tensor in the presence of a metric splitting.
In particular, if the metric splits as a product M = M1 × M2 g = g1 ⊕ g2 , then for
2
Xi ∈ T Mi unit we obtain 0 = K(X1 , X2 ) = |H(X1 , X2 )| , hence H = H1 ⊕ H2 .
Thus we may reduce to analyzing the case that (M, g) is irreducible.
Now fix an orthonormal basis {ei } for some point p ∈ M . Since ∇ is flat, by
parallel transport we can extend this to a global orthonormal frame on M . By the
identity
[ei , ej ] = −T∇+ (ei , ej ).
it follows directly that {ei } are Killing fields and also
0 = ∇ei H(ej , ek , el ) = −ei (h[ej , ek ], el i),
in other words the functions h[ei , ej ], el i are constant. Thus {ei } are the basis for
a Lie algebra. The corresponding simply-connected Lie group is a simple, compact
Lie group isometric to (M, g).
We next turn to low dimensional examples, and in the next proposition give a
classification of solutions in dimension n = 3.
Proposition 3.55. Let (M 3 , g, H) be a generalized Einstein three-manifold.
Then there exists λ ∈ R such that H = λdVg , and (M 3 , g) is isometric to a space
form with either zero sectional curvature if λ = 0, and positive sectional curvature
otherwise.
Proof. From the generalized Einstein equation we know that H is harmonic.
But since H ∈ Λ3 and M is three-dimensional this means that

H
0 = d ∗ H = d dV g
,
H
hence dVg ≡ λ, as claimed. Given this it follows that H 2 = 2λ2 g, and hence by the
vanishing of Rc+ we obtain
λ2
Rc = Rc+ + 41 H 2 = 2 g.
Hence g is an Einstein metric. Expressing the Ricci curvature in terms of an
orthonormal frame, one observes that R1212 + R1313 = R2121 + R2323 = R3131 +
R3232 , which implies that R1212 = R1313 = R2323 . Since this holds for an arbitrary
orthonormal frame it follows that the metric has constant sectional curvature, and
so g is a space form, as claimed.
Example 3.56. Note that in the previous proposition that a generalized Ein-
stein structure can be associated to a classical Einstein metric with nonzero Einstein
constant. In particular, consider the example of (S 3 , gS 3 ), the three sphere with
the round metric of constant sectional curvature 1. This metric is Einstein, in
particular RcgS3 = 2gS 3 . Now set H = 2dVg . This is a parallel three-form, and
a direct calculation shows that H 2 = 8g, and hence it follows that Rc+ ≡ 0, and
so the structure (gS 3 , H) is generalized Einstein. This example is actually Bismut
flat, and is left-invariant on the Lie group S 3 , in line with Theorem 3.54.
Example 3.57. Given (Mini , gi , Hi ), i = 1, 2 two generalized Einstein struc-

tures, we form a product from these structures as follows. First set M = M1 × M2 ,
with canonical projection maps πi : M → Mi . We furthermore set g = π1∗ g1 + π2∗ g2
and H = π1∗ H1 + π2∗ H2 . It is easily checked that this gives another generalized
Einstein structure. This allows us to form our first nontrivial example, namely
we take (S 3 , gS 3 , 12 dVgS3 ) and (S 1 , gS 1 , 0), both generalized Einstein (see Example
3.56), and so the resulting product structure is a nontrivial generalized Einstein
structure. This example plays a central role in the study of generalized Einstein
structures in complex geometry, as we will see in Chapter 8. Note also that not
only is the product metric π1∗ gS 3 +π2∗ gS 1 not an Einstein metric, but S 3 ×S 1 cannot
admit any Einstein metric by the Hitchin Thorpe inequality (cf. [98, 177]).
Question 3.58. A classical result of Alekseevskii-Kimelfeld [4] establishes that
a homogeneous Riemannian manifold M = G/K with zero Ricci curvature is locally
Euclidean, and so isometric to a product of a Euclidean space with a flat torus. An
elementary proof of this fact was given by Bérard-Bergery ([14]) which we briefly
recount here. Using Cheeger-Gromoll splitting [43, Theorem 5], the space splits as
a product of a compact homogeneous Ricci flat space with a flat Euclidean space.
Using the Bochner argument on the compact piece, it follows that all Killing fields
are parallel. Thus there is a global parallel frame and so the compact piece is flat.
Does this result extend to generalized setting? That is, given (M n , g, H) a homo-
geneous Riemannian manifold with H invariant and zero Bismut-Ricci curvature,
is the associated Bismut connection flat?
CHAPTER 4
Fundamentals of Generalized Ricci Flow
Having established fundamental properties of generalized geometry, and some

basic notions and examples of canonical structures in generalized geometry, it is
natural to ask the question,
Which manifolds admit canonical generalized geometries,
and how can we construct them?
As stated this question is quite broad, and we cannot expect satisfying answers
without further restrictions. In the context of Riemannian geometry, Hamilton
introduced a powerful tool, the Ricci flow, for approaching these questions in a
broad, principled, way. Ricci flow is an evolution equation for Riemannian metrics
taking the form
∂
g = − 2 Rc .
∂t
The basic philosophy underpinning the Ricci flow equation is at once elementary
and profound. This equation is meant to be interpreted as a heat equation for
Riemannian metrics, where one expects the curvature to smooth out and become
in a certain sense uniform, eventually yielding a canonical Riemannian metric, say
an Einstein metric, as a limit. The generalized Ricci flow seeks to adapt this same
philosophy to the setting of generalized geometry, and can be expressed in the
language of generalized geometry as
∂
G −1 G = − 2Rc(G).
∂t
As we will see, this is a kind of heat equation which deforms generalized metrics, at
least in principle, towards a canonical structure, say a generalized Einstein metric.
From this basic starting point we will show in the remainder of the text that the
generalized Ricci flow equation is rich with geometric structure, broad enough to
encompass meaningful geometric and topological questions, and yet tame enough
to yield global existence results.
4.1. The equation and its motivation

In this section we give various definitions and equivalent formulations of gen-
eralized Ricci flow, of increasing conceptual complexity. To begin we give the
equations first from the point of view of classical objects in Riemannian geometry.
4.1.1. Classical formulation.

Definition 4.1. Given M n a smooth manifold, H0 ∈ Λ3 T ∗ , dH0 = 0, we say
that a one-parameter family (gt , bt ) of pairs of Riemannian metrics and two forms
67
68 4. FUNDAMENTALS OF GENERALIZED RICCI FLOW
is a solution of generalized Ricci flow if, setting H = H0 + dbt , one has

∂
g = − 2 Rc + 21 H 2 ,
(4.1) ∂t
∂
b = − d∗g H.
∂t
As discussed in Chapter 3, this system first appeared in physics literature from
the study of renormalization group flow [34], with the first mathematical results
appearing in [138, 157]. From a naive point of view this flow is a natural coupling
of the Ricci flow equation with a degenerate heat equation for a two-form. To get
a feeling for the generalized Ricci flow, we recall some basic aspects of parabolic
equations. Consider the initial value problem for a section u of some smooth vector
bundle E over M with connection ∇,
∂u
= aij (t, x, u, ∇u)∇i ∇j u + f (t, x, u, ∇u),
(4.2) ∂t
u(0) = u0 .
Here of course the right hand side should be defined by some coordinate invariant
operators, and the equation (4.2) is to be interpreted in some coordinate chart.
Definition 4.2. The equation (4.2) is parabolic at (u, x, t) if
aij (x, t, u, ∇u) > 0
in the sense of bilinear forms.
The standard theory of parabolic partial diffferential equations gives that for
any initial data u0 on a compact manifold such that (4.2) is parabolic at all points
(u0 , x, 0), there exists ǫ > 0 and a solution to (4.2) on [0, ǫ). For more context and
further results in these areas we direct the reader to [61, 82, 126]. Due to the issue
of gauge invariance, the generalized Ricci flow is not a strictly parabolic equation.
However, the resemblance of generalized Ricci flow to a heat equation can be seen
in the next lemma.
Lemma 4.3. Fix M a smooth manifold, H0 ∈ Λ3 T ∗ , dH0 = 0, and (gt , bt )
a solution of generalized Ricci flow. Assuming d∗g b = 0, in harmonic coordinates
centered at a point (p0 , t0 ), the generalized Ricci flow is expressed as
∂ ∂ 2 gij
gij = g kl k l + Q1 (g, ∂g, H0 , ∂b),
∂t ∂x ∂x
∂ ∂ 2 bij
bij = g kl k l + Q2 (g, ∂g, H0 , ∂b) + Q3 (g, b, ∂ 2 g),
∂t ∂x ∂x
where the quantities Qi denote quadratic expressions in the arguments.
Proof. By [143, Lemma 49] we know that at the center of a harmonic coor-
dinate chart one has
∂ 2 gij
Rcij = − 21 g kl + Q(g, ∂g)
∂xk ∂xl
for some quadratic expression Q(g, ∂g) in first derivatives of g. As the term H 2 in
the definition of generalized Ricci flow is expressed as a quadratic in H0 and first
4.1. THE EQUATION AND ITS MOTIVATION 69
derivatives of b, the first claimed equation follows. As we have assumed d∗g b = 0,

we can furthermore express
∂
bij = − d∗g (H0 + db)
∂t
= ∆d b + Q(g, ∂g, H0 )
∂ 2 bij
= g kl + Q2 (g, ∂g, H0 , ∂b) + Q3 (g, b, ∂ 2 g),
∂xk ∂xl
as claimed.
Thus we see that in special coordinates for g, together with the divergence-free
condition for b, the generalized Ricci flow appears to be a strictly parabolic equation.
The choice of coordinates and imposition of d∗g b = 0 relate to the gauge-invariance
of the flow, explored more in §4.4, and exploited in the proof of short-time existence
of the flow in Chapter 5. For instance, the condition d∗g b = 0 corresponds to a choice
of gauge-fixing rendering b L2 -orthogonal to the infinitesimal action of inner B-field
symmetries, given by the image of T ∗ via (2.15), in the space of generalized metrics
(see Remark 2.42).
In a slight abuse of terminology, the one-parameter family of pairs (gt , Ht =
H0 + dbt ) will also be referred to as a solution of generalized Ricci flow. In the next
lemma we record the evolution equation for H.
Lemma 4.4. Given M n a smooth manifold, H0 ∈ Λ3 T ∗ , dH0 = 0, and (gt , bt )
a solution of generalized Ricci flow, one has

∂
− ∆d,gt H = 0.
∂t
Proof. Using the flow equations of (4.1) and the fact that H0 is closed we
compute
∂ ∂
H= (H0 + db) = −dd∗gt H = ∆d,gt H,
∂t ∂t
as required.
Remark 4.5. The tensor H 2 is positive semidefinite, thus it follows that a
solution to generalized Ricci flow is automatically a supersolution of Ricci flow, i.e.
∂
g ≥ − 2 Rc .
∂t
Many results have appeared in recent years concerning Ricci flow supersolutions (cf.
for instance [96, 115, 133, 175]), and thus all such results automatically apply
for generalized Ricci flow.
4.1.2. Generalized formulation. It is useful to express the generalized Ricci
flow directly in terms of the language of generalized geometry. To begin we give
an elementary lemma which expresses the generalized Ricci flow in terms of the
curvature of the Bismut connection.
Lemma 4.6. Fix M n a smooth manifold and H0 ∈ Λ2 T ∗ , dH0 = 0. Then
(gt , bt ) is a solution of generalized Ricci flow if and only if
∂
(g − b)(X, Y ) = − 2 Rc+ (X, Y ),
∂t
where Rc+ denotes the Ricci curvature tensor associated to the Bismut connection
∇+ = ∇ + 21 g −1 H.
Proof. This follows directly from the evolution equations (4.1) and the for-
mulas for the symmetric and skew-symmetric parts of the Bismut Ricci curvature
exhibited in Proposition 3.18.
These formulations make it clear that the generalized Ricci flow equations in-
volve precisely the data of a generalized metric on an exact Courant algebroid,
namely a Riemannian metric and a skew-symmetric two-form. Furthermore, it is
clear from Lemma 4.6 that the fixed points of the generalized Ricci flow are pre-
cisely Bismut Ricci-flat structures. It is thus natural to express the flow equations
directly in these terms, which is carried out next.
We fix a generalized metric G0 on an exact Courant algebroid E over the smooth
manifold M . Using the splitting determined by G0 (see Proposition 2.40), we can
identify E = T ⊕ T ∗ endowed with the H0 -twisted Courant bracket. Given now a
one-parameter family of generalized metrics Gt on E, via Proposition 2.38 we can
identify Gt with a one-parameter family of pairs (gt , bt ). Recall from Lemma 2.33
that we have the following explicit formula for the derivative of Gt along the family
−1
−1 ∂ g h g −1 kg −1
(4.3) G G= ,
∂t −k −hg −1
where
∂ ∂
g = h, b = k.
∂t ∂t
Notice that G −1 = G by definition of generalized metric. It is also important to
observe that (4.3) is written in the isotropic splitting induced by the metric Gt .
To state the next result, we denote Rc(G) = Rc(G, divG ) the generalized Ricci
tensor associated to a generalized metric G via its Riemannian divergence (see
Definition 3.31 and Definition 2.46).
Proposition 4.7. Given (E, [, ], h, i) → M an exact Courant algebroid, a one-

parameter family Gt of generalized metrics satisfies
∂
(4.4) G −1 G = − 2Rc(G)
∂t
if and only if the one-parameter family of pairs (gt , bt ) determined by Gt and the
initial splitting associated to G0 satisfies generalized Ricci flow.
Proof. Notice that Rc(G) defines a tangent vector to the space of generalized
metrics at G by Lemma 2.31 and Lemma 3.32. Then, (4.4) follows directly from
Lemma 2.35, Lemma 4.6, and Proposition 3.30.
Remark 4.8. The condition GRc(G) = −Rc(G)G in Proposition 3.30 shows

that we can alternatively write the generalized Ricci flow as (see [70, 162])
∂
G = [Rc(G), G].
∂t
4.2. EXAMPLES 71
4.2. Examples
Before delving into the analysis of generalized Ricci flow, we record some fun-
damental examples which give context to the analysis to follow and also indicate
the role the torsion tensor H plays in affecting the qualitative behavior of the flow.
We first record some terminology for certain kinds of entire solutions.
Definition 4.9. We say that a solution (gt , Ht ) to generalized Ricci flow is
(1) ancient if it is defined on M × (−∞, 0),
(2) immortal if it is defined on M × (0, ∞),
(3) eternal if it is defined on M × (−∞, ∞).
Example 4.10. Consider (S 3 , gS 3 ) the three-dimensional sphere with round
metric of constant sectional curvature 1. As noted in Example 3.56, this metric
satisfies RcgS3 = 2gS 3 . What is more, due to the scale invariance of the Ricci
tensor, for any λ > 0 one has RcλgS3 = 2gS 3 . Fix now a constant η ∈ R and let
H = ηdVgS3 . Note that for any pair of the form (g, H) = (λgS 3 , ηdVgS3 ) one has
2
that d∗g H = d∗λg 3 ηdVgS3 = 0, and H 2 = 2 λη2 gS 3 , hence
S
2

(4.5) RcB = Rcg − 41 H 2 − 12 d∗g H = 2 − 21 λη 2 gS 3 .
We now construct solutions to the generalized Ricci flow within this ansatz, i.e.
(gt , Ht ) = (λt gS 3 , ηt dVgS3 ). Using (4.5), we see that it will always be the case that
∂
∂t η = 0, and so our flow reduces to the evolution of λ, which we obtain from (4.1)
and (4.5) via

∂ ∂ 1 2 η02
λ gS 3 = (λgS 3 ) = −2 Rc + 2 H = −4 + 2 gS 3 ,
∂t ∂t λ
which implies
∂ η2
(4.6) λ = −4 + 02 .
∂t λ
Here the analysis splits into two cases, whose behavior is markedly different.
∂
In particular, in the case η0 = 0, we simply obtain ∂t λ = −4, and so the solution
λ0
to the flow exists on the time interval [0, 4 ), with explicit solution (gt , Ht ) =
((λ0 − 4t)gS 3 , 0). Note that, as H = 0 along the flow, the generalized Ricci flow
has reduced to the classical Ricci flow equation, and this example indicates one
of the basic qualitative behaviors of the Ricci flow, namely that the round sphere
shrinks homothetically at a constant rate until its diameter goes to zero at the finite
time λ40 . This solution in fact can be extended for all negative times, giving a basic
example of an ancient solution.
In the case η0 6= 0, i..e when the H field is present, the generalized Ricci
flow behaves quite differently. In particular, we see in this case that there is a
unique fixed point of (4.6), namely when λ = |η20 | . Furthermore, this fixed point
is attractive in the sense that for λ < |η20 | one has ∂t∂
λ > 0 and for λ > |η20 | one
∂
has ∂t λ < 0. Elementary calculus arguments show that for any choice of λ0 , the
solution to (4.6) exists on [0, ∞), and λt converges as t → ∞ to the value |η20 | . Thus
we see a striking difference in the qualitative behavior in this example depending
on the presence of H, where in some rough sense the presence of H attenuates
the otherwise singular behavior of the flow, turning what was once a finite time
singularity into a flow with infinite existence time and convergence at infinity. One
may argue that, in the case when H = 0, a simple rescaling in space and time
to normalize the volume along the flow suffices to recover global existence and
convergence at infinity to gS 3 , but nonetheless this example gives a sense of what
can be expected conjecturally in more general cases, where the overall behavior
is not as elementary. We note that, if λ − |η20 | > 0 initially, the solution extends
backwards in time to be an eternal solution, otherwise it will only be immortal.
Example 4.11. Consider (M 3 , gHyp ) a compact hyperbolic 3-manifold with
constant sectional curvature −1. Then RcgHyp = −2gHyp. Fix a constant η ∈ R
and let H = ηdVgHyp . Note that for any pair of the form (g, H) = (λgS 3 , ηdVgS3 )
2
one has that d∗g H = d∗λgHyp ηdVgHyp = 0, and H 2 = 2 λη2 gHyp , hence
2

(4.7) RcB = Rcg − 41 H 2 − 12 d∗g H = −2 − 21 λη 2 gHyp .
We now construct solutions to the generalized Ricci flow within this ansatz, i.e.
(gt , Ht ) = (λt gHyp , ηt dVgHyp ). Using (4.7), we see that it will always be the case
∂
that ∂t η = 0, and so our flow reduces to the evolution of λ, which we obtain from
(4.1) via

∂ ∂ 1 2 η02
λ gHyp = (λgHyp ) = −2 Rc + 2 H = 4 + 2 gHyp ,
∂t ∂t λ
which implies
∂ η2
λ = 4 + 02 .
∂t λ
It follows from elementary arguments that for arbitrary λ0 > 0 this solution exists
for [0, ∞), and λt is asymptotic to 4t in the sense that limt→∞ λtt = 4. These
solutions extend backwards to be immortal, but are not eternal solutions.
We now address the issue of how to appropriately renormalize this flow to
obtain some form of convergence. In particular, we set
gt e t = Ht ,
get = , H
t t
and define a new time parameter via s = log t. Then we compute
∂ ∂ gt ∂g gt e2 − e
ge = t = − = −2 Rc + 21 H 2 − ge = −2 Rceg + 21 H g,
∂s ∂t t ∂t t
where the last equality follows since the Ricci tensor is scale invariant, as is the
quantity H 2 . A similar calculation yields
∂ e e − H.
e dH e
H=∆
∂s
Thus this scaling has the effect of adding normalizing terms to the original equa-
tions. Since H was fixed under the original flow, it follows easily that H e → 0,
whereas since λtt approaches 4, we will obtain e g → 4gHyp.
Example 4.12. The presence of the torsion term H will of course not remove
all singular behavior. One only expects the term H to prevent the collapse of the
length of vectors v for which vyH 6= 0, a condition which is difficult to verify in
general along a solution to the flow, but which we can illustrate with a simple exam-
ple. In particular, as in Example 3.56, on S 3 , the structure (g, H) = (gS 3 , 21 dVgS3 )
4.2. EXAMPLES 73
is generalized Einstein, and so a fixed point of generalized Ricci flow. However,

consider M = S 3 × S 3 , with g = π1∗ gS 3 + π2∗ gS 3 , and H = 21 π1∗ (dVgS3 ). Elementary
arguments show that the generalized Ricci flow with this initial condition will be-
have as the product of the solutions on the factors of S 3 . In particular, the first
factor will remain fixed as it is generalized Einstein, and the second factor will
homothetically shrink to a point in finite time as described in Example 4.10.
Example 4.13. A classical singularity model of the Ricci flow is the shrinking
cylinder, which models regions of Ricci flow experiencing a neckpinch singularity.
Let M = S n × R, and consider the one-parameter family of metrics
gt = −2(n − 1)tgS n ⊕ ds2 ,
defined for t < 0. Since RcgSn = (n − 1)gS n and the metric ds2 is flat, it follows
that gt is a solution of Ricci flow. Thus this solution is ancient, and is meant to
model a collapsing neck.
The dynamics of necks under generalized Ricci flow can be quite different
depending on the dimension and the choice of H. First consider S 2 × R, let
H ≡ dVS 2 ∧ ds, fix constants φ, ψ > 0 and consider a metric g of the form
g = φgS 2 ⊕ ψds2 .
Note that the Ricci tensor for any such metric is
Rc = gS 2 .
Further direct computation shows that for these choices of g and H one has
H 2 = φ−1 ψ −1 gS 2 ⊕ φ−2 ds2 .
As H is harmonic with respect to any metric in the given ansatz, the generalized
Ricci flow can thus be reduced to the system of ODE
∂
φ = − 2 + 12 φ−1 ψ −1 ,
∂t
∂
ψ = 21 φ−2 .
∂t
We leave as an exercise to show that, for any positive initial values of φ and ψ,
the solution to this system of ODEs exists on a time interval of the form [0, ∞),
where
lim φ = 0, lim ψ = ∞.
t→∞ t→∞
In other words, the neck can still collapse, although it takes infinitely long to do so.
On the other hand if we consider S 3 × R, with H = dVS 3 and g = λgS 3 ⊕
ds2 , then the solution to generalized Ricci flow will split according to the product
structure, and the dynamics of the S 3 portion are the same as those described in
the case of S 3 in Example 4.10, and thus the neck stabilizes, yielding convergence
to a cylindrical metric.
To finish this section we analyze the generalized Ricci flow of all left-invariant
metrics on S 3 ∼ = SU(2). Before doing this we set up some general remarks on
the generalized Ricci flow of left-invariant structures on compact semisimple Lie
groups. First, as the unique left-invariant representative of any cohomology class
is harmonic with respect to any left-invariant metric, it follows that H will remain
fixed along any such solution. Furthermore, as indicated by the example of round
S 3 above, the initial cohomology class chosen for H can have a significant effect on
the qualitative behavior of the generalized Ricci flow. In fact, even for the case of
Ricci flow (H ≡ 0), there is not a general picture of the behavior of Ricci flow in
this setting. However, as described in Proposition 3.53, there are canonical Bismut-
flat structures given by the unique bi-invariant metric g together with torsion H =
±g([·, ·], ·).
Conjecture 4.14. Let G denote a compact semisimple Lie group with bi-
invariant metric g∞ . Given g0 a left-invariant metric and
H0 = ±g∞ ([·, ·], ·),
the generalized Ricci flow with initial condition (g0 , H0 ) exists on [0, ∞) and con-
verges to the Bismut-flat structure (g∞ , H0 ).
Example 4.15. We prove Conjecture 4.14 in the case of S 3 ∼ = SU(2). Recall
we constructed generalized Ricci flat structures on SU(2) × S 1 in Example 3.51,
using a left-invariant frame. Here we choose a slightly different frame, in particular
we choose left-invariant vector fields X1 , X2 , X3 such that
[X1 , X2 ] = −2X3 , [X2 , X3 ] = − 2X1 , [X3 , X1 ] = −2X2 .
Such a basis is called a Milnor frame, and we refer to [134] for further information
on the geometry of left-invariant metrics on Lie groups. This frame induces a global
coframe {µi }, and we can consider left-invariant metrics of the form
g = Aµ1 ⊗ µ1 + Bµ2 ⊗ µ2 + Cµ3 ⊗ µ3 .
The bi-invariant metrics occur exactly when A = B = C. An exercise shows that
the only nonvanishing components of the Ricci tensor are
2
Rc(X1 , X1 ) = A2 − (B − C)2 ,
BC
2
Rc(X2 , X2 ) = B 2 − (C − A)2 ,
CA
2
Rc(X3 , X3 ) = C 2 − (A − B)2 .
AB
Furthermore we can compute, up to a positive multiple,
H0 = µ 1 ∧ µ 2 ∧ µ 3 ,
and hence associated to g and H0 we have that the only nonvanishing components
of H02 are
2 2 2
H02 (X1 , X1 ) = , H02 (X2 , X2 ) = , H02 (X3 , X3 ) = .
BC CA AB
As remarked above, the generalized Ricci flow with initial condition (g0 = g, H0 )
will preserve Ht = H0 for all times. It follows from the above computations that
the flow reduces to the following system of ODE for the positive numbers A, B, C:
1
Ȧ = −4A2 + 4(B − C)2 + 1 ,
BC
1
Ḃ = −4B 2 + 4(C − A)2 + 1 ,
CA
1
Ċ = −4C 2 + 4(A − B)2 + 1 .
AB
4.3. MAXIMUM PRINCIPLES 75
Without loss of generality we can asssume A0 ≤ B0 ≤ C0 . A computation shows

that
d C −A
(C − A) = 4 B 2 − (A + C)2 + 1 .
dt ABC
This can be used to show that At ≤ Bt ≤ Ct for all times such that the solution is
defined. Furthermore, we observe that

Ȧ ≥ B −1 C −1 −4A2 + 1 ,
so a uniform positive lower bound for A is preserved for all times. We can also
compute

d C−A
(4.8) = − 8A−2 B −1 (C − A) (C + A − B) ≤ 0.
dt A
It follows that there is a uniform constant λ > 0 such that C ≤ λA. Using this
estimate we furthermore obtain
A 1 λ2
Ċ ≤ − 8 + 4 + ≤ −4 + 2 .
B AB C
It follows that there is a uniform upper bound for C. With the uniform lower bound
for A and upper bound for C in place, it follows that the solution to the ODE is
defined for all t > 0. Returning to (4.8), it follows that there is a uniform constant
δ > 0 so that

d C−A C −A
≤ −δ ,
dt A A
and the solution converges exponentially to A∞ = B∞ = C∞ , a multiple of the
bi-invariant metric.
It would be interesting to carry out a similar analysis for left-invariant solutions
of generalized Ricci flow in other classes of Lie groups, which help us to understand
qualitative properties of the equations. The case of nilpotent Lie groups has been
studied in [139] (see also [21, 59] in the setting of pluriclosed flow in Chapter 9).
4.3. Maximum principles

Moving beyond homogeneous examples, the most basic tool in analyzing second
order elliptic and parabolic partial differential equations in general is the maximum
principle, and these tools are crucial in understanding generalized Ricci flow. In this
section we record statements and some proofs of elliptic and parabolic maximum
principles on manifolds.
Proposition 4.16. Let (M n , g) be a compact connected Riemannian manifold
and suppose u ∈ C ∞ (M ) satisfies
Lu := ∆u + X · u ≤ 0,
where X ∈ Γ(T ) is a vector field. Then u is constant.
Proposition 4.17. Let (M n , gt ) be a compact manifold with a one-parameter
family of Riemannian metrics. Suppose u ∈ C ∞ (M × [0, T )) satisfies

∂
− ∆gt u ≤ 0.
∂t
Then
sup u ≤ sup u
M×{t} M×{0}
for all t ∈ [0, T ).
Proof. Fix ǫ > 0, and set uǫ := u − ǫ(1 + t). One obviously has that
supM×{0} uǫ < supM×{0} u =: λ. We will show that supM×{t} uǫ < λ, which
will finish the proposition by sending ǫ to zero. If this claim was false, there exists
a point such that uǫ ≥ λ. Since M is compact, we can choose (x, t) ∈ M × (0, T )
such that uǫ (x, t) = λ, and uǫ (y, s) ≤ λ for all y ∈ M , 0 ≤ s ≤ t. Hence (x, t) is
a spacetime maximum for uǫ on M × [0, t], and so it follows that ∆gt uǫ ≤ 0 and
∂
∂t uǫ ≥ 0 at (x, t). We conclude that, at (x, t) we have
∂
0≤ uǫ ≤ ∆gt uǫ − ǫ < 0,
∂t
a contradiction. Hence supM×{t} uǫ < λ for all t ∈ [0, T ), and the proposition
follows.
For various applications one needs to consider more general reaction-diffusion

equations beyond sub/supersolutions of the heat equation. We next formalize a
general class of equations which suffices for our purposes.
Definition 4.18. Given (M n , gt ) a smooth manifold with a smooth one-parameter
family of Riemannian metrics, Xt ∈ Γ(T ) a smooth one-parameter family of vector
fields and F : R × [0, T ) → R which is continuous in t and locally Lipschitz in x,
define the associated semi-linear heat operator via
∂
Lu := u − ∆gt u − Xt u − F (u, t).
∂t
Proposition 4.19. Let (M n , gt ) be a compact manifold with a one-parameter
family of Riemannian metrics. Suppose L is a semi-linear heat operator as in
Definition 4.18, and suppose u, v ∈ C 2 (M × [0, T )) satisfy
Lv ≤ Lu, sup (v − u) ≤ 0.
M×{0}
Then
sup (v − u) ≤ 0
M×{t}
for all t ∈ [0, T ).
Proof. We set w(x, t) = u(x, t) − v(x, t), and aim to show that w ≥ 0 at all
points. Fix a time τ ∈ [0, T ). Since the function F is locally Lipschitz in space and
the region M × [0, τ ] is compact, there exists a constant K such that
|F (u(x, t), t) − F (v(x, t), t)| ≤ K |u(x, t) − v(x, t)| .
for all (x, t) ∈ M × [0, τ ]. Now fix ǫ > 0 and let
wǫ (x, t) = u(x, t) − v(x, t) + ǫe2Kt .
4.3. MAXIMUM PRINCIPLES 77
Using the given differential inequalities we compute

∂wǫ
≥ ∆wǫ + hX, ∇wǫ i − F (u, t) + F (v, t) + 2Kǫe2Kt
∂t
≥ ∆wǫ + hX, ∇wǫ i − K |w| + 2Kǫe2Kt.
If ever wǫ = 0, then arguing as in Proposition 4.17 we can choose a point (x0 , t0 )
such that wǫ (x0 , t0 ) = 0 and (x0 , t0 ) realizes the minimum for wǫ on M × [0, t0 ].
Note then that at the point (x0 , t0 ), w = −ǫe2Kt0 , ∇wǫ = 0, ∆wǫ ≥ 0, ∂w ∂t ≤ 0.
ǫ
Thus we conclude
∂wǫ
0≥ ≥ −Kǫe2Kt0 + 2Kǫe2Kt0 ≥ Kǫe2Kt0 > 0,
∂t
a contradiction. Thus wǫ > 0 for all ǫ > 0 and thus w ≥ 0 on M × [0, τ ], finishing
the proof.
We end this section with the fundamental observation that if H vanishes at

the initial time of a solution to generalized Ricci flow, then it vanishes for all time
t > 0, and the one-parameter family of metrics evolves by the Ricci flow. This
justitifes the terminology of “generalized Ricci flow,” and yields the important
philosophical point that, without special assumptions on H, in the most general
setting the results must be extensions of those for Ricci flow. However as we have
already seen there are geometrically natural settings where the presence of H has
important qualitative effects on the behavior of the flow.
Proposition 4.20. Let M n be a smooth compact manifold and suppose (gt , Ht )
is a solution to generalized Ricci flow satisfying H0 ≡ 0. Then for all t such that
the flow is defined, Ht ≡ 0, and
∂
g = − 2 Rc,
∂t
that is, gt is a solution to Ricci flow.
Proof. Suppose the solution exists on [0, T ), fix some τ < T and let
K := sup |Rmg | .
M×[0,τ ]
Using Lemma 4.4 and the Bochner formula (cf. 5.20 below) we obtain the differen-
tial inequality

∂
− ∆ |H|2 ≤ CK |H|2 .
∂t
2 ∂

In particular, if we define Φ := e−CKt |H| , we easily obtain ∂t − ∆ Φ ≤ 0.
Applying the maximum principle (Proposition 4.17) we see that for all t ∈ [0, τ ]
one has
sup e−CKt |H|2 ≤ sup e−CKt |H|2 = 0,
M×{t} M×{0}
and so H vanishes on [0, τ ]. This argument applies for arbitrary τ < T and so H
vanishes identically wherever the flow is defined. The fact that gt is a solution to
Ricci flow now follows directly from (4.1).
4.4. Invariance group and solitons

A fundamental feature of many evolution equations in geometry is the presence
of an infinite dimensional invariance group. For instance, the Ricci tensor of a
Riemannian metric is diffeomorphism covariant, and so diffeomorphisms can act on
solutions to the Ricci flow, preserving the equation. More generally, one can pull
a solution to Ricci flow back by a family of time-dependent diffeomorphisms. It
is natural to consider this as equivalent to the original flow, as the corresponding
time slices are isometric. However, the resulting flow equation now includes a Lie
derivative term corresponding to the variation of the family of diffeomorphisms.
This circle of ideas is relevant to the generalized Ricci flow as well. As discussed in
§2.2, the relevant symmetry group includes B-field transformations. By letting a
time-dependent family of symmetries act on a solution to generalized Ricci flow one
produces a more general flow equation which includes modification by Lie derivative
terms as well as a closed two-form.
4.4.1. Gauge-fixed generalized Ricci flow. To begin we recall the deriva-

tion of the gauge-fixed Ricci flow equation from the Ricci flow. We say that g = gt
satisfies the gauge-fixed Ricci flow if there exists a one-parameter family of vector
fields Xt ∈ Γ(T ) such that egt = (φt )∗ gt is a solution of the Ricci flow, where φt is
the one-parameter family of diffeomorphisms generated by Xt
∂
Xt ◦ φt = φt .
∂t
Using the naturality of the of the Ricci tensor under diffeomorphisms, the gauge-
fixed Ricci flow is equivalent to
!
∂ ∂
ge = φt ∗ g − LXt g = −2φt ∗ Rc(g) = −2 Rc(e
g)
∂t ∂t
and we obtain the equation

∂
(4.9) g = −2 Rc +LXt g.
∂t
Following this line of thought, we can define a gauge-fixed version of generalized
Ricci flow.
Definition 4.21. Given E an exact Courant algebroid, a one-parameter family
of generalized metrics Gt satisfies the gauge-fixed generalized Ricci flow if there
exists a one-parameter family of generalized diffeomorphisms Ft ∈ Aut(E) such
that
Ge = (Ft )∗ G = Ft Gt Ft−1
is a solution of the generalized Ricci flow.
Proposition 4.22. Given H0 ∈ Λ3 T ∗ such that dH0 = 0, consider the exact
Courant algebroid E = (T ⊕ T ∗ , h, i , [, ]H0 , π). Then, a one-parameter family of
generalized metrics Gt satisfies the gauge-fixed generalized Ricci flow if and only if
there exists a family of vector fields Xt and Bt ∈ Λ2 satisfying
d (Bt + iXt H0 ) = 0,
4.4. INVARIANCE GROUP AND SOLITONS 79
such that the associated one-parameter family of pairs (gt , bt ) satisfies

∂
g = − 2 Rc + 12 H 2 + LX g
(4.10) ∂t
∂
b = − d∗g H − B + LX b.
∂t
Proof. Using the naturality of the generalized Ricci tensor Rc, we first observe
that the generalized Ricci flow equation is equivalent to
!
∂ ∂
Get
−1
Ge = Ft ∗ G −1 −1
G + G ζt · G = −2Ft ∗ Rc(G) = −2Rc(G) e
∂t ∂t
where ζt · G denotes the infinitesimal left action of the Lie algebra element
∂
ζt = Ft−1 Ft ∈ Lie Aut(E).
∂t
Thus, G = Gt is a solution of the gauge-fixed generalized Ricci flow if and only if
there exists a one-parameter family of infinitesimal automorphisms ζt ∈ Lie Aut(E)
such that
∂
(4.11) G −1 G = −2Rc(G) − G −1 ζt · G.
∂t
We use now E = (T ⊕ T ∗ , h, i , [, ]H0 , π) to write this condition explicitly. We have
(see Corollary 2.22)
Lie Aut(E) = {X + B | LX H0 = −dB} = {X + B | d(iX H0 + B) = 0}.
A one-parameter family ζt = Xt + Bt ∈ Lie Aut(E) can be integrated to
Ft = φt ◦ eB t ∈ Aut(E)
where φt is the one-parameter family of diffeomorphisms generated by Xt and
Z t
Bt = φ∗s Bs ds.
0
Now, if G = G(g, b) as in Proposition 2.38, it follows that
Ft∗ G(g, b) = G(φt ∗ g, φt ∗ (b + B t )),
and therefore
−1
−g LXt g −g −1 (LXt b − Bt )g −1
G −1 ζt · G = ,
LXt b − Bt LXt gg −1
where the previous formula is written in the isotropic splitting given by the metric
Gt . From this, we conclude that the gauge-fixed generalized Ricci flow is given by
∂
g = − 2 Rc + 12 H 2 + LXt g,
(4.12) ∂t
∂
b = − d∗g H − Bt + LXt b,
∂t
where H = H0 + db, as required.
Remark 4.23. Note that in the formulation above, Bt is not necessarily closed.
It is possible to rephrase the above gauge-fixed equation in terms of the action of
a closed B-field. Namely set
kt = −Bt − iXt H0 + d(iXt b).
Then
−Bt + LXt b = kt + iXt H0 + iXt db = kt + iXt H,
and we obtain the following alternative description of the flow:
∂
g = − 2 Rc + 12 H 2 + LXt g,
(4.13) ∂t
∂
b = − d∗g H + iXt H + kt .
∂t
We will at times refer to this as (Xt , kt )-gauge-fixed generalized Ricci flow. Observe
that in either formulation the evolution of the three-form is
∂ ∂
H = (H0 + db) = −dd∗g H + LXt H.
∂t ∂t
It turns out that it is possible to interpret the gauge-fixed flow naturally in
terms of the generalized Ricci tensor Rc+ associated to a particular choice of di-
vergence operator.
Proposition 4.24. Let E be an exact Courant algebroid over a smooth man-
ifold M . Let Xt be a one-parameter family of vector fields Xt on M . Then, a
one-parameter family Gt of generalized metrics satisfies
∂
G −1 G ◦ π− = −2Rc+ (G, div) ◦ π− ,
∂t
where div = divG + h2iXt gt , i, if and only if the one-parameter family of pairs (gt , bt )
determined by Gt and the initial splitting associated to G0 satisfies the (Xt , kt )-
gauge-fixed generalized Ricci flow (4.13) with kt = −d(iXt gt ).
Proof. Let ϕ ∈ T ∗ a one-form. From Proposition 3.30, the generalized Ricci
tensor Rc+ with divergence div = divG − h2ϕ, i is given by

Rc+ (σ− X, σ+ Y ) = Rc − 14 H 2 + 12 Lϕ♯ g − 21 d∗ H + 12 dϕ − 12 iϕ♯ H (X, Y ),
where we have used the identity

∇+ ϕ = 21 Lϕ♯ g + 21 dϕ − 21 iϕ♯ H.
Setting ϕ = −iXt gt and identifying the generalized Ricci tensor Rc+ with a map
Rc+ : V− → V+ , the statement follows combining equation (4.13) with Lemma
2.35.
For the case of a gradient vector field the divergence div = divG + h2iXt gt , i
is exact (see Definition 3.48) and we can express the gauge-fixed generalized Ricci
flow (4.13) as generalized Ricci flow with divergence operator.
Proposition 4.25. Let E be an exact Courant algebroid over a smooth mani-
fold M . Let ft be a one-parameter family of functions on M . Then, a one-parameter
family Gt of generalized metrics satisfies
∂
G −1 G = −2Rc(G, divµt ),
∂t
where µt = e−ft dVgt , if and only if the one-parameter family of pairs (gt , bt ) deter-
mined by Gt and the initial splitting associated to G0 satisfies the (−∇ft , 0)-gauge-
fixed generalized Ricci flow (4.13).
Proof. Recall from Definition 3.48 that divµt = divG − h2df, i. Thus, by
Proposition 4.24 and Lemma 3.32, Gt satisfies generalized Ricci flow with diver-
gence operator divµt as above if and only if (gt , bt ) satisfies (4.13) with (Xt , kt ) =
(−∇ft , 0).
4.4.2. Generalized Ricci solitons. In understanding geometric evolution
equations, a key role is played by the fixed points of the equation in determining
the possible existence and regularity behavior of the flow. Initially we would say
that the fixed points of the equation must satisfy Rc ≡ 0, and so are generalized
Ricci flat structures. More generally though, it is possible for the flow to evolve
along a one-parameter family of Courant symmetries, so that while the pairs (g, H)
are indeed evolving, at each time slice they are geometrically indistinguishable, and
these solutions are called solitons.
To begin we recall the derivation of the Ricci soliton equations. In particular,
Ricci solitons arise as special solutions of the gauge-fixed Ricci flow, upon impos-
ing two conditions: the first is that gt evolves in a self-similar way, that is, by
homotheties via the canonical dilation of a fixed metric
gt = (1 − 2tλ)g.
Here λ ∈ {−1, 0, 1} is a constant which determines the character of the soliton.
Second is that
−1
Xt = X
1 − 2λt
for a vector field X which is constant in time. From these conditions and (4.9), we
get the Ricci soliton equation
Rc + 21 LX g = λg,
which is expanding, steady, or shrinking depending on whether λ = −1, 0, 1, respec-
tively.
To adapt these ideas to generalized Ricci flow, we first observe that the space
of generalized metrics is not invariant under scaling. However, we can define a
canonical variation which corresponds to scaling the underlying metric.
Definition 4.26. Let λ ∈ {−1, 0, 1}. The canonical variation of a generalized
metric G on E is given by

0 (1 − 2tλ)−1 g −1
Gt,λ =
(1 − 2tλ)g 0
in the splitting determined by G. Note that

1 − 2tλ 0
GGt,λ =
0 (1 − 2tλ)−1
is the canonical dilation in the splitting determined by G.
Definition 4.27. Given (E, [, ], h, i) an exact Courant algebroid, we say that
G is a generalized Ricci soliton if there exists λ ∈ {−1, 0, 1} such that the canonical
variation Gt,λ solves the gauge-fixed generalized Ricci flow (4.11) with
−1
ζt = (X + B),
1 − 2λt
where X + B ∈ Lie Aut(E) is constant in time. We again call the soliton expanding,
steady, or shrinking depending on whether λ = −1, 0, 1, respectively.
Proposition 4.28. Given (E, [, ], h, i) an exact Courant algebroid, suppose G

is a generalized Ricci soliton. If λ 6= 0, then H ≡ 0 and the associated metric g
is an expanding or shrinking Ricci soliton. If λ = 0 then one has the equivalent
equations
0 = Rc − 14 H 2 + 21 LX g,
(4.14)
0 = 12 d∗g H − 21 B,
for B satisfying
d(B + iX H) = 0.
Thus, in particular,
0 = − 12 ∆d H + 21 LX H.
Proof. Applying the equations (4.12), we see that G is a generalized Ricci
soliton if there exists λ ∈ {−1, 0, 1} such that
1
λg = Rc − H 2 + 21 LX g,
4(1 − 2λt)2
1
0= d∗g H − B ,
1 − 2λt
for all t, where X + B ∈ Lie Aut(E). If λ 6= 0 then we obtain
λg = Rc + 12 LX g, H = 0,
that is, it is a Ricci soliton in the usual sense. The claimed equations (4.14) in
the case λ = 0 follow directly, and by taking the exterior derivative of the second
equation of (4.14) we obtain the final claimed equation.
Remark 4.29. Thus we see that in our formulation, shrinking and expanding
generalized Ricci solitons automatically have vanishing torsion. This arises in part
because there is no obviously natural way to introduce a scaling of the b-field
associated to a generalized metric. Nonetheless, in taking certain rescalings or
reparameterizations of generalized Ricci flow, it is reasonable to consider flows for
which the pairs (g, H) might both scale homothetically (cf. Definition 6.30 below
for nonsingular solutions). Although, we are only aware of rigidity results showing
that solitons for such flows automatically have H ≡ 0.
Remark 4.30. The dilation GGt,λ in Definition 4.26 has appeared in [54] under
the name of conformal change. Here we make the observation that, in order to be
invariant, this transformation depends upon a choice of isotropic splitting on E.
By adapting the point of view of Remark 4.23, it is possible to give a slightly
different formulation for the generalized Ricci soliton equations which involve the
action of a closed B-field instead. We state this as an equivalent definition, as it is
a consequence of the above discussion.
Definition 4.31. Given (E, [, ], h, i) an exact Courant algebroid, a generalized
metric G is a steady generalized Ricci soliton if there exists a vector field X and a
closed two-form k such that
0 = Rc − 41 H 2 + 21 LX g,
(4.15)
0 = 12 d∗g H + 21 iX H + k.
We furthermore say that the soliton is gradient if there exists a smooth function f
such that X = ∇f .
Question 4.32. Bryant has shown the existence of steady Ricci solitons with
rotational symmetry [28]. Are there examples of steady generalized Ricci solitons
with rotational symmetry and nontrivial torsion H?
One fundamental consequence of Perelman’s energy formula for Ricci flow is
that compact steady solitons for Ricci flow are automatically gradient. As we will
see below in Chapter 6, by adapting these energy functionals to generalized Ricci
flow, we can show that steady generalized Ricci solitons on compact manifolds are
automatically gradient, and moreover satisfy k = 0. It is an interesting question to
probe the possibilities for X and k in the noncompact setting. We end this section
with some further structural results on solitons. First, we show that generalized
Ricci solitons satisfy fundamental differential equations which generalize the fact
that Einstein manifolds automatically have constant scalar curvature using the
Bianchi identity.
Proposition 4.33. Let (M n , g, H, f ) be a gradient generalized Ricci soliton
with k = 0. Then
1 2
4 |H| + ∆f = 0,
(1) R −
(2) ∇ R − 125
|H|2 + |∇f |2 = 0,

1 2 2
(3) ∇ R − 12 |H| + 2∆f − |∇f | = 0.
Proof. The first equation follows by tracing the first equation of (4.15) with
g. For the second we begin by differentiating and applying (4.15) to yield

5 2 2 5 2
∇i R − 12 |H| + |∇f | = ∇i R − 12 ∇i |H| + 2∇i ∇j f ∇j f
(4.16)
2
= ∇i R − 125
∇i |H| + 12 H 2 − 2 Rc ij ∇j f.
But, using the Bianchi identity and property (1) we obtain
2
∇i R = 14 ∇i |H| − ∇i ∆f
2
= 14 ∇i |H| − ∇j ∇i ∇j f + Rcij ∇j f
2
(4.17) = 14 ∇i |H| + ∇j Rc − 41 H 2 ij + Rcij ∇j f
= 14 ∇i |H|2 + 21 ∇i R − 41 (div H 2 )i + Rcij ∇j f
2
= 21 ∇i |H| − 21 (div H 2 )i + 2 Rcij ∇j f,
where the last line follows by combining the ∇R terms and multiplying by 2. Also,
using Lemma 3.19 and the skew-symmetric piece of the soliton equation we obtain
2
− 21 (div H 2 )i = − 1
12 ∇i |H| + 21 (d∗ H)mn Himn
(4.18)
= − 1
12 ∇i |H|2 − 21 Hij
2
∇j f.
Plugging (4.17) and (4.18) into (4.16) yields the second claim. Taking twice the
gradient of (1) and subtracting (2) gives claim (3).
Remark 4.34. For the third equation in Proposition 4.33 it is possible to give
a more conceptual argument using the F functional to be introduced in Chapter
6. In particular, we know that a generalized steady soliton is a critical point for λ,
but by definition the relevant f extremizes the F -functional, thus by deforming f
1 2 2
by R − 12 |H| + 2∆f − |∇f | and perturbing the metric by scaling to preserve the
unit volume condition, the claim of Proposition 4.33 (3) follows from (6.6).
In the next result we prove that compact steady gradient Ricci solitons are
Einstein. Recall that, as a consequence of Perelman’s energy formula for Ricci
flow, compact steady solitons for Ricci flow are automatically gradient. Thus any
compact Ricci soliton is Einstein. However, there are compact steady generalized
Ricci solitons with nontrivial H and nontrivial f (cf. Remark 8.28). This shows
that the class of steady generalized Ricci solitons is strictly larger than the class of
generalized Ricci flat structures.
Proposition 4.35. Compact steady gradient Ricci solitons are Einstein.
Proof. In the case of H ≡ 0 we have the simplified soliton equation
Rc +∇2 f = 0.
Taking the divergence of this equation and using the Bianchi identity yields
0 = 12 ∇j R + ∇i (∇j ∇i f )
p
= 12 ∇j R + ∇j ∆f − Riji ∇p f
= 12 ∇j R + ∇j ∆f + Rcpj ∇p f.
Using the first item of Proposition 4.33 yields
0 = 12 ∇j R − ∇j R + Rcpj ∇p f.
Rearranging and taking another divergence yields

∆R = 2∇j Rcpj ∇p f
= h∇R, ∇f i + 2 Rc, ∇2 f
2
= h∇R, ∇f i − 2 |Rc|
≤ h∇R, ∇f i .
Since M is compact, R is constant by the strong maximum principle (Proposition
4.16).
4.5. Low dimensional structure

In this section we consider the generalized Ricci flow in dimensions n = 2, 3,
where the low dimensionality gives extra structure which renders the flow easier to
study. In dimension n = 2 there are no nontrivial three-forms, and thus we expect
that the flow must reduce to Ricci flow. We formalize this in the next proposition.
Proposition 4.36. Let M 2 be a Riemann surface, fix 0 = H0 ∈ Λ3 T ∗ = {0},
and let (gt , bt ) be a solution to generalized Ricci flow on M . Then (gt , bt ) satisfies
∂ ∂
(4.19) g = − 2 Rc = −Rg, b = 0.
∂t ∂t
In other words, gt is a solution of Ricci flow and bt ≡ b0 .
Proof. Since the manifold is two dimensional, Λ3 T ∗ = {0}, and so Ht ≡ 0 for
∂
all t. Comparing against the generalized Ricci flow system we see that ∂t b = 0 and
∂ 1
∂t g = −2 Rc. Also in two dimensions Rc = 2 Rg, finishing the lemma.
4.5. LOW DIMENSIONAL STRUCTURE 85
Remark 4.37. A complete picture of the Ricci flow on compact Riemann sur-
faces was first completed by Chow/Hamilton [49, 93]. In all cases, the volume-
normalized flow converges exponentially fast to a constant scalar curvature metric
g∞ . Later a different proof using energy methods from conformal geometry was
given by Struwe [174]. Using this result, and being slightly overpedantic, we could
observe that the generalized Ricci flow converges up to a generalized gauge trans-
formation to (g∞ , 0), where by an overall b-field action we can set b to zero.
In dimension n = 3, all three-forms are closed, and Λ3 T ∗ is a rank 1 bundle,
thus we can reduce the tensor H to an equivalent scalar function. We formalize
this and derive the relevant evolution equations in the next proposition.
Proposition 4.38. Let M 3 be a three-manifold, fix H0 ∈ Λ3 T ∗ and let (gt , bt )
be a solution of generalized Ricci flow on M . Define φt ∈ C ∞ (M ) by
Ht
φt = .
dVgt
Then
∂
g = − 2 Rc +φ2 g,
∂t
∂
φ = ∆φ + Rφ − 32 φ3 .
∂t
Proof. By direct computations in normal coordinates and expressing dVg =
dx1 ∧ dx2 ∧ dx3 one obtains
2
Hij = Hipq Hjrs g pr g qs = 2φ2 gij , |H|2 = 6φ2 .
Furthermore we can differentiate, using the generalized Ricci flow equation and
Lemma 5.2,
∂ ∂ H
φ=
∂t ∂t dVg

∆d H 1 ∂ H
= − 2 trg g
dVg ∂t dVg

2
= ∆φ + R − 41 |H| φ
= ∆φ + Rφ − 23 φ3 ,
where the penultimate line follows since, using that ⋆dVg = 1,
∆d H = −(d ⋆ d⋆)(φdVg ) = −d ⋆ dφ = − ⋆ d∗ dφ = ∆φdVg .

If we impose some symmetries on the underlying structure, it turns out that
we can realize other naturally occurring geometric evolution equations as special
cases of generalized Ricci flow.
Proposition 4.39. Let S 1 → P → M 2 be a principal S 1 bundle, and suppose
(gt , θt ) is a solution to
∂
g = − 2 Rc +2F 2 ,
(4.20) ∂t
∂
θ = − d∗g F.
∂t
Let gt = gt ⊕ θt ⊗ θt , and Ht = Ft ∧ θt . Then (g t , Ht ) is a solution of generalized

Ricci flow on P .
Proof. The proof relies on several basic computations, which are left as ex-
ercises. First, the Ricci tensor of a metric g = g ⊕ θ ⊗ θ on P is expressed in
natural block diagonal form as

Rc − 21 F 2 21 d∗ F
Rc = 1 ∗ 1 2 .
2d F 4 |F |
Furthermore, for H = F ∧ θ, we obtain
2
2F 0
H2 = 2 .
0 |F |
Putting these computations together and using the definition of g shows that g
satisfies the metric component of generalized Ricci flow.
A further computation shows that dg∗ (F ∧ θ) = d∗g F ∧ θ. Since M is two
dimensional we thus obtain
∂ ∂
H= F ∧ θ = −d∗g F ∧ θ + F ∧ (−d∗g F ) = −d∗g F ∧ θ = −dg∗ H,
∂t ∂t
as required.
The system of equations (4.20) is a natural coupling of the Ricci flow with the
Yang-Mills flow for a U (1) principal connection. This system is thus called Ricci
Yang-Mills flow, and was introduced in [173, 187] (up to a scalar factor on the
F 2 term which can be fixed by rescaling the fiber metric). We see here that in
the case of U (1) gauge group (or U (1) × · · · × U (1)), the flow is a special case of
generalized Ricci flow. This point of view is useful in understanding the relationship
of generalized Ricci flow to T -duality in Chapter 10. Further results on the global
existence and convergence properties of (4.20) appeared in [158, 159]. Some related
equations motivated by considerations in mathematical physics have appeared in
[62, 97]
Yet a different viewpoint on the equations (4.20) is as generalized Ricci flow on
a transitive Courant algebroid on M 2 obtained by reduction of the (twisted) exact
Courant algebroid T P ⊕ T ∗ P , as defined in [70]. This follows directly from the
fact that any three-form on M 2 identically vanishes, and suggests that generalized
Ricci flow is a robust geometric flow which is preserved under natural geometric
operations on Courant algebroids such as generalized reduction [31] or T-duality
(see Chapter 10). For stationary points of the flow, the preservation of the Ricci
flat condition under generalized reduction has been proved in [12]. We leave the
general case, concerning generalized Ricci flow, as an open question.
CHAPTER 5
Local Existence and Regularity
In this chapter we address basic aspects of the existence and regularity of

generalized Ricci flow. We saw in Lemma 4.3 that, with special gauge choices made,
the generalized Ricci flow appears to be a nonlinear parabolic system of equations.
Our first goal is to rigorously prove short-time existence of generalized Ricci flow
with arbitrary smooth initial data on compact manifolds, using the method of
gauge fixing. Given this we approach the question of how long the solution can
be extended, and what the obstructions are to global existence. We begin by
deriving evolution equations for the Riemannian and Bismut curvatures, which lead
to smoothing estimates in the presence of a bound on the Riemannian curvature
tensor. Using this we can show that the Riemannian curvature must blow up at
any finite time singularity of generalized Ricci flow, first established in [157]. We
end the chapter with a general discussion of the theory of compactness of sequences
of solutions of generalized Ricci flow, omitting most of the technical detail.
5.1. Variational Formulas

We record here elementary variational formulas related to Riemannian metrics,
the Riemannian curvature tensor, and the Bismut curvature tensor, necessary for
what follows.
Remark 5.1. We note here a notational convention we will use in several

places. In particular, given tensors A and B on a Riemannian manifold (M n , g),
we let A ⋆ B denote any tensor derived from A ⊗ B by raising and lowering indices
using g and taking contractions.
Lemma 5.2. Given (M n , g) a Riemannian manifold and gs a one-parameter

family of metrics such that
∂
gs = h, g0 = g,
∂s s=0
then
∂ 1
dVgs = 2 (trg h) dVg .
∂s s=0
p
Proof. Recall the local coordinate formula dVgs = det gij (s)dx1 ∧ · · · ∧ dxn .
Also, for a one-parameter family of positive definite symmetric matrices
As with
∂ ∂ −1
variation ∂s s=0
As = B, one has ∂s s=0 det(A s ) = det(A0 ) tr A0 B . Combining
87
88 5. LOCAL EXISTENCE AND REGULARITY
these facts yields

q
∂ ∂
dVgs = det gij (s)dx1 ∧ · · · ∧ dxn
∂s s=0 ∂s s=0

= √1 det gij tr g −1 h dx1 ∧ · · · ∧ dxn
2 det gij
1
= 2 (trg h) dVg ,
as required.

∂
gs = h, g0 = g,
∂s s=0
∂
H = dK, H0 = H,
∂s s=0
then
∂ 2
|H| = − 3 h, H 2 + 2 hdK, Hi .
∂s s=0
Proof. Computing in local coordinates and using the symmetries of H and g

we see that
∂ 2 ∂
|H| = g i1 j1 g i2 j2 g i3 j3 Hi1 i2 i3 Hj1 j2 j3
∂s s=0 ∂s s=0
i1 k
= − 3g hkl g lj1 g i2 j2 g i3 j3 Hi1 i2 i3 Hj1 j2 j3 + 2g i1 j1 g i2 j2 g i3 j3 (dK)i1 i2 i3 Hj1 j2 j3
= − 3 h, H 2 + 2 hdK, Hi ,
as required.

∂
gs = h, g0 = g,
∂s s=0
then
∂
(∇)kij = 21 g kl (∇i hjl + ∇j hil − ∇l hij ) ,
∂s s=0
∂ h i
l p p
Rijk = 21 g ql ∇i ∇k hjq − ∇i ∇q hjk − ∇j ∇k hiq + ∇j ∇q hik − Rijk hpq − Rijq hkp ,
∂s s=0
∂
Rjk = 21 g qi (∇i ∇j hkq + ∇i ∇k hjq ) − ∆hjk − ∇j ∇k trg h
∂s s=0
∂
R = − ∆ trg h + div div h − hh, Rci .
∂s s=0
5.1. VARIATIONAL FORMULAS 89
Proof. We fix a point p ∈ M and choose normal coordinates for g at p. Using

that ∂i gjk (p) = Γkij (p) = 0, we directly obtain
∂ ∂
(∇gs )kij = 1 kl
2g (∂i gjl + ∂j gil − ∂l gij )
∂s s=0 ∂s s=0
= 12 g kl (∂i hjl + ∂j hil − ∂l hij )
= 12 g kl (∇i hjl + ∇j hil − ∇l hij ) ,
as required.
For the variation of the curvature tensor we recall the coordinate formula and
differentiate, again using the vanishing of ∂i gjk (p) and Γkij (p), to yield
∂ ∂
l
Rijk = ∂i Γljk − ∂j Γlik + Γpjk Γlip − Γpik Γljp
∂s s=0 ∂s s=0

= 2 ∂i g ql (∇j hkq + ∇k hjq − ∇q hjk ) − ∂j g ql (∇i hkq + ∇k hiq − ∇q hik )
1
= 21 g ql [∇i ∇j hkq + ∇i ∇k hjq − ∇i ∇q hjk − ∇j ∇i hkq − ∇j ∇k hiq + ∇j ∇q hik ]

h i
p p
= 21 g ql ∇i ∇k hjq − ∇i ∇q hjk − ∇j ∇k hiq + ∇j ∇q hik − Rijk hpq − Rijq hkp ,
where the last line follows using the definition of the curvature tensor.
Given this, the variational formula for the Ricci tensor follows directly by con-
tracting. Finally we use this formula to obtain the variation of scalar curvature
via
∂ ∂
R= g jk Rjk
∂s s=0 ∂s s=0

= − g jp hpq g qk Rjk + 21 g jk g qi (∇i ∇j hkq + ∇i ∇k hjq ) − ∆hjk − ∇j ∇k trg h
= − ∆ trg h + div div h − hh, Rci ,
as required.
Lemma 5.5. Given (M n , g) a Riemannian manifold and (gs , Hs ) a one-parameter

family such that
∂ ∂
gs = h, g0 = g, Hs = dK, H0 = H,
∂s s=0 ∂s s=0
Furthermore set A = h + K. Then

∂ k 1 kl p p p

∇+ ij = 1 kl
2g ∇+
i Ajl + ∇+
j Ali − ∇+
l Aji + 2 g −H A
ij lp + H A
il jp + H A
jl pi
∂s s=0
∂ l l l l
˙+ ˙+ p ˙+
R+ ijk = ∇+
i ∇ − ∇+
j ∇ + Hij ∇ ,
∂s s=0 jk ik pk
∂ i i i
˙+ ˙+ p ˙+
R+ = ∇+
i ∇ − ∇+
j ∇ + Hij ∇ ,
∂s s=0 jk jk ik pk
i i i
∂ ˙+ ˙+ p ˙+
R+ = − h, Rc+ + g jk ∇+ i ∇ − ∇+
j ∇ + H ij ∇ .
∂s s=0 jk ik pk
Proof. We use the formula from Definition 3.12, expressed in components as

(Γ+ )kij = Γkij + 21 Hijl g kl , then apply the result of Lemma 5.4 to obtain
∂ ∂ 1
∇+ 1 kl
s = 2 g (∇i hjl + ∇j hil − ∇l hij ) + 2 Hijl g
kl
(5.1) ∂s s=0 ∂s s=0
= 21 g kl (∇i hjl + ∇j hil − ∇l hij ) + 21 dKijl g kl − 12 Hijl g kp hpq g ql ,
Using the definition of the Bismut connection, we see that
∇i hjk = ∂i hjk − Γlij hlk − Γlik hjl

= ∂i hjk − (Γ+ )lij − 21 Hijp g pl hlk − (Γ+ )lik − 21 Hikp g pl hjl
1 p 1 p
= ∇+
i hjk + 2 Hij hpk + 2 Hik hjp .
Hence
∇i hjl + ∇j hil − ∇l hij = ∇+ + +
i hjl + ∇j hil − ∇l hij
+ 21 g pq (Hijp hql + Hilp hjq + Hjip hql + Hjlp hiq − Hlip hqj − Hljp hiq )
= ∇+ + +
i hjl + ∇j hil − ∇l hij + g
pq
(Hilp hjq + Hjlp hiq ) .
Similarly we have
(dK)ijl = ∇i Kjl + ∇l Kij + ∇j Kli
= ∇+ + +
i Kjl + ∇l Kij + ∇j Kli

p
+ 12 Hij Kpl + Hilp Kjp + Hlip Kpj + Hljp Kip + Hjl
p p
Kpi + Hji Klp
p p p
= ∇+ + +
i Kjl + ∇l Kij + ∇j Kli + Hij Kpl + Hil Kjp + Hlj Kip
Plugging these computations into (5.1) yields the first equality.

Next, using the definition of the curvature tensor, we differentiate the general
coordinate formula to yield
(5.2)
∂ l ∂ n o
R+ ijk = ∂i (Γ+ )ljk − ∂j (Γ+ )lik + (Γ+ )lip (Γ+ )pjk − (Γ+ )ljp (Γ+ )pik
∂s s=0 ∂s s=0
= + l
∂i (Γ̇ )jk − ∂j (Γ̇+ )lik + (Γ̇+ )lip (Γ+ )pjk + (ΓB )lip (Γ̇+ )pjk
− (Γ̇+ )ljp (Γ+ )pik − (Γ+ )ljp (Γ̇+ )pik
+ p + p
= ∇+ + l + l + + l + l
i (Γ̇ )jk + (Γ )ij (Γ̇ )pk − ∇j (Γ̇ )ik − (Γ )ji (Γ̇ )pk
p
= ∇+ + l + + l + l
i (Γ̇ )jk − ∇j (Γ̇ )ik + Hij (Γ̇ )pk ,
+
as required. Taking the trace over i and l yields the variation for Rjk .
+ jk +
Lastly, using that R = g Rjk , we obtain
∂ ∂ +
R+ = g jk Rjk
∂s s=0 ∂s s=0

+ ∂ +
= − g hpq g qk Rjk
jp
+ g jk Rjk
∂s s=0
i i i
˙+ ˙+ p ˙+
= − h, Rc+ + g jk ∇+ i ∇ − ∇ +
j ∇ + H ij ∇ ,
jk ik pk
as required.
5.2. SHORT TIME EXISTENCE 91
5.2. Short time existence

In this section we establish the fundamental fact that for smooth initial data
on a compact manifold, there exists a smooth solution to the generalized Ricci flow
on some small time interval depending on the initial data. As we will show in
this section, the generalized Ricci flow equation is a degenerate parabolic equation,
and so the classical theorems concerning short-time existence of nonlinear parabolic
evolution equations do not directly apply. Many geometric evolution equations en-
counter this obstacle, with the degeneracies of the equation arising from an infinite
dimensional group of symmetries acting on solutions to the flow.
The resolution of this issue is via the “DeTurck method”, where the large
symmetry group is broken via the explicit introduction of extra terms tangent
to the group action which render the new equation strictly parabolic. As such,
one can now apply classical results to obtain a short-time solution to this flow. By
construction, the extra terms can be removed via an application of the group action,
producing finally a solution to the desired evolution equation. In the case of Ricci
flow, the relevant group is the diffeomorphism group of the underlying manifold.
As we have seen in §2.2, the symmetry group in generalized geometry is enlarged
to include b-field transformations, which indeed preserve the generalized Ricci flow.
Thus our proof exploits this enlarged group, but structurally follows the outline
described above precisely.
Theorem 5.6. Let (E, [, ], h, i) → M be an exact Courant algebroid over a
compact manifold M . Given G0 a generalized metric on E, there exists ǫ > 0 and
a unique smooth solution Gt to generalized Ricci flow with initial condition G0 on
[0, ǫ).
Proof. Step 1: Short time existence of gauge-fixed flow
The first step is to construct a solution of the gauge-fixed generalized Ricci
flow, specifically the metric equation of (4.13) and the induced equation on H,
for a particular choice of X and k = 0. Let g0 denote the Riemannian metric
determined by G0 . We will work in the splitting of E determined by G0 . Given a
metric g we define the vector field
X := X(g, g0 ) = trg (∇g − ∇g0 ) ,
which is expressed in components as

i i
X i = g pq (Γg )pq − (Γg0 )pq .
This vector field is precisely the one used in showing short-time existence of so-
lutions to Ricci flow, but we note here that it admits a natural interpretation in
generalized geometry as the difference of divergence operators associated to G and
G0 (cf. §2.4). Now define the differential operator
O1 (g, H) := −2 Rc + 12 H 2 + LX(g,g0 ) g.
We proceed to compute the symbol of O1 in phases. We fix one-parameter families
(gt , bt ) such that
d d
ds s=0 gs = h, ds s=0 H = dK.
We first observe that the variation of −2 Rc is already contained in Lemma 5.4.
Also, the dependence of the term H 2 on g is algebraic, and thus its linearization is
a zeroth order operator which we can ignore. To address the Lie derivative term,
we fix normal coordinates for g at some point p ∈ M we compute
d d k k

ds s=0 (LX(g,g0 ) g) = ds s=0 ∇i X gkj + ∇j X gki

d
= ds s=0
∂i X k gkj + Γ(g) ⋆ X + ∂j X k gki + Γ(g) ∗ X

= ∂i 12 g pq (∇p hqj + ∇q hpj − ∇j hpq ) + ∂j 12 g pq (∇p hqi + ∇q hpi − ∇i hpq )
= 12 g pq (∇i ∇p hqj + ∇i ∇q hpj + ∇j ∇p hqi + ∇j ∇q hpi − ∇i ∇j hpq − ∇j ∇i hpq )
= ∇i div hj + ∇j div hi − ∇i ∇i trg h + l.o.t,
where “l.o.t” denotes any term involving at most one derivative of h or K. Com-
bining this calculation with Lemma 5.4, we obtain that
d
(5.3) ds s=0 O1 (gs , H) = ∆d h + l.o.t.
On the other hand, examining O1 it is clear that the only dependence on b occurs
in the term H 2 , which is first-order in b, and hence one immediately obtains
d
(5.4) ds s=0 O1 (g, Hs ) = l.o.t.
Next, let
O2 (g, H) := ∆d H + LX(g,g0 ) H.
We compute, using the Bochner formula,
d
(5.5) ds s=0 O2 (g, Hs ) = ∆K + l.o.t.
Also, note that the variation of O2 with respect to g will be a certain second-
order differential operator, due to the second derivative terms appearing in ∆d H
and LX(g,g0 ) H. However, the precise form of this operator is not relevant to the
argument. In particular, combining this observation with (5.3)-(5.5) we obtain

h ∆ 0 h
σ D(g,H) (O1 , O2 ) = .
K ⋆ ∆ K
This shows that the system of equations
∂
g = O1 (g, H) = −2 Rc + 12 H 2 + LX(g,g0 ) g
(5.6) ∂t
∂
H = O2 (g, H) = ∆d H + LX(g,g0 ) H
∂t
is strictly parabolic. Applying the general result on existence of short-time solutions
to strictly parabolic evolution systems on compact manifolds, we conclude that
there exists ǫ > 0 and a unique solution to (5.6) on [0, ǫ).
Remark 5.7. An alternative proof of the short-time existence of generalized
Ricci flow could possibly be obtained directly from the gauge-fixed flow (4.10), for
X as above and B = dξt − iXt H0 for a suitable one-parameter family of one-forms
ξt ∈ T ∗ . This gauge fixing by inner symmetries of the exact Courant algebroid
is very natural, and we expect that it would provide a conceptual explanation of
Lemma 4.3.
Step 2: Gauge modification to solve generalized Ricci flow

5.2. SHORT TIME EXISTENCE 93
Fix G0 a generalized metric, with associated pair (g0 , H0 ). By the arguments

of Step 1, there exists ǫ > 0 and a solution to (5.6) on [0, ǫ). We must perform a dif-
feomorphism gauge transformation to properly gauge-fix the system. In particular,
we define a one-parameter family of diffeomorphisms φt via
∂
φt = − X(gt , g0 ) ◦ φt
(5.7) ∂t
φ0 = Id .
Now let
g t := φ∗t gt
H t = φ∗t H.
We first compute

∂ ∂ ∂
g = φ∗t gt + (φ∗t+s (gt ))
∂t t ∂t ∂s s=0

= φ∗t −2 Rc + 12 H 2 + LX(gt ,g0 ) g − L(φ−1 )∗ X(gt ,g0 ) (φ∗t gt )
t
2
= − 2 Rcg + 12 H ,
2
where the quantity H naturally employs the metric g. Next we compute

∂ ∗ ∂ ∂
H t = φt Ht + φ∗ Ht
∂t ∂t ∂s s=0 t+s

= φ∗t ∆d,g H + LX(g,g0 ) H − L(φ−1 )∗ X(gt ,g0 ) φ∗t (Ht )
t
= ∆d,g H.
We can now recover the 2-form potential b by direct integration. That is, we are
given b0 and define bt via
∂
bt = − d∗g t H.
∂t
It follows directly by comparing the evolution equations that H0 + dbt = H t for
all t. Thus we have proved that (g t , bt ) is a solution to generalized Ricci flow, as
claimed.
Step 3: Uniqueness via coupled harmonic map heat flow
A beautiful trick for establishing uniqueness of the Ricci flow equation reverses
the above procedure, by connecting solutions to the Ricci flow to the appropriately
gauge-fixed Ricci flow. The subtlety is to discover the appropriate gauge transfor-
mation, which arises by expressing the ODE defining φt in terms of the metric g,
after which it follows that φt is a solution to the harmonic map heat flow.
In particular, given (gt , bt ) a solution to generalized Ricci flow, let φt be the
one-parameter family of diffeomorphisms which is the unique solution to
∂
φt = ∆g t ,g0 φt
(5.8) ∂t
φ0 = Id .
Here ∆g t ,g 0 denotes the harmonic map Laplacian with source metric g t and target
metric g 0 . As this is a strictly parabolic equation and M is compact, solvability of
this equation for some short-time ǫ > 0 follows. A fundamental computation (cf.
[52] Remark 2.50) shows that, letting Id denote the identity map of M ,
(5.9) ∆g t ,g0 φt = ∆(φ−1 )∗ g Id = −X((φ−1 ∗

t ) g t , g 0 ).
t t ,g 0
Now set
gt = (φ−1 ∗
t ) gt
Ht = (φ−1 ∗
t ) Ht
A standard computation using the generalized Ricci flow equation and (5.9) shows
that gt satisfies the gauge-fixed generalized Ricci flow of (5.6).
i
To finish the uniqueness argument, we suppose (g it , bt ), i = 1, 2, are two so-
lutions to generalized Ricci flow with the same initial data. These have induced
i
torsions H t . By employing the gauge-fixing procedure described above, we pro-
duce two solutions (gti , Hti ) of the gauge-fixed system (5.6) with the same initial
data. This system is strictly parabolic, and enjoys uniqueness of solutions, thus
gt1 = gt2 , Ht1 = Ht2 for all times where either is defined. As discussed above, we can
now interpret the relevant diffeomorphisms used to gauge-fix, φit , as solutions to the
ODE (5.7). Since the two metrics gti agree, it follows that φ1t = φ2t for all relevant
1 2 ∂ 1 ∂ 2
times, and thus g 1t = g 2t , H t = H t for all times t. It follows then that ∂t bt = ∂t bt
1 2
for all times, and thus bt = bt for all t.
Remark 5.8. For various applications in geometry and mathematical physics

it would be desirable to study the generalized Ricci flow on complete, noncompact
spaces. In this setting even short-time existence is a delicate issue, and examples
such as a cylinder with arbitrarily small necks pinched at infinity suggest that a
general short-time existence result should not be attainable. With some form of
control over the initial metric, say bounded curvature, it seems likely that the flow
should admit short-time solutions, as is known in the case of Ricci flow [153].
Uniqueness of these solutions within the class of solutions with bounded curvature
was established in [44], (cf. [116]).
5.3. Curvature evolution equations

Fundamental to the study of generalized Ricci flow is to understand the evo-
lution of quantities determined by the associated curvature tensors. Though the
Bismut connection is in many ways the most relevant connection, the evolution
of the Riemannian curvature is relevant for some applications. As we will see,
while the key aspect of the proof of Theorem 5.6 was overcoming the lack of strict
parabolicity of the equation, all gauge-invariant quantities such as curvature, tor-
sion, and their derivatives, will obey strictly parabolic equations. The reason for
this difference is because of the extra differential identity satisfied by curvature,
namely the Bianchi identity, a manifestation of its gauge-invariance.
5.3.1. Riemannian curvature evolution.

5.3. CURVATURE EVOLUTION EQUATIONS 95
Lemma 5.9. Let (E, [, ], h, i) → M be an exact Courant algebroid. Given Gt a

solution to generalized Ricci flow, one has

∂
− ∆ Rmijkl = 2 (Qjikl − Qijkl − Qikjl + Qiljk )
∂t
− Rqjkl Rciq −Riqkl Rcjq −Rijql Rckq −Rijkq Rclq
2 2

+ 1
4 ∇i ∇k Hjl − ∇i ∇l Hjk − ∇j ∇k Hil2 + ∇j ∇l Hik
2 2
+ Rijkq Hql 2
+ Rijql Hkq .
where
(5.10) Qijkl := Rpijq Rpklq .
∂
Proof. We insert the evolution equation ∂t g = −2 Rc + 21 H 2 into the varia-
tional formula of Lemma 5.4, and compute the two terms separately. Formulaically
this takes the form
p
∂ ∂ ∂
Rmijkl = Rm gpl + Rmpijk g
(5.11) ∂t ∂t ijk ∂t pl
n p
o
= ∇2 Rc +∇2 H 2 ijk gpl + Rmpijk −2 Rcpl + 12 Hpl 2
.
First, the contribution of the Ricci tensor yields

(5.12)
p h i
p p
∇2 Rc g
ijk pl
= − ∇ ∇
i k Rc jl −∇ ∇
i l Rc jk −∇ ∇
j k Rc il +∇ ∇
j l Rc ik −Rijk Rc pl −Rijl Rc kp .
To see that the second derivative terms above are really ∆ Rm in disguise, we need
to use the Bianchi identity. In particular, using several applications of the Bianchi
identity, commuting derivatives and noting the definition of Q (5.10) we see
(5.13)
∆Rijkl = ∇p ∇p Rijkl
= − ∇p ∇j Rpikl − ∇p ∇i Rjpkl
= − ∇j ∇p Rpikl − Rjppq Rqikl − Rjpiq Rpqkl − Rjpkq Rpiql − Rjplq Rpikq
− ∇i ∇p Rjpkl − Ripjq Rqpkl − Rippq Rjqkl − Ripkq Rjpql − Riplq Rjpkq
= − ∇j ∇p Rklpi − Rcjq Rqikl + Rjpiq (Rplqk + Rpklq ) − Qjkil + Qjlik
− ∇i ∇p Rkljp + Ripjq (Rqlpk + Rqklp ) − Rciq Rjqkl + Qikjl − Qiljk
= ∇j ∇l Rpkpi + ∇j ∇k Rlppi − Rcjq Rqikl + Qjilk − Qjikl − Qjkil + Qjlik
+ ∇i ∇l Rpkjp + ∇i ∇k Rlpjp − Rciq Rjqkl + Qijkl − Qijlk + Qikjl − Qiljk
= − ∇j ∇l Rcki +∇j ∇k Rcli +∇i ∇l Rckj −∇i ∇k Rclj + Rcjq Riqkl + Rciq Rqjkl
+ 2 (Qijkl − Qjikl + Qikjl − Qiljk ) .
Next, plugging directly into the result of Lemma 5.4 we simply record
p 2
∇2 H 2 ijk gpl = 41 ∇i ∇k Hjl 2
− ∇i ∇l Hjk − ∇j ∇k Hil2 + ∇j ∇l Hik
2
(5.14) 2 2

−Rijkq Hql − Rijlq Hkq .
Combining (5.11)-(5.14) yields the claim.



∂
− ∆ Rc = 2 Rm ◦ Rc − 41 H 2 − Rc ◦ Rc − 14 H 2
∂t
h i
2
+ 41 ∆H 2 − ∇2 |H| + ∇ ◦ div H 2 .
Proof. We return to Lemma 5.4 to directly compute

∂ ∂ i
Rcjk = R
∂t ∂t ijk
i
= ∇2 Rc +∇2 H 2 ijk

(5.15) p
− 12 g qi Rijk p
−2 Rc + 21 H 2 pq + Rijq −2 Rc + 21 H 2 kp
i p
= ∇2 Rc +∇2 H 2 ijk + Rijk Rcip − Rcpj Rcpk
p
− 41 Rijk (H 2 )ip + 1
4 Rcpj (H 2 )kp .
Next we simplify, commuting derivatives and applying the second Bianchi identity,
(5.16)
i
∇2 Rc ijk
= g qi [−∇i ∇k Rcjq +∇i ∇q Rcjk +∇j ∇k Rciq −∇j ∇q Rcik ]
= ∆ Rcjk +∇j ∇k R − ∇j (div Rc)k − g qi ∇i ∇k Rcjq
p p
= ∆ Rcjk + 21 ∇j ∇k R − g qi ∇k ∇i Rcqj −g qi Rkij Rcpq −g qi Rkiq Rcjp
p
= ∆ Rcjk + 21 ∇j ∇k R − ∇k (div Rc)j − Rkij Rcip − Rcpk Rcpj
p
= ∆ Rcjk −Rkij Rcip − Rcpk Rcpj .
Also we simplify by commuting derivatives,
i
∇2 H 2 ijk = 41 g qi ∇i ∇k Hjq
2 2
− ∇i ∇q Hjk 2
− ∇j ∇k Hiq 2
+ ∇j ∇q Hik
2 2
= 14 ∆Hjk − 41 ∇j ∇k |H| + 14 ∇j (div H 2 )k + 41 g qi ∇i ∇k Hjq
2
2 2
= 41 ∆Hjk − 41 ∇j ∇k |H| + 14 ∇j (div H 2 )k + 41 g qi ∇k ∇i Hjq
2
(5.17) p p
+ 14 g qi Rkij (H 2 )pq + 41 g qi Rkiq 2
Hjp
h
2 2
= 41 ∆Hjk − ∇j ∇k |H| + ∇j (div H 2 )k + ∇k (div H 2 )j
i
p p
+g qi Rkij (H 2 )pq + g qi Rkiq 2
Hjp .
Combining (5.15)-(5.17) yields the result.


∂
− ∆ R = ∆R − 21 ∆ |H|2 + 12 div div H 2 + 2 Rc, Rc − 41 H 2 .
∂t
Proof. We leave as an exercise to derive this from Lemma 5.10. Here we

derive the result directly from the variation formula for scalar curvature in Lemma
5.4. Specifically, plugging the generalized Ricci flow equation into the variation
formula and applying Bianchi identities and Lemma 3.19 yields
∂
R = − ∆ trg −2 Rc + 21 H 2 + div div −2 Rc + 21 H 2 − Rc, −2 Rc + 21 H 2
∂t
2
= ∆R − 21 ∆ |H| + 12 div div H 2 + 2 Rc, Rc − 41 H 2 ,
as claimed.

Xk
∂
− ∆ ∇k Rm = ∇k+2 H 2 + ∇j Rm +H 2 ⋆ ∇k−j Rm,
∂t j=0

∂ 2 2
− ∆ ∇k Rm gt = − 2 ∇k+1 Rm gt + ∇k+2 H 2 ⋆ ∇k Rm
∂t
k
X
+ ∇j Rm +H 2 ⋆ ∇k−j Rm ⋆∇k Rm .
j=0
Proof. We compute, using the result of Lemma 5.9,
∂ k ∂
∇ Rm = {(∂ + Γ) . . . (∂ + Γ) Rm}
∂t ∂t
k−1
X
∂ ∂
= ∇j Γ ⋆ ∇k−1−j Rm + ∇k Rm
j=0
∂t ∂t
k−1
X
= ∇j ∇ Rc +H 2 ⋆ ∇k−1−j Rm + ∇k ∆ Rm + Rm⋆2 + Rm ⋆H 2 + ∇2 H 2
j=0
k
X
= ∇k ∆ Rm +∇k+2 H 2 + ∇j Rm +H 2 ⋆ ∇k−j Rm
j=0
k
X
k k+2 2

= ∆∇ Rm +∇ H + ∇j Rm +H 2 ⋆ ∇k−j Rm,
j=0
where the last line follows by commuting derivatives. This yields the first equation,
and the second follows as an elementary exercise, noting that the terms arising
from the variation of the inner product yield more terms of the form j = 0 in the
final sum.

Xk Xk
∂
− ∆ ∇k H = ∇j Rm ⋆∇k−j H + ∇j H 2 ⋆ ∇k−j H,
∂t j=0 j=1
k
X
∂ 2 2
− ∆ ∇k H gt
= − 2 ∇k+1 H + ∇j Rm ⋆∇k−j H ⋆ ∇k H
∂t j=0
k
X
+ ∇j H 2 ⋆ ∇k−j H ⋆ ∇k H.
j=1
Proof. We compute, using the result of Lemma 4.4 and the Bochner formula,
∂ k ∂
∇ H= {(∂ + Γ) . . . (∂ + Γ) H}
∂t ∂t
k−1
X
j ∂ k−1−j k ∂
= ∇ Γ⋆∇ H +∇ H
j=0
∂t ∂t
k−1
X
= ∇j ∇ Rc +H 2 ⋆ ∇k−1−j H + ∇k (∆H + Rm ⋆H)
j=0
k−1
X k
X

= ∇k ∆H + ∇j+1 Rm +H 2 ⋆ ∇k−1−j H + ∇j Rm ⋆∇k−j H
j=0 j=0
k
X k
X
= ∆∇k H + ∇j Rm ⋆∇k−j H + ∇j H 2 ⋆ ∇k−j H,
j=0 j=1
where in the last line we have reindexed one sum and commuted derivatives. This
proves the first formula, and again we leave the derivation of the second equation
as an exercise.
5.3.2. Bismut curvature evolution.
Lemma 5.14. Let (M n , g) be a smooth manifold with closed three-form H. Then
+
∆+ Rijkl = − ∇+ + + + + + + + + + + +
j ∇l Rcik +∇j ∇k Rcil +∇i ∇l Rcjk −∇i ∇k Rcjl
− Rc+ + + + + +
jq Rqikl − Rjpiq Rpqkl − Qjkil + Qjlik
+
− Ripjq +
Rqpkl − Rc+ + + +
iq Rjqkl + Qikjl − Qiljk
+ T1 (∇+ ∇+ (H 2 )) + T2 (∇+ (H ⋆ R+ )),

where
Q+ + +
ijkl = Rpijq Rpklq ,
X X
T1 (∇+ ∇+ H 2 )ijkl = ∇+ +
i ∇p Hkpq Hljq − ∇+ +
j ∇p Hkpq Hliq ,
σ(kpl) σ(kpl)
T2 (∇+ (H ⋆ R+ ))ijkl =

q + q + q +
− ∇+p Hij Rqpkl + Hpi Rqjkl + Hjp Rqikl

q q + q
+ ∇+j Hpk +
Rqlpi + Hlp Rqkpi + Hkl Rc+ + + +
qi − Rckq Hqli + Rpklq Hpqi + Rpkiq Hplq

+ +
+Rpliq Hqkp + Rplkq Hiqp − Rc+ lq Hikq

q q + q
+ ∇+i Hpk +
Rqljp + Hlp Rqkjp − Hkl Rc+ + + +
qj + Rckq Hqlj − Rpklq Hpqj − Rpkjq Hplq

+
−Rpljq +
Hqkp − Rplkq Hjqp + Rc+ lq Hjkq .
Proof. We work in local coordinates. Also, we drop the superscript on the

connection and curvature for notational simplicity, with all objects in this proof
associated to the Bismut connection. First we note that, using that H is closed we
can simplify formula (3.27) to
1
Rijkl − Rklij = 2 (∇k Hilj + ∇j Hkil + ∇l Hjki + ∇i Hljk )
= ∇k Hilj + ∇l Hjki − 21 (d∇ H)kilj
X
(5.18) = ∇k Hilj + ∇l Hjki + Hkip Hljp
σ(kil)
X
= ∇j Hkil + ∇i Hljk + Hjkp Hilp .
σ(jki)
Using Proposition 3.20 and (5.18) we obtain, with all objects associated to the
Bismut connection,
(5.19)
∆Rijkl = ∇p ∇p Rijkl
q q q
= ∇p −∇j Rpikl − ∇i Rjpkl − Hij Rqpkl − Hpi Rqjkl − Hjp Rqikl
= − ∇j ∇p Rpikl − Rjppq Rqikl − Rjpiq Rpqkl − Rjpkq Rpiql − Rjplq Rpikq
− ∇i ∇p Rjpkl − Ripjq Rqpkl − Rippq Rjqkl − Ripkq Rjpql − Riplq Rjpkq
q q q
− ∇p Hij Rqpkl + Hpi Rqjkl + Hjp Rqikl
 
X
= − ∇j ∇p Rklpi + ∇k Hpli + ∇l Hikp + Hkpq Hliq 
σ(kpl)
 
X
− ∇i ∇p Rkljp − ∇k Hplj − ∇l Hjkp − Hkpq Hljq 
σ(kpl)
− Rcjq Rqikl − Rjpiq Rpqkl − Qjkil + Qjlik
− Ripjq Rqpkl − Rciq Rjqkl + Qikjl − Qiljk
q q q
− ∇p Hij Rqpkl + Hpi Rqjkl + Hjp Rqikl .
We next analyze the highest order terms appearing in the first line of the final
equality above. First,
q q q
∇p Rklpi = − ∇l Rpkpi − ∇k Rlppi − Hpk Rqlpi − Hlp Rqkpi − Hkl Rqppi
q q q
= ∇l Rcki −∇k Rcli −Hpk Rqlpi − Hlp Rqkpi − Hkl Rcqi .
Next
q q q
∇p Rkljp = − ∇l Rpkjp − ∇k Rlpjp − Hpk Rqljp − Hlp Rqkjp − Hkl Rqpjp
q q q
= − ∇l Rckj +∇k Rclj −Hpk Rqljp − Hlp Rqkjp + Hkl Rcqj .
Next
∇p ∇k Hpli = ∇k ∇p Hpli − Rpkpq Hqli − Rpklq Hpqi − Rpkiq Hplq
= − ∇k d∗ Hli − Rpkpq Hqli − Rpklq Hpqi − Rpkiq Hplq
= ∇k (Rcli − Rcil ) − Rpkpq Hqli − Rpklq Hpqi − Rpkiq Hplq .
Next
∇p ∇l Hikp = ∇l ∇p Hikp − Rpliq Hqkp − Rplkq Hiqp − Rplpq Hikq
= − ∇l d∗ Hik − Rpliq Hqkp − Rplkq Hiqp − Rplpq Hikq
= ∇l (Rcik − Rcki ) − Rpliq Hqkp − Rplkq Hiqp − Rplpq Hikq .
Next
∇p ∇k Hplj = ∇k ∇p Hplj − Rpkpq Hqlj − Rpklq Hpqj − Rpkjq Hplq
= − ∇k d∗ Hlj − Rpkpq Hqlj − Rpklq Hpqj − Rpkjq Hplq
= ∇k (Rclj − Rcjl ) − Rpkpq Hqlj − Rpklq Hpqj − Rpkjq Hplq .
Lastly
∇p ∇l Hjkp = ∇l ∇p Hjkp − Rpljq Hqkp − Rplkq Hjqp − Rplpq Hjkq
= − ∇l d∗ Hjk − Rpljq Hqkp − Rplkq Hjqp − Rplpq Hjkq
= ∇l (Rcjk − Rckj ) − Rpljq Hqkp − Rplkq Hjqp − Rplpq Hjkq .
Inserting the above computations into (5.19) gives the result.

∂ k
+ + + p + p + p +
∇+ ij = − g kl ∇+ i Rc lj +∇ +
j Rc il −∇+
l Rc ij + g kl
H ij Rc pl −H il Rc pj −H jl Rc ip .
∂t
Proof. Using the result of Lemma 4.6 expressed in terms of the notation of
Lemma 5.5, we have
sym
Aij = − 2(Rc+ )ij + 2(Rc+ )skew
ij
= − Rc+ + + +
ij − Rcji + Rcij − Rcji
= − 2 Rc+
ji
Plugging this into the result of Lemma 5.5 gives the result.
5.4. SMOOTHING ESTIMATES 101
Proposition 5.16. Let (E, [, ], h, i) → M be an exact Courant algebroid. Given

Gt a solution to generalized Ricci flow, one has

∂ + +
−∆ Rijkl = R+ ⋆ R+ + ∇+ ∇+ (H 2 ) + ∇+ (H ⋆ R+ ).
∂t
Proof. Using Lemma 5.5 and the result of Lemma 5.15 we obtain
∂ ∂ + q
R+ ijkl = (R )ijk gql
∂t ∂t
q q q
p
= ∇+ i ∇˙+ − ∇ +
j ∇˙+ + H ij ∇
˙+ gql − 2(R+ )qijk (RcSym )ql
jk ik pk

+ + + q + q + q +
= ∇+
i − ∇+ + +
j Rclk +∇k Rcjl −∇l Rcjk + Hjk Rclq −Hjl Rckq −Hkl Rcqj
q
q q
− ∇+ j − ∇ +
i Rc +
lk +∇ +
k Rc +
il −∇ +
l Rc +
ik + H ik Rc +
lq −H il Rc +
kq −H kl Rc +
qi

p
+ Hij − ∇p Rckl +∇k Rclp −∇l Rckp + Hpk Rclr −Hpl Rckr −Hkl Rc+
+ + + + + + r + r + r
rp
− 2(R+ )qijk (RcSym )ql

19
X
=: Ai .
i=1
We first simplify
+ + +
A1 + A7 = g lp ∇+ + + + + +
j ∇i − ∇i ∇j Rckp = Rijkq Rcqp +Rijpq Rckq .
Next using Lemma 5.14 we have

+
A2 + A3 + A8 + A9 = ∆+ Rijkl + Rc+ + + + + +
jq Rqikl + Rjpiq Rpqkl + Qjkil − Qjlik
+
+ Ripjq +
Rqpkl + Rc+ + + +
iq Rjqkl − Qikjl + Qiljk
− T1 (∇2 (H 2 )) − T2 (∇(H ⋆ R)).

The claimed formula now follows, with the precise forms of the different expressions
implicit by the formulas above.
Question 5.17. A fundamental property of the Ricci flow is that the holonomy
of the Levi-Civita connection is preserved along the flow [95]. While the holonomy
can reduce in singular or infinite time limits, it will not reduce in finite time for
smooth solutions [117]. It is natural to ask if there is a similar behavior for the
generalized Ricci flow, namely, if there is a natural notion of holonomy in generalized
geometry which is preserved along the flow. Due to the lack of a canonical Levi-
Civita connection in generalized geometry (see Chapter 3), it is not even completely
clear what the correct notion of holonomy should be. A possible candidate is given
by the holonomy of the classical Bismut connection.
5.4. Smoothing estimates

A fundamental property of the heat equation on Rn is the concept of “smooth-
ing,” namely that a solution with low regularity becomes instantaneously smooth
and moreover obtains uniform estimates on all derivatives in terms of the elapsed
time. This is a very general feature of parabolic equations, where typically in the
nonlinear setting some kind of a priori control is needed before the smoothing ef-
fect can take over. As a nonlinear degenerate parabolic equation, the generalized
Ricci flow exhibits this behavior, where a bound on the curvature tensor implies a
smoothing effect for all derivatives of the curvature. This result sets the stage for
obtaining long time existence results for the flow, indicating that a bound on the
curvature tensor suffices to completely control the flow, which we prove in the next
section.
Theorem 5.18. Let (E, [, ], h, i) → M be an exact Courant algebroid with M
α
compact. Let Gt be a solution to generalized Ricci flow on [0, T0 ], 0 < T0 ≤ K
satisfying
n o
2
(5.20) sup |Rm| + |∇H| + |H| ≤ K.
M×[0,T0 ]
Given k ∈ N, there exists C = C(n, k, α) such that

CK
(5.21) sup ∇k Rm + ∇k+1 H + |H| ∇k H ≤ .
M×[0,T0 ] tk/2
Proof. We begin with the case k = 1. Consider the test function

2

2 2 2
F = t |∇ Rm| + ∇2 H + P |Rm| + Q |∇H| ,
where P ≥ 1 is to be chosen below. Combining Lemmas 5.12, 5.13, and the Cauchy-
Schwarz inequality yields
 
X1
∂ 2
− ∆ F = t −2 ∇2 Rm + ∇3 H 2 ⋆ ∇ Rm + ∇j Rm +H 2 ⋆ ∇1−j Rm ⋆∇ Rm
∂t j=0
 
X2 X 2
2
+ t −2 ∇3 H + ∇j Rm ⋆∇2−j H ⋆ ∇2 H + ∇j H 2 ⋆ ∇2−j H ⋆ ∇2 H 
j=0 j=1
2 2 2
+ (1 − 2P ) |∇ Rm| + (1 − 2P ) ∇ H

+ P ∇2 H 2 ⋆ Rm + Rm ⋆ Rm ⋆(Rm +H 2 )
 
1
X
j 1−j 2
+P  ∇ Rm ⋆∇ H ⋆ ∇H + ∇H ⋆ H ⋆ ∇H 
j=0
 
1
X
2
≤ t C ∇3 H |H| |∇ Rm| + C |∇ Rm| |Rm| + ∇j H 2 ∇1−j Rm |∇ Rm|
j=0
 
2
X 2
X 2
+ ∇j Rm ∇2−j H ∇2 H + ∇j H 2 ∇2−j H ∇2 H  − 2 ∇3 H 
j=0 j=1
2

2 3 2 2
− P |∇ Rm| − P ∇2 H + P ∇2 H 2 |Rm| + |Rm| + |Rm| |H|

2
+ P |Rm| |∇H| + |∇ Rm| |H| |∇H| + ∇H 2 |H| |∇H| .
Applying 5.20 and the arithmetic-geometric mean we obtain

∂ 2 2 2 P
− ∆ F ≤ |∇ Rm| (CKt − P ) + ∇ H CKt − + CP K 3 (1 + tK).
∂t 2
5.4. SMOOTHING ESTIMATES 103
α
Using that t ≤ K, we can choose P large with respect to universal constants and
α to obtain

∂
− ∆ F ≤ CK 3 .
∂t
α
Applying the maximum principle on this time interval we obtain, for t ≤ K,
sup F ≤ sup F + CK 3 t ≤ CK 2 (1 + α).

M×{t} M×{0}
This finishes the proof for the case k = 1.

We now prove the inductive step. Assuming the result holds for all values
p ≤ k − 1, define
2 2
k−1
X 2 2

k k k+1
F =t ∇ Rm + ∇ H + tj Pj ∇j Rm + ∇j+1 H .
j=0
We will derive the relevant differential inequality for F in stages. First of all we
apply Lemmas 5.12 and 5.13 to obtain, for times t > 0,
(5.22)

∂ 2 2
−∆ ∇k Rm + ∇k+1 H
∂t
k
X
2
≤ − 2 ∇k+1 Rm + ∇k+2 H 2 ⋆ ∇k Rm + ∇j Rm +H 2 ⋆ ∇k−j Rm ⋆∇k Rm
j=0
k+1
X k+1
X
2
− 2 ∇k+2 H + ∇j Rm ⋆∇k+1−j H ⋆ ∇k+1 H + ∇j H 2 ⋆ ∇k+1−j H ⋆ ∇k+1 H.
j=0 j=1
We now estimate some terms in (5.12), using the induction hypothesis. First
(5.23)
 
k
X
∇k+2 H 2 ⋆ ∇k Rm ≤ C  ∇k+2 H |H| + ∇k+1 H |∇H| + ∇j H ∇k+2−j H  ∇k Rm
j=2
1 k+2 2 2
2 2

≤ ∇ H + C |H|2 ∇k Rm + CK ∇k+1 H + ∇k Rm
2
K K
+ C (j−1)/2 (k+1−j)/2 ∇k Rm
t t
1 k+2 2 2 2
K3
≤ ∇ H + CK ∇k+1 H + ∇k Rm +C k .
2 t
Similarly we estimate
(5.24)
k+1
X
∇j Rm ⋆∇k+1−j H ⋆ ∇k+1 H
j=0
 
k−1
X
≤ C  ∇k+1 Rm |H| + ∇k Rm |∇H| + ∇j Rm ∇k+1−j H  ∇k+1 H
j=0
2 2
2 2

2
≤ 1
2 ∇k+1 Rm + C |H| ∇k+1 H + CK ∇k Rm + ∇k+1 H
K K 2
+C
j/2 (k−j)/2
∇k+1 H
t t
2
2 2
K3
≤ 21 ∇k+1 Rm + CK ∇k Rm + ∇k+1 H +C k .
t
We leave as an elementary exercise to estimate the two remaining terms of (5.22)
using the induction hypothesis via
(5.25)
k
X 2 2
K3
∇j Rm +H 2 ⋆ ∇k−j Rm ⋆∇k Rm ≤ CK ∇k Rm + ∇k+1 H +C k ,
j=0
t
k+1
X 2 2
K3
∇j H 2 ⋆ ∇k+1−j H ⋆ ∇k+1 H ≤ CK ∇k Rm + ∇k+1 H +C .
j=1
tk
Plugging (5.23) - (5.25) into (5.22) yields the differential inequality

(5.26)

∂ 2 2
−∆ ∇k Rm + ∇k+1 H
∂t
2 2
2 K3 2

≤ − ∇k+1 Rm − ∇k+2 H + CK ∇k+1 H + ∇k Rm. +C
tk
Using this differential inequality, we can choose the constants Pj by an elementary
induction argument to arrive at the differential inequality

∂
− ∆ F ≤ CK 3 ,
∂t
α α
for all times t ≤ K . Applying the maximum principle on the time interval [0, K ]
yields the estimate
sup F ≤ sup F + CK 3 t ≤ CK 2 (1 + α),
M×{t} M×{0}
which after rearranging yields the required estimate.
5.5. Results on maximal existence time

Having established the general short-time existence of solutions to generalized
Ricci flow on compact manifolds, the natural follow-up question is, how far can this
solution be extended? As we have seen in Example 4.10, the maximal interval of
5.5. RESULTS ON MAXIMAL EXISTENCE TIME 105
existence can certainly be finite. In this section we provide a basic criterion for de-
termining the singular time of a solution to generalized Ricci flow, namely showing
that at a singular time for generalized Ricci flow, the norm of the Riemann curva-
ture tensor must blow up. Verifying global existence of a solution to generalized
Ricci flow thus reduces to establishing a bound on the Riemann curvature tensor.
Lemma 5.19. Let M n be a smooth manifold. Given H ∈ Λ3 T ∗ and g a Rie-
mannian metric on M , one has
2 4
H2 g
≥ 1
n |H|g .
Proof. From the definition (3.23) it is clear that H 2 is a positive semidefinite

2
two tensor, and that one has trg H 2 = |H| . Using the orthogonal decomposition
◦
1
H2 = H2 + n trg H 2 g into tracefree and diagonal tensors, we obtain
◦ 2
2
H2 g
= H2 + 1
n trg H 2 g
◦ 2
2
= H2 + 1
n2 trg H 2 |g|2g
g
◦ 2
= H2 + 1
n |H|4g
g
1 4
≥ n |H|g ,
as required.
Proposition 5.20. Let (E, [, ], h, i) → M be an exact Courant algebroid. Let
α
Gt be a solution to generalized Ricci flow on [0, K ] such that
sup |Rm| ≤ K.
α
M×[0, K ]
α
Then there exists a constant C = C(n) such that for all t ∈ [0, K ],
2
sup t |H| ≤ C max{α, 1}.
M×{t}
Proof. Combining Lemma 4.4 with the Bochner formula and the estimate of
Lemma 5.19 yields the differential inequality

∂ ∂
|H|2 = 2 h∆d H, Hi − 3 g , H2
∂t ∂t

= 2 h∆H + Rm ⋆H, Hi − 3 −2 Rc + 21 H 2 , H 2
2 2 2
= ∆ |H| − 2 |∇H| + Rm ⋆H ⋆2 − 3
2 H2
≤ ∆ |H|2 − 2 |∇H|2 + CK |H|2 − 3
2n |H|4 ,
α
where C depends only on n. As a direct consequence we obtain, using that t ≤ K,

∂ 2 2 3 4
− ∆ t |H| ≤ (1 + CtK) |H| − 2n t |H|
∂t
3 2 4
≤ C max{α, 1} |H| − 2n t |H|

2 3 2
≤ |H| C max{α, 1} − 2n t |H| .
2
We note that at any point where t |H| ≥ 2C max{α,1}
3n =: C1 , the right hand side is
2
nonpositive. Since supM×{0} t |H| is obviously zero, it follows from the maximum
α
principle that the estimate supM×{t} ≤ C1 is preserved for t ∈ [0, K ].
Proposition 5.21. Let (E, [, ], h, i) → M be an exact Courant algebroid with

α
M compact. Suppose Gt is a solution to generalized Ricci flow on [0, K ] such that
sup |Rm| ≤ K.
α
M×[0, K ]
α
Then there exists a constant C = C(n) such that for all t ∈ [0, K ],
sup t |∇H| ≤ C max{α, 1}.
M×{t}
Proof. Combining Lemma 4.4 with the Bochner formula we obtain

∂ ∂ ∂
∇H = ∇ H +∇ H
∂t ∂t ∂t
1 2

= ∇ Rc − 4 H ⋆ H + ∇ (∆H + Rm ⋆H)
= ∆∇H + Rm ⋆∇H + (∇ Rm +∇H ⋆ H) ⋆ H.
A further calculation then yields, using Lemma 5.19,

∂ 2 ∂
|∇H| = 2 ∇H, ∇H
∂t ∂t
n ij kl o
− −2 Rc + 21 H 2 g kl g pq g rs + 3 −2 Rc + 21 H 2 g ij g pq g rs ∇i Hkpr ∇j Hlqs
= 2 h∆∇H + Rm ⋆∇H + (∇ Rm +∇H ⋆ H) ⋆ H, ∇Hi

+ Rm ⋆∇H ⋆2 − 21 (H 2 )ij g kl g pq g rs + 3(H 2 )kl g ij g pq g rs ∇i Hkpr ∇j Hlqs
2 2 2 2 2
≤ ∆ |∇H| − 2 ∇2 H + C |Rm| |∇H| + |∇ Rm| |H| |∇H| + C |H| |∇H|
α
where C depends only on n. As a direct consequence we obtain, using that t ≤ K,
and applying Proposition 5.20,

∂
− ∆ t |∇H|2 ≤ 1 + CtK + Ct |H|2 |∇H|2 + t |∇ Rm|2
∂t
2 2
≤ C max{α, 1} |∇H| + t |∇ Rm| .
2
We note that at any point where t |H| ≥ 2C max{α,1}
3n =: C1 , the right hand side is
nonpositive. Since supM×{0} t |H|2 is obviously zero, it follows from the maximum
2 α
principle that the estimate supM×{t} t |H| ≤ C1 is preserved for t ∈ [0, K ].
Lemma 5.22. Let M n be a smooth manifold, and suppose (gt , Ht ) is a smooth

one-parameter family of pairs on [0, T ), T < ∞, such that for any m1 , m2 ∈ N,
m2 m2

m1 ∂ m1 ∂
sup ∇ g + ∇ H < ∞.
M×[0,T ) ∂tm2 ∂tm2
Then there exists a smooth pair (gT , HT ) such that
(gt , Ht ) → (gT , HT )
∞
in the C topology.
5.5. RESULTS ON MAXIMAL EXISTENCE TIME 107
Proof. First we establish the existence of the limiting metric gT . Fix a vector
V ∈ Tp M , 0 ≤ t1 ≤ t2 < T and observe
Z t2
g(p, t2 )(V, V ) ∂g V V
log = , dt
g(x, t2 )(V, V ) t1 ∂t (p,t) |V | |V |
Z T
∂g
≤ dt
t1 ∂t (p,t) g(t)
≤ C(T − t1 ).
This shows that limt→T gt (V, V ) exists, and by polarization we obtain gT = limt→T gt .
With this uniform equivalence of g established we can obtain the limit HT . In
particular we fix V, X, Y ∈ Tp M , 0 ≤ t1 ≤ t2 < t and estimate
Z t2
∂H
H(p,t2 ) (V, X, Y ) − H(p,t1 ) (V, X, Y ) = (V, X, Y )dt
t1 ∂t (p,t)
Z t2
∂H
≤ |V |g(t) |X|g(t) |Y |g(t) dt
t1 ∂t (p,t) g(t)
≤ C(T − t1 ).
This shows limt→T H(V, X, Y ) exists, defining a continuous 3-form HT .
Showing strong convergence to and regularity of (gT , HT ) requires inductively
proving stronger estimates on the derivatives of g and H with respect to some fixed
background connection. These technical details are identical to that for Ricci flow,
and we refer the reader to ([51] Proposition 6.48) for the proof.
Theorem 5.23. Let (E, [, ], h, i) → M be an exact Courant algebroid. Given
G0 a generalized metric, let T ∈ R>0 ∪ {∞} denote the maximal extended real
number such that the solution to generalized Ricci flow with initial condition G0
exists smoothly on [0, T ). If T < ∞, then
lim sup sup |Rm| = ∞.
t→T M×{t}
Proof. We argue by contradiction, and suppose that

(5.27) lim sup sup |Rm| = K < ∞.
t→T M×{t}
Fix some 0 < δ < T . As the flow is smooth on [0, δ], for all m ∈ N there exists a
constant Cm such that

sup |∇m Rm| + ∇m+1 H + |H| |∇m H| ≤ Cm .
M×[0,δ]
Given t ∈ [δ, T ], by applying Theorem 5.18 on [t − δ, t] one obtains for a different

′
constant Cm ,

sup |∇m Rm| + ∇m+1 H + |H| |∇m H| ≤ Cm ′
Kδ −m/2 .
M×{t}
Thus there is a uniform bound on the curvature, torsion, and all their covariant
derivatives on [0, T ). With these uniform estimates in place, we can apply Lemma
5.22 to obtain the existence of a pair (gT , HT ) such that
lim (gt , Ht ) = (gT , HT ).
t→T
As M is compact, we may apply Theorem 5.6 to claim that there exists ǫ > 0 and
a smooth solution to generalized Ricci flow with initial condition (gT , HT ) on [0, ǫ).
Concatenating this solution with the original solution yields a smooth solution to
generalized Ricci flow on [0, T + ǫ), contradicting the maximality of T . Thus (5.27)
cannot hold, finishing the proof.
5.6. Compactness results for generalized metrics

In the study of generalized Ricci flow, as in any partial differential equation,
it is very useful to have a compactness principle for taking limits of sequences of
solutions. In the context of Riemannian geometry, the relevant limits are taken with
respect to the Cheeger-Gromov topology (cf. [42], [143] Chapter 10). Hamilton
later extended this definition to taking limits of one-parameter families of metrics
satisfying certain geometric bounds [94], with the goal of applying it to solutions
of Ricci flow. In this section we will briefly recall these notions, and sketch a
proof of an elementary extension of Hamilton’s result to one-parameter families
of generalized metrics, which will be instrumental in taking limits of solutions of
generalized Ricci flow in various contexts. We begin with a fundamental definition:
Definition 5.24. We say that a sequence {(Mk , gk , Ok )} of complete pointed
Riemannian manifolds converges in the C ∞ Cheeger-Gromov topology to a limit
(M∞ , g∞ , O∞ ) if there exists an exhaustion of M∞ by open sets Uk containing O∞
and a sequence of diffeomorphisms φk : Uk → Vk ⊂ Mk such that φk (O∞ ) = Ok ,
and φ∗k gk → g∞ uniformly on compact subsets in the C ∞ topology.
This definition highlights a fundamental point, which is that if we want to take
limits of metrics satisfying natural geometric conditions, i.e. conditions invariant
under the diffeomorphism group, then we must take the diffeomorphism invariance
explicitly into account and choose an appropriate diffeomorphism gauge before the
limits of the metric tensors can be taken in a classic sense. In view of our discussion
in Proposition 2.15 of the enlarged gauge group relevant to generalized geometry,
the following definition is inevitable.
Definition 5.25. We say that a sequence {(Ek , Mk , Gk , Ok )} of complete pointed
generalized Riemannian manifolds converges in the C ∞ generalized Cheeger-Gromov
topology to a limit (E∞ , M∞ , G∞ , O∞ ) if there exists an exhaustion of M∞ by open
sets Uk containing O∞ and a sequence of diffeomorphisms φk : Uk → Vk ⊂ Mk
covered by Courant algebroid automorphisms Fk such that
(1) φk (O∞ ) = Ok ,
(2) Fk−1 Gk Fk → G∞ uniformly on compact subsets in the C ∞ topology.
We observe that if gk and Hk denote the metric and three-forms canonically asso-
ciated to (Ek , Gk ) by Proposition 2.40, with g∞ , H∞ similarly defined, then
(1) φ∗k gk → g∞ uniformly on compact subsets in the C ∞ topology,
(2) φ∗k Hk → H∞ uniformly on compact subsets in the C ∞ topology.
Remark 5.26. A basic fact of Cheeger-Gromov convergence is invariance under
the diffeomorphism group. It follows from Propositions 2.14 and 2.40 that the
limiting process is invariant under the action of the enlarged symmetry group, in
particular under the action of closed B-fields. In particular, if {(Ek , Mk , Gk , Ok )}
converges to (E∞ , M∞ , G∞ , O∞ ), and Fk denotes a Courant automorphism of Ek ,
then the sequence {(Ek , Mk , Fk−1 Gk Fk , Ok )} also converges to (E∞ , M∞ , G∞ , O∞ ).
5.6. COMPACTNESS RESULTS FOR GENERALIZED METRICS 109
Definitions 5.24 and 5.25 give a notion of sequential convergence, which by a

standard method can be bootstrapped to define an actual topology on the space of
generalized Riemannian manifolds. We point to [148] for recent further progress
on understanding the moduli space of generalized metrics. However, the specifics
of this topology are not as directly useful here as the notion of convergence and
precompactness in this topology, which arises via the question of when a sequence of
generalized metrics admits a convergent subsequence. In the context of Riemannian
metrics, the fundamental question was answered by Cheeger [42]. We state here a
simplified version which suffices for our purposes, proved by Hamilton [94].
Theorem 5.27. ([94] Theorem 2.3) Given a sequence {(Mk , gk , Ok )} of com-
plete pointed Riemannian manifolds such that for all j ∈ N there exists Cj < ∞
such that
sup ∇jgk Rmgk gk
≤ Cj ,
Mk
and such that there exists δ > 0 so that

injgk (Ok ) ≥ δ,
then there exists a convergent subsequence.
Building upon this fundamental theorem, we prove a related theorem on com-
pactness for sequences of generalized metrics satisfying certain natural bounds.
Corollary 5.28. Given a sequence {(Ek , Mk , Gk , Ok } of complete pointed gen-
eralized Riemannian manifolds such that for all j ∈ N there exists Cj < ∞ such
that
n o
sup ∇jgk Rmgk g + ∇j Hk g ≤ Cj ,
k k
Mk
and such that there exists δ > 0 so that

injgk (Ok ) ≥ δ,
then there exists a convergent subsequence.
Proof. As the hypotheses of Theorem 5.27 are satisfied for the sequence {gk },
we can choose a subsequence such that the underlying Riemannian manifolds con-
verge in the C ∞ Cheeger-Gromov topology to a limiting Riemannian manifold
(M∞ , g∞ ). Using the derivative estimates for H, it follows from the Arzela-Ascoli
theorem that the sequence of pullbacks {φ∗k Hk } converges uniformly on compact
subsets of M∞ to a limiting three-form H∞ . As discussed in §2.1, this tensor H∞
defines a twisted Dorfman bracket on M∞ , which determines the exact Courant
algebroid E∞ . If we define G∞ on E∞ using g∞ and the canonical splitting of
T M∞ ⊕ T ∗ M∞ , then {(Ek , Gk , Ok )} converges to (E∞ , G∞ , O∞ ).
Remark 5.29. In Corollary 5.28, in addition to the topology of the limiting
space changing, the topology of the Courant bracket structure can also change. In
particular, fix M any smooth manifold which admits H a closed three-form such
that [H] 6= 0 ∈ H 3 (M, R). We can define generalized metrics G associated to g
and the canonical splitting on the twisted Courant algebroid defined by λH for
any λ ∈ R. Sending λ → 0 gives a sequence of generalized metrics with nontrivial
Ševera classes converging to a generalized metric on a Courant algebroid with trivial
Ševera class.
What is most relevant for studying generalized Ricci flow is to take limits not
just of generalized metrics, but of one-parameter families of generalized metrics.
The relevant notion from Riemannian geometry was introduced by Hamilton [94].
Given the clear relationship between Definitions 5.24 and 5.25, in the interest of
space we only state the definition in the context of generalized metrics.
Definition 5.30. We say that a sequence {(Ek , Mk , Gkt , Ok )} of one-parameter
families of complete pointed generalized Riemannian manifolds converges in the C ∞
t
generalized Cheeger-Gromov topology to a limiting one-parameter family (E∞ , M∞ , G∞ , O∞ ),
defined for t ∈ (α, ω), if there exists an exhaustion of M∞ by open sets Uk contain-
ing O∞ and a sequence of diffeomorphisms φk : Uk → Vk ⊂ Mk covered by Courant
automorphisms Fk such that
(1) φk (O∞ ) = Ok ,
(2) Fk−1 (Gk )t Fk → G∞ t
uniformly on compact subsets of M∞ × (α, ω) in the
∞
C topology.
As before, if gkt and Hkt denote the families of metrics and three-forms canonically
t
associated to (Ek , Gkt ) by Proposition 2.40, with g∞ t
, H∞ similarly defined, then
(1) φk gk → g∞ uniformly on compact subsets of M∞ × (α, ω) in the C ∞
∗ t t
topology,
(2) φ∗k Hkt → H∞ t
uniformly on compact subsets of M∞ × (α, ω) in the C ∞
topology.
Theorem 5.31. Let {(Ek , Mk , Gkt )} be a sequence of complete solutions to
generalized Ricci flow defined on time intervals Ik = (αk , ωk ) ⊂ R. Suppose
αk → α, ωk → ω, and for all j ∈ N there exists Cj < ∞ such that
n o
sup ∇jgk Rmgk g + ∇j Hk g ≤ Cj .
k k
Mk ×Ik
Furthermore, suppose there exists Ok ∈ Mk , t0 ∈ (α, ω), and δ > 0 so that

injgk (t0 ) (Ok ) ≥ δ,
for all sufficiently large k. Then there exists a subsequence of {(Ek , Mk , Gkt , Ok )}
converging to a pointed solution of generalized Ricci flow defined on the time interval
(α, β).
Proof. We give a brief sketch of the proof, leaving technical arguments as
lengthy exercises. First, the sequence of generalized Riemannian manifolds {(Ek , Mk , Gkt0 , Ok )}
satisfies the hypotheses of Corollary 5.28, and thus there is a subsequence converg-
ing to a limiting generalized Riemannian manifold (E∞ , M∞ , G∞ , O∞ ). Having
established convergence of the given time slices, one must use the generalized Ricci
flow equation together with the uniform derivative estimates to show convergence
for all times. The key point is to obtain uniform estimates for g and H on compact
subsets of spacetime. These estimates follow as in the proof of Lemma 5.22. The
uniform equivalence for the metrics is proved in Lemma 5.22, whereas for the higher
regularity estimates we refer to ([51] Proposition 6.48).
CHAPTER 6
Energy and Entropy Functionals
In the previous two chapters we motivated the generalized Ricci flow by show-
ing that it is a well-posed evolution equation with generalized Einstein metrics as
fixed points, and establishing fundamental analytic structure along the flow stem-
ming from its parabolic structure. For deeper applications it is natural to seek
monotone quantities along the flow which provide further control of the metric, as
well as convergence results for smooth long-time solutions. For the Ricci flow equa-
tion the discovery of such quantities was among the fundamental breakthroughs
of Perelman [140]. By the explicit incorporation of extra terms involving H, it is
possible to modify these constructions to yield energy and entropy functionals for
the generalized Ricci flow.
6.1. Generalized Ricci flow as a gradient flow

6.1.1. Energy Functional. As was discovered in [138] following Perelman’s
construction, the generalized Ricci flow is the gradient flow for the first eigenvalue
of a natural Schrödinger operator, directly generalizing Perelman’s energy mono-
tonicity for Ricci flow. The energy functional below already appeared in Chapter
3 as the generalized Einstein-Hilbert functional (cf. Definition 3.49).
Definition 6.1. Let M n be a smooth manifold. Given g a Riemannian metric,
H ∈ Λ3 T ∗ , dH = 0, and f ∈ C ∞ , let
Z
2 2
F (g, H, f ) = 1
R − 12 |H| + |∇f | e−f dV.
M
Furthermore, let
λ(g, H) = R inf F (g, H, f ).
{f | M
e−f dV =1}
Remark 6.2. Note that we have chosen to define the functional F using the
three-form H directly, as opposed to freezing a background H0 and defining the
functional using b ∈ Λ2 T ∗ , with the convention that H = H0 + db, as is done for
instance in [138].
Lemma 6.3. Given (M n , g, H) a smooth compact Riemannian manifold with
H ∈ Λ3 T ∗ , dH = 0, the infimum defining λ(g, H) is uniquely achieved, and is the
lowest eigenvalue of the Schrödinger operator
(6.1) − 4∆ + R − 1
12 |H|2 .
f
Proof. We note that we may reexpress the functional F in terms of u = e− 2
as
Z h i
2 2
F (g, H, u) = R− 1
12 |H| u2 + 4 |∇u| dV.
M
111
112 6. ENERGY AND ENTROPY FUNCTIONALS
R
If we minimize this functional over u satisfying M u2 dV = 1, a direct variational
computation with Langrange multipliers shows that u must satisfy

1 2
−4∆ + R − 12 |H| u = λu.
This minimizing function exists and is unique by ([145] XIII.12).
Remark 6.4. The Schrödinger operator (6.1), introduced in [138], has been re-
cently interpreted in [152] as an operator on half-densities to provide an alternative
definition of the entropy functional F in Definition 6.1.
6.1.2. Conjugate heat equation. Note that the definition of F requires
the extra parameter f , whereas the infimum in the definition of λ removes this
parameter. Obtaining evolution equations for λ is challenging to achieve directly
due to this infimum, since the function f achieving this infimum can change in a
discontinuous way as g and H change. For this reason it is more effective to work
with F directly. It is natural then to couple an evolution equation for the parameter
f to the generalized Ricci flow, the most natural choice being the conjugate heat
equation, which will preserve the unit volume condition.
Definition 6.5. Let (M n , gt , Ht ) be a solution to generalized Ricci flow. A
one-parameter family ut ∈ C ∞ (M ) is a solution to the conjugate heat equation if

∂ 2
(6.2) + ∆gt u = Ru − 14 |H| u.
∂t
We summarize this equation in terms of the conjugate heat operator ∗ , where
∂ 2
(6.3) ∗ := − − ∆gt + R − 41 |H| .
∂t
In particular, a function u solves the conjugate heat equation if and only if ∗ u = 0.
The operator ∗ is the L2 adjoint of the time-dependent heat operator (cf. Lemma
6.6 below)
∂
:= − ∆gt .
∂t
It is convenient to phrase many results below in terms of the generalized Ricci
flow in a general gauge (cf. Remark 4.23). Given a solution to (Xt , kt )-gauge-fixed
generalized Ricci flow, the relevant conjugate heat equation becomes

∂ 2
(6.4) + ∆gt − X u = Ru − 41 |H| u.
∂t
Observe that the sign on the left hand side of (6.2) indicates that, as a heat
equation, it is evolving in reverse time of the generalized Ricci flow. This is nat-
ural from the point of view of understanding monotonicity for F , as one wants to
construct a test function at some forward time, and pass it in a natural way to an
earlier time, where one has some geometric understanding. The next lemma shows
that the conjugate heat equation preserves the evolving f -weighted volume, and is
the L2 adjoint of the heat equation with respect to the evolving metric.
Lemma 6.6. Let (M n , gt , Ht ) be a solution to generalized Ricci flow.
d
R
(1) If φt is a solution of ∗ φ = 0, then dt M φdVg = 0.
6.1. GENERALIZED RICCI FLOW AS A GRADIENT FLOW 113
(2) Given φt , ψt ∈ C ∞ (M ) such that φ0 = φT = 0, one has

Z TZ Z TZ
(6.5) (φ) ψdV = φ (∗ ψ) dV.
0 M 0 M
Proof. For statement (1), differentiate directly using (6.2) and Lemma 5.2 to
see
Z Z
d ∂ 1 2
dt φdV = φ + φ −R + 4 |H| dV
M M ∂t
Z
= (−∆φ)dV
M
= 0.
Similarly we compute for general φ, ψ using the self-adjointness of the Laplacian,
Z Z
d ∂ ∂ 1 2
φψdV = φ ψ+φ ψ + φψ −R + 4 |H| dV
dt M M ∂t ∂t
Z
∂ ∂
= − ∆ φ ψ − φ − − ∆ + R − 14 |H|2 ψ dV
∂t ∂t
ZM Z
= (φ) ψdV − φ (∗ ψ) dV.
M M
Integrating this result over [0, T ] and applying the boundary condition yields the
result.
6.1.3. Monotonicity. We are now in a position to establish monotonicity of

the F -functional and also λ. We first require a general variational formula for F .
Lemma 6.7. Given (M n , g, H) a smooth manifold with H ∈ Λ3 T ∗ , dH = 0,
suppose gs , Hs , fs are one-parameter families such that
∂ ∂
g = h, g0 = g, H = dK, H0 = H
∂s s=0 ∂s s=0
∂
f = φ, f0 = f.
∂s s=0
Then
(6.6)
Z
d
F (gs , Hs , fs ) = − Rc + 41 H 2 − ∇2 f, h + 12 (−d∗ H − ∇f yH) , K
ds s=0 M
i −f
1 2 2 1
+ R − 12 |H| + 2∆f − |∇f | 2 trg h − φ e dV.
Proof. Combining Lemmas 5.2, 5.3 and 5.4 we obtain

(6.7)
Z
d
F (gs , Hs , fs ) = (−∆ trg h + div div h − hh, Rci
ds s=0 M
− 61 H 2 , h + 2 h∇φ, ∇f i − hh, ∇f ⊗ ∇f i
hdK, Hi + 1
4
−f
1 2 2 1
+ R − 12 |H| + |∇f | 2 trg h − φ e dV.
Integrating by parts yields the formulas

(6.8)
Z Z
−f

(−∆ trg h + div div h) e dV = trg h(−∆e−f ) + h, ∇2 e−f dV
M
ZM
2
= trg h(∆f − |∇f | ) + h, ∇f ⊗ ∇f − ∇2 f e−f dV.
M
Also
Z Z
− 61 hdK, Hi e−f dV = − 12 ∇i Kjk Hijk e−f dV
M M
Z

1
= 2 Kjk ∇i Hijk e−f dV
(6.9) ZM
1
= 2 Kjk (−d∗ Hjk − (∇f yH)jk ) e−f dV
Z M
= K, 12 (−d∗ H − ∇f yH) e−f dV.
M
Lastly we have
Z Z
2
(6.10) 2 h∇φ, ∇f i e−f dV = φ −2∆f + 2 |∇f | e−f dV.
M M
Inserting (6.8)-(6.10) into (6.7) and simplifying yields the result.
Finally we are ready to establish the monotonicity of F . The proof exploits the
diffeomorphism invariance of F in an essential way, and we refer back to Remark
4.23 for the notion of the general (Xt , kt )-gauge-fixed generalized Ricci flow.
Proposition 6.8. Let (M n , gt , Ht ) be a solution to the (Xt , kt )-gauge-fixed
generalized Ricci flow equations. Let ut denote a solution to the (Xt , kt )-gauge-
fixed conjugate heat equation, and let ft = − log ut . Then
Z h i
d 2
F (gt , Ht , ft ) = 2 Rc − 41 H 2 + ∇2 f + 21 |d∗ H + ∇f yH|2 e−f dV.
dt M
Proof. Since the functional F is gauge invariant, it suffices to prove the for-
mula for a specific choice of gauge. In particular, given ut and ft = − log ut as
in the statement, we can pull back g, H and f by time-dependent diffeomorphisms
to produce a solution of the (−∇f, 0)-gauge-fixed generalized Ricci flow. Using
equations (4.13) and (6.2) we see that this system of equations takes the form
∂
g = − 2 Rc − 21 H 2 + ∇2 f ,
∂t
∂
(6.11) H = − d d∗g H + ∇f yH ,
∂t
∂ 2
f = − ∆f − R + 14 |H| .
∂t
1 ∂ ∂
Observing that, for this variation, 2 trg ∂t g − ∂t f = 0, Lemma 6.7 implies the
result.
Remark 6.9. As a curiosity, observe that if gt , Ht , ft is a solution of (6.11),

then the volume form µt = e−ft dVgt is constant along the flow (cf. §3.7). Instead,
one could consider that µt evolves by the generalized scalar curvature in Definition
3.41 and then by Lemma 6.7 we still have monotonicity of F along the flow.
Corollary 6.10. Suppose M is compact and let (M, gt , Ht ) be a solution to
the (Xt , kt )-gauge-fixed generalized Ricci flow equations. Then for all t1 < t2 , one
has
λ(gt1 , Ht1 ) ≤ λ(gt2 , Ht2 ).
The case of equality holds if and only if (gt1 , Ht1 ) is a steady generalized Ricci
soliton with f realizing the infimum in the definition of λ. Moreover, the flow on
[t1 , t2 ] evolves by diffeomorphism pullback by the family generated by ∇f .
R
Proof. Choose ft2 such that M e−ft2 dVgt2 = 1, and F (gt2 , Ht2 , ft2 ) = λ(gt2 , Ht2 ),
which exists by Lemma 6.3. As M is compact, there exists a unique solution to
the conjugate heat equation on [t1 , t2 ] with boundary condition ut2 = −eft2 . Using
Proposition 6.8 we obtain
λ(gt1 , Ht1 ) ≤ F (gt1 , Ht1 , ft1 )
= F (gt2 , Ht2 , ft2 )
Z t2 Z h i
2 2
− 2 Rc − 41 H 2 + ∇2 f + 1
2 |d∗ H + ∇f yH| e−f dV
t1 M
= λ(gt2 , Ht2 )
Z t2 Z h i
2 2
− 2 Rc − 41 H 2 + ∇2 f + 1
2 |d∗ H + ∇f yH| e−f dV.
t1 M
The final term appearing above is manifestly nonpositive, so the claimed inequality
follows, with equality if and only if this term vanishes, which implies the soliton
equations hold.
Corollary 6.11. Given (M n , g, H) a compact steady generalized Ricci soliton

with vector field X and two-form k, then k = 0 and there exists f ∈ C ∞ (M ) such
that X = ∇f , i.e it is a gradient soliton.
Proof. As a steady soliton, the solution to generalized Ricci flow with initial
condition (g, H) evolves by pullback by a one-parameter family of diffeomorphisms.
Thus λ(gt , Ht ) is constant, and the result follows from Corollary 6.10.
6.1.4. Steady Harnack. The classical Harnack estimate for the heat equa-
tion yields a definite improvement in the oscillation of the solution over certain
domains in spacetime, and is a crucial point in establishing the regularity of this
equation. In the setting of Ricci flow Harnack-type estimates were discovered by
Perelman as pointwise versions of his energy and entropy monotonicity formulas,
and again play a crucial role in understanding the singularity formation of Ricci
flow. We extend these here to the setting of generalized Ricci flow.
Theorem 6.12. Let (M n , gt , Ht ) be a solution to generalized Ricci flow and
suppose ut satisfies the conjugate heat equation. Let f = − log u and let v = Su
where (cf. (3.39))
2 1 2
S = 2∆f − |∇f | + R − |H| .
12
Then
" #
2
1 1 ∗ 2
v = − 2 Rc − H 2 + ∇2 f
∗
+ |d H − ∇f yH| u.
4 2
Proof. Observe that f satisfies the equation

∂ 2 1 2
(6.12) f = − ∆f + |∇f | − R + |H| .
∂t 4
We compute

∂ 1 2 2 2 1 2
+ ∆ 2∆f = 2 2 Rc − H , ∇ f + ∆ −∆f + |∇f | − R + |H|
∂t 2 4

1 1
− div H 2 − ∇ |H|2 , ∇f + ∆∆f
2 4

2 2 2 1 2 1 2 2
= 4 Rc −H , ∇ f + 2∆ |∇f | − 2∆R + ∆ |H| + ∇ |H| − div H , ∇f .
2 2
Next

∂ 2 2 1 2
+ ∆ − |∇f | = − 2 ∇(−∆f + |∇f | − R + |H| ), ∇f
∂t 4

1
+ 2 Rc − H 2 , ∇f ⊗ ∇f + ∆ |∇f |2
2

2 1 2
= 2 ∇ ∆f − |∇f | + R − |H| , ∇f
4

1 2 2
+ H − 2 Rc, ∇f ⊗ ∇f − ∆ |∇f | .
2
Then we have

∂ 2 1 2 1 1
+ ∆ R = 2∆R + 2 |Rc| − ∆ |H| + div div H 2 − Rc, H 2 .
∂t 2 2 2
Next

∂ 1 2 1 3 2
+∆ − |H| = − 6 Rc − H 2 , H 2 + 2 h∆d H, Hi + ∆ |H|
∂t 12 12 2

1 2 1 1 1
= H − Rc, H 2 − h∆d H, Hi − ∆ |H|2 .
8 2 6 12
Combining these yields

∂ 1 1 1
+ ∆ S = ∆ |∇f |2 − ∆ |H|2 − h∆d H, Hi + div div H 2
∂t 12 6 2

1 2 1
+ ∇ |H| − div H 2 , ∇f + H 2 − 2 Rc, ∇f ⊗ ∇f
2 2
2
1 2
+ 2 Rc − H 2 + ∇2 f − 2 ∇2 f
4

2 1 2
+ 2 ∇ ∆f − |∇f | + R − |H| , ∇f .
4
Using Lemma 3.19 one has
1 1 1 2 1 2
div div H 2 − h∆d H, Hi − ∆ |H| = |d∗ H| ,
2 6 12 2
and

1 1D E
∇ |H|2 − div H 2 , ∇f = ∇ |H|2 , ∇f − hd∗ H, ∇f yHi .
2 3
Also one has
2 2
∆ |∇f | − 2 ∇2 f − 2 hRc, ∇f ⊗ ∇f i = 2 h∇∆f, ∇f i .
Therefore
2
∂ 1 1 ∗
(6.13) + ∆ S = 2 Rc − H 2 + ∇2 f + |d H − ∇f yH|2 + 2 h∇S, ∇f i .
∂t 4 2
It follows that

∂ ∂
+∆ v = + ∆ (Su)
∂t ∂t

∂ ∂
= +∆ S u+S + ∆ u + 2 h∇S, ∇ui
∂t ∂t
" #
2
1 2 2 1 ∗ 2
= 2 Rc − H + ∇ f + |d H − ∇f yH| u
4 2

2 1 2 2
+ S ∆f − |∇f | + R − |H| u + S −∆f + |∇f | u
4
" #
2
1 2 2 1 ∗ 2 1 2
= 2 Rc − H + ∇ f + |d H − ∇f yH| u + v R − |H| ,
4 2 4
as claimed.
One application of this fundamental calculation is a pointwise estimate for S

which can be interpreted as a kind of Harnack estimate for generalized Ricci flow.
Corollary 6.13. Let (M n , gt , Ht ) be a solution to generalized Ricci flow on a

compact manifold. With u and f defined as above, one has that sup S is a nonde-
creasing function of t.
Proof. We let τ = −t denote a backwards time parameter. Using Theorem

6.12 we compute
∂ ∂
S= − S
∂τ ∂t
2
1 1 2 1 2
= − 2 Rc − H 2 + ∇2 f − |d∗ H − ∇f yH| − RS + |H| S
4 2 4

∆v S 2 1 2
+ + ∆f − |∇f | + R − |H| u
u u 4
2 2
h∇v, ∇ui v |∇u| 1
= ∆S + 2 2
−2 3
− 2 Rc − H 2 + ∇2 f
u u 4
1 ∗
− |d H − ∇f yH|2
2
2
h∇u, ∇Si 1 1 ∗ 2
= ∆S + 2 − 2 Rc − H 2 + ∇2 f − |d H − ∇f yH| .
u 4 2
Thus, in terms of the backwards time parameter τ , the function S is a subsolution

of a parabolic equation, and so by the maximum principle (cf. Proposition 4.19),
the supremum of S is nonincreasing in τ , and is thus nondecreasing in t.
Remark 6.14. We observe that Proposition 6.8 follows immediately from The-
∂ 2
orem 6.12. In particular, since ∂t dV = −R + 14 |H| along a solution to generalized
Ricci flow, it follows from Theorem 6.12 that
( " # )
2
∂ 1 2 2 1 ∗ 2
(SudV ) = −∆(Su) + 2 Rc − H + ∇ f + |d H − ∇f yH| u dV.
∂t 4 2
Integrating this result over M yields Proposition 6.8.
6.2. Expander entropy and Harnack estimate

The quantities and results of §6.1 are designed to capture the behavior of steady
solitons, which are flow lines which evolve purely by diffeomorphisms. By appro-
priately modifying the solution to the conjugate heat equation, we can obtain an
entropy functional and Harnack inequality as in §6.1 which is fixed on expand-
ing solitons. We simplify the discussion however by deriving the main differential
inequality and observing the monotonicity and Harnack inequalities as corollaries.
These quantities were originally discovered in [157], generalizing Feldman-Ilmanen-
Ni’s adaptation [64] of Perelman’s quantities to the expanding setting.
Theorem 6.15. Let (M n , gt , Ht ) be a solution to generalized Ricci flow on

[T, T ′] and suppose ut is a solution to the conjugate heat equation. Let f+ be defined
e−f+
by u = n , and let
(4π(t−T ))2

2 1 2
v+ = (t − T ) 2∆f+ − |∇f+ | + R − |H| − f+ + n u.
12
6.2. EXPANDER ENTROPY AND HARNACK ESTIMATE 119
Then
!
2
1 g 1
∗ v+ = − 2(t − T ) Rc − H 2 + ∇2 f+ + + |d∗ H − ∇f+ yH|2 u
4 2(t − T ) 4
1 2
− |H| u.
6
Proof. We assume without loss of generality that T = 0. Let V := 2∆f+ −
2 1 2
|∇f+ | + R − 12 |H| . To compute the relevant evolution equation for V we first
observe that f+ satisfies
∂f+ 1 n
(6.14) = −∆f+ + |∇f+ |2 − R + |H|2 − .
∂t 4 2t
Note that this differs from the variation of f in the steady setting only by the
∂
addition of the constant term (see (6.12), the computation of ( ∂t + ∆)V is identical
to that in Theorem 6.12, thus we may jump to line (6.13) to yield
(6.15)
2
∂ 1 1 ∗ 2
+ ∆ V = 2 Rc − H 2 + ∇2 f+ + |d H − ∇f+ yτ | + 2 h∇V, ∇f+ i .
∂t 4 2
Now let W := tV − f+ + n. Using (6.15) we get
2
∂ 1 t ∗ 2
+ ∆ W = V + 2t Rc − H 2 + ∇2 f+ |d H − ∇f+ yH|
+
∂t 4 2
2 1 2 n
+ 2t h∇V, ∇f+ i − |∇f+ | + R − |H| −
4 2t
1 2 n
= 2∆f+ + 2R − |H| + + 2 h∇W, ∇f+ i
3 2t
2
1 t 2
2t Rc − H 2 + ∇2 f+ + |d∗ H − ∇f+ yH| .
4 2
2
1 g t 2
= 2t Rc − H 2 + ∇2 f+ + + |d∗ H − ∇f+ yH|
4 2t 2
1
+ |H|2 + 2 h∇W, ∇f+ i .
6
Finally we can compute

∂ ∂
+ ∆ v+ = + ∆ (W u)
∂t ∂t

∂ ∂
= +∆ W u+W + ∆ u + 2 h∇W, ∇ui
∂t ∂t
2
1 g t
= 2t Rc − H 2 + ∇2 f+ + + |d∗ H − ∇f+ yH|2
4 2t 2

1 2 1 2
+ |H| + 2 h∇W, ∇f+ i u + W Ru − |H| u − 2 h∇W, ∇f+ i u
6 4
2
1 g t ∗ 2 1 2
= 2t Rc − H 2 + ∇2 f+ + u+ |d H − ∇f+ yH| u + |H| u
4 2t 2 6
1 2
+ Rv+ − |H| v+
4
and the result follows.
Definition 6.16. Let (M n , g, H) be a smooth manifold with H ∈ Λ3 T ∗ , dH =

e−f+
0. Fix u ∈ C ∞ , u > 0, and define f+ via u = n . The expanding entropy
2 (4π(t−T ))
associated to this data is
Z
2 1 2
W+ (g, H, f+ , σ) := σ |∇f+ | + R − |H| − f+ + n udV.
M 12
Furthermore, let
µ+ (g, H, σ) := n R
inf n o W+ (g, H, f+ , σ) ,
f+ | M
e−f+ (4πσ)− 2 dV =1
and
ν+ (g, H) := sup µ+ (g, H, σ) .

σ>0
Corollary 6.17. Let (M n , gt , Ht ) be a solution to the (Xt , kt )-gauge-fixed

generalized Ricci flow equations. Let ut denote a solution to the (Xt , kt )-gauge-
e−f+
fixed conjugate heat equation, and define f+ via u = n . Then
2 (4π(t−T ))
d
W+ (g, H, f+ , t − T )
dt
Z " 2
! #
g 1 2 1 2
= (t − T ) 2 Rc − 41 H 2 + ∇2 f+ + + |d∗ H + ∇f+ yH| + |H| udV.
M 2(t − T ) 2 6
Proof. As in the proof of Proposition 6.8, it suffices to consider the flow in any
∂ 2
specific gauge, and in this case we pick the standard gauge. Since ∂t dV = R− 41 |H|
along a solution to generalized Ricci flow, it follows from Theorem 6.15 that
" ! #
2
∂ g 1 2 1 2
(v+ dV ) = (t − T ) 2 Rc − 14 H 2 2
+ ∇ f+ + + |d∗ H + ∇f+ yH| + |H| udV.
∂t 2(t − T ) 2 6
Integrating this result over M yields the result.
Corollary 6.18. Let (M n , gt , Ht ) be a solution to generalized Ricci flow on

a compact manifold, t ∈ [T, T ′ ]. Let ut be a solution to the conjugate heat equation
and define v+ as in Theorem 6.15. Then sup vu+ is nondecreasing in t.
6.2. EXPANDER ENTROPY AND HARNACK ESTIMATE 121
Proof. Without loss of generality assume T = 0. We compute with respect

to the backwards time parameter τ = −t,
∂ v+ ∂ v+
= −
∂τ u ∂t u
2
1 g 1 2 1 2
= − 2t Rc − H 2 + ∇2 f+ + − t |d∗ H − ∇f+ yH| − |H|
4 2t 2 6

v+ 1 2 v+ ∆v+ v+ 1 2
−R + |H| + + 2 ∆u + Ru − |H| u
u 4 u u u 4
v h∇v+ , ∇ui v+ |∇u|
2
+
=∆ +2 − 2
u u2 u3
2
1 g 1 2 1 2
− 2t Rc − H 2 + ∇2 f+ + − t |d∗ H − ∇f+ yH| − |H|
4 2t 2 6
v ∇u, ∇ vu+ 1 g
2
+
=∆ +2 − 2t Rc − H 2 + ∇2 f+ +
u u 4 2t
1 ∗ 1
− t |d H − ∇f+ yH|2 − |H|2 .
2 6
The result follows by the maximum principle as in Corollary 6.13.
Proposition 6.19. Let (M n , gt , Ht ) be a solution to generalized Ricci flow on

a compact manifold. The infimum in the definition of µ+ is attained by a unique
function f . Furthermore µ+ (gt , Ht , t − T ) is monotonically nondecreasing along
generalized Ricci flow, and is constant only on a generalized Ricci expander, i.e.
H ≡ 0 and g is a Ricci expander.
1
e−f+
Proof. Using the relationship u = n and then setting w = u 2 , we
(4π(t−T ) 2
can express, up to a constant shift,
Z
2 1 2
W+ (g, H, w, σ) = σ 4 |∇w| + Rw2 − |H| w2 + w2 log w2 dV.
M 12
This functional is coercive for the Sobolev space W 1,2 and convex. By [146] this
functional has a unique smooth minimizer. Thus at any time t0 we can choose the
smooth minimizer ut0 defining µ+ , and construct a solution to the conjugate heat
equation on [0, t0 ] with this final value. It follows using Corollary 6.17 that, at time
t0 , one has
d
µ+ (gt , Ht , t − T )
dt
d
= W+ (gt , Ht , ut , t − T )
dt "
Z 2
1 g
= 2 (t − T ) Rc − H 2 + ∇2 f+ +
M 4 2(t − T )

1 2 1 2
+ |d∗ H − ∇f+ yH| + |H| udV.
4 12
The monotonicity and rigidity statements follow easily from this formula, which
holds for the corresponding choice of u at all times t.
6.3. Shrinking Entropy and local collapsing

Having dealt with the steady and expanding soliton cases, it is natural to seek
a monotone entropy quantity which is fixed along a shrinking soliton. Following the
ideas of the previous sections, a natural entropy functional is suggested which is a
generalization of Perelman’s original entropy which includes H. However, as we will
see below, this entropy is not monotone along generalized Ricci flow. Nonetheless,
by assuming the existence of a certain kind of subsolution to the heat equation along
the solution, which exists in many applications, it is possible to further modify to
obtain a monotone quantity. In this setting we adapt Perelman’s ideas to obtain a
κ-noncollapsing result for solutions to generalized Ricci flow.
[0, T ] and suppose ut is a solution to the conjugate heat equation. Let f− be defined
e−f−
by u = n , and let
(4π(T −t)) 2

2 1 2
v− = (T − t) 2∆f− − |∇f− | + R − |H| + f− − n u.
12
Then
!
2
1 g 1
∗ v− = 2(T − t) Rc − H 2 + ∇2 f− − + |d∗ H − ∇f− yH|2 u
4 2(T − t) 4
1 2
− |H| u.
6
Proof. This is an elementary modification of Theorem 6.15, and we leave the
details as an exercise.
2
Remark 6.21. The presence of the final term − 61 |H| u above makes it more
difficult to use Theorem 6.20 without further hypotheses. Nonetheless we will
be able to modify this quantity further to obtain a monotone entropy in some
settings. The discovery of a monotone entropy quantity fixed on shrinking solitons
for generalized Ricci flow in full generality remains an important open problem.
Definition 6.22. Let (M n , g, H) be a smooth manifold with H ∈ Λ3 T ∗ , dH =
e−f−
0. Fix u ∈ C ∞ , u > 0, and define f− via u = n . The shrinking entropy
2 (4π(T −t))
associated to this data is
Z
1
W− (g, H, f− , τ ) := τ |∇f− |2 + R − |H|2 + f− − n udV.
M 12
Furthermore, let
µ− (g, H, τ ) := n R
inf n o W− (g, H, u, τ ) ,
f− | M
e−f− (4πτ )− 2 dV =1
and
ν− (g, H) := sup µ− (g, H, τ ) .
τ >0
The evolution equation for W− follows as an immediate application of Theorem

6.20, and will not be monotone in general. To modify this to obtain a monotone
energy we make a further definition.
6.3. SHRINKING ENTROPY AND LOCAL COLLAPSING 123
Definition 6.23. Let M n be a compact manifold, and suppose (gt , Ht ) is a

solution to generalized Ricci flow. We say that a one-parameter family of functions
φ is a torsion-bounding subsolution if φ ≥ 0 and
∂ 2
φ ≤ ∆φ − |H| .
∂t
The following basic lemma indicates the usefulness of a torsion-bounding sub-
solution.
Lemma 6.24. Let M n be a compact manifold. Given (gt , Ht ) a solution to
generalized Ricci flow, ut a solution of the conjugate heat equation, and φt a solution
to
∂
φ = ∆φ + ψ,
∂t
one has
Z Z
∂
φudV = ψudV.
∂t M M
Proof. We compute
Z Z h i
∂ 1 2
φudV = φ̇u + φu̇ + φu −R + 4 |H| dV
∂t M
ZM
= [(∆φ + ψ) u + φ∆u]
ZM
= ψudV,
M
as required.
Definition 6.25. Let (M n , gt , Ht ) be a solution to generalized Ricci flow on
a compact manifold, and suppose furthermore that φt is a smooth one-parameter
family of functions. Let the augmented entropy be
W− (g, H, φ, f− , τ )
Z
2 1 2 −n
:= τ |∇f− | + R − |H| − φ + f− − n (4πτ ) 2 e−f− dV.
M 12
Moreover, let
µ− (g, H, φ, τ ) = R inf n W− (g, H, φ, f− , τ )
{f− | M
(4πτ )− 2 e−f− =1}
Proposition 6.26. Let (M n , gt , Ht ) be a solution to generalized Ricci flow on

a compact manifold, and suppose that φt is a torsion-bounding subsolution. Then,
setting τ = T − t for some time T we have
∂
W− (g, H, φ, f− , τ )
∂t
Z " 2
#
1 1 τ ∗ 2 5 2 n
= 2τ Rc − H 2 + ∇2 f− − g + d H + i∇f− H + |H| (4πτ )− 2 e−f− dV.
M 4 2τ 6 6
Proof. This follows immediately by combining the results of Theorem 6.20

and Lemma 6.24.
Proposition 6.27. Let (M n , gt , Ht ) be a solution to generalized Ricci flow on a

compact manifold admitting a torsion-bounding subsolution φt . The infimum in the
definition of µ− is attained by a unique function f . Furthermore µ− (gt , Ht , T −t) is
monotonically nondecreasing along generalized Ricci flow, and is constant only on
a generalized shrinking Ricci soliton, i.e. H ≡ 0 and g is a shrinking Ricci soliton.
Proof. This proof is directly analogous to Proposition 6.19.
Thus in the presence of a torsion-bounding subsolution, we have constructed a
monotone entropy functional. To end this section we give one key consequence of
this result, namely that in the presence of a torsion-bounding subsolution, solutions
to generalized Ricci flow are not locally collapsing.
Definition 6.28. A solution (gt , Ht ) to generalized Ricci flow is locally col-
lapsing at T if there is a sequence tk → T and a sequence of metric balls Bk =
B(pk , rk , gtk ) such that rk2 /tk is bounded, |Rm| (gtk ) ≤ rk−2 in Bk , and
lim r−n Vol(Bk ) = 0.
k→∞ k
This property, roughly speaking, says that if we rescale around the given se-
quence so that the curvature is unit size, then the volume of unit balls around the
basepoint in the blowup sequence approaches zero, indicating that the manifold is
converging to a lower-dimensional space. If we want to apply the compactness the-
orems of §5.6 to a blowup sequence of generalized Ricci flows, then we need to rule
this behavior out. The augmented entropy monotonicity allows us to establish this
key estimate. We indicate in Chapter 9 (see Remark 9.47) some natural situations
where a torsion-bounding subsolution can be found and hence this estimate can be
applied.
[0, T ), T < ∞. Suppose there exists a torsion-bounding subsolution φt on [0, T ).
Then (gt , Ht ) is not locally collapsing at T .
Proof. Suppose there exists a sequence {(pk , tk , rk )} which exhibits local col-
lapsing as in Definition 6.28. We fix η : [0, ∞) → [0, 1] a nonincreasing function
such that η(s) = 1 for s ∈ [0, 12 ], η(s) = 0 for s ≥ 1, and |η ′ (s)| ≤ 4. Then let

Uk (x) := e−λk /2 η dgtk (pk , x)/rk .
These functions Uk are scaled cutoff functions for the sequence of collapsing balls.
We aim to use these as test functions in W− . Specifically, we define fk via e−fk /2 =
Uk , and will show that W− (gtk , Htk , fk , rk2 ) → −∞. Given this, by the monotonicity
of µ− (gt , Ht , φt , T − t), it follows that µ− (g0 , H0 , tk + rk2 ) → −∞, which contradicts
that µ(g0 , H0 , τ ) is finite for all τ , yielding a contradiction.
For fk to be a valid test function for µ, we need to ensure the unit volume
condition, which is determined by the scaling factor λk . We observe that the
necessary condition implies
Z
− n
1= Uk2 4πrk2 2 dV
M
Z
−λk
− n
=e 4πrk2 2 η 2 dgtk (pk , x)/rk dV
M
−λk
− n
≤e 4πrk2 2 Vol(Bk ).
6.4. COROLLARIES ON NONSINGULAR SOLUTIONS 125
Thus, by the hypothesis of local collapsing, we have that λk → −∞. To esti-

mate W− from above, we first observe using the assumed curvature bound, the
normalization of fk and dropping some negative terms,
W− (gtk , Htk , φtk , fk , rk2 )
Z h i
n
2 −2
≤ C + 4πrk rk2 |∇fk |2 + fk e−fk dV.
M
By using the lower bound on the Ricci curvature on Bk , the estimates on η, and
Bishop-Gromov relative volume comparison one has, setting ηk = η(dgtk (pk , ·)/rk ),
Z h i
n
2 −2
4πrk rk2 |∇fk |2 + fk e−fk dV
M
Z h i
− n
= 4πrk2 2 e−λk rk2 |∇ηk |2 − 2ηk2 ln ηk dV + λk
M
Vol(B(pk , rk )) − Vol(B(pk , rk /2))
≤C + λk
Vol(B(pk , rk /2))
≤ C + λk .
Thus W− (gtk , Htk , φtk , fk , rk2 ) → −∞, as required.
6.4. Corollaries on nonsingular solutions

The energy and entropy functionals above play an important role in under-
standing singularities of generalized Ricci flow. The simplest application is to un-
derstanding global solutions with some asymptotic control at infinite time.
Definition 6.30. Given (M n , gt , Ht ) a solution to generalized Ricci flow, we
say that the solution is nonsingular if there exists Λ ∈ {−1, 0, 1} such that the
corresponding rescaled solution of
∂ 1
g = − 2 Rc + H 2 + Λg,
∂t 2
∂
H = ∆d H + ΛH,
∂t
exists on [0, ∞) and satisfies
sup |Rmg |g < ∞.
M×[0,∞)
Remark 6.31. (1) This definition differs slightly from the classical def-
inition of nonsingular solution which asks for bounded curvature along
the normalization which fixes the volume. In many cases nonsingular so-
lutions in this sense will approach constant scalar curvature, rendering
them nonsingular in our sense. However, in general the two notions are
different.
(2) Solitons are examples of nonsingular solutions.
(3) From the point of view of analysis, it is natural to modify the flow of
H as we have done, so that the resulting family (gt , Ht ) is a spacetime
scaling of the corresponding unnormalized flow. Note however that this
operation will of course change the cohomology class of H, thus altering
the underlying Courant algebroid structure.
Theorem 6.32. Suppose (M n , gt , Ht ) is a nonsingular solution of generalized

Ricci flow with Λ = 0, −1. Then exactly one of the following two statements holds:
(1) lim supt→∞ injgt = 0.
(2) There exists a sequence (pj , tj ), tj → ∞ such that (M n , gtj , Htj , pj ) con-
verges in the generalized Cheeger-Gromov topology to a generalized Ricci
n
soliton (M∞ , g∞ , H∞ , p∞ ). In the case Λ = −1, G∞ is a generalized ex-
panding soliton, i.e. H∞ ≡ 0 and g∞ is an expanding Ricci soliton with
negative scalar curvature.
Proof. If the first case does not hold, there exists δ > 0 and a sequence of
points (pj , tj ), tj → ∞ such that injgtj (pj ) ≥ δ. Our goal is to modify this sequence
using the energy and entropy monotonicities to obtain the claimed limit. At this
point we restrict the proof to the case Λ = 0 and apply the monotonicity of the F -
functional. In the case Λ = −1 one must use the W+ functional, with the vanishing
of H following from the extra |H|2 term in the monotonicity formula of Corollary
6.17. We leave the details to the reader. We first observe that by choosing the
function ft ≡ log Vol(gt ), it follows that
Z
2
λ(gt , Ht ) ≤ F (gt , Ht , ft ) = Vol(gt )−1 1
R − 12 |H| dV ≤ C,
M
where the last line follows since the curvature is uniformly bounded along the
flow. As λ is now monotonically nondecreasing and bounded above, we expect
convergence to a critical point provided we can establish regularity of the flow.
Given one of the times tj above, as in the proof of Corollary 6.10 we construct
a solution ft to the conjugate heat equation where
Z
e−ftj +1 dV = 1, F (gtj +1 , Htj +1 , ftj +1 ) = λ(gtj +1 , Htj +1 ).
M
Applying the argument of Corollary 6.10 then yields
λ(gtj , Htj ) ≤ λ(gtj +1 , Htj +1 )
Z tj +1 Z h i
2 2
− 2 Rc − 41 H 2 + ∇2 f + 1
2 |d∗ H + ∇f yH| e−f dV.
tj M
Without loss of generality we can assume tj+1 ≥ tj + 1, and then by summing

the above inequalities we obtain, taking care with the definition of f on each time
interval and using that λ is bounded above,
X∞ Z tj +1 Z h i
2 2
2 Rc − 41 H 2 + ∇2 f + 21 |d∗ H + ∇f yH| e−f dV < ∞.
j=1 tj M
It follows that the terms in the sum must go to zero, and then in particular there
exists tj ≤ t′j ≤ tj + 1 such that
Z h i
2 2
(6.16) lim 2 Rc − 41 H 2 + ∇2 f + 21 |d∗ H + ∇f yH| e−f dV = 0.
j→∞ M
It remains to take a smooth limit of (M n , gtj , Htj , ftj , pj ). To obtain this,
first observe that for all t ≥ 1, we can apply Theorem 5.18 on [t − 1, t] to conclude
uniform estimates on all derivatives of curvature and of H. Since we have a uniform
estimate on ∂g
∂t , the first part of the argument of Lemma 5.22 applies to yield uniform
equivalence of all metrics gt , tj ≤ t ≤ tj + 1. Since injgtj (pj ) ≥ δ, it follows that
6.4. COROLLARIES ON NONSINGULAR SOLUTIONS 127
there exists δ ′ > 0 so that injgt′ (pj ) ≥ δ ′ . We next sketch an argument that the
j
functions ft′j satisfy estimates of the form
sup ∇kgt′ ft′j ≤ C(k, R).

BR (pj ,gt′ ) j
gt′
j j
First, given the uniform curvature and derivative estimates, at any point we can
work on the universal cover of the unit ball, which is then uniformly equivalent to
a Euclidean ball by Jacobi field estimates. Since utj +1 = e−ftj +1 by construction
realizes the infimum of λ, it satisfies the elliptic equation (cf. Lemma 6.3)

1 2
−4∆ + R − 12 |H| − λ utj +1 = 0
Since all derivatives of curvature and H are bounded and ∆ is uniformly elliptic,
we can apply elliptic regularity to obtain estimates on all derivatives of utj +1 on
any unit ball as described above. Moreover, the conjugate heat equation ∗ u = 0
2
is uniformly parabolic, and the linear term R − 14 |H| has uniform estimates, so we
can invoke parabolic regularity results to obtain uniform control over all derivatives
of ut , tj ≤ t ≤ tj +1. We can now apply Theorem 5.31 to conclude that the sequence
of pointed solutions to generalized Ricci flow determined by {(M n , gt , Ht , pj ), t ∈
[tj , tj+1 ]} admits a convergent subsequence, and that the associated functions f as
defined above converge as well. It follows from (6.16) that this limit is a generalized
Ricci soliton.
Remark 6.33. (1) In the first case above where the injectivity radius
goes to zero everywhere, the manifolds are collapsing with bounded cur-
vature. Convergence results for the Ricci flow which yield geometric limits
in this setting were proved by Lott [131, 132]. Further rigidity results
for the generalized Ricci flow appeared in [84].
(2) The limit space M∞ need not be diffeomorphic to M . An elementary
example occurs for hyperbolic three-manifolds (cf. Example 4.11), where
the limit of the unrescaled flow will be flat R3 with H ≡ 0 (noting that
while H is fixed in time as a tensor its gt -norm goes to zero along the flow).
As described there the appropriately rescaled flow will converge back to
the hyperbolic metric with H ≡ 0. Taking products of this example with
a Bismut-flat structure on a semisimple Lie group yields examples with
a nontrivial limit for the unrenormalized flow, where the topology of the
base manifold has changed.
(3) A similar result can be obtained in the case Λ = 1, under the further
assumption of the existence of a torsion-bounding subsolution, using the
shrinking entropy.
Remark 6.34. Perelman’s energy and entropy quantities are also related to
certain formal metric constructions on the whole spacetime of a solution to Ricci
flow. A related spacetime geometry approach for generalized Ricci flow was taken
in [45], yielding certain canonical solitons associated to any solution.
CHAPTER 7
Generalized Complex Geometry
In this chapter we introduce foundational concepts of generalized complex ge-

ometry, following fundamental works of Hitchin [99] and Gualtieri [88]. A core
conceit in this subject is the unification of complex and symplectic structures into
a single, natural object most clearly defined using the language of Courant alge-
broids. We begin with the basic definitions and indicate how complex and sym-
plectic structures fit in as examples of generalized complex structures, and also
illustrate nontrivial examples. We also express the geometry of pluriclosed Her-
mitian structures in this language, and discuss their relation to Courant algebroids
in the holomorphic category. We then turn to generalized Kähler geometry, first
motivating and defining these structures in terms of generalized geometry, then
discussing the equivalence to the classical description, which first arose in work of
Gates-Hull-Roček on supersymmetric σ-models [74]. We provide examples and a
discussion of the associated Poisson geometry and conclude by exhibiting natural
classes of variations of generalized Kähler structure.
7.1. Linear generalized complex structures

Before beginning our study of generalized complex structures we briefly recall
the conceptual buildup of complex structures on manifolds. Given M 2n a smooth
manifold, an almost complex structure on M is an endomorphism J of the tangent
bundle covering the identity map satisfying
J 2 = − Id .
The pair (M 2n , J) is called an almost complex manifold. The restriction to an even

dimensional manifold is necessary for the existence of an endomorphism which
squares to − Id.
√ The endomorphism J gives a splitting of the complexified tangent
bundle into ± −1 eigenbundles, and it is natural to ask when these subbundles
are closed under the Lie bracket. This is measured precisely by the vanishing of
the Nijenhuis tensor of J, defined by
N (X, Y ) = [JX, JY ] − [X, Y ] − J[JX, Y ] − J[X, JY ].
We say that J is integrable if N ≡ 0. In this case we say that (M 2n , J) is a

complex manifold. It follows from the Newlander-Nirenberg Theorem [137] that
the vanishing of the Nijenhuis tensor is equivalent to the existence of a complex
coordinate atlas, i.e. complex coordinate charts covering the manifold with biholo-
morphic transition maps. The story of generalized complex structures follows a
similar path, first defining almost complex structures on T ⊕ T ∗ , then analyzing
the geometric implications of the integrability of the associated splitting under the
Dorfman bracket.
129
130 7. GENERALIZED COMPLEX GEOMETRY
7.1.1. Definition and examples. To begin we give the definition of a linear

generalized complex structure, which is a natural extension of the notion of com-
plex structure to the generalized tangent bundle, further preserving the natural
symmetric inner product. As for generalized metrics, we can also motivate this
concept from the point of view of structure groups. Whereas a classical complex
structure reduces the structure group of the tangent bundle to U (n), a generalized
complex structure will be a reduction of the structure group of T ⊕ T ∗ to U (n, n).
Following our discussion in §2.3.1, we fix a pair of real vector spaces V and W
which fit into a short exact sequence
/ V∗ π
(7.1) 0 /W /V / 0.
We assume that W is endowed with a metric h, i of split signature (n, n) such that
the image of the first arrow in (7.1), given by 12 π ∗ : V ∗ → W ∗ ∼
= W , is isotropic.
Definition 7.1. A generalized complex structure on W is an endomorphism
J of W such that
(1) J 2 = − Id.
(2) J is orthogonal with respect to the symmetric inner product h, i on W .
When W = V ⊕ V ∗ and h, i is given by (2.1), we will simply say that J is a
generalized complex structure on V .
Of course, any generalized complex structure on W defines a generalized com-
plex structure on V upon a choice isotropic splitting of the sequence (7.1). Nonethe-
less, in our discussion we will often find generalized complex structures interacting
with generalized metrics, which determine a preferred isotropic splitting, and the
abstract definition above will be useful.
We next discuss two central examples which will inform the entire discussion,
illustrating that classical complex structures as well as symplectic structures admit
interpretations as generalized complex structures.
Example 7.2. Let W = V ⊕ V ∗ and h, i given by (2.1). Let J : V → V be a
complex structure, and define

−J 0
JJ := ,
0 J∗
where the block diagonal structure comes from the natural decomposition of V ⊕V ∗
and J ∗ α is the natural induced action of J on V ∗ , i.e. J ∗ α(v) = α(Jv). The
equation JJ2 = − Id is immediate, and we can furthermore compute
hJJ (X + ξ), JJ (Y + η)i = h−JX + J ∗ ξ, −JY + J ∗ ηi
= 1
2 (−J ∗ ξ(JY ) − J ∗ η(JX))
1
= 2 (ξ(Y ) + η(X))
= hX + ξ, Y + ηi .
Thus JJ defines a generalized complex structure on V .
Example 7.3. Let ω ∈ Λ2 V ∗ be a symplectic structure on V . Observe that ω
defines a canonical map ω : V → V ∗ using interior product, i.e ω(v) := iv ω. By
assumption this map is invertible, and we denote the inverse map as ω −1 : V ∗ → V .
7.1. LINEAR GENERALIZED COMPLEX STRUCTURES 131
Now set

0 −ω −1
Jω := .
ω 0
The equation Jω2 = − Id is immediate, and we can furthermore compute
hJω (X + ξ), Jω (Y + η)i = −ω −1 ξ + ω(X), −ω −1 η + ω(Y )

= 12 −ω(X, ω −1 η) − ω(Y, ω −1 ξ)
1
= 2 (η(X) + ξ(Y ))
= hX + ξ, Y + ηi .
Thus Jω defines a generalized complex structure on V .
Example 7.4. Given vector spaces Vi with generalized complex structures Ji ,
i = 1, 2, one defines the direct product J = J1 ⊕ J2 . A simple check shows that
this defines a generalized complex structure on V1 ⊕ V2 . Using this construction in
conjunction with the two previous examples, one immediately obtains generalized
complex structures which are not determined solely by a complex or symplectic
structure.
Example 7.5. An important generalization of these basic examples exploits
b-field symmetries. Let V be a vector space with generalized complex structure
J . Given b ∈ Λ2 (V ∗ ), let Jb := eb J e−b . An elementary calculation shows that
Jb2 = − Id. Moreover, since b-field transformations are orthogonal for h, i by Lemma
2.13, as is J , it follows immediately that Jb is orthogonal with respect to h, i,
showing that Jb defines a new generalized complex structure. Observe that the
different choices of b ∈ Λ2 (V ∗ ) can be regarded as different isotropic splittings of
the sequence (7.1) and therefore Jb is, in a sense, equivalent to J .
Remark 7.6. Since every even-dimensional vector space V admits both com-
plex and symplectic structures, it follows from the above examples that they admit
generalized complex structures as well. In fact, if V admits a generalized complex
structure J , then V must be even dimensional. To see this, fix X + ξ ∈ V ⊕ V ∗
a vector which is null for the neutral inner product. As J is a complex structure
on V ⊕ V ∗ orthogonal with respect to the neutral inner product, it follows that
J (X + ξ) is null, and orthogonal to X + ξ. These two vectors span a J -invariant
isotropic plane S, and so we can choose a new vector X ′ +ξ ′ orthogonal to this plane
and obtain that J (X ′ + ξ ′ ) is null, and orthogonal to S and X ′ + ξ ′ . Repeating
this inductively we eventually construct a maximal isotropic subspace, which has
even dimension. Since the inner product has signature (n, n) with n = dim V , it
follows that V is even dimensional.
Using this fact we can give the characterization of generalized complex struc-
tures in terms of reduction of structure groups. Given W as in (7.1) we denote by
O(2n, 2n) := O(W ) the group of linear isometries of W .
Lemma 7.7. If W as in (7.1) admits a generalized complex structure, then
dim V = 2n. Furthermore, a generalized complex structure on W is equivalent to a
choice of maximal compact subgroup U (n, n) ≤ O(2n, 2n).
Proof. The first part follows from Remark 7.6 by identifying W = V ⊕ V ∗ via
a choice of isotropic splitting of (7.1). As for the second part, endomorphisms which
preserve J will lie in GL(2n, C), and so a choice of generalized complex structure
on W determines a maximal compact O(2n, 2n) ∩ GL(2n, C) = U (n, n).
7.1.2. Linear Dirac structures. To expand upon the examples above, we

will show that a generalized complex structure is characterized in terms of Dirac
structures. For simplicity, we will assume that W = V ⊕ V ∗ with h, i given by (2.1).
Definition 7.8. Given V a vector space, a Dirac structure on V is a subspace
L ⊂ V ⊕ V ∗ which is maximally isotropic for the pairing h, i.
Example 7.9. Both V and V ∗ are clearly isotropic for h, i, and have the max-
imal possible dimension for an isotropic subspace, dim V , and hence define Dirac
structures.
Example 7.10. Fix U ⊂ V a subspace, and consider L = U ⊕Ann(U ) ⊂ V ⊕V ∗ ,
where Ann(U ) denotes the annihilator of U in V ∗ . It follows directly by construction
that L is isotropic. Since dim Ann(U ) = codim U , it follows that L has the maximal
dimension, and so defines a Dirac structure.
Example 7.11. Let b : V → V ∗ be a linear map such that the associated
element of V ∗ ⊗ V ∗ , also denoted b, is skew-symmetric. Then graph(b) ⊂ V ⊕ V ∗
is isotropic since for X, Y ∈ V ,
1
hX + b(X, ·), Y + b(Y, ·)i = 2 (b(X, Y ) + b(Y, X)) = 0.
Since the dim(graph(b)) = dim V , it follows that graph(b) is a Dirac structure.
Example 7.12. The two previous examples can be unified by the following
construction, which yields all possible Dirac structures. Fix U ⊂ V any subspace,
fix φ ∈ Λ2 U ∗ , and let

L(U, φ) = X + ξ ∈ U ⊕ V ∗ | ξ|U = φ(X, ·) .
Analogous to the case of graph(b) above, we observe that for X +ξ, Y +η ∈ L(U, φ),
1 1
hX + ξ, Y + ηi = 2 (ξ(Y ) + η(X)) = 2 (φ(X, Y ) + φ(Y, X)) = 0.
Hence L(U, φ) is isotropic. Note that for each X, the ξ such that X +ξ ∈ L(U, φ) are
prescribed on U , but arbitrary otherwise, and so the space of such ξ has dimension
equal to codim(U ). Thus dim(L) = dim(U ) + codim(U ) = dim(V ), and so L is
maximal.
Proposition 7.13. Every maximal isotropic subspace L ⊂ V ⊕ V ∗ is of the
form L(U, φ).
Proof. Fix L a maximal isotropic and let U = πV L, where πV denotes the

canonical projection onto V . Since L is isotropic, it follows that L ∩ V ∗ = Ann(U ).
Since in general one has a canonical identification U ∗ = V ∗ / Ann(U ), we may define
a map φ : U → U ∗ via

φ(e) = πV ∗ πV−1 (e) ∩ L ∈ V ∗ / Ann(U ).
To check that this is well-defined, fix e+η, e+µ ∈ L, so that πV (e+η) = πV (e+µ) =
e. It follows that η − µ ∈ L ∩ V ∗ , thus πV ∗ (η − µ) ∈ Ann(U ), finishing the claim.
Also the map is skew-symmetric with respect to the neutral inner product since,
using that L is isotropic,
hπV ∗ (X + η), πV (Y + µ)i + hπV ∗ (Y + µ), πV (X + η)i
= η(Y ) + µ(X) = hX + η, Y + µi = 0.
The tensor φ is canonically identified with a section of Λ2 U ∗ , and it follows from
the construction that L = L(U, φ).
We give next our characterization of generalized complex structures in terms

of complex Dirac structures on (V ⊕ V ∗ ) ⊗ C.
Proposition 7.14. A generalized complex structure on V is equivalent to a
choice of maximal isotropic complex subspace L ⊂ (V ⊕ V ∗ ) ⊗ C satisfying L ∩ L =
{0}.
√
Proof. Given J a generalized complex structure, let L be the −1-eigenspace
of the induced operator on (V ⊕ V ∗ ) ⊗ C. To show that L is isotropic, fix x, y ∈ L
and use the orthogonality property of J to obtain
√ √
hx, yi = hJ x, J yi = −1x, −1y = − hx, yi .
√
Since J is a real endomorphism, it follows immediately that L is the − −1-
eigenspace for J , implying moreover that L ∩ L = {0}, and that√L is maximal.
Conversely,
√ given such an L one defines J as multiplication by −1 on L and
− −1 on L. An elementary calculation shows that this defines a real endomor-
phism, which by construction certainly satisfies J 2 = − Id. Moreover, to check
orthogonality, since L is isotropic and an eigenspace for J it suffices to check or-
thogonality for x ∈ L, y ∈ L, in which case
√ √
hJ x, J yi = −1x, − −1y = hx, yi ,
as required.
7.1.3. Spinor formulation. Generalized complex structures can also be ele-

gantly described in terms of spinors. This point of view will also be useful in dealing
with integrability issues for generalized complex structures on manifolds. We fix a
vector space V and consider the pairing h, i on V ⊕ V ∗ given by (2.1).
Definition 7.15. Let CL(V ⊕ V ∗ ) denote the Clifford algebra defined by the
relation
v · v = hv, vi , v ∈ V ⊕ V ∗.
Furthermore define
Spin(V ⊕ V ∗ ) = {v1 . . . vk | vi ∈ V ⊕ V ∗ , hvi , vi i = ±1, k even} .
The action of V ⊕ V ∗ on the exterior algebra Λ∗ V ∗ , given by
(7.2) (X + ξ) · ρ = iX ρ + ξ ∧ ρ,
extends to a natural action of CL(V ⊕ V ∗ ) on Λ∗ V ∗ .
Lemma 7.16. The operation (7.2) defines an algebra representation of CL(V ⊕
V ∗ ) on Λ∗ V ∗ .
Proof. We compute
(X + ξ) · (X + ξ) · ρ = iX (iX ρ + ξ ∧ ρ) + ξ ∧ (iX ρ + ξ ∧ ρ)
= (iX ξ) ρ
= hX + ξ, X + ξi ρ,
as required.
As a spin representation, the exterior algebra decomposes into a direct sum
Λ∗ V ∗ = Λeven V ∗ ⊕ Λodd V ∗
corresponding to irreducible representations. Note that we have the double cover
homomorphism
ψ : Spin(V ⊕ V ∗ ) → SO(V ⊕ V ∗ )
defined by
ψ(x)(v) = x · v · x−1 ,
for x ∈ Spin(V ⊕ V ) and v ∈ V ⊕ V ∗ . Using this map, a Lie algebra element of
∗
the form

A 0
(7.3) ζ= ∈ so(V ⊕ V ∗ )
B −A∗
acts on Λ∗ V ∗ via
ζ · ρ = −B ∧ ρ − A∗ ρ + 1
2 tr Aρ.
Here, A ∈ End(V ) and B ∈ Λ2 V ∗ , and we have used the natural identification
so(V ⊕ V ∗ ) = Λ2 (V ⊕ V ∗ ) = End(V ) ⊕ Λ2 V ∗ ⊕ Λ2 V.
Observe that elements of the form (7.3) are precisely those which preserve the flag
(7.1), and will appear later as infinitesimal Courant algebroid automorphisms (see
Proposition 2.15).
Exponentiating ζ above we obtain spin group elements corresponding to B-field
transformations, given by
eB · ρ = e−B ∧ ρ = (1 − B + 21 B ∧ B + . . .) ∧ ρ.
We also obtain spin group elements corresponding to linear automorphisms A ∈
GL+ (V ) ⊂ GL(V ) with positive determinant,
√
A · ρ = det A(A−1 )∗ ρ.
This indicates that, as a GL+ (V ) representation, the spinors decompose as
S = Λ∗ T ∗ ⊗ (det V )1/2 .
We turn next to the study of linear Dirac structures on V in terms of spinors.
Definition 7.17. Given ρ ∈ Λ∗ V ∗ , define the associated null space by
Lρ := {v ∈ V ⊕ V ∗ | v · ρ = 0} .
Note that Lρ is isotropic, since for v1 , v2 ∈ Lρ , we have
1
hv1 , v2 i ρ = 2 (v1 v2 + v2 v1 ) · ρ = 0,
thus hv1 , v2 i = 0. The spinor ρ is called pure if Lρ is maximal, that is, has dimension
equal to dim V .
The next result shows that every maximal isotropic subspace L ⊂ V ⊕ V ∗ is

represented by a unique line KL ⊂ Λ∗ V ∗ of pure spinors.
Proposition 7.18. Every maximal isotropic subspace L = L(U, φ) is deter-
mined by a pure spinor line KL ⊂ Λ∗ V ∗ . If θ1 , . . . , θk is a basis of Ann(U ) and
B ∈ Λ2 V ∗ is such that i∗ B = −φ, where i : U → V is the inclusion, then the pure
spinor line KL representing L(U, φ) is generated by
ρ = e B θ1 ∧ . . . ∧ θk .
The integer k is called the type of the maximal isotropic.
Proof. By Proposition 7.13, any maximal isotropic L = L(U, φ) may be ex-
pressed as the B-field transform of L(U, 0) for B as in the statement. It is not diffi-
cult to see that the pure spinor line with null space L(U, 0) is precisely det Ann(U ).
Thus, the statement follows from
KL = Ke−B L(U,0) = eB KL(U,0) = eB det Ann(U ).
We refer to ([48] Chapter III) for further details.
Consider now a generalized complex structure J on V and the associated com-
plex Dirac structure (see Proposition 7.14)
L ⊂ (V ⊕ V ∗ ) ⊗ C.
Regarding Λ∗ V ∗ ⊗ C as a representation of the complex Clifford algebra of V ⊕ V ∗
via (7.2), the analogue of Proposition 7.18 applies, and we obtain a complex pure
spinor line KL ⊂ Λ∗ V ∗ ⊗ C generated by
√
ρ = eB+ −1ω
θ1 ∧ . . . ∧ θk
for B, ω ∈ Λ2 V ∗ and θ1 , . . . , θk a basis of L ∩ (T ∗ ⊗ C). To provide a convenient
characterization of the condition
L ∩ L = {0}
in Proposition 7.14, let us recall the definition of the Mukai pairing,
(, ) : Λ∗ V ∗ ⊗ Λ∗ V ∗ → det V ∗ .
Given α, β ∈ Λ∗ V ∗ this is defined by
(α, β) = (αT ∧ β)top ,
where α → αT is the antiautomorphism of the Clifford algebra determined by the
tensor map v1 ⊗ . . . ⊗ vk → vk ⊗ . . . ⊗ v1 and ()top indicates taking the top degree
component of the form. The Mukai pairing extends C-linearly to Λ∗ V ∗ ⊗ C and we
have the following result, which we state without proof (see [48]).
Lemma 7.19. Maximal isotropics L, L′ ⊂ (V ⊕ V ∗ ) ⊗ C satisfy L ∩ L′ = {0} if
and only if their pure spinor representatives ρ, ρ′ satisfy
(ρ, ρ′ ) 6= 0.
The following characterization of linear generalized complex structures is now
a direct consequence of Proposition 7.14 and Lemma 7.19. As a matter of fact,
this was the original point of view on generalized Calabi-Yau manifolds taking by
Hitchin in his seminal paper [99].
Proposition 7.20. A generalized complex structure on V is equivalent to a

choice of complex pure spinor line K ⊂ Λ∗ V ∗ ⊗ C such that any pure spinor repre-
sentative ρ satisfies
(ρ, ρ) 6= 0.
7.2. Generalized complex structures on manifolds

Having defined the appropriate linear algebraic structure in the previous sec-
tion, we now extend this definition to manifolds following a common two-step pro-
cess. First we define “almost generalized complex structures” as smooth sections
of End(T ⊕ T ∗ ) satisfying the appropriate algebraic conditions. We then define a
natural notion of integrability for such structures, and discuss how classical inte-
grability conditions for complex and symplectic structures fit into this picture. For
the most recent developments on the subject we refer to [10, 11, 41] and references
therein.
7.2.1. Almost generalized complex structures.

Definition 7.21. Given M 2n a smooth manifold, an almost generalized com-
plex structure on M is a section J ∈ Γ(End(T ⊕ T ∗ )) such that Jp defines a linear
generalized complex structure on Tp ⊕ Tp∗ for all p ∈ M . More generally, given an
exact Courant algebroid E over M , an almost generalized complex structure on E is
a section J ∈ Γ(End E) such that Jp defines a linear generalized complex structure
on Ep for all p ∈ M .
Following Lemma 7.7, we can regard an almost generalized complex structure
on M as a reduction of the O(2n, 2n)-bundle of frames of the generalized tangent
bundle to the maximal compact U (n, n). Alternatively, by Proposition 7.14, J is
equivalent to a maximal isotropic subbundle L ⊂ (T ⊕T ∗)⊗C satisfying L∩L = {0}.
Yet another point of view on this structure is provided by spinors. By the discussion
in §7.1.3, the spinor bundle for T ⊕ T ∗ corresponds to
S = Λ∗ T ∗ ⊗ | det T |1/2 ,
where | det T |1/2 denotes the line bundle of half-densities on M . Since | det T |1/2 is
trivializable, it will be more convenient to work with the twisted spinor bundle
S = Λ∗ T ∗
endowed with the Clifford action (7.2). By the previous discussion, and also Propo-
sition 7.20, we obtain the following result giving different formulations of the notion
of almost generalized complex structure.
Proposition 7.22. Given M 2n a smooth manifold, a choice of almost gener-
alized complex structure J is equivalent to
• A reduction of structure for the O(2n, 2n)-bundle (T ⊕T ∗ , h, i) to the group
U (n, n).
• A maximal isotropic subbundle L ⊂ (T ⊕ T ∗ ) ⊗ C satisfying L ∩ L = {0}.
• A pure spinor line sub-bundle K ⊂ Λ∗ T ∗ ⊗ C such that any pure spinor
representative ρ satisfies
(ρ, ρ) 6= 0 ∈ det T ∗ .
7.2. GENERALIZED COMPLEX STRUCTURES ON MANIFOLDS 137
The previous result generalizes the fundamental fact that a complex structure
can be thought of as a tensor field, a reduction of structure groups, or in terms of
certain subbundles. Notice that (ρ, ρ) ∈ det T ∗ defines an orientation independent
of the choice of ρ, giving a global orientation on the manifold.
Example 7.23. Continuing the discussion in Example 7.2, suppose M 2n admits
an almost complex structure J, and then define an almost generalized complex
structure on M via

−J 0
JJ = .
0 J∗
√
It is clear by definition that the −1-eigenbundle takes the form
L = T 0,1 ⊕ T1,0
∗
.
The pure spinor line in Proposition 7.22 corresponds to Λn,0 and is generated,
locally, by a choice of non-vanishing (n, 0)-form.
Example 7.24. Continuing the discussion in Example 7.3, suppose M admits
an almost symplectic form ω, and then define an almost generalized complex struc-
ture on M via

0 −ω −1
Jω = .
ω 0
Given X ∈ T ⊗ C, note that
√
Jω √ X =
−1X
,
− −1ω(X) ω(X)
√
and hence such vectors lie in the −1-eigenspace of Jω . Since the complex dimen-
sion of√the space of such vectors is n, it follows that such vectors determine the
entire −1-eigenspace L, i.e.
√
L = e− −1ω T.
Observe that L is given by a B-field transformation of the tangent bundle, which
is a Dirac structure with pure spinor line generated by 1 ∈ (Λ∗ T ∗ ) ⊗ C. Thus,
by naturality of the Clifford action with respect to orthogonal transformations it
follows that the pure spinor line in Proposition 7.22 is generated in this case by the
global section
√ √
e− −1ω · 1 = e −1ω ∈ (Λ∗ T ∗ ) ⊗ C.
7.2.2. Integrability. We now turn to the question of integrability of almost
generalized complex structures. Recall that the classical
√ definition of integrabil-
ity of an almost complex structure J asks for the −1-eigenbundle to be closed
under the Lie bracket. For√integrability of generalized complex structures we will
ask for involutivity of the −1-eigenbundle under the Dorfman bracket. As this
eigenbundle is an example of an almost Dirac structure by Proposition 7.14, we
briefly work in greater generality and record fundamental definitions and results
concerning Dirac structures on exact Courant algebroids. Our exposition follows
[55, 88] closely.
M . An almost Dirac structure on E is a subbundle L ⊂ E which is maximally
isotropic with respect to the inner product h, i. We say that L is an integrable

Dirac structure if L is closed under the Dorfman bracket, i.e.
[L, L] ⊂ L.
Remark 7.26. It is important to note that, since L is isotropic with respect
to the symmetric pairing, it follows from (2.4) that the Courant and Dorfman
brackets agree on sections of L. Thus it is equivalent to ask that L is closed under
the Courant bracket. To save on notation we will refer to brackets of sections of L
as [e1 , e2 ], where this can be equivalently chosen as either the Courant or Dorfman
bracket.
This integrability condition admits a tensorial characterization, which helps in
verification. To begin we make a definition.
Definition
7.27. Given L an almost Dirac structure on E, define a map TL :
Γ L⊗3 → R via
TL (e1 , e2 , e3 ) := h[e1 , e2 ], e3 i .

Lemma 7.28. Given L an almost Dirac structure on E, TL ∈ Γ Λ2 L∗ ⊗ L∗ .
Proof. We must check that the map is C ∞ (M )-linear. It is obviously C ∞ (M )-
linear in the third position, and due to the skew-symmetry in the first two indices
it suffices to check C ∞ -linearity in either. To that end fix {ei } ∈ Γ(L) and f ∈
C ∞ (M ), and compute using Lemma 2.7,
TL (e1 , f e2 , e3 ) = h[e1 , f e2 ], e3 i
= hf [e1 , e2 ] + (Xf )e2 − he1 , e2 i df, e3 i
= f TL (e1 , e2 , e3 ),
where the last line follows using that L is isotropic. The lemma follows.
Proposition 7.29. Given L an almost Dirac structure, L is integrable if and
only if TL = 0.
Proof. If L is closed under the Courant bracket, then since L is isotropic
it follows directly that TL = 0. Conversely, suppose TL = 0. Then given any
e1 , e2 ∈ Γ(L), it follows by hypothesis that [e1 , e2 ] ∈ L⊥ . But since L is maximal
isotropic, it follows that L⊥ = L, and so L is closed under the bracket.
Furthermore, we can characterize the integrability of Dirac structures in terms
of the description in Proposition 7.13.
Proposition 7.30. Given M a smooth manifold and H ∈ Λ3 T ∗ , dH = 0,
consider the exact Courant algebroid (T ⊕ T ∗ , h, i , [, ]H ). A Dirac structure L(U, φ)
is integrable if and only if
(1) U is closed under the Lie bracket
(2) iY iX (H + dφ) = 0 for all X, Y ∈ Γ(U ).
Proof. Following the main computation of Proposition 2.14, we obtain that
[X + iX φ, Y + iX φ] = eφ [X, Y ] + iY iX (H + dφ) .
The result follows in a straightforward way from this equation.
With these preliminaries in hand we can give the definition of integrability of

an almost generalized complex structure. Following the previous discussion, we do
this in the generality of arbitrary exact Courant algebroids.
√
M . An almost generalized complex structure J on E is integrable if the −1-
eigenbundle L ⊂ E ⊗ C is an integrable Dirac structure. An integrable almost
generalized complex structure J on E will be called a generalized complex structure.
In particular,
√an almost generalized complex structure on a smooth manifold is
integrable if the −1-eigenbundle L ⊂ (T ⊕ T ∗ ) ⊗ C is an integrable Dirac structure
for the standard Dorfman bracket. It is not difficult to see that the analogue of
Proposition 7.22 holds for almost generalized complex structures on exact Courant
algebroids. We leave the details as an exercise.
Example 7.32. Continuing the discussion in Example 7.23, suppose M ad-
mits an almost complex
√ structure J, and consider the almost generalized complex
structure JJ . The −1-eigenbundle takes the form
L = T 0,1 ⊕ T1,0
∗
,
and it directly follows from this that L is closed under Courant bracket if and only
if T 0,1 is closed under the Lie bracket, which is in turn equivalent to the classical
integrability of J.
Example 7.33. Continuing the discussion in Example 7.24, suppose M admits
an almost symplectic form ω, and consider the almost generalized complex structure
Jω . Fix smooth complex vector field X, Y ∈ T ⊗ C and observe, using an obvious
extension of the proof of Proposition 2.14 to the case of complex coefficients,
√ √ √ √
[X − −1ω(X), Y − −1ω(Y )] = [e− −1ω (X), e− −1ω (Y )]
√ √
= e− −1ω [X, Y ] + −1iY iX dω.
It follows that L is integrable if and only if dω = 0.
It is obvious from Definition 7.31 that if J is an integrable generalized complex
structure on an exact Courant algebroid E, then conjugation by a Courant algebroid
automorphism provides a new generalized complex structure. We state this for
emphasis:
Lemma 7.34. Let J be a generalized complex structure on an exact Courant
algebroid E. Then, for any Courant algebroid automorphism F : E → E, F ◦J ◦F −1
is also a generalized complex structure.
Example 7.35. Let J be an integrable generalized complex structure on M
and suppose B ∈ Λ2 T ∗ , dB = 0. Then JB := eB J e−B is a generalized complex
structure by Proposition 2.21 and Lemma 7.34.
The prior class of examples together indicate the naturality of the circle of
definitions yielding the concept of generalized complex structure, as it beautifully
encodes complex and symplectic structures both algebraically and in terms of inte-
grability. To further indicate the manner in which generalized complex structures
unify complex and symplectic structure, we give an example of a one-parameter
family of generalized complex structures connecting one of complex type to one of
symplectic type.
Example 7.36. Fix (M 4n , g, I, J, K) a hyperKähler manifold. Using the nota-

tion of the previous examples, consider JI and JωJ , where ωJ denotes the Kähler
form associated to (g, J). These are generalized complex structures by the previous
examples. Now consider the one-parameter family
Jt = sin tJI + cos tJωJ , t ∈ [0, π2 ].
From the quaternion relation IJ = −JI one can check directly that the endomor-
phisms JI and JωJ anticommute, after which it directly follows that Jt2 = − Id. It
remains to check the integrability of Jt . To that end let Bt = (tan t)ωK , which is a
closed form defined for t ∈ [0, π2 ). A calculation using the hyperKähler conditions
shows that

0 −(sec tωJ )−1 B
Jt = e−B e ,
sec tωJ 0
thus for t < π2 , Jt is expressed as a B-field transformation of an integrable gener-
alized complex structure, and is hence integrable by Lemma 7.34.
Following the line of thought in Proposition 7.22, we close this section with the
characterization of integrability for almost generalized complex structures in terms
of spinors. The preliminaries for this result amount to expressing the Dorfman
bracket as a derived bracket. Let us consider the general setup of an exact Courant
algebroid E over a smooth manifold M . Consider the Clifford bundle CL(E) given
by the relation
v · v = hv, vi.
For a choice of isotropic splitting σ : T → E, inducing an isomorphism E = ∼ T ⊕T∗
with twisted bracket [, ]H for some closed three-form H, the bundle of polyforms
S = Λ∗ T ∗
is an irreducible Clifford module for CL(E), via the Clifford action
(X + ξ) · ρ = iX ρ + ξ ∧ ρ.
For a different choice of isotropic splitting σ ′ , we can write σ ′ = eB ◦ σ for B ∈
Γ(Λ2 T ∗ ), and one has
(7.4) (e−B (X + ξ)) · eB ∧ ρ = eB ∧ ((X + ξ) · ρ).
Thus, the definition of S is independent of the choice of σ. Consider the twisted de
Rham differential
dH : Γ(S) → Γ(S)
acting on the space of polyforms by
(7.5) dH ρ = dρ − H ∧ ρ.
Using formula (7.4), it is an exercise to check that dH is independent of the choice
of isotropic splitting, defining a canonical map
/0 : Γ(S) → Γ(S).
d
Via the decomposition of the exterior algebra as a spin representation, the spinor
bundle S inherits a Z2 -grading
S = Λeven T ∗ ⊕ Λodd T ∗ .
/0 has odd degree, as does the operator

With respect to this grading, the operator d
given by Clifford multiplication α → v · α for any element v ∈ E. Recall the
definition of the super-commutator of operators on a Z2 -graded vector space
[L1 , L2 ] = L1 L2 − (−1)deg L1 deg L2 L2 L1 .
Similar to how the usual de Rham differential is related to the Lie bracket, we have
the following formula for the Dorfman bracket.
Lemma 7.37. For any v1 , v2 ∈ Γ(E) and ρ ∈ Γ(S), we have
/0 , v1 ·], v2 ·]ρ = [v1 , v2 ] · ρ.
[[d
Proof. Choose an isotropic splitting σ : T → E and identify E ∼ = (T ⊕
T ∗ , h, i , [, ]H ) and d
/0 = dH . Then, for example, setting v1 = X ∈ T and v2 = Y ∈ T
we have
/0 , v1 ·], v2 ·]ρ = [dH , v1 ·](v2 · ρ) − v2 · ([dH , v1 ·]ρ)
[[d
= dH (v1 · v2 · ρ) + v1 · (dH (v2 · ρ)) − v2 · (dH (v1 · ρ)) − v2 · v1 · dH ρ
= d(iX iY ρ) + iX d(iY ρ) − iY d(iX ρ) − iY iX dρ
− H ∧ iX iY ρ − iX (H ∧ (iY ρ) + iY (H ∧ (iX ρ)) + iY iX (H ∧ ρ)
= LX (iY ρ) − iY (LX ρ)
− iX H ∧ iY ρ + iY H ∧ iX ρ − H ∧ iY iX ρ + iY iX (H ∧ ρ)
= i[X,Y ] ρ + (iY iX H) ∧ ρ
= [v1 , v2 ]H · ρ
We leave the remaining cases as an exercise.
We are ready now to characterize the integrability of almost generalized com-
plex structures in terms of the associated complex spinor line. We will use the third
characterization of these objects from Proposition 7.22.
Proposition 7.38. Let J be an almost generalized complex structure on an
exact Courant algebroid E, with pure spinor line K ⊂ Λ∗ T ∗ ⊗ C. Then, J is
integrable if and only if for any local trivialization ρ of K there exists a local section
v ∈ E such that
(7.6) /0 ρ = v · ρ.
d
Proof. Choose an isotropic splitting σ : T → E and identify E ∼ = (T ⊕
∗ /0 = dH . Let L ⊂ E ⊗ C be the complex Dirac structure as-
T , h, i , [, ]H ) and d
sociated to J , as in Proposition 7.22. Given v1 , v2 ∈ L, applying Lemma 7.37, we
calculate
[v1 , v2 ] · ρ = −v2 · v1 · dH ρ = (v1 · v2 − 2 hv1 , v2 i) · dH ρ = v1 · v2 · dH ρ.
Assuming now (7.6), we have
[v1 , v2 ] · ρ = v1 · v2 · v · ρ = v1 · (2 hv, v2 i − v · v2 ) · ρ = 0,
and therefore [v1 , v2 ] ∈ Lρ = L. Conversely, if [L, L] ⊂ L, then
(7.7) v1 · v2 · dH ρ = 0, for any v1 , v2 ∈ L.
Observe that the Dirac structure L induces a filtration
KL = F0 ⊂ F1 ⊂ . . . F2n = Λ∗ T ∗ ⊗ C
where n = dim M and Fk is the subbundle annihilated by products of k + 1 sections

of L. The subbundle F1 decomposes in even and odd degree parts as
F1 = F0 ⊕ E · F0 ,
and since dH is of odd degree, we see from (7.7) that dH ρ ∈ E · KL as claimed.
7.2.3. Associated Poisson structures. Generalized complex geometry turns
out to have many connections to Poisson geometry. As a first glimpse of this here
we will recall the fundamental notions of Poisson geometry and one way in which
they appear in generalized complex geometry. In §7.4 below we will see further
interactions in the case of generalized Kähler geometry.
Definition 7.39. Given M n a smooth manifold, a Poisson bracket on M is a
map
{·, ·} : C ∞ (M ) × C ∞ (M ) → C ∞ (M )
which satisfies
(1) R-linearity: {aφ, bψ} = ab{φ, ψ},
(2) skew-symmetry: {φ, ψ} = −{ψ, φ},
(3) Leibniz rule: {φψ, η} = φ{ψ, η} + ψ{φ, η},
(4) Jacobi identity: {φ, {ψ, η}} + {ψ, {η, φ}} + {η, {φ, ψ}} = 0.
Example 7.40. The most fundamental example of a Poisson bracket, and more-
over the reason for defining this structure in the first place, comes from classical
mechanics. In particular on R2n with variables pi , q i , one defines
X n
∂φ ∂ψ ∂φ ∂ψ
{φ, ψ} = − i .
i=1
∂pi ∂q i ∂q ∂pi
Remark 7.41. By properties 1 and 2 one can show that for a given function f ,
the map {f, ·} is a derivation of C ∞ (M ), i.e. a vector field, and that furthermore
a Poisson bracket determines a section
P ∈ Λ2 T,
which recovers the Poisson bracket via
{φ, ψ} = P (dφ, dψ).
This allows to define the rank of the Poisson structure at a point x ∈ M , as the
rank of the induced map Px : Tx∗ → Tx . In particular, for the example above one
obtains
X n
∂ ∂
P = ∧ i
i=1
∂p i ∂q
which has constant rank 2n. Using P , for any smooth function φ we can obtain the
associated Hamiltonian vector field via
Xφ = P (dφ).
Using these different descriptions we can obtain equivalent descriptions of the inte-
grability condition. In view of this lemma below we will also refer to the tensor P
as a Poisson structure.
Lemma 7.42. Given a smooth manifold M , let P ∈ Λ2 T with associated bracket
{φ, ψ} := P (dφ, dψ). Then the following are equivalent:
(1) For any torsion-free connection ∇, one has
P li ∇l P jk + P lj ∇l P ki + P lk ∇l P ij = 0.
(2) {, } satisfies the Jacobi identity.

(3) For any φ, ψ, one has
X{φ,ψ} = [Xφ , Xψ ].
Proof. Exercise.
Proposition 7.43. Given M 2n and J a generalized complex structure on M ,

define P ∈ End(T ∗ , T ) ∼
= T ⊗ T via
P (ξ) = πJ ξ,
where π : T ⊕ T ∗ → T is the anchor map. Then P is a Poisson structure.
Proof. First we check that P is skew-symmetric, following from

hη, P ξi = hη, πJ ξi = hη, J ξi = − hJ η, ξi = − hπJ η, ξi = − hP η, ξi .
Next we check the integrability. To begin we give an alternate description of P .
Given φ ∈ C ∞ (M ) observe that we can uniquely express
dφ = X + ξ + X + ξ,
where X + ξ ∈ L. Observe that X + X = 0. With this we define

√
Xφ = 2 −1X = −2Im(X), ξφ = −2Im(ξ).
Observe furthermore using the eigenspace decomposition that

J (dφ) = J X + ξ + X + ξ
√ √
= −1 (X + ξ) − −1 X + ξ
= Xφ + ξφ .
It thus follows that the Poisson bracket associated to P is determined by the vector
fields Xφ . In particular, for φ, ψ ∈ C ∞ (M ) we set
{φ, ψ} = Xφ ψ.
To prove the integrability, we will show that
(7.8) X{φ,ψ} = [Xφ , Xψ ],
from which the Jacobi identity follows using Lemma 7.42. We express
dφ = X + ξ + X + ξ, dψ = Y + η + Y + η,
√ √
as above, with Xφ = 2 −1X, Xψ = 2 −1Y . Using that L is involutive we see
that the bracket simplifies as
√ 1
Z + µ := 2 −1[X + ξ, Y + η] = √ [Xφ , Xψ ] + LXφ η − iXψ dξ ∈ L,
2 −1
where the answer lies in L since L is involutive. The vector field component Z is
pure imaginary, hence Z + Z = 0, whereas

µ + µ = LXφ (η + η) − iXψ dξ + dξ
= LXφ dψ − iXψ ddφ
= diXφ dψ
= d{φ, ψ}.
Hence (Z, µ) is the canonical section of L associated to the smooth function {φ, ψ},
and so 7.8 follows.
Building on this observation, we can construct a new class of examples by

explicitly combining complex and Poisson structures.
Example 7.44. Suppose (M 2n , J) is a complex manifold, and let σ ∈ Λ2,0 T
denote a holomorphic
√ Poisson structure. Let Q denote the real part of σ, i.e.
σ = JQ + −1Q, and let

−J Q
JQ = .
0 J∗
Using integrability of J and the Poisson structure, it is possible to show that this
defines an integrable generalized complex structure. This example is elucidated in
[89, Proposition 5], where it is moreover shown that this is a kind of canonical form
for a certain type of generalized complex structure.
7.3. Courant algebroids and pluriclosed metrics

In this section we recall the basic notion of pluriclosed Hermitian metrics (also
known as strong Kähler with torsion or SKT metrics). We interpret them in the
language of generalized geometry and prove some basic results on their structure.
We adopt here the point of view taken in [71, 72, 90] (cf. also [106]). Further
results on SKT metrics and their relationship to generalized geometry appear in
[38].
7.3.1. Pluriclosed metrics. We start by recalling the classical definition of

pluriclosed metrics, and some basic facts on their structure.
Definition 7.45. Let (M 2n , J) be a complex manifold. A Riemannian metric
g is called Hermitian if
g(JX, JY ) = g(X, Y )
for all X, Y ∈ T M . Associated to a Hermitian metric is the Kähler form
ω(X, Y ) = g(JX, Y ).
It follows from the Hermitian condition that ω ∈ Λ1,1

R . The triple (M
2n
, g, J) is a
Hermitian manifold.
When dealing with several complex structures on the same smooth manifold,
we will often use the notation ωJ = g(·, J·).
7.3. COURANT ALGEBROIDS AND PLURICLOSED METRICS 145
Definition 7.46. Let (M 2n , J) be a complex manifold. A Hermitian metric g

on M is called pluriclosed if
ddc ω = 0.
√
Here dc = −1(∂ − ∂) is the conjugate differential, and in particular dc ω =
−dω(J, J, J). We say that the metric is Kähler if furthermore dω = 0.
We have a local description of pluriclosed metrics similar to the local ddc -lemma
for Kähler metrics.
Lemma 7.47. Let (M 2n , J) be a complex manifold and suppose ω ∈ Λ1,1
R satisfies
c
dd ω = 0. Given p ∈ M , for any local complex coordinates defined near p there
exists α ∈ Λ1,0 defined in this coordinate chart so that
ω = ∂α + ∂α.
Proof. We give a sketch ignoring the relevant domains under consideration.

The pluriclosed condition tells us that ∂ω ∈ Λ2,1 is holomorphic, and thus in a local
coordinate chart one can apply the ∂-Poincaré lemma to obtain µ ∈ Λ2,0 such that
∂ω = ∂µ. It follows by a straightforward computation that d (ω − µ − µ) = 0. As
ω − µ − µ is real, it follows from the d-Poincaré lemma that there exists a ∈ Λ1R
such that ω − µ − µ = da. Expressing a = α + α, it follows easily that α ∈ Λ1,0 is
the required potential.
Example 7.48. Let K denote a compact semisimple Lie group with dim K =
2n. Let g denote the unique bi-invariant metric on K. By a classical result of
Samelson [149], we can choose a complex structure on the corresponding Lie algebra
k compatible with g, and we obtain a left-invariant complex structure J. We claim
that (g, J) is pluriclosed with
−dc ω(X, Y, Z) = H(X, Y, Z) = g([X, Y ], Z).
Using the invariance properties and the integrability of J we compute
−dc ω(X, Y, Z) = dω (JX, JY, JZ)
X
= − ω ([JX, JY ], JZ)
σ(X,Y,Z)
X
= − g ([JX, JY ], Z)
σ(X,Y,Z)
X
= − g (J[JX, Y ] + J[X, JY ] + [X, Y ], Z)
σ(X,Y,Z)
X
= − 3H(X, Y, Z) + 2 g ([JX, JY ], Z)
σ(X,Y,Z)
c
= − 3H(X, Y, Z) + 2d ω(X, Y, Z).
Rearranging yields −dc ω = H. It is left as an exercise to show dH = 0.
Example 7.49. Consider the compact four-dimensional manifold M = SU (2)×
U (1) as in Example 3.51. Using the same notation, for every non-zero constant
x ∈ R we define a left-invariant complex structure on M by
Jx e4 = xe1 , Jx e 2 = e 3 .
We leave as an exercise to check that Jx is integrable. Then, the metric g = gk,x

defined in Example 3.51 is Hermitian and the associated Kähler form is
ωk,x = −kxe14 + ke23 .
Furthermore, we have that
dcx ωk,x = −dωk,x (Jx , Jx , Jx ) = kxe234 (Jx , Jx , Jx ) = −ke123
and therefore g is pluriclosed. We will find this family of pluriclosed structures in
a different disguise in Theorem 8.26.
A fundamental result of Gauduchon shows that all compact complex surfaces
admit pluriclosed metrics:
Theorem 7.50. ([77]) Let (M 2n , g, J) be a Hermitian manifold. There exists
a unique u ∈ C ∞ (M ) such that
Z h ∧n−1 i
udV = 0, ∂∂ e2u ω = 0.
M
More specifically, in dimension n = 2 this theorem says that on compact complex

surfaces every Hermitian metric has a conformally related metric which is pluri-
closed, and this metric is unique up to scale in the conformal class. In higher
dimensions pluriclosed metrics need not exist. In [65] examples of compact com-
plex nilmanifolds with invariant complex structure are given which do not admit
compatible pluriclosed metrics.
7.3.2. Courant algebroids and liftings. The relation between pluriclosed

metrics and generalized geometry has typically appeared in relation to generalized
Kähler structures [90] (see also [38]). Following [72], here we shall adopt a dif-
ferent viewpoint which uncovers gauge-theoretical aspects of pluriclosed metrics in
interaction with Courant algebroids. Connecting to the classical perspective, we are
interested in pluriclosed structures on a manifold endowed with closed three-form,
as given in the following definition.
Definition 7.51. Given M 2n a smooth manifold and H0 ∈ Λ3 T ∗ , dH0 = 0,
a pluriclosed structure is a triple (g, b, J) consisting of a Riemannian metric g, a
two-form b, and an integrable complex structure J such that
(1) The metric g is compatible with J,
(2) One has
−dc ω = H0 + db.
Of course, the choice of closed three-form H0 determines an exact Courant
algebroid E over M . Our goal in this section is to describe pluriclosed structures
in terms of the geometry of E. We fix an exact Courant algebroid E over a smooth
manifold M endowed with an integrable complex structure J.
Definition 7.52. A lifting of T 0,1 to E is an isotropic, involutive subbundle
ℓ ⊂ E ⊗ C mapping isomorphically to T 0,1 under the C-linear extension of the
anchor map π : E ⊗ C → T ⊗ C.
By a choice of isotropic splitting on E, we can give an explicit characterization
of liftings.
Lemma 7.53. Given H0 ∈ Λ3 T ∗ , dH0 = 0, consider the exact Courant algebroid

(T ⊕ T ∗ , h, i , [, ]H0 ). Then, a lifting of T 0,1 is equivalent to γ ∈ Ω1,1+0,2 satisfying
(7.9) (H0 + dγ)1,2+0,3 = 0.
More precisely, given γ satisfying (7.9), the lifting is
(7.10) ℓ = {eγ (X 0,1 ), X 0,1 ∈ T 0,1 },
and, conversely, any lifting is uniquely expressed in this way.
Proof. An isotropic subbundle ℓ ⊂ T ⊕ T ∗ ⊗ C mapping isomorphically to
T X under π is necessarily of the form (7.10) for a suitable γ ∈ Λ1,1+0,2 . By
0,1
Proposition 2.14 we have

[eγ X 0,1 , eγ Y 0,1 ]H0 = eγ [X 0,1 , Y 0,1 ] + iY 0,1 iX 0,1 (H0 + dγ).
Then, ℓ is involutive if and only if (7.9) holds.
As we will see in §7.3.3, a lifting relates to the complex Courant algebroid E ⊗C
as a Dolbeault operator relates to a smooth complex vector bundle, in the sense
that it will enable us to construct a Courant algebroid in the holomorphic category
out of E ⊗ C. With this idea in mind, and taking into account that we have a
canonical real structure
E ⊂ E ⊗ C,
it is natural to look for a Chern correspondence. This has been obtained in [72] in
a much more general setup, and we summarise here the result for the case of our
interest.
Definition 7.54. A horizontal lift of T to E is given by a subbundle W ⊂ E
such that
rk W = dim M, and W ∩ T ∗ = {0}.
By definition, a horizontal lift W ⊂ E is equivalent to a (possibly degenerate) real
symmetric 2-tensor g on X and an isotropic splitting σ : T → E such that
(7.11) W = {σ(V ) + g(V ) | V ∈ T }.
In particular, it induces a closed three-form H and an isomorphism E ∼
= (T ⊕
T ∗ , h, i , [, ]H ).
Lemma 7.55. Any lifting ℓ ⊂ E ⊗ C of T 0,1 determines a unique horizontal lift
W ⊂ E such that
(7.12) ℓ = {e ∈ W ⊗ C | π(e) ∈ T 0,1 }.
Furthermore, such W determines uniquely a two-form ω = g(J·, ·) ∈ Λ1,1
R , and
H ∈ Λ3 T ∗ , dH = 0, such that
H = −dc ω.
Therefore, in particular, ddc ω = 0.
Proof. We choose an isotropic splitting σ0 : T → E inducing an isomorphism
E ∼= (T ⊕ T ∗ , h, i , [, ]H0 ) for some closed three-form H0 . By Lemma 7.53, ℓ is
uniquely determined by γ ∈ Λ1,1+0,2 via (7.10). On the other hand, (7.11) implies
that there exists a uniquely defined real two-form b ∈ Λ2 T ∗ such that
W = eb {X + g(X) | X ∈ T }.
The isotropic condition for (7.12) implies that g is a symmetric tensor of type (1, 1).
Denote the associated (1, 1) form by
ω = g(J·, ·) ∈ Λ1,1
R .
Then, condition (7.12) implies

√
eγ (T 0,1 ) = eb+ −1ω
(T 0,1 ),
and therefore √
b1,1 + −1ω 1,1 = γ 1,1 , b0,2 = γ 0,2 .
Take now σ0 to be the √
isotropic splitting determined by W . Then, b = 0 and
H0 = H, and hence γ = −1ω, and thus (see (7.9))
√
H 0,3+1,2 = − −1 ∂ω.
Since ω and H are real, this yields the result.
The previous lemma can be regarded as an analogue of the Chern correspon-
dence for exact Courant algebroids. In particular, the horizontal subspace W ⊂ E,
and hence the associated real (1, 1)-form ω, can be thought of as a connection-like
object. The pluriclosed condition on ω can be regarded as a Bianchi identity for
W . The interpretation of two-forms as connections is reminiscent of higher gauge
theory (see [72] for further insights on this relation). We close this subsection with
a characterization of pluriclosed structures in terms of Courant algebroids.
Theorem 7.56. Let (M, J) be a complex manifold, and H0 ∈ Λ3 T ∗ , dH0 = 0.
Then, a pluriclosed structure of the form (g, b, J) is equivalent to a lifting
ℓ ⊂ (T ⊕ T ∗ ) ⊗ C
of T 0,1 for the twisted Courant algebroid (T ⊕ T ∗ , h, i , [, ]H0 ), such that the real
symmetric two-tensor obtained via Lemma 7.55 is positive definite.
Proof. Given a lifting ℓ as in the statement, by Lemma 7.53 and the proof of
Lemma 7.55 we have that g is a Hermitian metric on M and
√
(H0 )0,3+1,2 + ∂(b1,1 + −1ω) + db0,2 = 0.
Since ω, b, and H0 are real, this yields −dc ω = H0 + db. The converse is left as an
exercise.
7.3.3. Metrics on holomorphic Courant algebroids. Motivated by the
Chern correspondence in Lemma 7.55, we show next how a lifting of T 0,1 determines
a Courant algebroid in the holomorphic category, following [90]. We will then go on
and show that a pluriclosed structure yields a natural notion of Hermitian metric
on these objects, following [71, 160].
Definition 7.57. Let (M, J) be a complex manifold. A holomorphic Courant
algebroid is a holomorphic vector bundle E → M together with a nondegenerate
holomorphic symmetric bilinear form h, i, a bracket [, ] on holomorphic sections of
E, and a holomorphic bundle map π : E → T 1,0 satisfying the axioms in Definition
2.8 (in the holomorphic category). We will say that E is exact if it fits into an exact
sequence of holomorphic vector bundles
∗ π∗ π
0 −→ T1,0 −→ E −→ T 1,0 −→ 0.
A complete classification of exact holomorphic Courant algebroids was obtained

by Gualtieri in [90]. Let us summarize the result which we will use.
Theorem 7.58. Let (M, J) be a complex manifold. Denote by Λ2,0 cl the sheaf of
closed (2, 0)-forms on M . Then, the set of isomorphism classes of exact holomor-
2,0
phic Courant algebroids on M is bijective to the first Čech cohomology H 1 (Λcl ).
Furthermore, there is a vector space isomorphism
3,0+2,1
∼ Ker d : Λ → Λ4,0+3,1+2,2
(7.13) H 1 (Λ2,0
cl ) = .
Im d : Λ2,0 → Λ3,0+2,1
The sheaf Λ2,0
cl in the previous classification result plays an important role in
various developments in mathematical physics (see [73] and references therein). A
one-cocycle for this sheaf determines local data for gluing the model T 1,0 ⊕ T1,0
∗
by
means of holomorphic B-field transformations. On the other hand, a representative
of a cohomology class in the right hand side of (7.13) yields a convenient global
description of a holomorphic Courant algebroid.
Definition 7.59. Let (M, J) be a complex manifold. Given H ∈ Λ3,0+2,1 ,
dH = 0, we denote by
EH = T 1,0 ⊕ T1,0
∗
the exact holomorphic Courant algebroid with Dolbeault operator

∂(X + ξ) = ∂X + ∂ξ − iX H 2,1 ,
symmetric bilinear form
hX + ξ, X + ξi = ξ(X),
bracket given by
[X + ξ, Y + η]H = [X, Y ] + iX ∂η + ∂(η(Y )) + iY iX H 3,0 ,
and anchor map π(X + ξ) = X. We leave as an exercise to check that EH so
defined satisfies the axioms in Definition 7.57.
Given now an exact Courant algebroid E over M and a lifting ℓ ⊂ (T ⊕ T ∗ ) ⊗ C
of T 0,1 (see Definition 7.52), it is a formality to associate a holomorphic Courant
algebroid, namely,
ℓ → Eℓ := E−2√−1∂ω
where ω is determined by ℓ as in Lemma 7.55. More invariantly, Eℓ can be obtained
from ℓ via a natural reduction procedure applied to E ⊗C, which explains the choice
of normalization for ∂ω (see [90] for details). This illustrates further the fact that
ℓ should be regarded as a Dolbeault operator on E ⊗ C, while W in Lemma 7.55
plays the role of its Chern connection.
Taking now the holomorphic point of view, we introduce the notion of a metric
on E (cf. [71, 72]).
Definition 7.60. Let (M, J) be a complex manifold endowed with an exact
holomorphic Courant algebroid E. A metric on E is given by a triple (E, ℓ, ϕ),
where
(1) E is an exact Courant algebroid over M ,
(2) ℓ ⊂ E ⊗ C is a lifting of T 0,1 with ω > 0,
(3) ϕ : Eℓ → E is an isomorphism of holomorphic Courant algebroids inducing
the identity on M .
We next unravel the previous definition in terms of the model in Definition

7.59. The following result illustrates the important role played by (2, 0)-forms in
the theory of pluriclosed structures.
Lemma 7.61. Let (M, J) be a complex manifold, and H0 ∈ Λ3,0+2,1 , dH0 = 0.

Then, there is one to one correspondence between the set of metrics on EH0 and
n √ o
ω + β | ω > 0, dβ = H0 + 2 −1∂ω ⊂ Λ1,1 R ⊕Λ
2,0
.
Proof. By Lemma 7.55, the pair (E, ℓ) determines ω > 0 and an isomorphism
E ∼
= (T ⊕ T ∗ , h, i , [, ]−dc ω ). By definition, Eℓ = E−2√−1∂ω and, by Theorem 7.58,
the isomorphism ϕ : Eℓ → EH0 corresponds to
ϕ = e−β : T 1,0 ⊕ T1,0

∗
→ T 1,0 ⊕ T1,0
∗
√
for β ∈ Λ2,0 satisfying dβ = H0 + 2 −1∂ω. We refer to [72] for further details.
We close this section indicating how the fairly abstract Definition 7.60 produces
a Hermitian metric on the holomorphic vector bundle underlying E, in the classical
sense. Let E be an exact Courant algebroid endowed with a metric (E, ℓ, ϕ). By
Lemma 7.55 the pair (E, ℓ) determines a pluriclosed Hermitian metric g. Define a
Hermitian metric G on Eℓ = E−2√−1∂ω by

gij 0
G= .
0 g lk
Via the isomorphism ϕ : Eℓ → E this construction yields a natural Hermitian metric

(ϕ−1 )∗ G on E, which by direct computations is
!
−1 ∗ gij + βik β jl g lk βip g lp
(ϕ ) G = .
β jp g pk g lk
This metric has an associated Chern connection, and associated Chern curvature.
As we will discuss in §9.6, there a striking relationship between the curvature of
the Chern connection associated to G and the curvature of the Bismut connection
associated to the underlying pluriclosed structure, an idea which is briefly discussed
in [18, Theorem 2.9]. This will play an important role in the higher regularity for
the pluriclosed flow in Chapter 9 (cf. [160, 108]).
7.4. Generalized Kähler geometry

In this section we will give two equivalent definitions of generalized Kähler struc-
ture. This concept first arose in mathematical physics literature [74], expressed in
terms of classical Hermitian geometry and not exploiting the structures of general-
ized geometry. Later Gualtieri [88] rederived this definition motivated by natural
reductions of the structure group of T ⊕ T ∗ . Further works appeared in mathe-
matical physics [103, 104, 105, 127, 128, 129] which influence our discussion
below.
7.4. GENERALIZED KÄHLER GEOMETRY 151
7.4.1. Definitions and equivalences. We will begin with the classical for-
mulation of generalized Kähler geometry in terms of biHermitian geometry, a con-
dition not obviously motivated from the point of view of classical complex and
Kähler geometry.
Definition 7.62. Given M 2n a smooth manifold and H0 ∈ Λ3 T ∗ , dH0 = 0, a
generalized Kähler structure is a quadruple (g, b, I, J) consisting of a Riemannian
metric g, b ∈ Λ2 T ∗ , and two integrable complex structures I, J, such that
(1) The metric g is compatible with I and J.
(2) One has
−dcI ωI = H0 + db = dcJ ωJ .
As the tensor b can be absorbed by redefining H0 , at times we will refer to a
generalized Kähler structure only using the data (g, I, J).
Remark 7.63. We note that the definition allows for the possibility that [H0 ] 6=
0. In some literature, including [88], this is referred to as a twisted generalized
Kähler structure, an extra distinction we will not make.
We next give the definition of Gualtieri [88], expressed in terms of generalized
complex geometry, and motivated from the point of view of reduction of structure
groups. On a 2n-manifold, a choice of complex structure gives a reduction of the
structure group to GL(n, C), whereas a Kähler metric is a reduction to the maximal
compact subgroup U (n). Similarly, as explained in Lemma 7.7, the presence of
a single generalized complex structure, together with the neutral inner product,
reduces the structure group of E to U (n, n). By seeking a further reduction to
the maximal compact subgroup U (n) × U (n), it is possible to derive the following
definition.
M 2n . A generalized Kähler structure is a pair (J1 , J2 ) of generalized complex
structures such that
(1) [J1 , J2 ] = 0,
(2) G = −J1 J2 defines a generalized metric.
The fundamental theorem of Gualtieri shows that Definitions 7.62 and 7.64 are
equivalent. Before showing this we record some further basic structure associated
to generalized Kähler structures as per Definition 7.64. In particular, given J1 , J2
a generalized Kähler structure, the generalized metric G = −J1 J2 determines an
isotropic splitting of E, and therefore an isomorphism E ∼= (T ⊕ T ∗ , [, ]H , h, i , π) for
3 ∗
a uniquely determined√H ∈ Λ T , dH = 0. We also obtain two maximal isotropic
subspaces L1 , L2 , the −1-eigenbundles of J1 , J2 , such that
(T ⊕ T ∗ ) ⊗ C = L1 ⊕ L1 = L2 ⊕ L2 .
Furthermore,
√ as J1 and J2 commute, L1 itself must admit a decomposition into
the ± −1-eigenspaces for J2 , yielding
+ −
(T ⊕ T ∗ ) ⊗ C = L+ −
1 ⊕ L1 ⊕ L1 ⊕ L1 .
As intersections of isotropic, integrable subspaces, each of the four summands above

is itself isotropic and integrable. We are now ready to prove the equivalence.
Theorem 7.65 ([88]). Given M 2n a smooth manifold and H0 ∈ Λ3 T ∗ , dH0 =

0, a quadruple (g, b, I, J) as in Definition 7.62 determines a generalized Kähler
structure (J1 , J2 ) on (T ⊕ T ∗ , [, ]H0 , h, i , π) such that

1 1 0 I ±J −(ωI−1 ∓ ωJ−1 ) 1 0
(7.14) J1,2 = .
2 b 1 ωI ∓ ωJ − (I ∗ ± J ∗ ) −b 1
Conversely, given a generalized Kähler structure (J1 , J2 ) on an exact Courant

algebroid E, it determines, up to a choice of isotropic splitting with associated
H0 ∈ Λ3 T ∗ , dH0 = 0, a quadruple (g, b, I, J) as in Definition 7.62 such that (7.14)
holds.
Proof. To begin we define the relevant maps connecting the two definitions,
then prove that the integrability conditions are the same. We begin by mapping
a generalized Kähler pair (J1 , J2 ) on an exact Courant algebroid identified with
(T ⊕T ∗, [, ]H0 , h, i , π), to a biHermitian structure (g, b, I, J). Let G = −J1 J2 denote
the generalized metric associated to this pair. As described in §2.3, associated to
G is an eigenspace decomposition E = V+ ⊕ V− of E. Observe that since V+ is
positive definite for the neutral inner product, whereas T is a null space, it follows
that
π : V± → T
are isomorphisms, and so we may interpret structures on V± as structures on T

using this map.
In particular, the neutral symmetric inner product defines an inner product
g on T , which is positive definite by construction. Similarly the neutral skew-
symmetric inner product defines a two-form b. This is a manifestation of the fact
that V± are the graphs of b ± g as in Proposition 2.38. To construct the complex
structures, first note that since [J1 , J2 ] = 0 we get a fourfold decomposition of E
according to the eigenspaces of the Ji , and from this it follows easily that V± are
preserved by each Ji . We can transport J1 from both V± to obtain the two almost
complex structures I and J. Since J1 is orthogonal with respect to the neutral
inner product, it follows that I and J are compatible with g.
To finish the proof we must check the integrability conditions, in particular
that I and J are integrable and that dcI ωI = H0 + db = −dcJ ωJ . To check the
√
integrability of I, we must check that the −1-eigenbundle TI1,0 inside T ⊗ C is
closed under the Lie bracket. By definition I is obtained by the action of J1 on V+ ,
using the isomorphism obtained by π. It follows that the preimage of TI1,0 in V+
1,0
must be L1 ∩ V+ ⊗ C = L+ +
1 . As remarked above L1 is integrable, which implies TI
is closed under Lie bracket. The proof for integrability of J is similar. Lastly we
check the conditions on the torsion. In particular, using the explicit descriptions
of the associated metric, one can compute an explicit description of the two Dirac
structures L±1 ⊂ V± ⊗ C, specifically that
√
L+
1 = {X + (b − −1ωI )X| X ∈ TI1,0 },
(7.15) √
L−
1 = {X + (b + −1ωJ )X| X ∈ TJ1,0 }.
Having described these Dirac structures explicitly as graphs, by Proposition 7.30,

we see that since L±1 are both integrable, we obtain the equations (cf. Lemma 7.53)
√
iY iX H0 + d(b − −1ωI ) = 0, for all X, Y ∈ TI1,0
√
iY iX H0 + d(b + −1ωJ ) = 0, for all X, Y ∈ TJ1,0
Considering
√ the (2, 1) piece of the first equation above yields (H0 +√db)2,1 =
−1∂I ωI . Conjugating and using that b is real yields (H0 + db)1,2 = − −1∂ I ωI ,
and hence H0 + db − dcI ωI = 0, and similarly we can obtain H0 + db + dcJ ωJ = 0.
We conclude dcI ωI = H0 + db = −dcJ ωJ , as required (cf. Theorem 7.56).
To reverse this construction, fix (g, b, I, J) as above. Using g and b we determine
a generalized metric G via equation (2.20). With the metric in hand, we have access
to the various projection operators onto V± , which allows us to rebuild the Ji by
hand from the way we extracted I and J above. In particular, one must have
−1 −1
J1 = π|V+
Iππ+ + π|V−
Jππ−
−1 −1
J2 = π|V+
Iππ+ − π|V−
Jππ− .
Unwinding these definitions is a lengthy computation left as an exercise. The
answer is expressed cleanly in terms of the Kähler forms, defined by
ωI (X, Y ) = g(IX, Y ), ωJ (X, Y ) = g(JX, Y ).
Using these one obtains (7.14).
To finish the proof we must establish the integrability of J1,2 , in other words
the Courant involutivity of L1 , L2 , and it suffices to check L1 . To begin we note
that, as the algebraic aspects of the correspondence have been established, it follows
that the Dirac structures L± 1 ⊂ V± ⊗ C take the form of (7.15). Moreover, each
of L±
1 is also integrable by Proposition 7.30 as above. It remains to show that the
bracket of a pair of sections taken one each from L± + −
1 remains in L1 ⊕ L1 . To that
+ −
end we fix X + (b + g)X ∈ Γ(L1 ), Y + (b − g)Y ∈ Γ(L1 ), and compute
[X + (b + g)X, Y + (b − g)Y ] = [X, Y ] − LX (gY ) − LY (gX)
+ d(g(X, Y )) + b([X, Y ]) + iY iX H.
To understand the Lie bracket term we will use the two natural Bismut connections
associated to the generalized Kähler structure, namely
∇IX Y, Z = h∇X Y, Zi − 21 dcI ωI (X, Y, Z),
∇JX Y, Z = h∇X Y, Zi − 12 dcJ ωJ (X, Y, Z).
These are connections compatible with I and J respectively, equivalent to the
connections ∇± of Definition 3.12, whose properties are discussed more fully in
§8.1. Using these connections and letting H = −dcI ωI = dcJ ωJ , we can express
[X, Y ] = ∇X Y − ∇Y X

= ∇JX + 12 g −1 HX Y − ∇IY − 21 HY X
= ∇JX Y − ∇IY X.
A further computation shows that one has
−LX (gY ) − LY (gX) + d(g(X, Y )) = −g(∇X Y + ∇Y X).
Combining these observations now yields

[X + (b + g)X,Y + (b − g)Y ]
= ∇JX Y − ∇IX Y − g(∇X Y + ∇Y X) + b([X, Y ]) + iY iX H
= ∇JX Y − ∇IY X + b(∇JX Y − ∇IY X) − g(∇JX Y + ∇IY X)

= − ∇IY X + (b + g)∇IY X + ∇JX Y + (b − g)∇JX Y .
Noting that, by construction, X ∈ TI1,0 , Y ∈ TJ1,0 , it follows that ∇IY X ∈ TI1,0 and
∇JX Y ∈ TJ1,0 , and thus the expression above is a section of L+ −
1 ⊕ L1 , finishing the
proof.
Remark 7.66. As is clear from the proof of Theorem 7.65, given a generalized
+ −
Kähler structure the subspace L1 (resp. L1 ) is a lifting of TI0,1 (resp. TJ0,1 ) in the
sense of Definition 7.52. This point of view was taken in [90] to analyze generalized
Kähler structures in terms of Morita equivalence for holomorphic Dirac structures,
and has recently been exploited in [17] (cf. Remark 7.79).
7.4.2. Poisson structure. A remarkable fact about generalized Kähler struc-
tures is that there is a further refinement of the Poisson tensors associated to the
generalized complex structures. These play a strong role in determining the local
geometry of a GK structure, and are moreover crucial to constructing certain nat-
ural deformation classes of GK structure (cf. [85, 102]). The Poisson tensor σ
below was originally discovered in [6], [144] in the four-dimensional case, and [101]
in general dimensions.
Definition 7.67. Given (M 2n , g, I, J) a generalized Kähler manifold, let
σ := 21 [I, J]g −1 .
Already from Proposition 7.43 we know that a generalized Kähler structure
will come equipped with Poisson tensors arising from the associated generalized
complex structures, and one arrives at the tensor σ above by considering the type
decomposition of these tensors according to the complex structures I and J. In
particular, this tensor arises as the (2, 0) − (0, 2) projection of the given Poisson
tensors, with respect to either I or J:
I J
σ = Iπ2,0+0,2 πJ1 = Jπ2,0+0,2 πJ2 ,
as follows from a short computation using the Gualtieri map (7.14). To establish
that σ is indeed a Poisson structure, we first need to establish that it is the real
part of a holomorphic (2, 0) tensor with respect to both I and J.
Lemma 7.68. Let (V 2n , J) be a real vector space with almost complex structure
J. Given H ∈ Λ3 V ∗ ,

(π3,0+0,3 H)ijk = 41 Hijk − Jjp Jkq Hipq − Jip Jkq Hpjq − Jip Jjq Hpqk

(π2,1+1,2 H)ijk = 43 Hijk + 14 Jjp Jkq Hipq + Jip Jkq Hpjq + Jip Jjq Hpqk .
Proof. Exercise.
Proposition 7.69. Given (M 2n , g, I, J) a generalized Kähler manifold, one
has that σ is a Poisson tensor, and furthermore
σ ∈ ΛI2,0+0,2 (T M ) ∩ ΛJ2,0+0,2 (T M ), ∂ I σI2,0 = 0, ∂ J σJ2,0 = 0.
Proof. It already follows as discussed above that σ has the claimed algebraic
structure, but we check this directly using the definition of σ. Given α, β ∈ T ∗ M ,
using that I anticommutes with the commutator [I, J], and that g is compatible
with I we obtain
σ(Iα, Iβ) = (Iβ)([I, J]g −1 (Iα))

= β(I[I, J]g −1 Iα)
= − β([I, J]Ig −1 Iα)
= − β([I, J]g −1 α)
= σ(α, β).
Thus σ ∈ ΛI2,0+0,2 (T M ), and an identical argument yields σ ∈ ΛJ2,0+0,2 (T M ).

To check the holomorphicity, we first note that the (2, 0) projection can be
computed as
√
σI2,0 (α, β) = σ(α, β) + −1σ(Iα, β).
Thus to√compute ∂ I σI2,0 it suffices to take the I-Chern derivative in the direction
of X + −1IX. We again refer to §8.1 for fundamental properties of the relevant
Chern connections associated to (g, I) and (g, J). Taking the real part of this it
suffices to check that

0=g ∇C,I
X [I, J] + I∇ C,I
IX [I, J] Y, Z

= g I∇X J − (∇X J)I + I I∇C,I
C,I C,I C,I
IX J − (∇IX J)I Y, Z
(7.16)
= − g (∇C,I
X J)Y, IZ − g (∇ C,I
X J)IY, Z

C,I C,I
− g (∇IX J)Y, Z + g (∇IX J)IY, IZ .
Now we note, using (8.1),
D E
(∇C,I
X J)(Y ), Z
= (∇C,I − ∇C,J )X J)(Y ), Z

= (∇C,I − ∇C,J )X )(JY ) − J(∇C,I − ∇C,J )X )(Y ), Z
= (∇C,I − ∇C,J )X )(JY ) + (∇C,I − ∇C,J )X )(Y ), JZ
1
= 2 (H(X, IJY, IZ) − H(X, Y, JZ) + H(X, IY, IJZ) − H(X, JY, Z)) .
Plugging this into (7.16) means that we must check

0 = − (H(X, IJY, IIZ) − H(X, Y, JIZ) + H(X, IY, IJIZ) − H(X, JY, IZ))
− (H(X, IJIY, IZ) − H(X, IY, JZ) + H(X, IIY, IJZ) − H(X, JIY, Z))
− (H(IX, IJY, IZ) − H(IX, Y, JZ) + H(IX, IY, IJZ) − H(IX, JY, Z))
+ (H(IX, IJIY, IIZ) − H(IX, IY, JIZ) + H(IX, IIY, IJIZ) − H(IX, JIY, IZ))
= H(X, IJY, Z) + H(X, Y, JIZ) − H(X, IY, IJIZ) + H(X, JY, IZ)
− H(X, IJIY, IZ) + H(X, IY, JZ) + H(X, Y, IJZ) + H(X, JIY, Z)
− H(IX, IJY, IZ) + H(IX, Y, JZ) − H(IX, IY, IJZ) + H(IX, JY, Z)
− H(IX, IJIY, Z) − H(IX, IY, JIZ) − H(IX, Y, IJIZ) − H(IX, JIY, IZ)
16
X
= Ai .
i=1
Using that H is of type (2, 1)+ (1, 2) (cf. Lemma 7.68), one observes that A1 + A4 +
A9 +A12 = 0, A2 +A3 +A14 +A15 = 0, A5 +A8 +A13 +A16 = 0, A6 +A7 +A10 +A11 =
0, finishing the claim.
Now we can show that σ is a Poisson tensor. First note that P = πJ1 is a
Poisson tensor, and we have the type decomposition
P = PI2,0 + PI1,1 + PI0,2 = σI2,0 I + PI1,1 + σI0,2 I.
Since P is Poisson, item (1) of Lemma 7.42 holds for P , choosing the Levi-Civita
connection of an arbitrarily chosen metric. The (3, 0) piece of that equation has
terms of the form σI2,0 ⋆ ∇σI2,0 and also of the form PI1,1 ⋆ ∇σ 2,0 . Since σI2,0 is
I-holomorphic, the terms of the form PI1,1 ⋆ ∇σ 2,0 vanish, thus it follows that item
(1) of Lemma 7.42 holds for σI2,0 . Thus σI2,0 is Poisson, as is σ. A similar argument
holds for σJ2,0 .
7.4.3. Examples. We give here some fundamental examples of generalized
Kähler structure which will guide the discussion to follow.
Example 7.70. If (M 2n , g, J) is Kähler, it is clear that this also represents a
generalized Kähler structure with H0 = 0, b = 0 and I = ±J. Note from (7.14)
one has that one of the associated generalized complex structures comes from the
complex structure I = J as in Example 7.32, whereas the other is associated to the
Kähler form ω as in Example 7.33. Also, for this structure one has σ = 0.
Example 7.71. Let (Mi2ni , gi , Ki ), i = 1, 2 be two Kähler manifolds. Let
M = M1 × M2 , g = π1∗ g1 + π2∗ g2 . Note that each Ki defines in a natural way
an endomorphism of T Mi ⊂ T M , and thus we may define

K1 0 K1 0
I= , J= .
0 K2 0 −K2
It is easy to check that these define integrable complex structures on M , which more-
over are both compatible with g. One easily computes that dωI = dπ1∗ ω1 + dπ2∗ ω2 =
0 and dωJ = dπ1∗ ω1 − dπ2∗ ω2 = 0 (note the sign change due to the sign change in
the definition of J), and so this defines a generalized Kähler structure. Here again
the underlying pairs (g, I) and (g, J) are Kähler, but we will show in §7.4.4 below
how to deform this example in general to produce non-Kähler structures.
Example 7.72. Let (M 4n , g, I, J) be a hyperKähler manifold. Setting H0 = 0,

it follows from the properties of hyperKähler structures that (g, 0, I, J) defines a
generalized Kähler structure. Note that even though the underlying Hermitian
structures are Kähler, the associated generalized complex structures are not the
standard ones as in Example 7.70. Furthermore, observe that in this setting one
has using the quaternionic relations
−1
σ = 21 [I, J]g −1 = Kg −1 = ωK .
In particular, the Poisson structure σ defines a nondegenerate pairing on T ∗ . In
this sense this example and Example 7.71 represent the two extremes of generalized
Kähler geometry, and their local structure and geometry vary considerably, as we
will see in the remainder of this chapter.
Example 7.73. Let K denote a compact semisimple Lie group. Let g de-
note the unique bi-invariant metric on K. Fixing any complex structure on the
corresponding Lie algebra k compatible with g, we obtain left- and right-invariant
complex structures IL and IR . We claim that (g, IL , IR ) is a generalized Kähler
structure with
H(X, Y, Z) = g([X, Y ], Z).
As computed in Example 7.48, we know that dcIL ωIL = H. Using that the right
Lie algebra is anti-isomorphic to the left Lie algebra, it follows that dcIR ωIR = −H,
finishing the proof.
7.4.4. Generalized Kähler structures with vanishing Poisson struc-
ture. In this subsection we explore further the local structure of generalized Kähler
manifolds in the case σ = 0, which is equivalent to [I, J] = 0. We set up and prove
a basic structure theorem (cf. [7] Theorem 4) which yields a splitting of the tangent
bundle in this setting, which in a very rough sense indicates that generalized Kähler
geometry with σ = 0 behaves locally like a twisted product of Kähler structures.
Theorem 7.74. Let (M 2n , g, I, J) be a generalized Kähler manifold with
[I, J] = 0.
Let Π = IJ. Then the ±1-eigenspaces of Π are g-orthogonal I- and J-holomorphic
foliations on whose leaves g restricts to a Kähler metric.
Proof. Let T± denote the eigenspace decomposition of Π, whose eigenvalues

are ±1 as it is a real endomorphism which squares to the identity. We note that
T± = ker(Π ∓ Id) = ker(I ± J). But ker(I ± J) is the image of I ∓ J, thus T±
coincide with the images of the associated Poisson tensors
σ1 = πJ1 π∗ = (I − J)g −1 , σ2 = πJ2 π∗ = (I + J)g −1 ,
which are thus integrable. Since g is compatible with both I and J, the projection
Π is orthogonal, and hence T± are orthogonal. We refer to the restrictions of ωI to
these subbundles ωI± , similarly g ± , ωJ± .
The complex structures preserve T± , and thus we obtain a refinement T+ ⊗ C =
A ⊕ A, T− ⊗ C = B ⊕ B, where
A = TI1,0 ∩ TJ0,1 , B = TI1,0 ∩ TJ1,0 .
As intersections of integrable distributions, A and B are both integrable. To show

A is a holomorphic subbundle we show it is preserved by the ∂ operator of I. Using
the decompositions above, given X ∈ Γ(TI0,1 ) and Z ∈ Γ(A),
∂ X Z = [X, Z]1,0 = [X, Z]A + [X, Z]B = [X, Z]A + [XA , Z]B + [XB , Z]B .
Since XA , Z ∈ Γ(T+ ) and T+ is integrable, the term [XA , Z]B vanishes. Also, since
A ⊕ B = TJ0,1 is again involutive, it follows that the term [XB , Z] vanishes, finishing
the claim that A is a holomorphic subbundle.
Lastly, since I = J along the leaves of T− , it follows that dcI = dcJ , and ωI− = ωJ− ,
thus dcI ωI− = dcJ ωJ− , and hence the restriction of ωI− to the leaves of T− is closed.
Similarly, along the leaves of T+ we have I = −J, and hence dcI = −dcJ , while
ωI+ = −ωJ+ . It follows that dcI ωI+ = dcJ ωJ+ and thus the restriction of ωI+ to the
leaves of T+ is closed.
This theorem yields a splitting of the holomorphic tangent bundle T 1,0 = T+1,0 ⊕
T−1,0 , yielding in turn a four-fold splitting
TC = T+1,0 ⊕ T−1,0 ⊕ T+0,1 ⊕ T−0,1 .
By duality we also obtain a splitting of the cotangent bundle, which we indicate as
TC∗ = Λ1,0 1,0 0,1 0,1
+ ⊕ Λ− ⊕ Λ+ ⊕ Λ− .
Going further, this yields a splitting of all form spaces Λp,q . In particular, we define
∧p ∧r ∧q ∧s
1,0
Λp,q
r,s := Λ+ ∧ Λ1,0
− ∧ Λ0,1
+ ∧ Λ0,1 − .
Using the natural projection operators, it follows that there is a further refinement
of the exterior derivative d = ∂ + ∂ into
d : Λp,q p+1,q
r,s → Λr,s ⊕ Λp,q+1
r,s ⊕ Λp,q p,q
r+1,s ⊕ Λr,s+1
d = ∂+ + ∂ + + ∂− + ∂ − ,
where
∂+ = πp+1,q,r,s ◦ d, ∂ + = πp,q+1,r,s ◦ d,
∂− = πp,q,r+1,s ◦ d, ∂ − = πp,q,r,s+1 ◦ d.
A basic example of these structures occurs by taking products of Kähler mani-
folds and changing the orientations of some factors, as in Example 7.71. More gen-
erally we can take quotients of such examples to yield nontrivial, i.e. non-Kähler,
examples. The first such occurs on Hopf surfaces.
Definition 7.75. By definition, a Hopf surface is a compact complex sur-
face whose universal cover is biholomorphic to C2 \{0}. The surface is primary if
π1 (M ) = Z, and is otherwise secondary. For primary Hopf surfaces the fundamental
group is generated by
γ(z1 , z2 ) = (αz1 , βz2 + λz1m )
where 0 < |α| ≤ |β| < 1, and
(α − β m ) λ = 0.
The map γ is obviously a biholomorphism, and thus one obtains a well-defined
complex structure on the quotient, which by an exercise is always diffeomorphic
to S 3 × S 1 . The surface is class 1 if λ = 0, and is class 0 otherwise. Lastly, we say

that the Hopf surface is diagonal if |α| = |β|.
For diagonal Hopf surfaces, a certain Hermitian metric, called the Boothby/Hopf
metric, will play a key role. Let
gE (X, Y )
(7.17) g(X, Y )(z1 ,z2 ) = 2 2,
|z1 | + |z2 |
where gE is the standard Euclidean metric on C2 . This metric is Hermitian with

respect to the standard complex structure on C2 . It is pluriclosed, and its curvature
and torsion properties are further explored in Proposition 8.25 below.
Example 7.76. Consider a diagonal Hopf surface as described above. First

note that, if we set

0 −1
J= ,
1 0
then the standard complex structure I on C2 \{0} can be expressed with respect to
the standard coordinate basis {∂x1 , ∂y1 , ∂x2 , ∂y2 } as the block diagonal matrix

J 0
I= .
0 J
We now define a second complex structure by changing the orientation of the z2

plane, analogously to Example 7.71. In particular, let

J 0
J= .
0 −J
This is easily checked to be integrable, and moreover the map γ(z1 , z2 ) = (αz1 , βz2 ),
the generator of the group of Deck transformations for the Hopf surface, is certainly
holomorphic for the complex structure J as well as I. It follows that both I and J
descend to the quotient surface.
We claim that if g denotes the metric of (7.17), then the triple (g, I, J) is
generalized Kähler. The compatibility of g with I and J is immediate. Note that I
and J commute, and so already we obtain the decomposition d = ∂+ +∂ + +∂− +∂ − ,
and we observe that
√ √
dcI = −1(∂ I − ∂I ) = −1(∂ − + ∂ + − ∂− − ∂+ ),
√ √
dcJ = −1(∂ J − ∂J ) = −1(∂ − − ∂ + − ∂− + ∂+ ).
Thus we compute
!
√ 1
dcI ωI = −1(∂ − + ∂ + − ∂− − ∂+ ) 2 2 (dz1 ∧ dz 1 + dz2 ∧ dz 2 )
|z1 | + |z2 |
√
−1
= 2 2 ((z2 dz 2 − z 2 dz2 )dz1 ∧ dz 1 + (z1 dz 1 − z 1 dz1 ) ∧ dz2 ∧ dz 2 )
(|z1 | + |z2 | )2
and
!
√ 1
dcJ ωJ = −1(∂ − − ∂ + − ∂− + ∂+ ) (dz1 ∧ dz 1 − dz2 ∧ dz 2 )
|z1 |2 + |z2 |2
√
−1
= ((z 2 dz2 − z2 dz 2 ) ∧ dz1 ∧ dz 1 − (z1 dz 1 − z 1 dz1 ) ∧ dz2 ∧ dz 2 )
(|z1 |2 + |z2 |2 )2
= − dcI ωI ,
verifying that dcI ωI = −dcJ ωJ . We leave the computation that ddcI ωI = 0 to Propo-
sition 8.25. This example is notable since while every generalized Kähler manifold
with [I, J] = 0 admits an integrable splitting of the tangent bundle, the manifold
itself need not split as a product according to these leaves. Note that the S 1 leaves
are the projection of radial lines through the origin, the tangent space to which
does not lie in either T± .
A key structural feature of Kähler metrics is the √ ∂∂-lemma, which says that
locally every Kähler metric can be described as ω = −1∂∂f for a potential func-
tion f . It turns out that generalized Kähler metrics with vanishing Poisson tensor
also admit local scalar potentials [164], and potential functions determine natural
global deformations of generalized Kähler structures of this type [7].
Lemma 7.77. Let (M 2n , g, I, J) be a generalized Kähler manifold with [I, J] =
0. Given p ∈ M there exist local complex coordinates (z, w) such that T+ =
∂
span{ ∂z∂ i }, T− = span{ ∂w i } and a smooth function f defined in this coordinate
chart so that
√
ωI = −1 ∂+ ∂ + − ∂− ∂ − f.
Furthermore, given f ∈ C ∞ (M ), suppose
√
ωIf := ωI + −1 ∂+ ∂ + − ∂− ∂ − f > 0,
and let g f := ωIf (I, ·). Then (M 2n , g f , I, J) is a generalized Kähler structure.
Proof. Using Theorem 7.74 we can choose local complex coordinates (z, w)
near p as claimed in the statement. To derive f we first note that since d+ ω+ = 0, on
each w ≡ const
√ plane we can apply the usual ∂∂-lemma to obtain a function ψ+ (z)
such that −1∂+ ∂ + ψ+ = ω+ on that plane. Since ω+ is smooth, we can moreover
choose these on√ each slice so that the resulting function ψ+ (z, w) is smooth, and
thus satisfies −1∂+ ∂ + ψ+ = ω+ on all of√U . Arguing similarly on the z ≡ const
planes we obtain a function ψ− such that −1∂− ∂ − ψ− = ω− everywhere on U .
Using that ω is pluriclosed we obtain
√ √
0 = −1∂+ ∂ + ω− + −1∂− ∂ − ω+ = −∂+ ∂ + ∂− ∂ − (ψ+ + ψ− ) .
We next claim that, as an element in the kernel of the operator ∂+ ∂ + ∂− ∂ − , the
function ψ+ + ψ− , can be expressed as
(7.18) ψ+ + ψ− = λ1 (z, z, w) + λ1 (z, z, w) + λ2 (w, w, z) + λ2 (w, w, z).
To see this we first note that for φ such that ∂+ ∂ + ∂− ∂ − φ = 0, any component
of ∂− ∂ − φ can be expressed as the real part of a ∂ + -holomorphic function, so
(∂− ∂ − φ)wi wj = µij (w, w, z) + µij (w, w, z), where the indices on the µ refer to
the fact that each component of the ∂− ∂ − -Hessian can be expressed this way. It
√
follows that ∆− φ := −1φ,wi wi is the real part of a ∂ + -holomorphic function.
Applying the Greens function on each z-slice it follows that φ can be expressed
as the real part of a ∂ + -holomorphic function, up to the addition of an arbitrary
∂ − -holomorphic function. Thus (7.18) follows.
We√ claim that f = ψ+ −λ2 −λ2 is the required
√ potential function. In particular,
since −1∂+ ∂ + λ2 + λ2 = 0 it follows that −1∂+ ∂ + f = ω+ . Also, we compute
using (7.18),
√ √
− −1∂− ∂ − f = − −1∂− ∂ − −ψ− + λ1 + λ1
√
= −1∂− ∂ − ψ−
= ω− .
The first claim follows.
To show the second claim we first observe that
√
ωJf := ωJ + −1 ∂+ ∂ + + ∂− ∂ − f.
It is clear by construction that d± (ωIf )± = d± (ωJf )± = 0. Also, since I = J along
the leaves of T− , we have (dcI )− = (dcJ )− , while (ωIf )+ = −(ωJf )+ , so
dcI (ωIf )+ = (dcI )− (ωIf )+ = −(dcJ )− (ωJf )+ = −dcJ (ωJf )+ .
A similar computation shows that dcI (ωIf )− = −dcJ (ωJf )− . It is elementary to check
ddcI ωIf = 0, finishing the proof.
Question 7.78. The natural global analogue of this scalar reduction is still
open. Given (M 2n , g, I, J) a compact generalized Kähler manifold such that [I, J] =
0, suppose one has φ1 , φ2 ∈ Λ2 T ∗ such that
(1) φi ∈ Λ1,1 1,1
I ∩ ΛJ ,
(2) d+ φ+ = 0 = d− φi− ,
i
(3) ddcI φi = 0,
(4) [φ1 ] = [φ2 ],
where the final equality is an equation of Aeppli cohomology classes (cf. Definition
9.14) on (M 2n , I) (equivalently J). Conjecturally there exists f ∈ C ∞ (M ) such
that
√
φ1 = φ2 + −1 ∂+ ∂ + − ∂− ∂ − f.
Remark 7.79. Lemma 7.77 shows that a generalized Kähler metric is deter-
mined by a local potential function in the special case [I, J] = 0. These local
potentials conjecturally exist in full generality, with suggestive results appearing in
[17, 129] dealing with most cases.
7.4.5. Generalized Kähler manifolds with nondegenerate Poisson struc-
ture. The other extreme case of generalized Kähler geometry occurs when the
Poisson structure σ defines a nondegenerate pairing on T ∗ . In this setting we can
consider the inverse tensor Ω = σ −1 . We make this definition for an arbitrary gen-
eralized Kähler structure, where one interprets it as existing on the locus where σ
is invertible.
Definition 7.80. Let (M 2n , g, I, J) be a generalized Kähler structure, with
associated Poisson tensor σ as in Definition 7.67. We set
Ω = σ −1 ,
wherever this is well-defined. Observe as an immediate consequence of Proposition

7.69 that
Ω ∈ ΛI2,0+0,2 T ∗ ∩ ΛJ2,0+0,2 T ∗ , ∂ I Ω2,0 = 0, ∂ J ΩJ2,0 = 0.
If M is a manifold where Ω is globally defined, then the pairing σ is everywhere
nondegenerate, and we call the generalized Kähler structure itself nondegenerate.
As it turns out there are further natural symplectic structures associated to a
generalized Kähler structure of this type.
Lemma 7.81. Let (M 2n , g, I, J) be a nondegenerate generalized Kähler struc-
ture. Then
−1
F± = g (I ± J)
define symplectic forms on M . Furthermore, one has dcI ωI = db = −dcJ ωJ where
b = F+ (I − J) .
Proof. Since σ = 12 [I, J]g −1 is nondegenerate, the endomorphism [I, J] =

1
2 (I − J)(I + J) is invertible, and thus so are I ± J. By Proposition 7.43 the
generalized complex structures J1/2 induce Poisson tensors, which by (7.14) are
given by (I ± J)g −1 . These are also nondegenerate, and thus the tensors F± =
−1
g (I ± J) as in the statement are well-defined symplectic forms. Now set b =
F+ (I − J). The claimed equation dcI ωI = db follows by differentiating ωI + bI =
−2F+ and using that dF+ = 0, with dcJ ωJ = −db following similarly.
We next construct a natural class of deformations using Ω-Hamiltonian dif-

feomorphisms first appearing in [6] (cf. [16, 17, 85, 91, 102] for further de-
velopments), indicating a fundamental difference between generalized Kähler and
classical Kähler geometry, namely that the basic deformations occur in a nonlinear
space. The starting point is to reduce the construction of nondegenerate general-
ized Kähler structures purely in terms of the holomorphic symplectic structures. In
fact every nondegenerate generalized Kähler structure arises from the description
below.
Proposition 7.82. Suppose (I, ΩI ) and (J, ΩJ ) are two holomorphic symplec-
tic structures on M satisfying
(1) Re(ΩI ) = Re(ΩJ ),
I
(2) π1,1 (−Im(ΩJ )) is positive definite.
Then, setting
g = −2Im(ΩJ )I,
we have that (g, I, J) is a nondegenerate generalized Kähler structure with Ω =
Re(ΩI ).
Proof. Since the real parts of ΩI and ΩJ agree, we can express

√ √
ΩI = Ω + −1IΩ, ΩJ = Ω + −1JΩ.
where Ω is a real symplectic form on M satisfying Ω ∈ ΛI2,0+0,2 ∩ ΛJ2,0+0,2 . Let

F := 2 Im(ΩI ) − Im(ΩJ ) = 2(IΩ − JΩ).
This is a real symplectic form which, by condition (2), tames the complex structure
I. We thus obtain
(7.19) F (X, IY ) = g(X, Y ) + b(X, Y ),
as the decomposition of −IF into its symmetric part, which is positive, and a skew-
symmetric part b. Now observe, using that Ω is type (2, 0) + (0, 2) with respect to
I, that
F I = 2(IΩI − JΩI) = 2(Ω + JIΩ) = JF.
It is an exercise to show using this that g is J-invariant, and furthermore
(7.20) F (X, JY ) = g(X, Y ) − b(X, Y ).
Having constructed the metric and shown that it is biHermitian, we turn to the
integrability condition.
I J
First, let ωI = π1,1 F and ωJ = π1,1 F denote the Kähler forms of (g, I) and
(g, J), respectively. Note that we can re-express (7.19) and (7.20) as
F = ωI + Ib = ωJ − Jb.
Since F is closed we obtain
dωI = −dIb ∈ ΛI2,1+1,2 , dωJ = dJb ∈ ΛJ2,1+1,2 .
Carrying out the pairing with the complex structure and using the type decompo-
sition, it follows that
dcI ωI = db, dcJ ωJ = −db,
as required.
With this description in place we are now ready to exhibit a natural class of
deformations of nondegenerate generalized Kähler structures.
Proposition 7.83. Let (M 2n , g, I, J) be a nondegenerate generalized Kähler
manifold. Let ft ∈ C ∞ (M ) be a one-parameter family of smooth functions M , and
let Xft be the one-parameter family of Ω-Hamiltonian vector fields associated to ft ,
i.e.
dft = − Xft y Ω.
Let φt be the one-parameter family of diffeomorphisms of M generated by Xf . Then
for all t such that
I

π1,1 −Im(Ωφ∗t J ) > 0,
the triple (I, φ∗t J, Ω) are the complex structures and symplectic structure associated
to a unique nondegenerate generalized Kähler structure.
Proof. For the given nondegenerate generalized Kähler structure, we first
construct the (2, 0) pieces of Ω via
√ √
ΩI = Ω + −1IΩ, ΩJ = Ω + −1JΩ.
For φt as in the statement, let (ΩJ )t = φ∗t (ΩJ ). The pair (φ∗t J, φ∗t ΩJ ) is certainly
holomorphic symplectic, and since φt is Ω-Hamiltonian, it follows that
√ √
(ΩJ )t = φ∗t (Ω + −1JΩ) = Ω + −1(φ∗t J)Ω.
Thus Re((ΩJ )t ) = Re(ΩI ) for all t, and the result follows from Proposition 7.82.
Remark 7.84. (1) Observe that while the complex structure J is acted
on by a diffeomorphism, the metric gt will be expressed as per Proposition
7.82) via
gt = −2Im(Ωφ∗t J )I,
which is not a diffeomorphism pullback, and so these deformations are in
general nontrivial for g.
(2) This proposition is directly analogous to the ∂∂-lemma in Kähler geome-
try, whereby a single scalar function naturally determines a deformation
of Kähler structure. However, as remarked above, the scalar potentials
drive a one-parameter family of Hamiltonian diffeomorphisms, and so we
do not in general have an explicit description of gt in terms of ft alone,
rather it is determined by the whole path fs , 0 ≤ s ≤ t.
(3) Even when starting at a hyperKähler structure, as long as the functions f
are not constant the deformation arising from Proposition 7.83 will yield
structures which are not hyperKähler. This can be checked by showing
1
that the angle function p = − 4n tr(IJt ) becomes nonconstant, ruling out
that the structure is hyperKähler.
7.4.6. Generalized Kähler structures with Poisson tensor of mixed
rank. In general, the rank of the associated Poisson tensor σ can vary across the
manifold. Our first example of this is a different generalized Kähler structure on
diagonal Hopf surfaces, where the Poisson tensor σ is nondegenerate away from a
set of real codimension 2.
Example 7.85. Let (g, I) denote the metric and complex structure as in Ex-
ample 7.76. Whereas in that example the complex structure J was easily obtained
by changing orientation on the z2 plane, here the second complex structure J is
quite different. Consider the complex 1-forms on M ,
µ1 = z 1 dz1 + z2 dz 2 ,
µ2 = z 1 dz2 − z2 dz 1 .
The complex structure J is defined by declaring TJ0,1 = ker µ1 ∩ker µ2 . This is indeed
integrable, as direct computations show that dµ1 ∧ µ1 ∧ µ2 = dµ2 ∧ µ1 ∧ µ2 = 0.
The plane spanned by µ1 , µ2 is easily checked to be isotropic with respect to g −1 ,
hence g is compatible with J. Further direct computations yield
√
−1
ωJ = 2 2 (µ1 ∧ µ1 + µ2 ∧ µ2 ) .
(|z1 | + |z2 | )2
Using this the integrability condition dc ωI = −dc ωJ can be checked by a tedious
but straightforward computation.
We can interpret the standard complex structure I analogously via TI0,1 =
ker dz1 ∩ker dz2 . This makes clear that on the elliptic curve {z2 = 0}, one has I = J,
whereas along the elliptic curve {z1 = 0}, one has I = −J. It follows that σ has rank
4 away from these two elliptic curves, where it has rank 0. Furthermore, observe
that the Hopf metric is compatible with the standard hypercomplex structure on
C2 , but that the complex structure J is not part of this hypercomplex structure.
By adapting the methods of the previous subsection we can produce deforma-
tions of this structure by taking the vanishing locus of σ into account explicitly.
Proposition 7.86. Let (M 4n , g, I, J) be a generalized Kähler manifold such

that σ is generically nondegenerate, and let
K = {p ∈ M | rank σ < 2n}.
Let ft be a one-parameter family of smooth functions on M \K, and let Xft be the
one-parameter family of Ω-Hamiltonian vector fields associated to ft , i.e.
dft = −Xft yΩ.
Suppose
(1) Xft extends to a smooth one-parameter family of vector fields X eft defined
on M
(2) ddcJ ft extends to a smooth one-parameter family of tensor fields Wt defined
on M .
Let φt be the one-parameter family of diffeomorphisms generated by X eft . Then for
all t such that
I

π1,1 −Im(Ωφ∗t J ) > 0,
the triple (I, φ∗t J, Ω) are the complex structures and symplectic structure associated
to a unique nondegenerate generalized Kähler structure defined on M \K, which
extends smoothly to a generalized Kähler structure on M .
Proof. To show that the metric is well-defined across K we must compute
the variation of the tensor Ft = 2(Im(ΩI ) − Im(φ∗t ΩJ )). Using the definitions above
and the Cartan formula we see
∂
Ft = − 2LXft (JΩ) = −2ddcJ ft = −2Wt .
∂t
The above computation a priori only makes sense away from K, but since the right
hand side is defined smoothly across K we take the equation to be the definition
of the variation of g and b, using that F I = g + b away from K. The preservation
of the generalized Kähler conditions holds on the complement of K by Proposition
7.83. Since this set is dense, it follows that the resulting structure is generalized
Kähler, as claimed.
Remark 7.87. In [102], Hitchin constructed nontrivial generalized Kähler
structures on del Pezzo surfaces. In particular, given (M 4 , J) del Pezzo with Kähler
metric g, one chooses a holomorphic Poisson tensor σJ , and then the construction
follows as in Proposition 7.86, deforming the (Kähler) generalized Kähler structure
2
(g, J, J) using the function f = log σJ−1 . One has to check the behavior of f
near the vanishing locus of σJ to show that the associated diffeomorphisms φt are
well-defined. Using that c1 > 0 it follows that J, φ∗t J and
J

ωJ := π1,1 −Im Ωφ∗t J > 0
define a non-Kähler generalized Kähler structure.
CHAPTER 8
Canonical Metrics in Generalized Complex

Geometry
Having defined the basic objects and deformations in pluriclosed and gener-
alized Kähler geometry, we now turn to the question of motivating and defining
notions of canonical pluriclosed and generalized Kähler structures. We begin by
taking the classical point of view, defining connections associated to a given Her-
mitian manifold, then analyzing their torsion and curvature. We use this to define
an Einstein-type equation for certain Hermitian metrics, and then show that these
structures define generalized Ricci solitons. We end the chapter by giving examples
and classification results for these structures.
8.1. Connections, Torsion, and Curvature

8.1.1. Hermitian connections.
Definition 8.1. Let (M 2n , g, J) be a Hermitian manifold. We say that a
connection ∇ on T M is Hermitian if it is compatible with both g and J, i.e.
∇g ≡ 0, ∇J ≡ 0.
For a Kähler manifold, owing to the integrability of J and the fact that dω = 0,
it is an exercise to show that the Levi-Civita connection is Hermitian. When these
integrability conditions are weakened however, this will no longer be the case, and
one must modify the Levi-Civita connection to obtain one which preserves both g
and J. While many Hermitian connections are present in general (cf. [79]), we will
focus on the two most relevant to generalized Ricci flow.
Definition 8.2. Let (M 2n , g, J) be a Hermitian manifold. Let ω(X, Y ) =
g(JX, Y ) define the Kähler form, and let ∇ denote the Levi-Civita connection
associated to g. The Chern connection is defined by
∇C 1
X Y, Z = h∇X Y, Zi − 2 dω(JX, Y, Z).
The Bismut connection is defined by

∇B 1 c
X Y, Z = h∇X Y, Zi − 2 d ω(X, Y, Z).
Also, it is useful to express the Chern connection in terms of the Bismut torsion.
Using that dc ω(X, Y, Z) = −dω(JX, JY, JZ) we obtain
(8.1) ∇C 1 c
X Y, Z = h∇X Y, Zi + 2 d ω(X, JY, JZ).
These two connections admit characterizations in terms of natural conditions

on the torsion of possible Hermitian connections.
Lemma 8.3. Let (M 2n , g, J) be a Hermitian manifold.
167
168 8. CANONICAL METRICS IN GENERALIZED COMPLEX GEOMETRY
(1) The Chern connection is the unique Hermitian connection such that
T 1,1 ≡ 0,
that is, the (1, 1) piece of the torsion, interpreted as a section of Λ2 T ∗ ⊗T ,

vanishes.
(2) The Bismut connection is the unique Hermitian connection with torsion
tensor T satisfying
T g ∈ Λ3 T ∗ .
Proof. Exercise.
In what follows below we will write
T = ∂ω ∈ Λ2,1
for the (2, 1) piece of the torsion tensor of the Chern connection, while
H = −dc ω ∈ Λ3
denotes the torsion of the Bismut connection.
Remark 8.4. Note that of course the connection ∇B is the same as the Bismut
connection ∇+ of Definition 3.12, with H = −dc ω. We keep the notation ∇B in
the context of Hermitian geometry to emphasize the distinction between it and the
Chern connection, and also since a ± duality appears in relation to many other
constructions as well. Furthermore, in the context of generalized Kähler geometry,
we see that the two Bismut connections ∇± associated to H = −dcI ωI = dcJ ωJ are
in fact the Bismut connections associated to the two pluriclosed structures (g, I)
and (g, J). Here again we will adopt specialized notation and name the associated
Bismut connections ∇I and ∇J , i.e.
∇IX Y, Z = h∇X Y, Zi − 12 dcI ωI (X, Y, Z) = h∇X Y, Zi + 21 H(X, Y, Z),

∇JX Y, Z = h∇X Y, Zi − 12 dcJ ωJ (X, Y, Z) = h∇X Y, Zi − 21 H(X, Y, Z).
These connections already appeared in Chapter 7.
8.1.2. The Lee form. In generalized geometry, understanding the torsion is

of central importance, and in the setting of complex geometry we have a natural
contraction of the torsion tensor, classically called the Lee form. This tensor plays
a central role in understanding the structure of pluriclosed flow and generalized
Kähler-Ricci flow, entering as a modification of the underlying divergence operator.
Definition 8.5. Let (M 2n , g, J) be a Hermitian manifold. The associated Lee

form is defined by
θ = −d∗ ω ◦ J.
Lemma 8.6. Let (M 2n , g, J) be a Hermitian manifold. Then

(1) θ(X) = 12 H(JX, ei , Jei ),
(2) If n = 2, then θ = ⋆H.
8.1. CONNECTIONS, TORSION, AND CURVATURE 169
Proof. We first compute in local coordinates, using that J is parallel with

respect to the Bismut connection,
∇i ωjk = ∇i (gjl Jkl )
= (∇i − ∇B l
i )Jk gjl
p l
= 1
2 Hik l
Jp − Hip Jkp gjl .
Contracting with respect to the indices i and j and using −d∗ ω = trg ∇ω, yields
−d∗ ω(X) = 12 H(ei , Jei , X),
from which item (1) follows. Item (2) is left as an exercise.
8.1.3. Curvature formulas. As we will see below, the curvatures of the Levi-
Civita, Bismut, and Chern connections are all relevant to understanding canonical
metrics in Hermitian geometry, and relationships between them are key. We record
some relevant formulas here.
Definition 8.7. Let (M 2n , g, J) be a Hermitian manifold. We let RB,C denote
the (3, 1)-curvature tensors of the Bismut and Chern connection, respectively. That
is,

RB,C (X, Y )Z = ∇B,C
X ∇B,C
Y Z − ∇B,C
Y ∇B,C B,C
X Z − ∇[X,Y ] Z.
We furthermore adopt the standard notation for the (4, 0)-curvature tensor,
RB,C (X, Y, Z, W ) = RB,C (X, Y )Z, W .
It is an elementary exercise to show that
RB ∈ Λ2 ⊗ Λ1,1 , RC ∈ Λ1,1 ⊗ Λ1,1 ,
where RC is type (1, 1) in the first indices due to the torsion having no (1, 1)
component.
As with the Riemannian curvature tensor, there is only one natural trace (up
to sign), yielding the Ricci tensor. We have already derived a formula for the
Bismut Ricci curvature tensor in Proposition 3.18. In the complex setting, due to
the complex structure many different traces are possible, and several are relevant
to the sequel.
Definition 8.8. Let (M 2n , g, J) be a Hermitian manifold. Let
ρB,C (X, Y ) := 1
2 RB,C (X, Y )Jei , ei ,
SB,C (X, Y ) := 1
2 RB,C (Jei , ei )X, Y ,
where {ei } is any orthonormal basis for the tangent space at a given point. The ten-
sors ρB,C are referred to as the (first) Bismut/Chern Ricci curvatures, and SB,C are
the (second) Bismut/Chern Ricci curvatures. Furthermore we define the associated
Bismut/Chern scalar curvatures
sB,C = ρB,C (Jei , ei ) = SB,C (Jei , ei ).
The tensors ρB,C admit a different interpretation which makes clearer their
topological significance. Since the connections ∇B,C preserve J, it is clear that they
induce connections on Λn,0 T , the anticanonical line bundle associated to (M 2n , J).
The two forms ρB,C are the curvature tensors of these connections, and as such are
closed and determine representatives of πc1 (M, J). Furthermore, as noted above,
for Kähler manifolds both the Chern and Bismut connections agree with the Levi-
Civita connection, and moreover the tensors ρ and S are equal to each other, and
satisfy
ρ(X, Y ) = S(X, Y ) = Rc(JX, Y ) = − Rc(X, JY ).
In the general non-Kähler setting the tensors ρB,C and SB,C are all different, but
useful for different purposes. In the lemmas below we derive relationships between
these different Ricci-type tensors needed for the sequel (cf. [107] for related formu-
las, although with different conventions).
Lemma 8.9. Let (M 2n , g, J) be a pluriclosed structure. Then
ρB (X, Y ) = dd∗ ω(X, Y ) + ρC (X, Y ),
(8.2)
ρB (X, Y ) = − RcB (X, JY ) − ∇B
X θ(JY ).
Proof. To compute the first item of (8.2) it suffices to compute the induced
connection on the canonical bundle. To that end we compute, using (8.1) and
Lemma 8.6,
(∇B − ∇C )X Jei , ei = − 21 dc ω(X, Jei , ei ) − 12 dc ω(X, JJei , Jei )
= − dc ω(X, Jei , ei )
= 2d∗ ω(X).
Using this and the definitions of ρB,C gives the first item of (8.2). For the second
item of (8.2) we first use the Bianchi identity for RB with H = −dc ω, specifically
combining (3.24) and (3.26), to obtain
X
RB (X, Y, Z, U )
σ(X,Y,Z)
X
= H(H(X, Y ), Z, U ) + (∇B
X H)(Y, Z, U )
σ(X,Y,Z)
X
= {H(H(X, Y ), Z, U ) − 2g(H(X, Y ), H(Z, U ))} + (∇B
U H)(X, Y, Z)
σ(X,Y,Z)
X
= (∇B
U H)(X, Y, Z) − g(H(X, Y ), H(Z, U )).
σ(X,Y,Z)
1
Adding back 2 times (3.26) and using dH = 0 yields
(8.3)
 
X X
RB (X, Y, Z, U ) = 1
2
 (∇B B 
X H)(Y, Z, U ) + (∇U H)(X, Y, Z) .
σ(X,Y,Z) σ(X,Y,Z)
Tracing the left hand side of this over Z and U with respect to J and using that
RB takes values in (1, 1) forms, we obtain
RB (X, Y, ei , Jei ) + RB (Y, ei , X, Jei ) + RB (ei , X, Y, Jei )
= − 2ρB (X, Y ) − RB (Y, ei , JX, ei ) − RB (ei , X, JY, ei )
= − 2ρB (X, Y ) + RcB (Y, JX) − RcB (X, JY ).
8.1. CONNECTIONS, TORSION, AND CURVATURE 171
Also, the right hand side of (8.3), when traced over Z and U with respect to J
yields

1
2 (∇B B B B
X H)(Y, ei , Jei ) + (∇Y H)(ei , X, Jei ) + (∇ei H)(X, Y, Jei ) + (∇Jei H)(X, Y, ei )
= 21 (∇B 1 B
X H)(Y, ei , Jei ) + 2 (∇Y H)(ei , X, Jei )
= ∇B B
X θ(JY ) − ∇Y θ(JX),
where the last line follows from Lemma 8.6. Putting these computations together
yields
(8.4) −2ρB (X, Y ) + RcB (Y, JX) − RcB (X, JY ) = −∇B B
Y θ(JX) + ∇X θ(JY ).
On the other hand we can contract (3.22) to obtain

RcB (Y, JX) + RcB (X, JY ) = RmB (ei , Y, JX, ei ) + RmB (ei , X, JY, ei )
= − RmB (ei , Y, X, Jei ) − RmB (ei , X, Y, Jei )
= RmB (X, Jei , ei , Y ) − RmB (ei , Y, X, Jei )
(8.5)
= 21 ∇B 1 B
X H(Jei , ei , Y ) − 2 ∇Jei H(X, ei , Y )
− 21 ∇B 1 B
ei H(Y, X, Jei ) + 2 ∇Y H(ei , X, Jei )
= − ∇B B
X θ(JY ) − ∇Y θ(JX).
Combining (8.4) and (8.5) gives the second item of (8.2).

B
Proposition 8.10. Let (M 2n , g, J) be a pluriclosed structure. Let d∇ denote
the exterior derivative associated to the Bismut connection, and let
(8.6) Q(X, Y ) = hXyT, Y yT i ,
where T denote the torsion of the Chern connection. Then
ρ1,1
B = SC − Q,
(8.7) ρ1,1 1 2 1
B (·, J·) = Rc − 4 H + 2 Lθ ♯ g,
2,0+0,2 B
ρB (·, J·) = − 12 d∗ H + 21 d∇ θ.
Proof. The first item is left as an exercise. For the second item we compute
using Lemma 8.9,
ρ1,1
B (X, JY ) =
1
2 (ρB (X, JY ) − ρB (JX, Y ))

= 1
2 RcB (X, Y ) + ∇B B B
X θ(Y ) + Rc (JX, JY ) + ∇JX θ(JY ) .
Using line (8.5) with X replaced by JX yields

RcB (JX, JY ) = RcB (Y, X) − ∇B B
JX θ(JY ) + ∇Y θ(X).
Plugging this in above yields

ρ1,1
B (X, JY ) = RcB (X, Y ) + RcB (Y, X) + ∇B
1
2
B
X θ(Y ) + ∇Y θ(X)

= Rc − 41 H 2 + 12 Lθ♯ g (X, Y ),
where the last line follows using Proposition 3.18 and the skew-symmetry of the
torsion of ∇B .
To show the last claim we again use Lemma 8.9 and line (8.5) and compute
2,0+0,2
ρB (X, Y ) = 12 ρB (X, Y ) − ρB (JX, JY )

= − 21 RcB (X, JY ) + ∇B B B
X θ(JY ) + Rc (JX, Y ) + ∇JX θ(Y )

= 12 RcB (Y, JX) − RcB (JX, Y ) + ∇B B
Y θ(JX) − ∇JX θ(Y )
B
= 12 d∗ H(JX, Y ) − 21 d∇ θ(JX, Y ),
where the last line follows using Proposition 3.18. The final claimed formula follows
easily from this.
8.2. Canonical metrics in complex geometry

In this section we will give a define a notion of canonical pluriclosed metric us-
ing the curvature of the associated Bismut connection. Recall that in Chapter 3 we
began by investigating the curvature of the Bismut connection, then determining
how the Bismut Ricci curvature fits into generalized geometry, eventually arriving
at the generalized Einstein-Hilbert functional which has generalized Einstein struc-
tures as critical points. We proceed by using the Bismut curvature associated to
a pluriclosed structure to define a kind of Einstein-type condition we call Bismut
Hermitian-Einstein. We will quickly see that these structures are in fact generalized
Ricci solitons. The relationship of this condition to generalized complex geometry
will become clear through the rest of this chapter and the next. In particular, in
the setting of generalized Kähler geometry we will see the relationship between the
Bismut Ricci curvature and the spinor formulation of the associated generalized
complex structures. Also in Chapter 9 (cf. Remark 9.53) we will see that Bismut
Hermitian-Einstein structures naturally define an associated Hermitian metric on
T 1,0 ⊕ T1,0
∗
which is Hermitian Yang-Mills.
8.2.1. Bismut Hermitian-Einstein metrics.

Definition 8.11. Let (M 2n , g, J) be a pluriclosed structure. We say that it is
Bismut Hermitian-Einstein if
ρB = 0.
Remark 8.12. Bismut Hermitian-Einstein structures have been studied be-
fore in mathematics and physics literature (cf. for instance [107] and references
therein). We resist the terminology ‘Bismut Ricci flat’ as this is reserved for the
vanishing of the Ricci tensor associated to the Bismut connection ∇+ as in Chap-
ter 3. Furthermore, as we will see in Proposition 8.14 below these structures are
actually solitons.
More generally it is possible to allow for an Einstein constant in this definition,
and consider either the equation ρB = λω or ρ1,1 B = λω. The first automatically
implies ω is Kähler if λ 6= 0 by taking exterior derivative. No known solutions to
the second equation are known which are not already Kähler, and in fact rigidity
results are known (cf. [166, Proposition 3.5]), in line with Proposition 4.28 above.
Even if one considers the more general equation ρ1,1 B = f ω for a smooth function
f , then at least on a compact√ manifold
the function f is forced to be constant.
To see this, apply ω n−2 ∧ −1∂∂ to both sides of the equation and use that
√
−1∂∂ρ1,1
B = 0 to obtain a strictly elliptic linear equation for f , which forces f to
8.2. CANONICAL METRICS IN COMPLEX GEOMETRY 173
be constant by the maximum principle. For these reasons we focus here only on
the equation ρB = 0.
Lemma 8.13. Let (M 2n , J) be a complex manifold. The Bismut Hermitian-
Einstein equation is a strictly elliptic equation for a pluriclosed metric g.
Proof. Using the first equation of Proposition 8.10, we can rewrite the (1, 1)
piece of the Bismut Hermitian-Einstein equation as
SC − Q = 0.
Since Q is a first-order differential operator in g, and ω is zeroth order in g, we
only need to consider the quantity SC . In local complex coordinates the Chern
connection has Christoffel symbols
Γkij = g lk ∂i gjl ,
and then one directly computes the curvature
√ √ √
(SC )ij = −1g lk Rklij
C
= − −1g lk gjm ∂l Γm lk ⋆2
ki = − −1g ∂l ∂k gij + ∂g ,
which is manifestly a strictly elliptic operator for g, and the lemma follows.
The next proposition indicates the key linkage between the Bismut Hermitian-
Einstein condition and the generalized Ricci flow, namely that Bismut Hermitian-
Einstein structures are automatically steady generalized Ricci solitons.
Proposition 8.14. Let (M 2n , g, J) be Bismut Hermitian-Einstein. Then (g, H)
is a steady generalized Ricci soliton with X = θ♯ , B = dθ − iθ♯ H.
Proof. Pairing the (1, 1) part of the Bismut Hermitian-Einstein equation with
J and using the second equation of (8.7) gives
Rc − 41 H 2 + 21 Lθ♯ g = 0,
which is the metric component of the steady generalized Ricci soliton equation with
2,0+0,2
vector field X = θ♯ . Using the vanishing of ρB and the third equation of (8.7)
we obtain
B
0 = d∗ H − d∇ θ = d∗ H − dθ + iθ♯ H.
Thus we have verified the equations (4.14) of Proposition 4.28 for X = θ♯ and
B = dθ − iθ♯ H, as claimed.
8.2.2. Bismut Ricci curvature in generalized Kähler geometry. In this
subsection we derive explicit formulas for the Bismut Ricci curvature of a general-
ized Kähler manifold. A central role is played by a natural scalar quantity associated
to a generalized Kähler structure:
Definition 8.15. Let (E, [, ], h, i) be an exact Courant algebroid. Suppose
(J1 , J2 ) is a generalized Kähler structure defined by pure spinors ψ1 , ψ2 . The
Bismut Ricci potential is
(ψ1 , ψ 1 )
Φ = log ,
(ψ2 , ψ 2 )
where (, ) is the Mukai pairing (cf. §7.1.3).
We note that Gualtieri defined a special class of generalized Kähler structure

called ‘generalized Calabi-Yau metric geometry’ [88, Definition 6.40] which in our
notation is expressed
Φ ≡ constant.
As we will see below, in general this quantity determines both the Lee forms and
the Bismut Ricci curvature of a given generalized Kähler structure. Furthermore,
in Chapter 9 we will show that in a natural sense this scalar quantity determines
the generalized Kähler-Ricci flow lines.
To begin our computations, we recall the fundamental transgression formula,
whose proof is left as an exercise.
Proposition 8.16. Given (M 2n , J) a complex manifold, E → M a holomor-
phic vector bundle, and Hermitian metrics h1 , h2 on E with associated Chern cur-
vature tensors Ω1 , Ω2 , one has
√ det h1
tr (Ω1 − Ω2 ) = − −1∂∂ log .
det h2
We first the scalar reduction for the Kähler-Einstein equation, and how it fits into
generalized Kähler geometry.
Remark 8.17. Suppose (M 2n , g, J) is a Kähler manifold. As noted above
in this case the Bismut and Chern connections coincide. Using the transgression
formula, in local complex coordinates we can express
√
ρC = − −1∂∂ log det g.
If M admits a holomorphic volume form Ω ∈ Λn,0 , one instead obtains the global
formula
√ ωn
(8.8) ρC = − −1∂∂ log .
Ω∧Ω
In case M is compact, by the maximum principle we see that ρC ≡ 0 if and only if
ωn
≡ const.
Ω∧Ω
This equation admits a natural interpretation in terms of generalized Kähler geom-
etry. First, we can set I = J and then it follows directly that (g, I, J) is generalized
Kähler. Recall that the generalized complex structures associated to a Kähler
structure interpreted this way are the structures JJ ,√Jω from Examples 7.32, 7.33,
naturally induced by the two global spinors Ω and e −1ω . Furthermore, the scalar
equation above can be expressed as equality of the two associated spinor norms, up
to scaling. That is, we see that ρC ≡ 0 if and only if
ωn (ω, ω)
Φ = log = log ≡ const.
Ω∧Ω Ω, Ω
Moreover, we obtain the global expression
ρIB = ρC = − 21 dJdΦ,
with the same equation holding with the roles of I and J exchanged. This same
formula will determine the Bismut Ricci tensor in general, as we will see in examples
below.
We now turn to the case of commuting generalized Kähler structures, adopting

the notation of §7.4.4. Before deriving the formula for the Bismut Ricci curvature
we make an observation about the associated Lee forms which we need in the sequel.
Lemma 8.18. Let (M 2n , g, I, J) be a generalized Kähler manifold such that
[I, J] = 0. Letting θI , θJ denote the Lee forms with respect to (g, I) and (g, J), then
one has
θI = θJ .
Proof. First we choose adapted complex coordinates for the complex struc-
ture I as in Lemma 7.77. These coordinates can be furthermore chosen so that
+ −
∂zi ωjk = ∂wi ωjk = 0. It follows using a computation of the Christoffel symbols in
these coordinates that
√
(θI )k = (Id∗ ωI )k = −1 g zj zi ∂zj (ωI+ )zi k + g wj wi ∂wj (ωI− )wi k .
To compute θJ , we use the same coordinates, and note that ωJ+ = −ωI+ , ωJ− = ωI− ,
whereas the action of J agrees with that of I on T− , and differs by a sign on T+ .
It follows that
√
(θJ )k = −1 −g zj zi ∂zj (ωJ+ )zi k + g wj wi ∂wj (ωJ− )wi k = (θI )k .

2n
Definition 8.19. Let (M , g, I, J) be a generalized Kähler manifold with
[I, J] = 0. Define the split canonical bundles by
K± = Λrank T± (Λ1,0
± ).
As holomorphic line bundles over M , these have first Chern classes with respect
to the complex structure I, and we refer to these as c± 1 := c1 (K± , I). Given
a Hermitian metric g on M , we obtain induced metrics on K± , and therefore
representatives of c±
1 . We refer to these curvature representatives as
ρ± = ρC (g, K± , I).
These curvature tensors are real two-forms of type (1, 1) with respect to I. It will
be useful for us to refer to the different components of these curvature operators
with respect to the decomposition induced by T = T+ ⊕ T− . In particular we set
ρ+
± = πΛ1,1 T±
+
∗ρ , ρ−
± = πΛ1,1 T±
−
∗ρ .
Proposition 8.20. Let (M 2n , g, I, J) be a generalized Kähler manifold with

[I, J] = 0. Then
(8.9) (ρIB )1,1 = ρ+ + − −
+ − ρ− + ρ− − ρ+ .
Furthermore, given h = h+ ⊕ h− a Hermitian metric, define

P (h) := ρ+ (h+ ) − ρ− (h+ ) + ρ− (h− ) − ρ+ (h− ),
where ρ(h± ) denotes the representative of c1 (K± , I) as above, and ρ± (h± ) denote
the projections as in Definition 8.19. Given e
h=eh+ ⊕ eh− another Hermitian metric,
one has
e
√ (ω h )∧k ∧ (ω−h l
)
(8.10) h) = − −1 ∂+ ∂ + − ∂− ∂ − log e+
P (h) − P (e .
h h
(ω+ ) ∧ (ω− )∧l
∧k
Proof. The proof of (8.9) is left as a fairly lengthy but straightforward exer-
1,1
cise. The simplest way is to derive a general formula for ρB in arbitrary complex
coordinates, choose split coordinates as per Lemma 7.77, and then compare against
the usual formula for the curvature of a determinant line bundle. See [165, Propo-
sition 3.2] for details.
To check (8.10), we use the definitions as well as the transgression formula for
the first Chern class (Proposition 8.16) to compute
n o n o
P (h) − P (eh) = ρ+ (h+ ) − ρ+ (eh+ ) − ρ− (h+ ) − ρ− (e h+ )
n o n o
+ ρ− (h− ) − ρ− (e
h− ) − ρ+ (h− ) − ρ+ (e h− )
( ) ( )
√ h ∧k √ h ∧k
(ω+ ) (ω+ )
= πΛ1,1∗ − −1∂∂ log e − πΛ1,1∗ − −1∂∂ log e
T
+ (ω+h )∧k T
−
h )∧k
(ω+
( ) ( )
√ h ∧l √ h ∧l
(ω− ) (ω− )
+ πΛ1,1∗ − −1∂∂ log e − πΛ1,1∗ − −1∂∂ log e
T
− (ω− h )∧l T
+ (ω− h )∧l
e
√ (ω h )∧k ∧ (ω−h l
)
= − −1 ∂+ ∂ + − ∂− ∂ − log e+ ,
h h
(ω+ ) ∧ (ω− )∧l
∧k
as claimed.
Remark 8.21. This curvature formula reveals an interesting local description

of the Bismut Ricci curvature in adapted local complex coordinates. In particular,
implicit in the computations underlying Proposition 8.20 is the local coordinate
description
√ det ω+
(ρIB )1,1 = − −1 ∂+ ∂ + − ∂− ∂ − log ,
det ω−
where ω± are the restrictions of the Kähler form ωI to Λ1,1 T±∗ as above. This is an
interesting extension
√ of the classical description of the Ricci curvature of a Kähler
manifold as − −1∂∂ log det ω.
The formula can in fact be globalized in case the split canonical bundles admit
global sections. We do this next, summarizing the computations above.
Proposition 8.22. Let (M 2n , g, I, J) be a generalized Kähler manifold with

[I, J] = 0. Suppose there exist global nonvanishing holomorphic sections

Ω± ∈ Λrank T± Λ1,0± .
Then one has

∧l
ω− ∧ Ω+ ∧ Ω+
Φ = log ∧k
,
ω + ∧ Ω− ∧ Ω−
and
(θI − θJ )♯ = σdΦ,
ρIB = − 12 dJdΦ.
Proof. A computation using (7.14) shows that the two associated generalized
complex structures are generated by the pure spinors
√ √
−1ω−
ψ1 = e ∧ Ω+ , ψ2 = e− −1ω+
∧ Ω− .
It follows that Φ takes the claimed form. We note that by Lemma 8.18, we have
θJ − θI = 0, and also σ = 0 by definition, so the claimed torsion formula follows
immediately. (We have chosen to express the result of Lemma 8.18 in this seem-
ingly overcomplicated way to draw a similarity to the nondegenerate case below).
Furthermore, by Proposition 8.20 it follows that
√ ω−∧l
∧ Ω+ ∧ Ω+
(ρIB )1,1 = − −1 ∂+ ∂ + − ∂− ∂ − log ∧k
ω + ∧ Ω− ∧ Ω−
1 1,1
= − 2 (dJdΦ)I .
Checking the (2, 0) + (0, 2) piece of the Bismut Ricci curvature formula is left as an
exercise.
Next we address the nondegenerate setting, where it turns out that once again
the quantity Φ determines several key formulas relevant to the Lee forms and Bismut
Ricci curvature, and in fact will determine the generalized Kähler-Ricci flow lines
in this setting (cf. §9.3).
Proposition 8.23. Let (M 2n , g, I, J) be a nondegenerate generalized Kähler
manifold. Then one has
det(I + J)
Φ = log .
det(I − J)
and
♯
(θI − θJ ) = σdΦ,
Proof. We recall as in Lemma 7.81 that for a nondegenerate generalized

Kähler structure both I ± J are invertible, and the Poisson tensors g −1 (I ± J)
associated to the generalized Kähler structures (cf. Proposition 7.43) are also non-
degenerate, yielding symplectic forms
F± = −g(I ± J)−1 .
Furthermore, it is an exercise to check
√
that the spinors defining the associated
generalized complex structures are e −1F± , thus yielding the claimed formula for
Φ. The torsion and Bismut Ricci curvature formulas are direct but lengthy compu-
tations and we leave the details to [8] (n.b. the conventions for σ there differ from
those here).
By comparing the discusion of the Ricci curvature in the Kähler setting, and
the results of Proposition 8.22 and 8.23, the general formula is now apparent. We
give the statement here, the proof of which can be obtained by a combination
of the different cases above. This can be thought of as a generalization of the
classical transgression formula for the Ricci curvature on a Kähler manifold to the
generalized Kähler setting.
Proposition 8.24. Let (E, h, i , [, ]) be an exact Courant algebroid, and suppose

(J1 , J2 ) is a generalized Kähler structure such that each Ji is globally defined by a
/
d0 -closed pure spinor ψi . Then
(θI − θJ )♯ = σdΦ,
8.3. Examples and rigidity results

We begin this section by explicitly describing the geometry of the Hopf/Boothby
metric (7.17), previously also encountered in Example 3.51 . We will show that it is
in fact isometric to gS 3 ⊕ gS 1 , the product of canonical spherical metrics on S 3 and
S 1 , and is Bismut Hermitian-Einstein. Remarkably, in complex dimension 2 this is
the only compact non-Kähler example of a Bismut Hermitian-Einstein metric, as we
will prove in Theorem 8.26 below (cf. [75, 78]). Even more remarkably, this metric
is generalized Kähler, in two different ways! As exhibited in Example 7.76, the
Hopf metric is compatible with the complex structure determined by changing the
orientation of one of the complex planes on C2 \{0}, yielding a generalized Kähler
structure. Also, in Example 7.85 we obtained another complex structure compati-
ble with the Hopf metric which yields a generalized Kähler structure nondegenerate
away from the two elliptic curves z1 = 0, z2 = 0.
Proposition 8.25. Consider the metric g defined on C2 \{0} via

gE (X, Y )
(8.11) g(X, Y )(z1 ,z2 ) = 2 2,
|z1 | + |z2 |
where gE is the standard Euclidean metric on C2 . Let γ(z1 , z2 ) = (αz1 , βz2 ), |α| =
|β| > 1 be the generator of the group of Deck transformations defining a diagonal
Hopf surface (S 3 × S 1 , J) as in Definition 7.75. Then g is γ-invariant, and so
descends to the quotient to define a metric, also denoted g, such that
√
(1) −1∂∂ω = 0,
(2) ∇θ = 0,
(3) RcB = 0,
(4) ρB = 0.
Proof. The invariance under γ is easily computed using |α| = |β|. Next we
can compute
√
−1
ω= 2 2 (dz1 ∧ dz 1 + dz2 ∧ dz 2 ) ,
|z1 | + |z2 |
hence
√
−1 2 2
dω = − 2 d |z1 | + |z2 | ∧ (dz1 ∧ dz 1 + dz2 ∧ dz 2 )
|z1 |2 + |z2 |2
√
−1
= − 2 [(z 2 dz2 + z2 dz 2 ) ∧ dz1 ∧ dz 1 + (z 1 dz1 + z1 dz 1 ) ∧ dz2 ∧ dz 2 ]
2 2
|z1 | + |z2 |
8.3. EXAMPLES AND RIGIDITY RESULTS 179
Thus one obtains
dc ω = − dω(J1 , J1 , J1 )
1
= − 2 [(z 2 dz2 − z2 dz 2 ) ∧ dz1 ∧ dz 1 + (z 1 dz1 − z1 dz 1 ) ∧ dz2 ∧ dz 2 ] .
2 2
|z1 | + |z2 |
Thus lastly
2
2 2
ddc ω = 3 d |z1 | + |z2 | ∧
2 2
|z1 | + |z2 |
[(z 2 dz2 − z2 dz 2 ) ∧ dz1 ∧ dz 1 + (z 1 dz1 − z1 dz 1 ) ∧ dz2 ∧ dz 2 ]
4
+ 2 [dz1 ∧ dz 1 ∧ dz2 ∧ dz 2 ]
2 2
|z1 | + |z2 |
= 0,
yielding property (1).

If we let π denote the radial projection of C2 onto S 3 , then these formulas
also exhibit that H = −dc ω = 2π ∗ dVgS3 , i.e. twice the standard volume form on
S 3 . Observing that the metric on the universal cover C2 \{0} is isometric to the
standard cylinder (S 3 × R, gS 3 ⊕ dt2 ), it follows that H is parallel. Furthermore,
using Lemma 8.6 it follows that θ = ⋆H = dt, which is also directly computable
from the formula for dc ω above. Thus θ is also parallel. As in Example 3.56, it
also folllows that Rc = 2gS 3 , while H 2 = 8gS 3 . Referring to Proposition 3.18 we
see that RcB = 0, and it then follows from Proposition 8.10 that ρB = 0, finishing
the proof.
Theorem 8.26. Let (M 4 , g, J) be a compact complex manifold with puriclosed

metric g such that ρ1,1 4
B ≡ 0. Then ρB ≡ 0, and (M , g, J) is biholomorphically
isometric to either
(1) A Calabi-Yau surface (M 4 , J) with Kähler-Ricci flat metric g
(2) A quotient of a diagonal Hopf surface (S 3 × S 1 , J), with Hopf metric g.
Proof. Using the curvature identities of Proposition 8.10, we see that the
condition ρ1,1
B ≡ 0 implies the equations
0 = Rc − 41 H 2 + 12 Lθ♯ g,
(8.12)
0 = dd∗ H + 21 Lθ♯ H.
However, since the real dimension is 4 we may apply Lemma 8.6 to conclude that
⋆θ = H and thus θ♯ yH = 0. Thus dd∗ H = 0, and by pairing with H and integrating
by parts we conclude d∗ H = 0. Thus H is harmonic, and using again that ⋆θ = H it
follows that θ is harmonic. Using the Bochner formula for 1-forms and integrating
by parts we conclude that

Z
0= h∆d θ, θi dVg
M
Z
= h∇∗ ∇θ + Rc(θ, ·), θi dVg
M
(8.13) Z h i
2
= |∇θ| + 14 H 2 − 21 Lθ♯ g (θ, θ) dVg
ZM h i
2
= |∇θ| − 12 Lθ♯ g(θ, θ) dVg .
M
However we further observe, using that the Lee form is divergence-free since the
metric is pluriclosed,
Z Z
Lθ♯ g(θ, θ)dVg = (∇i θj + ∇j θi ) θi θj dVg
M M
Z
=2 ∇i θj θj θi dVg
M
Z
2
= ∇i |θ| θi dVg
M
Z
2
= − |θ| ∇i θi dVg
M
= 0.
Plugging this into (8.13) we conclude that ∇θ = 0. If θ vanishes, then the metric
is Kähler, and Ricci flat, yielding case (1). Assuming θ 6= 0, we have found a
nontrivial parallel vector field. By the de Rham decomposition theorem, it follows
that the universal cover with the pullback Hermitian structure splits isometrically
as a product N × R. It follows that H = ⋆θ is a multiple of the volume form on
N , and so examining the curvature equation of 8.12, it follows that the slice has
constant positive Ricci curvature, and is thus isometric to a multiple of the round
3-sphere. It follows that the original structure (M 4 , g) is isometric to a quotient of
(S 3 × S 1 , gS 3 ⊕ λgS 1 ) for an appropriate constant λ.
In particular, it follows that (M 4 , J) is a Hopf surface. To obtain the precise
complex structure, we note that, as a Hopf surface, it is first of all covered by a
primary Hopf surface, and the universal cover is biholomorphic to C2 \{0}. We lift
the metric, which is now cylindrical, to this space. The unit vector tangent to the
∂
S 1 factor lifts to a global coordinate vector field ∂t . Setting r = et , we obtain the
∂
radial vector field r ∂r , as well as an associated flat metric g0 = r−2 g. Furthermore,
the generator of the fundamental group takes the form
γ(s, p) = (as, ψp),
where (s, p) ∈ R × S 3 , and ψ ∈ SO(4), as the map must preserve the cylindrical
metric. Since γ is also a biholomorphism, it follows that in fact ψ ∈ U (2). The map
ψ will be conjugate to an element of the maximal torus in U (2), in other words is
conjugate to a map
(z1 , z2 ) → (eiµ1 z1 , eiµ2 z2 ).

By conjugating as required we assume ψ itself takes this form. It follows that,

expressed in complex coordinates,
γ(z1 , z2 ) = (αz1 , βz2 ), α = aeiµ1 , β = aeiµ2 , |α| = |β| ,
as required.
Remark 8.27. It is possible to obtain this rigidity result in part using the en-
ergy functional of §6.1. In particular, using Proposition 8.14 we know that (g, H)
defines a generalized Ricci soliton with associated vector field θ♯ . Since M is com-
pact it follows from Corollary 6.11 that (g, H) is in fact a steady gradient soli-
ton. We have thus obtained two expressions for RcB , and comparing them yields
Lθ♯−∇f g ≡ 0. Since the Lee form is divergence-free one can trace this to obtain
∆f = 0, hence f is a constant. It follows that θ♯ is a Killing field, meaning the
symmetric part of ∇θ vanishes. But dθ = 0 as well, hence ∇θ = 0, after which the
argument proceeds as above.
Remark 8.28. It was recently shown [166] that the non-diagonal, class 1, Hopf
surfaces admit pluriclosed metrics which are steady generalized Ricci solitons in the
sense of Definition 4.27. Note that by Theorem 8.26 these surfaces cannot admit
Bismut Hermitian-Einstein metrics. These solitons are conjecturally the global at-
tractors for the pluriclosed flow on these surfaces. Also in [171] it was shown that in
fact these solitons admit two distinct compatible generalized Kähler structures, one
with vanishing Poisson tensor and another with generically nondegenerate Poisson
tensor. These metrics form a continuous family extending the examples on diagonal
Hopf surfaces discussed above.
It is possible to show rigidity results in higher dimensions with additional topo-
logical hypotheses. We begin with a general rigidity theorem which is a simplified
version of a result of Gauduchon (cf. [76]). First we recall the definition of Bott-
Chern cohomology.
Definition 8.29. Let (M 2n , J) be a complex manifold. Define the real (1, 1)
Bott-Chern cohomology via
n o
Ker d : Λ1,1
R → Λ3R
1,1
HBC,R := √ .
−1∂∂f | f ∈ C ∞
Proposition 8.30. Let (M 2n , g, J) be a compact Hermitian manifold such that
sC ≥ 0, with sC > 0 at some point. Then c1 (M, J) 6= 0 as an element of Bott-Chern
cohomology.
Proof. Let ω e = e2u ω be a Gauduchon metric conformally related to ω, which
exists by Theorem 7.50. If c1 (M, J) √ = 0 in Bott-Chern cohomology, then there
exists f ∈ C ∞ (M ) such that ρC (g) = −1∂∂f . It then follows that
Z Z Z
√
0= −1∂∂f ∧ ω e n−1 = e2(n−1)u ρC (g) ∧ ω n−1 = e2(n−1)u sC ω n > 0,
M M M
a contradiction.
Remark 8.31. We note for instance that, since the Hopf metric (7.17) has
sC > 0 everywhere, this proposition in particular implies that c1 (M, J) is nonzero
in Bott-Chern cohomology, while of course it is zero in de Rham cohomology, which
vanishes entirely.
Proposition 8.32. Let (M 2n , g, J) be a compact Bismut Hermitian-Einstein

manifold. If c1 (M, J) = 0 in Bott-Chern cohomology, then g is Kähler Calabi-Yau.
Proof. Using Proposition 8.10, we see that a Bismut Hermitian-Einstein man-
ifold satisfies
0 = ρ1,1
B = SC − Q.
2 2
Taking the trace then yields 0 = sC − |T | , and hence sC ≥ |T | ≥ 0. Since we
have assumed c1 (M, J) = 0, it follows directly from Proposition 8.30 that sC ≡ 0,
2
and hence |T | ≡ 0. The metric g is hence Kähler Calabi-Yau, as required.
Remark 8.33. In Proposition 9.64 below we use the strong maximum principle
to improve Proposition 8.32 to cover steady solitons.
Remark 8.34. In the nondegenerate generalized Kähler setting, the tensor
(Ω2,0
I )
∧n
is a nonvanishing holomorphic section of the canonical bundle, and so
c1 (M, J) = 0. Thus in this setting any Hermitian generalized Einstein metric is
automatically Kähler Calabi-Yau (cf. [8]). In fact, due to the presence of the
holomorphic symplectic form Ω2,0I , it follows from the Beauville-Bogomolov-Yau
theorem [13, 19, 186] that in fact (M 4n , g, I) is part of a hyperKähler structure.
Example 8.35. A higher-dimensional, simply-connected example of a Bismut
Hermitian-Einstein structure is given by a Calabi-Eckmann manifold. For the gen-
eral construction one considers Cn \{0} × Cm \{0}, and consider the action of C
given by
γ(z, w) = (eγ z, eαγ w),
where α ∈ C\R is a fixed non-real complex number. This action is free and proper,
and the quotient space is diffeomorphic to S 2n−1 × S 2m−1 √.
We consider the special case of n = m = 2, α = −1. Thus the manifold
is S 3 × S 3 , and is the total space of a T 2 fibration over CP1 × CP1 . This is the
product of the standard Hopf fibrations πi : S 3 → CP1 . Let ξi denote the canonical
vector fields associated to this fibration on the two factors, with µi the associated
canonical connections satisfying dµi = πi∗ ωF S . Note that the complex structure
further satisfies Jξ1 = ξ2 . We consider the Kähler form associated to the product
of round metrics, namely
ω = π1∗ ωF S + π2∗ ωF S + µ1 ∧ µ2 .
It follows that
dω = dµ1 ∧ µ2 − µ1 ∧ dµ2 = π1∗ ωF S ∧ µ2 − µ1 ∧ π2∗ ωF S ,
hence
H = −dc ω = dω(J, J, J) = π1∗ ωF S ∧ µ1 − π2∗ ωF S ∧ µ2
It is clear then that the geometric structure is that of a product of two copies of
the S 3 factor of the Hopf metric, and furthermore that dH = 0. It now follows by
computations similar to those in Proposition 8.25 that
Rc − 41 H 2 = 0.
Furthermore, it follows from Lemma 8.6 that θ = − 12 (µ1 + µ2 ). Thus θ♯ is a Killing
field which further preserves H, so by Proposition 8.10 it follows that ρ1,1
B = 0. We
2,0+0,2
leave as an exercise to further check that ρB = 0.
Question 8.36. Do Calabi-Eckmann manifolds for different choices of n, m,

and α admit Bismut Hermitian-Einstein metrics? Do they admit steady generalized
Ricci solitons?
Question 8.37. How rigid are higher dimensional non-Kähler Bismut Hermitian-
Einstein metrics? In complex dimension n = 3, is it possible to classify solutions
with nonzero, (anti)self-dual torsion?
Question 8.38. A natural class of Bismut Hermitian-Einstein manifolds has
been introduced in [71, Section 2.3]. These equations impose the stronger condi-
tions that the holonomy of the Bismut connections reduces to the special unitary
group and dθ = 0, and admit a natural interpretation in generalized geometry in
+ +
terms of the operators D− and D/ introduced in Chapter 3. A complete classifi-
cation of this geometry on complex surfaces has been obtained in [71, Proposition
2.11], but the higher-dimensional case is still open. It would be interesting to obtain
a classification of solutions of these stronger equations in three complex dimensions,
as a first step towards the classification of non-Kähler Bismut Hermitian-Einstein
three-folds.
A fundamental result of Samelson [149] shows that every compact Lie group
of even dimension admits left-invariant complex structures which are compatible
with a bi-invariant metric.
Proposition 8.39. Let G be an even-dimensional connected Lie group with
a bi-invariant metric g and left-invariant complex structure J compatible with g.
Then (G, g, J) is Bismut flat.
Proof. As noted in Proposition 3.53, the Levi-Civita connection of a bi-
invariant metric takes the form
∇X Y = 12 [X, Y ].
On the other hand, the computation of Example 7.48 shows that for a left-invariant
complex structure, one has H = −dc ω = g([X, Y ], Z). Thus ∇J is flat by Proposi-
tion 3.53.
Definition 8.40. A Samelson space is a Hermitian manifold (G, g, J) where G
is a simply-connected Lie group with g a bi-invariant metric, and J a left-invariant
complex structure on G compatible with g.
As it turns out, Samelson spaces are the only possible simply-connected Her-
mitian Bismut-flat manifolds, and moreover the possible fundamental groups of
compact Samelson spaces are quite restricted.
Definition 8.41. Let (G1 = G0 × Rk , g, J) be a Samelson space. Given ρ :
Zk → Isom(G) a homomorphism, define a free and properly discontinuous action
of Γρ ∼
= Zk on G1 via
γ(g, x) = (ρ(γ)g, x + γ).
The manifold M = G1 /Γρ inherits a complex structure and a Bismut flat Hermitian
metric, and is called a local Samelson space.
Theorem 8.42. ([183]) Let (M n , g, J) be a compact Bismut flat Hermitian
manifold. Then there exists a finite unbranched cover M ′ of M which is a local
Samelson space.
Remark 8.43. Theorem 8.42 can be used to give a precise description of all
compact Bismut flat Hermitian manifolds of complex dimension n = 2, 3. In par-
ticular, in complex dimension n = 2 one recovers the standard Hopf surfaces as
described in Theorem 8.26. Of course the result of Theorem 8.26 is stronger, giv-
ing a classification among the a priori larger class of Bismut-Ricci flat metrics.
In
complex dimension 3, the universal cover is either a product of C2 \{0} × C, with
metric a product of the Hopf metric (7.17) and the flat metric, or a central Calabi-
Eckmann manifold, defined as SU (2) × SU (2) with left-invariant complex structure
and bi-invariant metric.
Question 8.44. What further restrictions apply if we ask that a Bismut flat
pluriclosed metric is also part of a generalized Kähler structure? Note that as
observed in §7.4.3, the Hopf metric is part of a generalized Kähler structure in two
distinct ways. Furthermore, the Calabi-Eckmann metrics are compatible with a
commuting-type generalized Kähler structure, with the second complex structure
given by changing the orientation on the T 2 fiber.
Question 8.45. Note that the final stage of the proof of Theorem 8.26 identifies
the possible complex structures compatible with a particular flat Bismut connec-
tion. It is natural to ask this question in higher dimensions: given an even dimen-
sional Bismut flat structure (M 2n , g, H), can we describe all complex structures J
(necessarily left-invariant) which are compatible with g, and satisfy ∇B J ≡ 0?
CHAPTER 9
Generalized Ricci Flow in Complex Geometry
In Chapter 7 we introduced the fundamental concepts of generalized com-

plex geometry, including pluriclosed and generalized Kähler metrics. Furthermore,
Chapter 8 yielded a definition of canonical metric in pluriclosed geometry related to
the generalized Einstein condition. In this chapter we link together these ideas with
the generalized Ricci flow by showing that, when coupled to evolution equations for
the associated complex structures, the generalized Ricci flow preserves pluriclosed
and generalized Kähler geometry.
The presence of the extra structure coming from complex geometry renders
these flows decidedly more tractable than the full generalized Ricci flow. To begin
we show how to reduce the pluriclosed flow to a flow of (1, 0)-forms, and give a scalar
reduction for the generalized Kähler-Ricci flow in certain settings, generalizing the
classical reduction of the Kähler-Ricci flow to a parabolic complex Monge-Ampère
equation. Building on this we show smooth regularity of pluriclosed flow in the
presence of uniform parabolicity bounds, generalizing Calabi/Yau’s C 3 estimate for
the Monge-Ampère equation. Using these estimates we end the chapter by showing
global existence and convergence results for the pluriclosed flow and generalized
Kähler-Ricci flow and discuss their implications.
9.1. Kähler-Ricci flow

The natural starting point for discussing geometric flows in complex geometry
is Kähler-Ricci flow. Though we will not spend much time discussing the deep,
detailed theory of Kähler-Ricci flow here, it is instructive to review the rudimentary
aspects of the theory as a way of setting the stage for the more general flows we
discuss below. We refer the reader to [22, 155] for more detailed accountings of
the fundamental theory of Kähler-Ricci flow.
9.1.1. Basic properties.
Definition 9.1. Let (M 2n , J) be a complex manifold. A one parameter family

of Kähler metrics gt is a solution of Kähler-Ricci flow if the associated Kähler forms
ωt satisfy
∂
(9.1) ω = − ρC .
∂t
Noting that ρC is a closed (1, 1) form, one can expect the equation above to preserve
the Kähler conditions. Before proving this we observe the evolution equation for
the underlying Riemannian metric.
185
186 9. GENERALIZED RICCI FLOW IN COMPLEX GEOMETRY
Lemma 9.2. Let (M 2n , gt , J) be a solution to Kähler-Ricci flow. Then the

metric gt satisfies
∂
g = − Rc .
∂t
Proof. Using Lemma 8.9, applied here to a Kähler metric so that ρC = ρB
and θ = 0, we obtain
∂ ∂
g(X, Y ) = ω(X, JY ) = −ρC (X, JY ) = − Rc(X, Y ),
∂t ∂t
as required.
We next establish the short-time existence of solutions to Kähler-Ricci flow.
Note that this does not follow from Lemma 9.2 and the short-time existence of
solutions to Ricci flow, as we do not yet know that the Ricci flow should preserve
the Kähler conditions, which is rather a consequence of proving the short-time
existence of solutions to (9.1).
Proposition 9.3. Given (M 2n , J) a compact complex manifold and g0 a Kähler
metric, there exists a unique maximal solution gt of Kähler-Ricci flow with initial
condition g0 on [0, T ) for some T ∈ R ∪ {∞}. In particular, the solution to Ricci
flow with initial condition g0 is a solution to Kähler-Ricci flow.
Proof. We can compute in local complex coordinates, for a Kähler metric g,
ρC (ω)ij = g qp ∂q ∂p ωij + ∂ω ⋆ ∂ω.
It follows that equation (9.1), restricted to the open cone of positive tensors in
the linear space of closed (1, 1)-forms, is strictly parabolic. Thus we can apply
the standard theory for strictly parabolic equations to conclude the existence of
a maximal solution ωt defined on [0, T ). By Lemma 9.2, the associated family of
Riemannian metrics gt is a solution of Ricci flow with initial condition g0 , which is
unique by Theorem 5.6.
9.1.2. Kähler cone and formal existence time. To better understand the
qualitative picture of the long time existence and singularity formation of Kähler-
Ricci flow, we study the flow at the formal level of cohomology. We note that
Kähler metrics will define classes in Bott-Chern cohomology as in Definition 8.29.
Within this cohomology space lies the Kähler cone, which is the set of classes in
H 1,1 represented by Kähler metrics.
Definition 9.4. Let (M 2n , J) be a complex manifold. Define the Kähler cone
by
n o
1,1
K := [ψ] ∈ HBC,R | ∃ ω ∈ [ψ], ω > 0 .
Example 9.5. Let (M 2 , J) be a compact Riemann surface. Any metric g

compatible with J is Kähler as the Kähler form ω is automatically closed since
dω ∈ Λ3 (M ) ∼
= {0}. The form ω induces an orientation on M , and it follows that
1,1
HBC,R ∼
= R. It is clear by construction that λ[ω] admits the Kähler metric λω for
√
λ ≥ 0. Alternatively, for λ < 0, any form η = λω + −1∂∂f must satisfy
Z Z Z
√
η= λω + −1∂∂f = λ ω = λ Vol(g) ≤ 0.
M M M
9.2. PLURICLOSED FLOW 187
R
But if η was the Kähler form associated to a metric h one would have M η =
Vol(h) > 0. It follows that the Kähler cone is precisely the half-line {λ[ω] | λ > 0}.
The structure of the Kähler cone plays a fundamental role in understanding
the singularity formation of the Kähler-Ricci flow, and first of all gives an upper
bound for the possible smooth existence time of the flow.
Lemma 9.6. Let (M 2n , g0 , J) be a complex manifold with Kähler metric. Let
τ ∗ (g0 ) := sup{t ≥ 0 | [ω0 ] − tc1 ∈ K},
and let T denote the maximal smooth existence time for the Kähler-Ricci flow with
initial condition g0 . Then T ≤ τ ∗ (g0 ).
Proof. Let ωt denote the one-parameter family of Kähler forms evolving by

Kähler-Ricci flow with initial condition ω0 . It follows easily that [ωt ] = [ω0 ] − tc1 .
If the flow existed smoothly for some time t > τ ∗ (g0 ), then in particular there
exists a smooth positive definite metric in [ω0 ] − tc1 , contradicting the definition of
τ ∗ (g0 ).
In view of this lemma, the natural question arises which is, given a Kähler
metric g0 , is τ ∗ (g0 ) the maximal existence time of the flow? In other words, is the
flow smooth until the Kähler class reaches the boundary of the Kähler cone? The
answer is yes, and this is proved in Theorem 9.62 below.
9.2. Pluriclosed flow

Given the rich geometric and analytic structure of Kähler-Ricci flow, one can
ask if it can be extended to a natural flow of Hermitian, non-Kähler metrics. Though
the world of Hermitian, non-Kähler geometry is large and diverse, in keeping with
the thrust of this text we will focus our attention on the parts of the subject most
closely linked to generalized complex geometry.
9.2.1. Basic properties.

Definition 9.7. Let (M 2n , J) be a complex manifold, and fix H0 ∈ Λ3 T ∗ , dH0 =
0. A one parameter family of pluriclosed metrics (gt , bt ) is a √
solution of pluriclosed
flow if the associated Kähler forms ωt and (2, 0)-forms βt = −1b2,0 satisfy
∂ ∂
(9.2) ω = − ρ1,1
B , β = −ρ2,0
B .
∂t ∂t
The evolution equation for ω first appeared in [167]. We note that the equation
for βt is not strictly speaking necessary to determine the metric or torsion, since
H = −dc ω and ρB is closed. For this reason we will at times refer to a family of
Kähler forms ωt alone as a solution to pluriclosed flow. Nonetheless, the equation
for β plays an important role in deriving estimates and determining the relationship
to generalized geometry. Indeed, as we have different ways of describing pluriclosed
structures in terms of the underlying Riemannian metric, the Kähler form, or ob-
jects in generalized geometry, so pluriclosed flow has different faces when expressed
in terms of these underlying objects, each revealing different important analytic
and geometric properties of the flow. In particular, pluriclosed flow turns out to
be a gauge-modified version of generalized Ricci flow (up to a reparameterization
of time by a factor of 2). Another interesting aspect of the flow (9.2) is that, if we
fix H̃0 ∈ Λ3,0+2,1 , dH̃0 = 0, then the condition
∂ωt + dβt = H̃0
is preserved along the flow. Thus, pluriclosed flow can be regarded as a natural
flow for metrics on a fixed holomorphic Courant algebroid, as defined in §7.3.3.
Proposition 9.8. Let (M 2n , gt , J) be a solution to pluriclosed flow. Then
(1) The associated Kähler forms ωt satisfy
∂ ∗
ω = −∂∂ω∗ ω − ∂∂ ω ω − ρC (ω).
∂t
(2) The associated Kähler forms ωt satisfy
∂
(9.3) ω = −SC + Q,
∂t
where Q is as in (8.6).
(3) The associated pairs (gt , bt ) satisfy
∂
g = − Rc + 14 H 2 − 12 Lθ♯ g,
(9.4) ∂t
∂
b = − 21 d∗ H + 12 dθ − 21 iθ♯ H.
∂t
In particular, pluriclosed flow is a solution of (−θ♯ , dθ)-gauge-fixed gener-
alized Ricci flow (cf. 4.13).
(4) The associated generalized metrics G = G(gt , bt ) on (T ⊕ T ∗ , h, i , [, ]H0 , π)
satisfy
∂
G −1 G ◦ π− = −Rc+ (G, div) ◦ π− ,
∂t
where div = divG −2 hθ, i.
Proof. The first equivalence follows directly by projecting the first equation
from Lemma 8.9 onto the (1, 1) component. The second formulation follows directly
from Proposition 8.10. For the third formulation, the evolution equation for the
metric follows directly from Proposition 8.10, where the evolution equation for b
follows from the third equation of Proposition 8.10 and the computations of Propo-
sition 8.14. As for the last formulation, it follows from (3) and direct application
of Proposition 4.24.
Corollary 9.9. Let (M 2n , gt , J) be a solution to pluriclosed flow, and let φt
denote the one-parameter family of diffeomorphisms generated by 12 θt♯ . Then for all
t such that gt and φt are defined, let
gt′ = φ∗t gt , Ht′ = φ∗t Ht , Jt′ = φ∗t J
Then (gt′ , Ht′ ) is a solution of generalized Ricci flow, and Jt′ satisfies
∂ ′
J = 12 Lθ♯ J ′ = 21 ∆J ′ + Rm′ ⋆H ′ + H ′ ⋆ H ′ .
∂t
Proof. The claim that (gt′ , Ht′ ) is a solution of generalized Ricci flow is a direct
consequence of (9.4) and the naturality properties of Lie derivatives. The claim that
∂ ′ 1 ′
∂t J = 2 Lθ ♯ J also follows by definition. The final formula is an exercise using
properties of the Lee form on complex manifolds with pluriclosed metric.
Remark 9.10. The fact that the Lie derivative operator Lθ♯ J is actually a
heat operator in disguise is an interesting feature of pluriclosed flow, which plays
an essential role in the generalized Kähler-Ricci flow to come. An explicit formula
for the lower-order terms Rm ⋆H and H ⋆ H was computed in [169, Proposition
3.1], although it is quite unwieldy. It would be interesting to derive estimates for
the pluriclosed flow directly from this formula.
Remark 9.11. In view of Corollary 9.9, it is natural at times to express pluri-
closed flow explicitly as a solution to generalized Ricci flow. The system of equations
indicated there is
∂
g = − Rc + 41 H 2 ,
∂t
∂
(9.5) b = − 12 d∗g H,
∂t
∂
J = 12 Lθ♯ J,
∂t J
At times we will refer to solutions to this system also as “pluriclosed flow,” even
though it is related to the original definition by a nontrivial gauge transformation.
Remark 9.12. Item (2) of Proposition 9.8 expresses pluriclosed flow in terms
of the second Ricci curvature of the Chern connection, denoted SC in Definition
8.8, and a certain quadratic expression Q in the torsion of the Chern connection.
Thus the pluriclosed flow fits into a general family of geometric flows of Hermitian
metrics introduced in [168], which take the form
∂
ω = − SC + Q,
∂t
where Q is an arbitrary quadratic expression in the torsion. This general family
of equations is referred to as Hermitian curvature flow. Fundamental analytic
properties such as short-time existence, smoothing estimates, and stability results
were shown for this family of flows in [168]. Furthermore, a certain expression for
Q was identified from the point of view of seeking a fixed-point equation for the
flow which arises as the Euler-Lagrange equation for a Hilbert-type functional in
the context of Hermitian geometry.
Given the rich variety within Hermitian geometry, it seems natural to expect
that different choices of Q may be better suited for different questions. In particular,
recent work of Ustinovskiy [181] identifies a particular choice of Q such that the
resulting flow preserves nonnegative holomorphic bisectional curvature, yielding
a natural extension of the classic Frankel conjecture. This flow is also related to
recent work on the Hull-Strominger system [63]. Also Lee [123] has shown that with
Q = 0 the flow can be used to show that compact complex manifolds admitting a
metric of non-positive bisectional curvature and nonpositive Chern-Ricci curvature
which is negative at one point have ample canonical bundle. In a different direction,
flowing a Hermitian metric by the Chern-Ricci curvature directly has been explored
in [83, 179].
9.2.2. Short-time existence. The proposition below establishes short-time
existence for pluriclosed flow for smooth initial data on compact manifolds.
Proposition 9.13. Given M a smooth compact manifold and (g, H, J) a pluri-
closed structure, there exists a unique maximal solution (gt , Ht , J) to pluriclosed
flow with this initial condition on [0, T ) for some T ∈ R ∪ {∞}. Also, if g0 is
Kähler then Ht ≡ 0, and gt is the unique solution to Kähler-Ricci flow with initial
condition g.
Proof. Using Proposition 9.8 and the expression for the Chern curvature in
complex coordinates we see that the pluriclosed flow can be expressed as
∂
ω = − (SC )ij + Qij
(9.6) ∂t ij
= g qp ∂p ∂q ωij + ∂ω ⋆ ∂ω.
This is a strictly parabolic equation on the open cone of positive tensors in the linear
space of pluriclosed (1, 1)-forms, and so by the theory of strictly parabolic PDE we
can obtain unique short-time solution on compact manifolds. If g0 is Kähler, then
let e
gt denote the unique solution of Kähler-Ricci flow with initial data g0 , whose
existence is guaranteed by Proposition 9.3. As the metrics e gt are all Kähler, it
follows easily by Lemma 8.9 that ρC (e gt ) = ρ1,1
B (e
g t ), and thus get is a solution of
pluriclosed flow as well. Since solutions to pluriclosed flow are unique on compact
manifolds, it follows that gt = get is the solution to Kähler-Ricci flow.
9.2.3. A positive cone and conjectural existence for pluriclosed flow.

Given the short-time existence of solutions to pluriclosed flow, we are again faced
with the basic question of, what is the maximal smooth existence time, and what
happens to the metric at that time? As these flows are special cases of generalized
Ricci flow, the discussion of §5.5 certainly applies here, showing that the Riemann-
ian curvature must blow up at a singular time. However due to the extra rigidity
arising from the presence of complex structures, one expects sharper statements. In
this section we formulate a conjecture on the maximal existence time for pluriclosed
flow. Later in this chapter we verify these conjectures in various special cases.
Definition 9.14. Let (M 2n , J) be a complex manifold. Define the real (1, 1)-
Aeppli cohomology via
n √ o
Ker −1∂∂ : Λ1,1
R → Λ2,2
R
1,1
HA,R := .
∂α + ∂α | α ∈ Λ1,0
Furthermore, define the (1, 1)-Aeppli positive cone via
n o
1,1 1,1
PA := [ψ] ∈ HA,R | ∃ ω ∈ [ψ], ω > 0 .
In other words, this cone consists precisely of the (1, 1)-Aeppli cohomology classes
which contain pluriclosed metrics.
Remark 9.15. We observe here that in complex manifolds there is a natural
map
1,1 1,1
ι : HBC,R → HA,R ,
In particular we can apply this map to the first Chern class, yielding a well-defined
1,1
class in HA,R . In what follows we will not make a notational distinction between
the first Chern class thought of in these two different contexts, as we will mean the
1,1
element of HA,R unless otherwise specified.
Lemma 9.16. Let (M 2n , J) be a complex manifold, and suppose ωt is a one-

parameter family of metrics satisfying the pluriclosed flow equation. Then [ωt ] =
[ω0 ] − tc1 .
Proof. Fix a background Hermitian metric with Kähler form η, and note that
[ω0 − tρC (η)] = [ω0 ] − tc1 . We use the transgression formula of Proposition 8.16 to
observe that
Z t
∂
ωt = ω0 + ωds
0 ∂s
Z t
∗
= ω0 + −∂∂ ∗ ω − ∂∂ ω − ρC (ω) ds
0
Z t
∗ ∗ √ ωtn
= ω0 + −∂∂ ω − ∂∂ ω + −1∂∂ log n − ρC (η) ds
0 η
= ω0 − tρC (η) + ∂β + ∂β,
where
Z t √
∗ −1 ωn
β := −∂ ω − ∂ log n ds.
0 2 η
Hence [ωt ] = [ω0 − tc1 (η)] = [ω0 ] − tc1 , as claimed.
Lemma 9.17. Let (M 2n , g0 , J) be a complex manifold with pluriclosed metric.
Let
1,1
τ ∗ (g0 ) := sup{t ≥ 0 | [ω0 ] − tc1 ∈ PA },
and let T denote the maximal smooth existence time for the pluriclosed flow with
initial condition g0 . Then T ≤ τ ∗ (g0 ).
Proof. The proof is directly analogous to that of Lemma 9.6.
What follows is the main conjecture guiding the study of pluriclosed flow. While
Lemma 9.17 indicates the elementary fact that the maximal existence time for the
flow can be no larger than τ ∗ (g0 ), Conjecture 9.18 indicates that the flow is actually
smooth up to time τ ∗ (g0 ), i.e. that it equals the maximal existence time.
Conjecture 9.18. Let (M 2n , g0 , J) be a compact complex manifold with pluri-
closed metric. The unique maximal smooth solution to pluriclosed flow exists on
[0, τ ∗ (g0 )).
This conjecture first appeared in [170], inspired by a theorem of Tian-Zhang
(Theorem 9.62 below). We will give a proof of this theorem as well as proofs
of Conjecture 9.18 in various special cases below. As a guide for the nature of
singularity formation it is useful to have a characterization of the necessary and
sufficient conditions for cohomology classes to lie in the appropriate positive cone.
In the remainder of this section we will record a theorem characterizing this cone
in the case of complex surfaces.
Lemma 9.19. Let (M 4 , ω, J) be a compact complex surface with pluriclosed
metric, and let
BR1,1 = {µ ∈ Λ1,1 1
R | ∃ a ∈ Λ , µ = da}.
There is an exact sequence

√
0→ −1∂∂Λ0R → BR1,1 → R,
where the final map above is given by the L2 inner product with ω.
Proof. Exactness in the first two positions is clear. To check exactness in the
third place, fix µ ∈ BR1,1 satisfying
Z
0= hµ, ωi dV.
M
√
It follows from the maximum principle that the kernel of the L2 adjoint of trω −1∂∂
consists
√ of only the constant functions. Thus trω µ is orthogonal to the cokernel of
trω −1∂∂, so by the Fredholm alternative we can find u such that
√
trω −1∂∂u = trω µ.
√
Thus −1∂∂u − µ is exact, and also anti-self-dual since its inner product with
ω
√ vanishes. Thus it is harmonic, and exact, and hence vanishes, yielding µ =
−1∂∂u, as required.
In view of this lemma, we define
BR1,1
Γ= .
i∂∂Λ0R
This is identified with a subspace of R, via the L2 inner product with ω as in
Lemma 9.19. Note that if (M 4 , J) is Kähler then the ∂∂-lemma holds and so
Γ = {0}. Let γ0 denote a positive generator of Γ, which is well-defined since the
space of pluriclosed metrics on M is connected (cf. [176] for further detail). This
1,1
form γ0 plays a key role in the characterization of the positive cone PA in the next
theorem.
Theorem 9.20. ([170] Theorem 5.6, cf. [29, 30, 119]) Let (M 4 , J) be a
1,1
complex non-Kähler surface. Suppose φ ∈ Λ1,1 is pluriclosed. Then [φ] ∈ PA if
and only if
R
(1) RM φ ∧ γ0 > 0
(2) D φ > 0 for every effective divisor with negative self intersection.
1,1
Question 9.21. Is there a natural characterization of PA in higher dimen-
sions?
9.3. Generalized Kähler-Ricci flow

According to Definition 7.62, a generalized Kähler structure can be interpreted
as a metric which is pluriclosed with respect to two distinct complex structures,
satisfying the further relation that dcI ωI = −dcJ ωJ . Given this it is natural to ask
whether pluriclosed flow will preserve the generalized Kähler condition. The answer
is yes, but with a crucial and subtle caveat, namely that the complex structures must
evolve as well, a feature not common to geometric evolution in complex geometry.
Given that we have seen natural deformations of generalized Kähler structure using
certain Hamiltonian diffeomorphisms acting on one of the complex structures in
§7.4.5, this is actually quite natural in this setting. We will first describe the
biHermitian formulation of the resulting flow, called generalized Kähler-Ricci flow,
9.3. GENERALIZED KÄHLER-RICCI FLOW 193
and give the proof that it does indeed preserve generalized Kähler geometry (cf.
[169]). Following that we will describe the flow in terms of generalized metrics and
complex structures.
9.3.1. BiHermitian formulation.
Definition 9.22. Let M be a smooth manifold, we say that a one-parameter
family of generalized Kähler structures (gt , Ht , It , Jt ) is a solution of generalized
Kähler-Ricci flow if
∂
g = − Rc + 41 H 2 ,
∂t
∂
b = − 12 d∗g H,
(9.7) ∂t
∂
I = 12 Lθ♯ I,
∂t I
∂
J = 12 Lθ♯ J.
∂t J
Observe that this system of equations is in a certain sense two copies of the pluri-
closed flow system (9.5) running simultaneously. Indeed this was how it was dis-
covered, as we will indicate in Theorem 9.23 below. This formulation, while it has
a pleasing symmetry to it, has the disadvantage that both complex structures are
evolving. When it comes time to estimate solutions of this system, it will be to
our advantage to freeze one of these complex structures by a gauge transformation.
Of course it is not possible to fix both, since I and J are evolving by (in general)
distinct gauge transformations. Gauge-fixing the flow to freeze I and comparing
against Proposition 9.8 yields
∂ ∂
g = − Rc + 41 H 2 − 21 Lθ♯ g, ωI = − (ρIB )1,1
∂t I ∂t
∂ ∂
(9.8) b = − 12 d∗g H + 12 dθI − 21 iθ♯ H, ←→ βI = − (ρIB )2,0
∂t I ∂t
∂ ∂
J = 21 L(θ♯ −θ♯ ) J. J = 12 L(θ♯ −θ♯ ) J.
∂t J I ∂t J I
√ 2,0
where βI = −1bI . We will refer to this system as generalized Kähler-Ricci flow
in the I-fixed gauge. Of course one can also study the flow in the J-fixed gauge as
well.
9.3.2. Short-time existence.
Theorem 9.23. Given M a smooth compact manifold and (g, H, I, J) a gener-
alized Kähler structure, there exists a unique maximal solution (gt , Ht , It , Jt ) to gen-
eralized Kähler-Ricci flow with this initial condition on [0, T ) for some T ∈ R∪{∞}.
Moreover, the pair (gt , Ht ) is the unique solution to generalized Ricci flow with ini-
tial condition (g, H).
Proof. We know from the definition of generalized Kähler structure that
(g, H, I) is a pluriclosed structure. Thus by Proposition 9.13 we obtain a solu-
tion (gt , Ht , I) to pluriclosed flow on a maximal time interval [0, T ). This can be
gauge-fixed as in Remark 9.11 to obtain (gt , Ht , It ) a solution to pluriclosed flow
in the form of (9.5). Similarly, the triple (g, −H, J) is a pluriclosed structure, and
so we obtain (e e t , Jt ) a solution to gauge-fixed pluriclosed flow, where (e
gt , H e t ) is
gt , H
the unique solution to generalized Ricci flow with initial condition (g, −H). We use
the notation ge, H e to emphasize that we do not yet know the relationship between
(gt , Ht ) and (e e t ).
gt , H
To link these two families, we make a trivial yet crucial observation: if (gt , Ht ) is
a solution to generalized Ricci flow, so is (gt , −Ht ). This follows since the evolution
equation for H is linear in H, whereas the evolution equation for g is quadratic in
H. In other words,
∂
g = − Rc + 41 H 2 = − Rc + 14 (−H)2 ,
∂t
∂ 1
(−H) = − 2 (∆d H) = 21 ∆d (−H).
∂t
Therefore, (egt , −He t ) is the unique solution to generalized Ricci flow with initial
condition (g, H), and thus (e e t ) = (gt , Ht ). Examining the construction it
gt , −H
follows that (gt , Ht , It , Jt ) is a one-parameter family of generalized Kähler structures
which satisfies (9.7).
9.3.3. Generalized geometry formulation.

Definition 9.24. Let (E, [, ], h, i) → M be an exact Courant algebroid. We say
that a one-parameter family of generalized Kähler structures (Jt1 , Jt2 ) is a solution
of generalized Kähler-Ricci flow if
∂ 1 I
J = [J1 , e−ρB J1 ],
(9.9) ∂t
∂ 2 I
J = [J2 , e−ρB J2 ],
∂t
where ρIB denotes the Bismut-Ricci form of the associated pluriclosed structure
(g, I).
We must show that this formulation of generalized Kähler-Ricci flow agrees
with that arising in equations (9.8). The key points are a variational formula and
a buildup of fundamental curvature identities for generalized Kähler manifolds.
Lemma 9.25. Given (E, [, ], h, i) → M an exact Courant algebroid, fix Kt ∈
Λ2 T ∗ a one-parameter family of two-forms. A one-parameter family of generalized
Kähler structures (J1t , J2t ) satisfies
∂
Ji = [Ji , eK Ji ],
∂t
if and only if the associated one-parameter family of biHermitian structures (gt , bt , It , Jt )
satisfies
ġ = − 21 [K, I], ḃ = − 21 (KI + IK) ,
ω̇I = − 21 [K, I]I, ω̇J = − 12 (KIJ + IJK) ,
I˙ = 0, J˙ = 1 [I, J]g −1 K.
2
Proof. This follows by differentiating the explicit expressions for Ji in equa-

tion (7.14), and then comparing against the formulas for [Ji , eK Ji ]. We leave the
details as an exercise (cf. [81]).
Lemma 9.26. Let (M 2n , g, I, J) be a generalized Kähler structure. Let ∇I , ∇J

denote the Bismut connections associated to the pluriclosed structures (g, I) and
(g, J), with associated curvature tensors RI , RJ . Then
RI ∈ Λ1,1 1,1
J ⊗ ΛI , RJ ∈ Λ1,1 1,1
I ⊗ ΛJ .
Proof. We give the proof for RI , the case of RJ being directly analogous.
Since ∇I preserves I it is easy to show that RI is of type (1, 1) with respect to I
in the last two indices. Using Remark 8.4, we know that if we set H = −dcI ωI then
∇I = ∇+ , ∇J = ∇− in the notation of Definition 3.12. Thus using Proposition
3.21 we see
RI (JX, JY, Z, W ) = R+ (JX, JY, Z, W )

= R− (Z, W, JX, JY )
= R− (Z, W, X, Y )
= R+ (X, Y, Z, W )
= RI (X, Y, Z, W ),
as required.
Lemma 9.27. Let (M 2n , g, I, J) be a generalized Kähler structure. Then

ρIB (X, [I, J]Y ) = g Lθ♯ −θ♯ J X, Y .
J I
Proof. As an exercise one can show that for a Hermitian manifold (M 2n , g, J)

one has
(9.10) g((LX J)Y, Z) = g((∇X J)Y − ∇JY X + J∇Y X, Z),
where ∇ denotes the Levi-Civita connection. We will compute the symmetric and
skew parts of the identity separately. First note, using that ρIB ∈ Λ1,1
J by Lemma
9.26,
ρIB ([I, J]X, JY ) + ρIB ([I, J]JY, X)

= ρIB (IJX − JIX, JY ) − ρIB (X, IJJY − JIJY )
= ρIB (IJX, JY ) − ρIB (IX, Y ) + ρIB (X, IY ) − ρIB (JX, IJY )
= − ρIB (JY, IJX) + ρIB (Y, IX) + ρIB (X, IY ) − ρIB (JX, IJY ).
Applying Lemma 8.9 we further compute

ρIB (JY, IJX) − ρIB (Y, IX) − ρIB (X, IY ) + ρIB (JX, IJY )
= RcI (JY, JX) + ∇IJY θI (JX) − RcI (Y, X) − ∇IY θI (X)
− RcI (X, Y ) − ∇IX θI (Y ) + RcI (JX, JY ) + ∇IJX θI (JY )
= RcJ (JY, JX) + RcJ (JX, JY ) − RcJ (X, Y ) − RcJ (Y, X)
+ ∇IJY θI (JX) − ∇IY θI (X) − ∇IX θI (Y ) + ∇IJX θI (JY )
= − ρJB (JY, X) − ρJB (JX, Y ) − ρJB (X, JY ) − ρJB (Y, JX)
− ∇JJY θJ (JX) − ∇JJX θJ (JY ) + ∇JX θJ (Y ) + ∇JY θJ (X)
+ ∇IJY θI (JX) − ∇IY θI (X) − ∇IX θI (Y ) + ∇IJX θI (JY )
= ∇X (θJ − θI )(Y ) + ∇Y (θJ − θI )(X)

− ∇JX θJ − θI (JY ) − ∇JY θJ − θI (JX)

= g Lθ♯ −θ♯ J X, JY + g Lθ♯ −θ♯ J JY, X .
J I J I
where the last line follows by comparing against (9.10).

To address the skew-symmetric piece we first compute, again using that ρIB ∈
1,1
ΛJ ,
ρIB ([I, J]X, Y ) − ρIB ([I, J]Y, X)
= ρIB ((IJ − JI)X, Y ) + ρIB (X, (IJ − JI)Y )
= ρIB (IJX, Y ) + ρIB (IX, JY ) + ρIB (X, IJY ) + ρIB (JX, IY )
2,0+0,2 2,0+0,2
= 2 ρIB (JX, IY ) + ρIB (IX, JY ) .
Now we compute using Lemma 8.9, Proposition 8.10 and equation (9.10),
2,0+0,2 2,0+0,2
2 ρIB (JX, IY ) + ρIB (IX, JY )
I I
= − d∗ HI (JX, Y ) + d∗ HI (JY, X) + d∇ θI (JX, Y ) − d∇ θI (JY, X)
I I
= d∗ HJ (JX, Y ) − d∗ HJ (JY, X) + d∇ θI (JX, Y ) − d∇ θI (JY, X)
J J
= − 2(ρJB )2,0+0,2 (X, Y ) + d∇ θJ (JX, Y ) + 2(ρJB )2,0+0,2 (Y, X) − d∇ θJ (JY, X)
I I
+ d∇ θI (JX, Y ) − d∇ θI (JY, X)
= − dθJ (JX, Y ) + dθJ (JY, X) + dθI (JX, Y ) − dθI (JY, X)

= g Lθ♯ −θ♯ J X, Y − g Lθ♯ −θ♯ J Y, X .
J I J I
Adding together the two pieces above and rearranging yields the result.
Proposition 9.28. Let (E, [, ], h, i) → M be an exact Courant algebroid. The

family (Jt1 , Jt2 ) is a solution of generalized Kähler-Ricci flow as in (9.9) if and
only if the associated family of biHermitian structures (gt , Ht , It , Jt ) is a solution
of generalized Kähler-Ricci flow in the I-fixed gauge as in (9.8).
Proof. We first assume that equations (9.9) hold. Using Lemma 9.25 with
K = −ρIB one notes that I is fixed and one easily derives the evolution equation
ω̇I = −(ρIB )1,1 . By Lemma 9.27 it follows that

J˙ = − 1 [I, J]g −1 ρIB = L 1
2 ♯ ♯ J.
2 (θJ −θI )
as required. Thus we have verified the evolution equations of (9.8). Note that the
induced evolution equation for b is not strictly necessary in defining the generalized
Kähler-Ricci flow from the biHermitian point of view. Nonetheless this torsion po-
tential (cf. §9.5) plays an important role in obtaining estimates for even pluriclosed
flow. Assuming equations (9.8) and further imposing ḃ = −{ρIB , I}, equations (9.9)
follow in a similar way using Lemma 9.25.
9.3.4. Evolution of the Poisson tensor. As described in §7.4.2, associated
to every generalized Kähler structure we have the real Poisson structure σ. A basic
question is how this tensor evolves along a solution to generalized Kähler-Ricci
flow. Since the I-(2, 0) piece of σ is I-holomorphic by Proposition 7.69, the flow in
the I-fixed gauge can move σ at most within the finite dimensional space of such
holomorphic Poisson structures. As it turns out, it remains completely fixed, a
consequence of the curvature identities of the previous subsection.
Proposition 9.29. Let (M 2n , gt , I, Jt ) be a solution to generalized Kähler-
Ricci flow in the I-fixed gauge. Let σt = 12 gt−1 [I, Jt ] be the one-parameter family of
associated Poisson structures. Then
σt ≡ σ0 .
Proof. Using the evolution equations in the I-fixed gauge (9.8) we obtain
∂ ∂ 1
σ= [I, J]g −1
∂t ∂t2
1 −1 −1 I 1,1 −1
= 2 [I, L 1 ♯ ♯ J]g + [I, J]g (ρB ) (·, I·)g
2 (θJ −θI )
= 0,
where the last line follows using Lemma 9.27.
Remark 9.30. Note that the tensor σ scales as g −1 , thus σ will change along
different normalizations of the flow. In particular, considering the normalized flows
as in Definition 6.30, it follows easily from Proposition 9.29 that the Poisson tensors
σt will satisfy
σt = e−Λt σ0 .
For instance, on CP 2 we expect the normalized pluriclosed flow (with Λ = 1) to
exist globally and converge to the Fubini-Study metric, and this is verified in some
cases. Thus for the generalized Kähler structures on CP 2 with nontrivial Poisson
tensor described in Remark 7.87, the normalized flow causes σ to exponentially
decay to zero, yielding Kähler structure for any smooth limit.
9.3.5. Positive cone in commuting generalized Kähler geometry. In
this subsection we define a cohomological cone for generalized Kähler metrics with
vanishing Poisson tensor, extending the idea of the Kähler cone in Definition 9.4
and specializing the positive cone in the pluriclosed setting in Definition 9.14. To
begin we record a preliminary definition which captures all the properties of being
a generalized Kähler metric in the commuting setting, aside from the positivity
condition.
Definition 9.31. Let (M 2n , g, I, J) be a generalized Kähler manifold with

[I, J] = 0, and let Π = IJ. Given φI ∈ Λ1,1
I,R , let
φJ = −φI (Π·, ·) ∈ Λ1,1

J,R .
We say that φI is formally generalized Kähler if

dcI φI = − dcJ φJ ,
ddcI φI = 0.
It is a basic exercise using the computations of §7.4.4 to show that, for a

smooth function f , the tensor
√
−1 ∂+ ∂ + − ∂− ∂ − f
is formally generalized Kähler. Using these facts we can give the notion of positive
cone in this setting.
Definition 9.32. Let (M 2n , g, I, J) be a generalized Kähler manifold such that

[I, J] = 0. Let
n o
φI ∈ Λ1,1
I,R | φI is formally generalized Kähler
H := √ .
−1 ∂+ ∂ + − ∂− ∂ − f | f ∈ C ∞
Furthermore, define the generalized Kähler cone by
K := {[φ] ∈ H | ∃ ω ∈ [φ], ω > 0}.
Definition 9.33. Let (M 2n , g0 , I, J) be a generalized Kähler manifold satisfy-

ing [I, J] = 0. Let h± denote Hermitian metrics on T±C M , and let P = P (h± ) as in
Proposition 8.20. Then set
τ ∗ (g0 ) := sup{t > 0 | [ω0 − tP ] ∈ K}.
Akin to Lemma 9.17, certainly τ ∗ represents the maximum possible existence

time for a solution to pluriclosed flow in this setting. In analogy with the Theorem of
Tian-Zhang [178] for Kähler-Ricci flow, and as a specialization of the corresponding
general conjecture for pluriclosed flow [170] we make the conjecture that τ ∗ does
indeed represent the smooth maximal existence time.
Conjecture 9.34. Let (M 2n , g0 , I, J) be a generalized Kähler manifold sat-
isfying [I, J] = 0. Then the solution to generalized Kähler-Ricci flow with initial
condition g0 exists on [0, τ ∗ (g0 )).
Question 9.35. Is it possible to give a natural characterization of the gener-
alized Kähler cone K in terms of cohomological conditions and subvarieties?
Question 9.36. Recently, in [66] the notion of blowup is extended to certain
cases in the context of pluriclosed metrics. Also, in [40, 182], the idea of blowup
has been extended to generalized complex and generalized Kähler manifolds. A
natural but broad question to ask is, what is the relationship of these blowups to
pluriclosed flow and generalized Kähler-Ricci flow? Do they inevitably produce
finite time singularities of the flow?
9.4. REDUCED FLOWS 199
9.4. Reduced flows

A fundamental observation in the analysis of Kähler-Ricci flow is the reduction
of the flow to a parabolic partial differential equation for a potential function. In
particular, as we show in Proposition 9.37 below, on any compact time interval in-
side [0, τ ∗ (g0 )), the flow is determined by a function evolving by a kind of parabolic
Monge-Ampère equation. As pluriclosed metrics are determined locally by a (1, 0)-
tensor (cf. Lemma 7.47), we cannot expect a scalar reduction for pluriclosed flow,
but rather a reduction to a flow of (1, 0)-forms, and this is carried out in Proposi-
tion 9.38 (cf. [170], with refinements in [160]). For generalized Kähler-Ricci flow,
in Chapter 7 we saw examples of variations of generalized Kähler metrics deter-
mined by one function, and so we can again hope for a reduction of the flow to a
scalar PDE. This is carried out below in some settings, although the nature of the
scalar reduction varies considerably depending on the rank of the Poisson structure
underlying the solution (cf. [8], [160], [163], [165]).
9.4.1. Scalar reduction of Kähler-Ricci flow.
Proposition 9.37. Let (M 2n , g0 , J) be a compact Kähler manifold, and fix
τ < τ ∗ (g0 ). There exists a volume form Ω and a one-parameter family ωt of Kähler
metrics, 0 ≤ t ≤ τ , such that ωt ∈ [ω0 ] − tc1 , and if ut ∈ C ∞ (M ) satisfies
√ n
∂ ωt + −1∂∂u
(9.11) u = log , u(0) = 0,
∂t Ω
√
then the one-parameter family of metrics ωtu := ωt + −1∂∂u is the unique solution
to Kähler-Ricci flow with initial condition ω0 .
Proof. Fix some background volume form√Ω. e Since τ < τ ∗ by hypothesis, we
e
may fix f ∈ C (M ) such that ω0 − τ ρC (Ω) + −1∂∂f > 0. Setting Ω = e−f /τ Ω,
∞ e
this now reads ω0 −τ ρC (Ω) > 0. Now set ωt = ω0 −tρC (Ω), which certainly satisfies
ωt ∈ [ω0 ] − tc1 . Now suppose ut satisfies (9.11). By construction it is clear that
ω0u = ω0 . Furthermore we compute using the transgression formula
∂ u ∂ √
ωt = ωt + −1∂∂u
∂t ∂t
√ n
√ ωt + −1∂∂u
= − ρC (Ω) + −1∂∂ log
Ω
= − ρC (ωtu ),
as required.
9.4.2. 1-form reduction of pluriclosed flow.
Proposition 9.38. Let (M 2n , g0 , J) be a compact manifold with pluriclosed
metric, and fix τ < τ ∗ (g0 ). There exists a volume form Ω, µ ∈ Λ1,0 , and a one-
parameter family ωt of pluriclosed metrics, 0 ≤ t ≤ τ , satisfying ωt ∈ [ω0 ] − tc1 ,
such that if αt ∈ Λ1,0 satisfies
∂α ∗ √
−1 (ωtα )n µ
(9.12) = −∂ ωtα ωtα − 2 ∂ log − , α(0) = 0,
∂t Ω τ
where
ωtα := ωt + (∂α + ∂α),
then the one-parameter family of metrics ωtα is the unique solution to pluriclosed
flow with initial condition ω0 .
Proof. Fix some background volume form Ω. Since τ < τ ∗ by hypothesis, we
may fix µ ∈ Λ1,0 such that ωτ := ω0 − τ ρC (Ω) + ∂µ + ∂µ > 0. Now set ωt =
t τ −t
τ ωτ + τ ω0 , which certainly satisfies ωt ∈ [ω0 ] − tc1 , interpreted as an equation
of Aeppli cohomology classes. Now suppose αt satisfies (9.12). By construction it
is clear that ω0α = ω0 . Furthermore we compute
∂ α ∂
ωt = ωt + ∂α + ∂α
∂t ∂t
n
1 ∗ α
√
−1 (ωtα ) µ
= − ρC (Ω) + (∂µ + ∂µ) + ∂ −∂ ωtα ωt − 2 ∂ log −
τ Ω τ
√ α n

(ω ) µ
+ ∂ −∂ω∗ tα ωtα + 2−1 ∂ log t −
Ω τ
√ α n
∗ (ω )
= − ∂∂ω∗ tα ωtα − ∂∂ ωtα ωtα + −1∂∂ log t − ρC (Ω)
Ω
∗
= − ∂∂ω∗ tα ωtα − ∂∂ ωtα ωtα − ρC (ωtα )
= − ρ1,1 α
B (ωt ),
as required.
Remark 9.39. We note that the local potential α in Proposition 9.38 also
recovers the flow for the (2, 0)-form β. In particular, if the initial condition satisfies
(9.13) ∂ω0 + dβ0 = H̃0 ,

for H̃0 ∈ Λ , dH̃0 = 0, and we set βt = β0 + ∂ α + t µτ then condition (9.13)
3,0+2,1
is preserved along the flow and furthermore

∂ ∂ µ ∗
β=∂ α+ = −∂∂ gtα ωtα = −ρ2,0
B ,
∂t ∂t τ
as required. Thus, the one-form reduction in Proposition 9.38 can be regarded as
a natural flow for metrics on a fixed holomorphic Courant algebroid, as defined in
§7.3.3 (cf. [71]).
9.4.3. Scalar reduction of generalized Kähler-Ricci flow in commut-
ing case.
Proposition 9.40. Let (M 2n , g0 , I, J) be a compact generalized Kähler mani-
fold satisfying [I, J] = 0. Given τ < τ ∗ (g0 ), there exist partial volume forms Ω± and
a one-parameter family of generalized Kähler metrics ωt such that if ut ∈ C ∞ (M )
satisfies
√ rank T+
∂ (ω+ )t + −1∂+ ∂ + u ∧ Ω−
u = log √ rank T−
,
(9.14) ∂t Ω+ ∧ (ω− )t − −1∂− ∂ − u
u(0) = 0,
then the one-parameter family of metrics
√
ωtu := ωt + −1 ∂+ ∂ + − ∂− ∂ − u
is the unique solution to generalized Kähler-Ricci flow with initial condition ω0 .
9.4. REDUCED FLOWS 201
Proof. Since τ < τ ∗ (g0 ), if we fix e

h a Hermitian metric, then by definition
[ω0 − τ P (e
h)] ∈ K,
thus there exists f ∈ C ∞ such that
√
ω0 − τ P (e
h) + −1 ∂+ ∂ + − ∂− ∂ − f > 0.
Using the transgression formula for P (equation 8.10), if we set h = ef e
h+ ⊕ e
h− ,
then this formula reads
ω0 − τ P (h) > 0.
It follows that ωt := ω0 − tP (h) is positive for all 0 ≤ t ≤ τ . Define Ω± =
h rank T±
(ω± ) , and then we can compute for u satisfying (9.14),
√ k
∂ u √ (ω+ )t + −1∂+ ∂ + u ∧ Ω−
ωt = − P (h) + −1 ∂+ ∂ + − ∂− ∂ − log √ l ,
∂t Ω+ ∧ (ω− )t − −1∂− ∂ − u
= − P (ωtu )
= − ρ1,1 u
B (ωt ),
as required.
Remark 9.41. This proposition indicates that in the commuting setting, the
GRKF in the I-fixed gauge will fix the complex structure J as well (see also Lemma
8.18). By contrast, we will see in §9.4.4 below that, in the nondegenerate setting,
the evolution of the complex structures determines the entire structure.
9.4.4. Scalar reduction of generalized Kähler-Ricci flow in nondegen-
erate case. As explained in Proposition 7.83, a natural class of deformations of
generalized Kähler structure in the nondegenerate setting arises via families of Ω-
Hamiltonian diffeomorphisms. The proposition below shows that the generalized
Kähler-Ricci flow in the nondegenerate setting is a deformation of this kind, with
the family of diffeomorphisms driven by the Bismut-Ricci potential.
Proposition 9.42. Suppose (gt , I, Jt ) is a smooth solution of the generalized
Kähler-Ricci flow in the I-fixed gauge with σ0 nondegenerate, and let
det(I + Jt )
Φt = log .
det(I − Jt )
Let φt denote the flow of the Ω-Hamiltonian vector field
Xt := σdΦt .
Then the induced family of generalized Kähler structures (gφt , I, Jφt ) obtained via
Proposition 7.83 coincides with (gt , I, Jt ).
Proof. By Proposition 8.23, Xt = (θJt − θIt )♯ , where θIt and θJt are the Lee
forms along the GKRF and ♯ denotes gt−1 . It thus follows that
∂
(Jφt − Jt ) = 0,
∂t
showing that Jφ∗t = Jt as they equal J at t = 0. Also, as φt are Ω-Hamiltonian, it
follows that
Im(φ∗t ΩJ ) = Im((φ∗t J)Ω) = Im(Jt Ω).
Taking (1, 1)-part with respect to I gives gt = gφt , as required.
Remark 9.43. In many ways Proposition 9.42 serves as a natural analogue of

the scalar reduction of Kähler-Ricci flow and generalized Kähler-Ricci flow in the
case [I, J] = 0. We have shown that the flow is determined at each time by a single
scalar function, but it remains to be seen if the geometry can be described globally
by a single function (cf. Remark 7.79).
9.5. Torsion potential evolution equations

A key challenge in understanding solutions to the pluriclosed flow is to control
the torsion tensor. As the 1-form potential α for the flow derived in Proposition 9.38
determines the metric tensor, it of course also determines the torsion tensors as well.
Specifically, as explained in Remark 9.39, the tensor ∂α serves as a holomorphic
potential for the torsion of the Chern connection,
T = ∂ω,
and is instrumental in bounding the torsion tensor. Somewhat miraculously, this
torsion potential satisfies a very clean evolution equation, derived in Proposition
9.45 below. Here and in the remainder of this chapter, to save on notation, unless
otherwise specified all connections will be Chern connections associated to a given
time-dependent Hermitian metric.
Lemma 9.44. Let (M 2n , gt , J) be a solution to pluriclosed flow, and suppose
βt ∈ Λp,0 is a one-parameter family satisfying
∂
β = ∆C β + µ,
∂t
where µt ∈ Λp,0 . Then

∂ C 2 2 2
−∆ |β| = − |∇β| − ∇β − p Q, trg β ⊗ β + 2ℜ hβ, µi .
∂t
Proof. We directly compute using the given evolution equation for β and the
pluriclosed flow equation (9.3),
∂ 2 ∂ j i ii
|β| = g . . . g j p ip βi1 ...ip βi1 ...ip β j i ...j p
∂t ∂t
∂g D C
E
(9.15) = −p , trg β ⊗ β + ∆C β + µ, β + β, ∆ β + µ
∂t
D C
E
= p SC − Q, trg β ⊗ β + ∆C β, β + β, ∆ β + 2ℜ hβ, µi .
A standard computation using the Leibniz rule shows that
2
∆C |β|2 = ∆C β, β + β, ∆C β + |∇β|2 + ∇β .
Also, using that β has pure type (p, 0), a computation using the definition of the
Chern curvature tensor shows that
h C
i
∆C − ∆ β = g sr (∇r ∇s − ∇s ∇r ) βi1 ...ip
i1 ...ip
p
X
= g sr Ωtrsik βi1 ...ik−1 tik+1 ...ip
k=1
p
X
= (SC )tik βi1 ...ik−1 tik+1 ...ip .
k=1
9.5. TORSION POTENTIAL EVOLUTION EQUATIONS 203
It follows that
D C
E
β, ∆ β = β, ∆C β − p hSC , trg (β ⊗ β)i .
Plugging these computations into (9.15) yields the result.
Proposition 9.45. Let (M 2n , gt , J) be a solution to pluriclosed flow. Fix back-
ground data ĝt , Ω, µ as in Proposition 9.38 and a solution αt to (9.12). Then

∂ α
− ∆C ∂α = − trgα ∇g Tĝ + ∂µ,
∂t

∂ 2 2 2
− ∆C |∂α| = − |∇∂α| − ∇∂α − 2 Q, tr ∂α ⊗ ∂α
∂t
D α
E
− 2ℜ trgα ∇g Tĝ + ∂µ, ∂α .
Proof. The first equation follows directly from a lengthy computation using
(9.12), see [160] for details. The second equation follows from the first and Lemma
9.44.
Proposition 9.46. Let (M 2n , gt , J) be a solution to pluriclosed flow. Fix back-
ground data ĝt , Ω, µ = 0 as in Proposition 9.38 and a solution αt to (9.12). Suppose
furthermore that there exists η ∈ Λ2,0 such that
(9.16) ∂ ω̂t = ∂ ω̂0 = ∂η.
Let φ = ∂α − η. Then

∂ C
−∆ φ = 0,
∂t
(9.17)
∂
− ∆C |φ|2 = − |∇φ|2 − |Tgt |2 − 2 Q, φ ⊗ φ .
∂t
Proof. We observe that
(trgα ∇gα Tĝ )ij = gαqp ∇p ∂ ω̂ijq = gαqp ∇p ∇q ηij = ∆gα ηij .
Thus using Proposition 9.45 and the assumption that µ = 0 we obtain
∂ ∂
φ= ∂α = ∆gt ∂α − trg ∇g Tĝ = ∆gt (∂α − η) = ∆gt φ.
∂t ∂t
This yields the first claimed equation, and then the second follows from Lemma
9.44 and the fact that
∇φ = ∂φ = ∂ (∂α − η) = −Tgt .

Remark 9.47. We observe that, in the setting of Proposition 9.46, the hypoth-
esis of (9.16) will automatically be satisfied if the initial torsion ∂ω0 has vanishing
Čech cohomology class in H 1 (Λ2,0cl ), so that the pluriclosed metrics are naturally
defined on the trivial holomorphic Courant algebroid √ (cf. Theorem 7.58). One
special case occurs when (M 2n , J) satisfies the −1∂∂ lemma. In this case we
observe that ∂ ω̂0 represents the zero cohomology class in H∂1,2 , and thus by the
√
−1∂∂ lemma, it represents the zero cohomology class √ in the Bott-Chern
√ cohomol-

1,2
ogy HBC . Thus there exists ξ ∈ Λ1,0 such that ∂ ω̂0 = −1∂∂ξ = ∂ − −1∂ξ . It
is easy to see from the construction of Proposition 9.38 that since µ = 0, we have
√ 2
∂ ω̂t = ∂ ω̂0 = ∂ − −1∂ξ . It furthermore follows in this setting that |φ| serves
as a torsion-bounding subsolution in the sense of §6.3.
9.6. Higher regularity from uniform parabolicity

In Theorem 5.23 we showed that the Riemann curvature tensor must blow up
at any finite time singularity of generalized Ricci flow. The key point is that in
the presence of a bound on the Riemann curvature one can obtain the higher-
order estimates of all derivatives of curvature from Theorem 5.18. For the case
of pluriclosed flow it is possible to significantly strengthen the characterization of
finite time singularities using the special structure of generalized complex geometry.
We first recall that in the setting of Kähler-Ricci flow we can reduce to the par-
abolic complex Monge-Ampère equation using Proposition 9.37. Assuming uniform
equivalence of the evolving metric with a background, one has a strictly parabolic,
fully nonlinear, convex PDE, with coefficients in L∞ . If these coefficients were
slightly more regular, i.e. in C α , then a standard bootstrapping argument using
Schauder estimates would yield C ∞ regularity of the flow. Thus improving the
coefficients from L∞ to C α becomes a crucial step. These C 3 estimates for the
potential function follow directly by applying the maximum principle to a precisely
chosen geometric quantity, namely the difference between the evolving Chern con-
nection and a background (cf. [33, 186]), which gives the needed regularity. This
estimate can also follow in this setting following the classical works of Evans-Krylov
[60, 118].
One faces a similar challenge in the setting of pluriclosed flow, where one seeks
regularity of the flow assuming uniform equivalence of the metric tensor with some
background. Comparing against equation (9.6), one then sees that the metric sat-
isfies a uniformly parabolic equation with L∞ coefficients and quadratic first-order
nonlinearity. By a combination of rescaling and Schauder estimates, it is possible to
show that a C α estimate for the metric would suffice to obtain full C ∞ estimates.
Thus obtaining an a priori C α estimate for the metric in the presence of upper and
lower bounds on the metric becomes a crucial barrier. It is hard to initially guess
what the right quantity to study is to obtain this estimate, as the pluriclosed flow is
a system of equations, and for parabolic systems this regularity is known in general
to be false. Moreover, we note that even in the restricted setting of generalized
Kähler-Ricci flow, the scalar reduction in (9.14) is a nonconvex scalar PDE, and
thus the classical methods will not apply.
Nonetheless, using ideas from generalized geometry, it is possible to obtain this
regularity. As mentioned above, for Kähler-Ricci flow the appropriate quantity to
study is the difference of Chern connections of the flowing metric and a background.
For the pluriclosed flow, it turns out that there is a clean evolution equation for
the difference of Chern connections associated to certain generalized Hermitian
metrics on T 1,0 ⊕ T1,0
∗
defined using the underlying Hermitian metric on T 1,0 and
an associated torsion potential. We refer the reader to ([165, 172, 160, 108]) to
see the conceptual buildup leading to the relatively simple formulation below. We
first record the relevant regularity statement expressed using classical objects, then
an equivalent version expressed using generalized metrics which is the version we
will prove.
9.6. HIGHER REGULARITY FROM UNIFORM PARABOLICITY 205
Definition 9.48. Fix a complex manifold (M 2n , J), and let g, e

g denote Her-
mitian metrics. Let
g) := ∇C
Υ(g, e C
g − ∇e
g
denote the difference of the Chern connections associated to g and ge. Furthermore,
let
k
X 2
fk (g, e
g) := ∇jg Υ(g, e
g) 1+j
.
j=0
The quantity fk is a natural measure of the (k + 1)-st derivatives of the metric

which scales as the inverse of the metric.
Theorem 9.49. Let (M 2n , J) be a compact complex manifold. Suppose gt is a
solution to the pluriclosed flow on [0, τ ), τ ≤ 1, with αt a solution to the (ĝt , Ω, µ)-
reduced flow (see Proposition 9.38). Suppose there exist constants λ, Λ such that
λg0 ≤ gt ≤ Λg0 , |∂α|2 ≤ Λ.
Given k ∈ N there exists a constant C = C(n, k, g0 , ĝ, Ω, µ, λ, Λ) such that
sup tfk (gt , h) ≤ C.
M×{t}
As discussed above, this theorem is proved using a certain generalized metric

on T 1,0 ⊕ T1,0
∗
. In particular, given (M 2n , J) a complex manifold, fix g a pluriclosed
metric and β ∈ Λ2,0 . Using these we define an associated generalized metric (cf.
§7.3)
!
gij + βik β jl g lk βip g lp
(9.18) G= .
β jp g pk g lk
This is a Hermitian metric on T 1,0 ⊕ T1,0
∗
, and thus comes equipped with a Chern
C
connection, which we will denote ∇G . Our estimates are phrased naturally in terms
of these connections, so we make a definition analogous to Definition 9.48:
Definition 9.50. Fix a complex manifold (M 2n , J), let g denote a Hermitian
e denote Hermitian metrics on T 1,0 ⊕ T ∗ . Let
metric, and let G, G 1,0

e := ∇G − ∇ e ∈ Γ T ⊗ End(T ⊕ T1,0
Υ(G, G) C C 1,0 1,0 ∗
)
G
e This is a nat-
denote the difference of the Chern connections associated to G and G.
ural tensor which measures the first derivatives of G with respect to a background
e Furthermore, let
choice G.
k
X 2
∇jg,G Υ(G, G)
1+j
e :=
fk (G, G) e .
g,G
j=0
The connection ∇g,G denotes the Chern connection on T 1,0 ⊗ End(T 1,0 ⊕ T1,0 ∗
)
1,0
induced naturally by the Chern connection of g on T and the Chern connnection
associated to G on End(T 1,0 ⊕ T1,0
∗
). The quantity fk is a natural measure of the
(k + 1)-st derivatives of the generalized metric which scales as the inverse of the
metric.
Theorem 9.51. Let (M 2n , J) be a compact complex manifold, and suppose

(ωt , βt ) is a solution of pluriclosed flow (9.2) on [0, τ ), τ ≤ 1. Let Gt denote the
one-parameter family of associated generalized Hermitian metrics, fix G e another
generalized Hermitian metric, and suppose
e ≤ G ≤ ΛG.
Λ−1 G e
e such that
Given k ∈ N there exists C = C(n, k, Λ, G)
e ≤ C.
sup tfk (G, G)
M×{t}
For the proof, which comprises the remainder of this section, we will assume
that our solution to pluriclosed flow satisfies
∂ω0 + ∂β0 = 0,
which implies that this equation holds at all times as discussed above. This is
not necessary for the proof but simplifies the exposition (cf. Remark 9.47). The
proof Theorem 9.51 relies principally on a very clean evolution equation for G
along pluriclosed flow. To state this we note that given G as in (9.18), we have an
associated curvature tensor ΩG ∈ Λ1,1 ⊗ End(T 1,0 ⊕ T1,0∗
). Akin to Definition 8.8,
we define the tensor
S G = trω ΩG ∈ End(T 1,0 ⊗ T1,0
∗
).
Amazingly, while the evolution equations for g and β along pluriclosed flow are
determined by the curvature of the Bismut connection, the evolution equation for
the metric G is determined by the curvature of its associated Chern connection in
a very simple way:
Proposition 9.52. Given (M 2n , J) a complex manifold, suppose (ωt , βt ) is
a solution of pluriclosed flow (9.2). Let Gt denote the one-parameter family of
associated generalized metrics as in (9.18). Then
∂
G−1 G = − SG.
∂t
Proof. We leave this lengthy but straightforward computation as an exercise,√
see [108] for details. The principal fact is that, by the proof of Lemma 7.61, e2 −1β
can be regarded as defining an homomorphism of holomorphic Courant algebroids
√
e2 −1β
: E−2√−1∂ω → T 1,0 ⊗ T1,0
∗
,
where the right hand-side is equipped with the trivial structure. Thus, for the
calculation of the Chern connection of G we can assume β = 0, provided that
we work with the twisted holomorphic structure in Definition 7.59. A lengthy
calculation then shows that
1,1
G ρB ρ2,0
B g −1
S G= −1 .
g −1 ρ0,2
B −g −1 ρ1,1
B g

Remark 9.53. The tensor S G as defined above is the same as that appearing in
the Hermitian-Yang-Mills equations [113]. In that setting one chooses a background
Hermitian metric ω on a given complex manifold, then solves for a Hermitian metric
h on an associated holomorphic vector bundle E such that Sω,h = trω Ωh = λ IdE ,
9.6. HIGHER REGULARITY FROM UNIFORM PARABOLICITY 207
where Ωh is the curvature of the canonically associated Chern connection. In our

setting both metrics ω and h = G are determined by the same underlying pluriclosed
metric on the base manifold M . Nonetheless, in this way the Bismut Hermitian-
Einstein condition can be interpreted as S G = 0. The existence of Hermitian-Yang-
Mills metrics is associated to the conditions of slope stability [57, 180], and one
wonders what role these obstructions can play in understanding the existence and
uniqueness of Bismut Hermitian-Einstein metrics.
Lemma 9.54. Given (M 2n , J) and (ωt , βt ) a solution to pluriclosed flow (9.2)

with associated generalized metric Gt and background generalized metric G,e one has

∂ e = − |Υ|2−1 −1 + g nm GBA Ω e
− ∆C trG G g ,G ,G e mnAB .
∂t
Proof. Using Proposition 9.52 we compute
∂ e = ∂ (GBA G
e ) = GCA GBD S Ge
trG G AB DC AB .
∂t ∂t
Furthermore,
e = g nm ∇m ∇n (GBA G
∆ trG G e )
AB
e ΥC
= − g nm GBA ∇m (GAC nB )

C
nm BA
=g G e D e C e C
GDC ΥmA ΥnB + GAC (ΩnmB − ΩnmB )
e
= |Υ|2g−1 ,G−1 ,Ge + GCA GBD SDC G nm BA e
AB − g G ΩmnAB .
Combining the two equations above gives the result.
Proposition 9.55. Given (M 2n , J) and (ωt , βt ) a solution to pluriclosed flow

(9.2), with associated generalized metric Gt and background generalized metric G, e
one has

∂ e +Υ⋆Υ⋆Ω e +Υ⋆∇ e Ω.
e
− ∆C |Υ|2 = −|∇Υ|2 − |∇Υ + T · Υ|2 + T ⋆ Υ ⋆ Ω
∂t
Proof. We work in local complex coordinates, and adopt the convention that
lower-case letters will be used for indices ranging over T 1,0 and capital letters will
be used to represent indices ranging over all T 1,0 ⊕ T1,0
∗
. A general calculation for
the variation of the Chern connection associated to a Hermitian metric yields
∂ B ∂
Υ = ∇i GB .
∂t iA ∂t A
Using Proposition 9.52 we obtain
∂ B B
(9.19) Υ = −∇i SA .
∂t iA
Using this, we compute

∆C ΥB kl
iA = g ∇l ∇k ΥiA
B

= g kl ∇l ΩkiA
B eB
−Ω kiA
(9.20) = g kl ∇i ΩB
klA
+ g kl Tilp ΩB
kpA
eB
+ g kl ∇l Ω ikA
kl p B kl e e B
= ∂t ΥB
iA − g Til ΩpkA + g ∇l ΩikA

+ g kl Υqli Ω
e B + ΥD
qkA
e
Ω B
lA ikD − Υ B eD
Ω
lD ikA .
Next we observe the commutation formula

C B B
∆ ΥjA = g lk ∇l ∇k ΥjA

lk B g q B C B B C
(9.21) = g ∇k ∇l ΥjA − (Ω )klj ΥqA − ΩklA ΥjC + ΩklC ΥjA
B B B C
= ∆C ΥjA − (S g )qj ΥqA − SA
C B
ΥjC + SC ΥjA .
Combining (9.20), (9.21), the evolution equations of Proposition 9.52 and the pluri-
closed flow equation we obtain

∂ 2 ∂ ji CA B D
|Υ|g−1 ,G−1 ,G = g G GBD ΥiA ΥjC
∂t ∂t
D LA B D
g
= − g jk −Skl + Qkl g li GCA GBD ΥB iA ΥjC − g G
ji CM G
−SML G ΥiA ΥjC
B D
g p B
+ g ji GCA −SBD G
ΥiA ΥjC + g ji GCA GBD ∆C ΥB kl
iA + g (T )il ΩpkA
h i D
−g kl ∇ e B − g kl Υq Ω
e lΩ e B + ΥMlA
e B − ΥB
Ω lM
eM
Ω ΥjC
ikA li qkA ikM ikA

C D g q kl e e D
+ g ji GCA GBD ΥB kl D
iA ∆ ΥjC − g (T )kj ΩlqC + g ∇k ΩljC

q
−g kl Υkj Ω e D + ΥL Ωe
kC jlL
D
− Υ
D
Ωe
kL jlC
L
.
qlC
We furthermore compute
D
∆C |Υ|2g−1 ,G−1 ,G = g kl g ji GCA GBD ∇l ∇k (ΥB
iA ΥjC )
= h∆C Υ, Υi + hΥ, ∆C Υi + |∇Υ|2 + |∇Υ|2

D D
= (S g )qj ΥqC ΥB ji CA
iA g G
Λ
GBD + SC ΥjΛ ΥB ji CA
iA g G GBD
Λ C
− SΛD ΥjC ΥB ji CA
iA g G GBD + h∆C Υ, Υi + hΥ, ∆ Υi + |∇Υ|2 + |∇Υ|2 .
Combining the two equations above yields

∂
− ∆C |Υ|2 = − |∇Υ|2 − |∇Υ|2
∂t

D p D q
+ g ji g kl GCA GBD −Qik ΥB Υ
lA jC + T il ∇Υ B
Υ
pkA jC
− T kj ∇Υ D
Υ B
lqC iA
+T ⋆Υ⋆Ω e +Υ⋆∇
e +Υ⋆Υ⋆Ω e Ω.
e
9.7. METRIC EVOLUTION EQUATIONS 209
Then, observing that the second through fifth terms above form a perfect square
we arrive at

∂ e +Υ⋆Υ⋆Ω e +Υ⋆∇ e Ω,
e
− ∆C |Υ|2 = −|∇Υ|2 − |∇Υ + T · Υ|2 + T ⋆ Υ ⋆ Ω
∂t
as claimed.
Proof of Theorem 9.51. To begin we prove the case k = 1. Fix a constant
A > 0 and let
2 e
Φ = t |Υ| + A trG G.
Combining Proposition 9.55 and Lemma 9.54 we obtain

∂ C e +Υ∗Υ∗Ω e +Υ∗∇
eΩe
−∆ Φ = t −|∇Υ|2 − |∇Υ + T · Υ|2 + T ∗ Υ ∗ Ω
∂t

2 e
+ |Υ| + A −|Υ|2g−1 ,G−1 ,Ge + g nm GBA Ω mnAB
2
≤ |Υ| (1 + tC − A) + CA
≤ CA,
where the last line follows using that t ≤ 1 and choosing A = C + 1. Applying the
maximum principle we obtain
sup Φ ≤ C(1 + t) ≤ C.
M×{t}
2
Since Φ ≥ |Υ| , the case k = 1 follows.
Having established the case k = 1, there are two ways to finish the general proof.
Beginning with (9.19) one can derive natural heat equations for all derivatives ∇j Υ,
and derive smoothing estimates for these quantities in a manner similar to Theorem
5.18. Alternatively one can use the scale invariance of the quantity tfk and apply a
rescaling argument as carried out in [108]. We leave the details to the reader.
9.7. Metric evolution equations

Theorem 9.49 shows that solutions to pluriclosed flow can be extended smoothly
as long as the equation remains uniformly parabolic. To measure this parabolicity
requires obtaining uniform estimates on the metric tensor along the flow. In this
section we derive reaction-diffusion equations for various quantities which measure
the metric evolving by pluriclosed flow against a background metric. We will apply
the maximum principle to these evolution equations in the next section to establish
global existence results for the pluriclosed flow.
Lemma 9.56. Let (M 2n , gt , J) be a solution to pluriclosed flow, and let h denote
another Hermitian metric on (M, J). Then

∂ det g 2
− ∆C log = |T | − trg ρC (h).
∂t det h
Proof. We compute in local complex coordinates. Here and below we will
use the expression for the tensor SC in terms of local complex coordinates, which
follows from a short computation,
(9.22) (SC )ij = − g qp gij,pq + g qp g sr gis,p grj,q .
Using this and equation (9.3 for pluriclosed flow we obtain

∂ det g ∂
log = g ji g
∂t det h ∂t ij
h i
= g ji g qp gij,pq − g qp g sr gis,p grj,q + Qij
h i
2
= g ji g qp gij,pq − g qp g sr gis,p grj,q + |T | .
Also

C det g qp det g
∆ log =g log
det h det h ,pq
h i
= g qp g ji gij,p − hji hij,p
,q
h i
= g qp g ji gij,pq − g jk gkl,q g li gij,p − hji hij,pq + hjk hkl,q hli hij,p .
Combining the above calculations and comparing against the transgression formula
of Proposition 8.16 yields

∂ C det g 2
−∆ log = |T | − trg ρC (h),
∂t det h
as required.


∂
−∆ C
trg h = − |Υ(g, h)|2g−1 ,g−1 ,h − hh, Qig + Ωh (g −1 , g −1 ).
∂t
Proof. We directly compute
∂ ∂ ji
trg h = g hij
∂t ∂t
jk ∂
= −g g g li hij
∂t kl
h i
= − g jk g li hij g qp gkl,pq − g qp g sr gks,p grl,q + Qkl .
On the other hand

h i
∆C trg h = g qp g ji hij
,pq
h i
qp
=g −g gkl,p g li hij + g ji hij,p
jk
,q
h
qp
=g g grs,q g gkl,p g hij − g gkl,pq g li hij + g jk gkl,p g lr grs,q g si hij
jr sk li jk
i
−g jk gkl,p g li hij,q − g jk gkl,q g li hij,p + g ji hij,pq .
9.7. METRIC EVOLUTION EQUATIONS 211
Combining the above calculations yields

∂
− ∆C trg h = − g qp g jr g sk g li hij grs,q gkl,p
∂t
h i
+ g qp g jk gkl,p g li hij,q + g jk gkl,q g li hij,p − g ji hij,pq − hh, Qig
2
= − |Υ(g, h)|g−1 ,g−1 ,h − hh, Qig + Ωh (g −1 , g −1 ).


∂ hh, Qig
− ∆C log trg h ≤ − + K trg h,
∂t trg h
where K = sup K(h), i.e. the supremum of all holomorphic bisectional curvatures.
Proof. We first observe the general fact that
2
∂ ∂ |∇f |
− ∆C log f = f −1
− ∆C f+ .
∂t ∂t f2
Using this together with Lemma 9.57 yields

∂
− ∆C log trg h
∂t
1 2

= − |Υ(g, h)|g−1 ,g−1 ,h − hh, Qig + Ωh (g −1 , g −1 )
trg h
2
|∇ trg h|
+
(trg h)2
= T1 + T2 + T3 + T4 ,
where we have labeled the four terms appearing on the right hand side. To estimate
the term T4 , we first pick special coordinates at some given point p ∈ M . In partic-
ular, we can choose complex coordinates for which hij = δij and g is diagonalized.
We furthermore observe then that, at the point p,

∇i trg h = ∂i g lk hkl
= − g lp ∂i gpq g qk hkl + g lk ∂i hkl

= − Υ(g, h)kip (g −1 h)pk .
Using this and the Cauchy-Schwarz inequality we can estimate

2
|∇ trg h|
T4 =
trg h
!−1
X
ii
= g g jj ∇j trg h∇j trg h
i
!−1 " #
X X X X
ii jj kk
= g g g Υkjk g ll
Υljl
i j k l
!−1 " #
X X Xh 1 1
i Xh 1 1 1
i 1
= g ii (g jj ) (g kk ) Υkjk
2 2 (g kk ) (g jj ) 2 (g ll ) 2 Υljl (g ll ) 2
2
i j k l
!−1   12 ! 21   21 ! 21
X X X X X
ii jj kk
≤ g  g g Υkjk Υkjk  g kk  jj ll
g g Υljl Υljl  g ll
i j,k k j,l l
≤ |Υ(g, h)|2g−1 ,h−1 ,g = T1 .

We also estimate, using the definition of A,
1
T3 = Ωh (g −1 , g −1 )
trg h
1
= (Ωh )ijkl g ji g lk
trg h
1
≤K h h g ji g lk
trg h ij kl
= K trg h.
Combining the above estimates, the lemma follows.
Lemma 9.59. Let (M 2n , gt , J) be a solution to pluriclosed flow. Then

∂ C 2
−∆ trh g = − |Υ(g, h)|g−1 ,h−1 ,g + trh Q − Ωh (h−1 gh−1 , g −1 ).
∂t
Proof. Using (9.22) we obtain
∂ ∂ ji
trh g = h g
∂t ∂t h ij i
= hji g qp gij,pq − g qp g sr gis,p grj,q + Qij
h i
= hji g qp gij,pq − g qp g sr gis,p grj,q + trh Q.
On the other hand
h i
∆C trh g = g qp hji gij
,pq
h i
qp
=g −h hkl,p hli gij + hji gij,p
jk
,q
h
qp
=g h hrs,q h hkl,p h gij − h hkl,pq hli gij + hjk hkl,p hlr hrs,q hsi gij
jr sk li jk
i
−hjk hkl,p hli gij,q − hjk hkl,q hli gij,p + hji gij,pq .
9.8. SHARP EXISTENCE AND CONVERGENCE RESULTS 213
Combining the above calculations yields

∂ C
−∆ trh g = − hji g qp g sr gis,p grj,q + trh Q
∂t
h
− g qp hjr hrs,q hsk hkl,p hli gij − hjk hkl,pq hli gij + hjk hkl,p hlr hrs,q hsi gij
i
−hjk hkl,p hli gij,q − hjk hkl,q hli gij,p
2
= − |Υ(g, h)|g−1 ,h−1 ,g + trh Q − g qp (Ωh )lk
pq gkl ,
as required.

∂ trh Q
− ∆C log trh g ≤ + K trg h,
∂t trh g
where K = − inf K(h), i.e. the negative of the infimum of all holomorphic bisec-
tional curvatures.
Proof. The proof follows that of Lemma 9.58 closely, building off of Lemma
9.59.
9.8. Sharp existence and convergence results

In this section we combine the results of the previous sections to establish global
existence and convergence results for the Kähler-Ricci flow, pluriclosed flow, and
generalized Kähler-Ricci flow, and discuss their consequences.
9.8.1. A general regularity result. We prove a technical tool underpinning
all of the global existence results to follow for the pluriclosed flow. In particular,
we show that for times τ < τ ∗ (g0 ), a lower bound on the metric suffices to obtain
full C ∞ regularity of the flow. The proof builds on the evolution equations of the
previous sections as well as the sharp higher order regularity estimate of Theorem
9.49.
Proposition 9.61. Let (M 2n , g0 , J) be a compact complex manifold with pluri-
closed metric. Fix τ < τ ∗ (g0 ). Suppose the solution to pluriclosed flow with initial
condition g0 exists on [0, τ ) and furthermore satisfies
gt ≥ δg0
for some constant δ, at all points (p, t) ∈ M × [0, τ ). Then gt has uniform C ∞
estimates on [0, τ ), and the solution to pluriclosed flow extends smoothly past τ .
Proof. We choose ĝt , Ω, µ, and a solution αt to (9.12) as in Proposition 9.38.
For simplicity we assume that these background objects satisfy the hypotheses of
Proposition 9.46, i.e µ = 0 and ∂ ω̂t = ∂ ω̂0 = ∂η, and we set φ = ∂α − η . The
general case requires estimating a few more background terms but is effectively the
same (cf. [160] Theorem 1.8). We estimate, using Lemma 9.56, Proposition 9.46
and the assumed lower bound for gt ,

∂ C det g 2 2
−∆ log + |φ| ≤ trg ρC (h) − |∇φ| − 2 hQ, φ ⊗ φi ≤ Cδ −1 .
∂t det h
By applying the maximum principle we obtain a time-dependent upper bound for

2 2
the metric as well as for |φ| and hence |∂α| . Applying Theorem 9.49 we obtain
∞
uniform C bounds on [0, τ ) and the result follows.
9.8.2. Kähler-Ricci flow.
Theorem 9.62. ([178]) Let (M 2n , ω0 , J) be a compact Kähler manifold. Then
the solution to Kähler-Ricci flow with initial condition ω0 exists smoothly on [0, τ ∗ (g0 )).
Proof. Fix any time τ < τ ∗ , and choose Ω, a family of background metrics
ωt , and ut a solution of the reduced flow (9.11), i.e.
√ n
∂ ωt + −1∂∂u
(9.23) u = log , u(0) = 0,
∂t Ω
as guaranteed by Proposition √ 9.37. It suffices to show that the one-parameter
family of metrics ωtu = ωt + −1∂∂u remains smooth on the time interval [0, τ ], as
the uniqueness of solutions to Kähler-Ricci flow then guarantees that the maximal
solution must exist on [0, τ ∗ (g0 )).
First, by a direct application of the maximum principle to (9.23) we obtain a
∂
uniform upper bound on ∂t u at maximum points of u, and thus a uniform upper
bound for u on [0, τ ]. Arguing similarly we obtain a lower bound as well. Next we
obtain a uniform upper and lower bound for the volume form. Let
(ω u )n
Φ1 := log t − Au,
Ω
where A is a constant to be determined. Combining Lemma 9.56 and (9.23) we
compute, bounding the background curvature ρC (Ω), using the a priori estimate
for u, and choosing A sufficiently large,

∂ ∂
− ∆C Φ = − trωtu ρC (Ω) − A u − trωtu (ωtu − ωt )
∂t ∂t

≤ B trωt ωt − A Φ1 + Au − n + trωtu ωt
u
≤ (B − A) trωtu ωt − AΦ1 + CA2 + An

≤ − AΦ1 + CA2 + An.
A direct application of the maximum principle yields an a priori upper bound for
(ωtu )n
Φ1 , and hence for log Ω . A directly analogous argument yields the lower bound.
We now obtain the lower bound for the metric. Fix a background metric h and let
Φ2 = log trωtu ωh − Au.
∂
Applying Lemma 9.57, (9.23), and the a priori estimate for ∂t u,

∂ C ∂ u
−∆ Φ2 ≤ K trωtu ωh − A u − trωtu (ωt − ωt )
∂t ∂t
≤ K trωtu ωh + AC − Aδ trωtu ωh
≤ (K − Aδ) trωtu ωh + AC
Aδ
≤ − trωtu ωh + A,
2
where the constant δ is such that ωt ≥ δωh , and A is chosen sufficiently large
to obtain the final inequality. By the maximum principle, at a sufficiently large
maximum for Φ2 we obtain an a priori upper bound for trωtu ωh . It follows that Φ2
has an a priori upper bound and thus trωtu ωh is bounded above, giving an a priori
lower bound for the metric ωtu . The smooth existence now follows from Proposition
9.61, finishing the proof.
Note that Theorem 9.62 is confirmation of Conjecture 9.18 in the case the initial
metric is Kähler, and was the inspiration for making Conjecture 9.18 in the first
place.
9.8.3. Pluriclosed flow.
Theorem 9.63. Let (M 2n , J) be a compact complex manifold admitting a Her-
mitian metric h with nonpositive holomorphic bisectional curvature. Given g0 a
pluriclosed metric on M , the solution to pluriclosed flow with initial condition g0
exists on [0, ∞).
Proof. We first note that the formal existence time of the flow is infinite.
In particular, using that h has nonpositive holomorphic bisectional curvature, it
follows that
ρC (h)ij = hlk Ωijkl
is also nonpositive. Thus for any τ ≥ 0 we have ω0 − τ ρC (h) ≥ 0, and hence
[ω0 ] − tc1 ∈ P, thus the formal existence time is infinite, as claimed.
Using that h has nonpositive holomorphic bisectional curvature, and that Q ≥
0, we observe using Lemma 9.57 that

∂ 2
− ∆C trg h = − |Υ(g, h)|g−1 ,g−1 ,h − hh, Qig + Ωh (g −1 , g −1 )
∂t
≤ 0.
From the maximum principle we conclude that for all times t such that the flow is
defined, we have
sup trg h ≤ sup trg h.
M×{t} M×{0}
Having established a uniform lower bound for the metric for all times t, the result
follows from Proposition 9.61
In general, the convergence behavior at infinity in the setting of Theorem 9.63
can be fairly complicated. In the case (M 2n , J) admits a Kähler-Einstein metric
of negative scalar curvature, we should expect the normalized flow to converge to
this Kähler-Einstein metric. Similar results are shown in the case the background
metric has nonpositive curvature operator in [124]. The flow in these settings can
also exhibit collapsing behavior at infinity. For instance, if Σ denotes a Riemann
surface of genus g ≥ 2, consider the product manifold T 2 × Σ with the product
complex structure. The product of a flat metric on T 2 with a hyperbolic metric on
Σ has nonpositive holomorphic bisectional curvature. The normalized pluriclosed
flow (actually Kähler-Ricci flow) with this initial data will remain a product metric,
keeping the metric on Σ fixed while homothetically shrinking the metric on T 2 at
an exponential rate. In some cases the hypothesis of nonpositive curvature can be
weakened to a topological condition, and global existence and convergence results
can be obtained (cf. [161]).
In the case (M 2n , J) is a torus, there is a Hermitian metric with vanishing
Chern curvature, and the solution to (unnormalized) pluriclosed flow will converge
exponentially to a flat metric, as we prove in Theorem 9.65 below. First we need a

rigidity result for steady solitons generalizing Proposition 8.32.
Proposition 9.64. Let (M 2n , g, J) be a compact complex manifold with pluri-

closed metric g which is a steady generalized Ricci soliton. If c1 (M, J) ≤ 0 in
Bott-Chern cohomology, then g is Kähler Calabi-Yau.
Proof. As c1 (M, J) ≤ 0 in Bott-Chern cohomology, we may choose a Hermit-

ian metric h such that ρC (h) ≤ 0. As g is a steady soliton, the pluriclosed flow will
evolve by diffeomorphism pullback by a vector field X. Using Lemma 9.56 we then
obtain
det g ∂ det g det g det g
X · log = log = ∆C log + |T |2 − ρC (h) ≥ ∆C log + |T |2 .
det h ∂t det h det h det h
det g
Applying the strong maximum principle (Proposition 4.16), it follows that log det h
2
is constant, and hence |T | = 0, and so g is Kähler. Compact steady Ricci solitons
are Einstein by Proposition 4.35, hence g is Kähler Calabi-Yau, as claimed.
Theorem 9.65. Given ω0 a pluriclosed metric on a complex torus (T 2n , J), the

solution to pluriclosed flow with initial condition ω0 exists on [0, ∞), and converges
exponentially to a flat Kähler metric.
Proof. The global existence of the flow, as well as an a priori lower bound
for the metric, follows from Theorem 9.63. We claim that there exists a uniform a
priori upper bound for the metric for all times t. To see this we can choose h to be
the given flat metric, set ω̂t = ω̂0 , µ = 0, and obtain a solution αt to the reduced
2n
pluriclosed flow
√ equation (9.12) using Proposition 9.38. Furthermore, since T is
Kähler, the −1∂∂-lemma holds, and thus following Remark 9.47, we can verify
the hypotheses of Proposition 9.46, and obtain φ = ∂α−η satisfying (9.17). Setting
det g 2
Φ = log det h + |φ| , combining Lemma 9.56 with (9.17) yields

∂ C 2
−∆ Φ ≤ − |∇φ| − 2 Q, φ ⊗ φ ≤ 0.
∂t
Thus by the maximum principle we conclude
sup Φ ≤ sup Φ.
M×{t} M×{0}
It follows that the determinant of g is bounded above, and thus since there is an a
priori lower bound, it follows that there is an a priori upper bound for the metric.
It follows that we can apply Theorem 9.49 to obtain uniform C ∞ estimates for the
metric at all times as well.
Given these estimates, for any sequence {tj } → ∞ there exists a subsequence,
also denoted {tj }, such that {gtj } converges in C ∞ to a smooth limiting metric
g∞ . Corollary 6.10 implies that λ(gt , Ht ) is monotonically increasing and more-
over bounded above due to the C ∞ control of g, thus there exists λ such that
limt→∞ λ(gt , Ht ) = λ. Choose a sequence {e tj } such that λ − λ(getj , Hetj ) ≤ j −1 . Let
fetj +1 denote the unique minimizer of F (getj +1 , Hetj +1 , ·), and let u = e−ft denote the
−f
unique solution to the conjugate heat equation on [e tj , e
tj + 1] with uet +1 = e tej +1 .
j
It follows from Proposition 6.8 that

Z etj +1
−1 d
j ≥ F (gt , Ht , ft )dt
e
tj dt
Z etj +1 Z h i
2 2
= Rc − 14 H 2 + ∇2 f + 1
4 |d∗ H + ∇f yH| e−f dV dt.
e
tj M
It follows that there exists a sequence of times tj ∈ [e

tj , e
tj + 1] such that
Z
2 2 2
(9.24) lim Rc − 4 H + ∇ f gt + 4 |d H + ∇f yH|gt e−ftj dVgtj = 0.
1 2 1 ∗
j→∞ M j j
Applying the argument above, the sequence {gtj } admits a further subsequence
which converges in C ∞ to a limiting metric g∞ . Furthermore, the subsequence
can be chosen so that the functions {ftj }, which admit C ∞ estimates by elliptic
regularity, also converge to a limit f∞ . By (9.24), the limit (g∞ , H∞ , f∞ ) is a
steady generalized Ricci soliton. Since c1 (T 2n , J) = 0 in Bott-Chern cohomology,
it follows from Proposition 9.64 that the limit g∞ is actually Kähler-Einstein, with
f∞ constant. As all Kähler-Einstein metrics on T 2n are flat, we have established
subsequential convergence to a flat Kähler metric. To finish the proof of exponential
convergence of the entire flow we appeal to a general stability result for pluriclosed
flow [168, Theorem 1.2] near Kähler, Ricci-flat metrics, which in this setting implies
that the pluriclosed flow starting sufficiently close to a flat metric on a torus will
converge exponentially to a flat metric, as required.
9.8.4. Generalized Kähler-Ricci flow: Commuting case. We first prove

a sharp local existence theorem confirming Conjecture 9.34 for the generalized
Kähler-Ricci flow in the commuting case with some topological hypotheses on the
bundle splitting in place. As described in [165], this result covers almost all cases
of generalized Kähler-Ricci flow in the commuting case on complex surfaces, in-
cluding general type, properly elliptic, bi-elliptic, some ruled surfaces, and Inoue
surfaces. The statement below was shown for the case n = 2 in [165], using various
techniques special to that case.
Theorem 9.66. Let (M 2n , g0 , I, J) be a compact generalized Kähler manifold
satisfying [I, J] = 0. Suppose
(1) rank T−1,0 = 1,
(2) c1 (T−1,0 ) ≤ 0.
Then the solution to generalized Kähler-Ricci flow with initial condition g0 exists
on [0, τ ∗ (g0 )).
Proof. Choose a time τ < τ ∗ (g0 ), and choose background partial volume
forms Ω± = det(ωh )± , a family of background metrics ω̂t , and ut a solution of the
reduced equation (9.14) by Proposition 9.40. Note that, by applying the maximum
principle to (9.14), it follows that there exists a uniform constant C such that
inf u ≥ inf u − tC.
M×{t} M×{0}
In particular, on the compact time interval [0, τ ] there is a uniform lower bound for
u.
Next we obtain a lower bound for the metric. The first step is to estimate the
metric on T− . A general computation for the generalized Kähler-Ricci flow in this
setting, closely related to Lemma 9.56, shows that the partial volume forms satisfy

∂ C det g± 2
(9.25) −∆ log = 21 |T | − trg ρC (h± ),
∂t det h±
where h± are background Hermitian metrics on T± . As c1 (T−1,0 ) ≤ 0, we can choose
h− such that ρC (h− ) ≤ 0. For such h− it follows from the maximum principle
applied to (9.25) that
det g− det g−
(9.26) inf ≥ inf .
M×{t} det h− M×{0} det h−
Since the rank of T−1,0 is 1, we can rephrase this estimate as

sup trg− h− ≤ sup trg− h− .
M×{t} M×{0}
∂u
Next we obtain a lower bound for ∂t .
Using (9.25) we compute

∂ C ∂u ∂ C det g+ det h−
−∆ = −∆ log − log
∂t ∂t ∂t det h+ det g−
(9.27)
= trg (ρC (h− ) − ρC (h+ ))
= − trg+ P (h) + trg− P (h).
Now let Φ1 = (τ − t) ∂u∂t + u. Before computing an evolution equation for Φ we

observe the basic fact that
√
∆C u = trg −1∂∂u
√ √
= trω+ −1∂+ ∂ + u + trω− −1∂− ∂ − u
(9.28)
= trω+ (ωt − ω̂t )+ + trω− (ω̂t − ωtu )−
= k − 1 − trg+ ĝ+ + trg− ĝ− ,
where k is the rank of T+ . Using this and (9.27) we obtain

∂ C ∂ C ∂u
−∆ Φ1 = (τ − t) −∆ − ∆C u
∂t ∂t ∂t
= trg+ (ĝ+ − (τ − t) P (h)) − trg− (ĝ− − (τ − t) P (h))
(9.29)
+k−1
= trg+ (ĝτ )+ − trg− (ĝτ )− + k − 1
≥ − C,
where the final inequality uses the uniform lower bound on g− of (9.26). It follows
from the maximum principle that there exists a uniform constant C such that
(9.30) inf Φ1 ≥ inf Φ1 − tC.
M×{t} M×{0}
Together with the uniform estimate for u, this yields a uniform lower bound for ∂u
∂t
on [0, τ − ǫ] for any ǫ > 0.
Finally we are ready to establish the lower bound for g. Fix a large constant
A to be determined, and let Φ2 = log trg h − Au. Using Lemma 9.57 and 9.28, we
estimate

∂ ∂
− ∆C Φ2 ≤ K trg h − A u + A k − 1 − trg+ ĝ+ + trg− ĝ−
∂t ∂t
≤ (C − A) trg+ ĝ+ + CA
≤ CA,
where the second line uses the equivalence of ĝ and h, the lower bound for ∂u
∂t , and
the lower bound for g− , and the final line follows by choosing A sufficiently large.
It follows from the maximum principle that
sup trg h ≤ sup trg h + Ct.
M×{t} M×{0}
Thus we have established a uniform lower bound for the metric on all time inter-
vals of the form [0, τ − ǫ]. Proposition 9.61 implies that the solution has uniform
estimates and hence exists on such intervals. Letting τ → τ ∗ (g0 ) and ǫ → 0, we
obtain that the solution exists on [0, τ ∗ (g0 )), as required.
The proof of Theorem 9.66 makes clear a key difficulty in establishing a full
confirmation of Conjecture 9.34, namely the fact that the resulting scalar PDE of
(9.14) is not convex, which makes directly applying the classical Pogorelov method
impossible. This is the reason for the topological hypotheses of Theorem 9.66,
which give some a priori partial control over the metric, after which the Pogorelov
method can be applied. We note that a global viscosity solution in this setting was
established in [164]. We next establish convergence on tori.
Corollary 9.67. Given (T 2n , g0 , I, J) a generalized Kähler structure on the
torus satisfying [I, J] = 0, the solution to generalized Kähler-Ricci flow with initial
condition g0 exists on [0, ∞), and converges exponentially to a flat metric. In
particular, if gE denotes
√ a flat metric compatible
with I and J, given any function
u0 such that ωE + −1 ∂+ ∂ + − ∂− ∂ − u > 0, the solution to
E
√ k E l
∂ ω+ + −1∂+ ∂ + u ∧ (ω− )
u = log √ l
,
(9.31) ∂t (ω E )k ∧ ω E − −1∂− ∂ − u
+ −
u(0) = u0 ,
exists on [0, ∞) and converges to u∞ = const.
Proof. Since the solution to generalized Kähler-Ricci flow in the I-fixed gauge
is the same as the solution to pluriclosed flow (cf. §9.3.1), the claims of long time
existence and convergence to a flat metric follow directly from Theorem 9.65. Using
that I|T± = ∓J|T± , it is easy to see that there exist flat metrics compatible with both
I and J. Choosing one such, gE , and using this as the choice of background metric
in Proposition 9.40, we see that equation (9.31) is the resulting scalar reduction of
generalized Kähler-Ricci flow, and so the claims of global existence and convergence
follow.
9.8.5. Generalized Kähler-Ricci flow: Nondegenerate case.
Corollary 9.68. Given (T 2n , g0 , I, J) a generalized Kähler structure on the
torus such that [I, J] is nondegenerate, the solution to generalized Kähler-Ricci
flow with initial condition g0 in the I-fixed gauge exists on [0, ∞), and converges
exponentially to a triple (g∞ , I, J∞ ) such that g∞ is flat and (I, J∞ ) are part of a
hyperKähler sphere of complex structures compatible with g∞ .
Proof. Since the solution to generalized Kähler-Ricci flow in the I-fixed gauge
is the same as the solution to pluriclosed flow (cf. §9.3.1), the claims of long time
existence and convergence to a flat Kähler metric follow directly from Theorem
9.65. By hypothesis, the initial Poisson tensor σ = 12 g −1 [I, J] is nondegenerate,
and we denote its inverse by Ω. By Proposition 9.29 we know that σ, and hence Ω,
remain fixed along the flow, and so we can express Ω = 2g∞ [I, J∞ ]−1 . This form
is of type (2, 0) − (0, 2) with respect to both I and J∞ , with the respective (2, 0)
parts each holomorphic with respect to the corresponding complex structure by
Proposition 7.69. By the Beauville-Bogomolov-Yau decomposition [13, 19, 186],
the data (g∞ , I, J∞ ) are part of a hyperKähler structure.
This corollary shows the power of generalized Kähler-Ricci flow to prove classifi-
cation results for problems in complex and symplectic geometry. On the one hand,
we have shown contractibility of the space of nondegenerate generalized Kähler
structures on tori to the space of structures (g, I, J) with g flat, and I and J part of a
compatible hyperKähler structure. Using the interpretation of generalized Kähler-
Ricci flow in this setting in terms of positive Ω-Hamiltonian diffeomorphisms in
Proposition 9.42, we have determined the global topology of this nonlinear space
as well. In principle one expects generalized Kähler-Ricci flow to be able to yield
further results on the structure of natural spaces arising in symplectic geometry.
Despite the long time existence and convergence results described above, the
global existence and convergence properties of pluriclosed flow and generalized
Kähler-Ricci flow are still largely open. The main existence Conjecture 9.18 has
only been established in special cases, the key missing point being to establish uni-
form parabolicity of the equation in the positive cone. At the conjectural singular
time τ ∗ (g0 ) we expect a wide variety of behavior including convergence, blowdown
of complex subvarities, collapse to a lower dimensional space, etc. In the setting of
complex surfaces, a conjectural framework for the convergence properties of pluri-
closed flow was described in [167, 170, 156]. In principle the limiting structure
of pluriclosed flow can be used to detect the complex structure of the underlying
complex manifold, resulting in a kind of geometrization conjecture for complex
surfaces.
CHAPTER 10
T-duality
In this chapter we present a method to relate solutions of the generalized Ricci

flow equation based on a natural symmetry of generalized geometry known as T-
duality. Quite remarkably, these pairs of T-dual solutions of the flow may live in
topologically distinct manifolds and typically have very different geometric features.
The idea has its origins in the string theory literature [32, 111, 147]. The main
result is the preservation of the generalized Ricci flow under T-duality, and builds on
a fundamental result by Cavalcanti and Gualtieri [39], which states that topological
T-duality for principal torus bundles–as defined by Bouwknegt, Evslin and Mathai
[24]–induces a Courant algebroid isomorphism. One consequence of our discussion
will be a mathematical explanation of the dilaton shift in string theory, as originally
observed by Buscher [32]. Furthermore, we illustrate examples of global existence
and convergence of generalized Ricci flows which are T-dual to solutions of classical
Ricci flow.
10.1. Topological T-duality

The goal of this section is to introduce some basic aspects of topological T-
duality for principal torus bundles, as defined in [24]. First, let us briefly summarize
the definition following [39]. Let T k be a k-dimensional torus acting freely and
properly on a smooth manifold M , so that M is a principal T k -bundle over the
smooth manifold B := M/T k . We denote by t the Lie algebra of T k . We endow M
with a choice of T k -invariant cohomology class
k
τ ∈ H 3 (M, R)T .
If B happens to be non-compact, we can replace H 3 (M, R) by the cohomology with
compact support on M . Consider the fiber product M = M ×B M̂ and the diagram
M❅
q ⑦⑦⑦ ❅❅❅ q̂
⑦ ❅❅
⑦⑦ ❅❅
~⑦
⑦
M❆ M̂
❆❆ ⑥⑥
❆❆ ⑥⑥
p ❆❆❆ ⑥⑥
~ ⑥ p̂
⑥
B
Definition 10.1. We say that two pairs (M, τ ) and (M̂ , τ̂ ) as above are T-dual
if there exist invariant representatives H ∈ τ and Ĥ ∈ τ̂ such that
(10.1) q̂ ∗ Ĥ − q ∗ H = dB,
221
222 10. T-DUALITY
where B ∈ Γ(Λ2 T ∗ M ) is a T k × T̂ k -invariant 2-form such that

B : Ker dq ⊗ Ker dq̂ → R
is non-degenerate.
Remark 10.2. In the literature, one typically considers B to be unimodular, in
the sense that it takes integer values for invariant period-1 elements of Γ(Ker dq) and
Γ(Ker dq̂). This condition is useful for interpreting the T-dual bundles as fibrations
of mutually dual tori, but we will not use it explicitly in our discussion.
We present next the most basic example of topological T-duality, given by the
case where the base of our torus bundles is a point.
Example 10.3. Consider the case where the base is a point B = {•} and k = 1,
so that M ∼
= S 1 . We regard M as the quotient
M = R/hαi,
where hαi denotes the additive subgroup generated by 0 6= α ∈ R. Consider the
dual torus M̂ = R∗ /hαi∗ , where
hαi∗ = {u : R → R | u(α) ∈ Z} = {α−1 u | u ∈ Z∗ }.
Identifying R ∼
= R∗ we regard
M̂ = R/hα−1 i.
For a choice of coordinate x on R, we can take B = dx ∧ dx̂ in (10.1), which realizes
M and M̂ (endowed with the trivial de Rham class) as topological T-dual circles.
Observe that, by construction, B is unimodular in the sense of Remark 10.2.
To obtain more interesting examples, which illustrate the topological nature of
Definition 10.1, we derive further structure of T-dual pairs. Given such a pair, we
can choose special representatives of the given de Rham classes which still satisfy
(10.1). To see this, consider the natural exact sequence
TM
(10.2) 0→B×t→ → T B → 0.
Tk
This induces a filtration
k
(10.3) Γ(Λ• T ∗ B) ∼
= F 0 ⊂ F 1 ⊂ · · · ⊂ F • = Γ(Λ• T ∗ M )T
where F i = Ann(∧i+1 t). The choice of a connection θ on p : M → B induces an
isomorphism T M/T k ∼
= T B ⊕ t and gives a splitting for the exact sequence (10.2),
inducing isomorphisms
m
M
k k
(10.4) Γ(Λm T ∗ M )T ∼= Γ(Λj T ∗ B ⊗ ∧m−j t∗ )T .
j=0
Remark 10.4. Throughout this section, θ will denote a connection on a prin-

cipal torus bundle. Notice that in previous chapters we reserved this notation for
the Lee form of a Hermitian structure, but as we will not use the Lee form in this
chapter this should not lead to any confusion.
Lemma 10.5. Let (M, τ ) and (M̂ , τ̂ ) be T-dual. Then, there exist representa-
tives H ∈ τ and Ĥ ∈ τ̂ in F 1 of M and M̂ , respectively, which satisfy the conditions
in Definition 10.1.
10.1. TOPOLOGICAL T-DUALITY 223
Proof. Using the filtrations (10.3) applied to q and q̂, we can write the two-
form B in Definition 10.1 as
B = q ∗ B + q̂ ∗ B̂ + B 0 ,
k k k
×T̂ k
where B ∈ Γ(Λ2 T ∗ M )T , B̂ ∈ Γ(Λ2 T ∗ M̂ )T̂ and B 0 ∈ Γ(Λ2 T ∗ M )T satisfies
B | Ker dq⊗Ker dq̂ = B 0 | Ker dq⊗Ker dq̂
and furthermore B 0 is in F 1 for both fibrations q and q̂. Hence, we have that B 0
is non-degenerate on Ker dq ⊗ Ker dq̂ and
dB 0 = q̂ ∗ (Ĥ + dB̂) − q ∗ (H − dB).
Using the last equation and the fact that B 0 is in F 1 for the fibration q, we leave
as an exercise to see that H − dB is in F 1 for the fibration p, and similarly for
Ĥ + dB̂.
The condition H ∈ F 1 in the previous result means that, for a choice of con-
nection θ on M , we can write (see (10.4))
(10.5) H = p∗ h − hĉ ∧ θi
for suitable ĉ ∈ Γ(Λ2 T ∗ B ⊗ t∗ ) satisfying dĉ = 0 and h ∈ Γ(Λ3 T ∗ B). Here, h, i
denotes the natural duality pairing between t and t∗ . With this observation in
hand, we can present a result about the existence of T-duals due to Bouwknegt,
Hannabuss, and Mathai [25].
Proposition 10.6. Given a principal T k -bundle M over a base manifold B
k
and a closed three-form H ∈ F 1 ∩Γ(Λ3 T ∗ M )T representing an integral cohomology
class [H] ∈ H 3 (M, Z), there exists a T-dual for (M, [H]).
Proof. Choose a connection θ on p : M → B so that H is given by (10.5). The
condition [H] ∈ H 3 (M, Z) implies that the closed t∗ -valued two-form ĉ represents
an integral cohomology class in H 2 (B, t∗ ). Consider the dual torus T̂ k with Lie
algebra t∗ and let M̂ be a principal torus bundle over B with Chern class [ĉ]. Let
θ̂ be a connection on p̂ : M̂ → B such that dθ̂ = ĉ, and define
Ĥ = p̂∗ h − hc ∧ θ̂i
where c = dθ. We leave as an exercise to check that dĤ = 0 and that (M, [H])
and (M̂ , [Ĥ]) are T-dual, taking B = −hθ ∧ θ̂i in (10.1).
We are ready to present two basic examples of topological T-duality, for a
positive dimensional base (cf. Example 10.3). The first illustrates the topology
change on the manifold M under T-duality.
Example 10.7. Consider the Hopf fibration p : M = S 3 → B = S 2 with H = 0.
Choose a connection θ such that dθ = ωS 2 is the standard symplectic structure on
S 2 (suitably normalized). We claim that (M, 0) is T-dual to (S 2 × S 1 , [Ĥ]), where
Ĥ = −θ̂ ∧ ωS 2 , where θ̂ is the line element on S 1 . To see this, take B = −q ∗ θ ∧ q̂ ∗ θ̂
and calculate
dB = −q ∗ dθ ∧ q̂ ∗ θ̂ = −q ∗ p∗ ωS 2 ∧ q̂ ∗ θ̂ = q̂ ∗ Ĥ.
The next example shows that T-duality is sensitive to the choice of cohomology
class of the three-form H, as is implicit in the proof of Proposition 10.6.
224 10. T-DUALITY
Example 10.8. Consider again the Hopf fibration in the previous example, but
regarding M as the compact Lie group M = SU (2). Consider generators for the
Lie algebra su(2) = he1 , e2 , e3 i as in Example 3.51 and the Cartan three-form
H = −e123 .
We claim that (M, [H]) is self T-dual. Notice that e1 generates the circle action on
the fibers of p : M → S 2 and hence e1 is a connection on M with de1 = e23 = p∗ ωS 2 .
Consider
B = −q ∗ e1 ∧ q̂ ∗ ê1
where ê1 is a left-invariant one-form dual to the generator of the circle in M̂ =
SU (2). Then,
dB = −q ∗ p∗ ωS 2 ∧ q̂ ∗ ê1 + q ∗ e1 ∧ q̂ ∗ p̂∗ ωS 2 = p̂∗ Ĥ − p∗ H,
and hence B satisfies the conditions in 10.1. More generally, following the pre-
vious argument one can prove that any semi-simple Lie group equipped with the
cohomology class of its Cartan 3-form is self T-dual (see [39, Example 2.4]).
10.2. T-duality and Courant algebroids

Having described the most basic features about T-duality in the previous sec-
tion, we now recall the relevant implications of the T-duality relation for the theory
of generalized geometry [39]. For this, let us first interpret the data in Definition
10.1 in the language of Courant algebroids.
Definition 10.9. An equivariant Courant algebroid is given by a triple (E, M, T k ),
where T k is a k-dimensional torus acting freely and properly on a smooth mani-
fold M , so that M is a principal T k -bundle over a base M/T k = B, and E is an
exact Courant algebroid over M with a T k -action, lifting the T k -action on M and
preserving the Courant algebroid structure.
Effectively, given an equivariant Courant algebroid (E, M, T k ), since T k is com-
pact we can choose an equivariant isotropic splitting σ : T → E and hence the pre-
vious definition reduces to having the twisted Courant algebroid (T ⊕ T ∗ , h, i , [, ]H )
k
over M with a T k -action on T ⊕T ∗ and a T k -invariant three-form H ∈ Γ(Λ3 T ∗ M )T .
Thus, any triple (E, M, T k ) determines a T k -invariant Ševera class
k
τ = [E] ∈ H 3 (M, R)T ,
leading to a pair (M, τ ) as in Definition 10.1. Conversely, given a pair (M, τ ) as in
Definition 10.1 we can consider the equivariant Courant algebroid (T ⊕T ∗, h, i , [, ]H )
over M , endowed with the canonical T k -action on T ⊕T ∗, for a choice of T k -invariant
element H ∈ τ .
Our goal here is to explain a theorem due to Cavalcanti and Gualtieri [39],
which states that for any T-dual pairs one can associate an isomorphism of equi-
variant Courant algebroids, upon reduction to the base B. The preliminaries for
this result are contained in §7.2, where the Dorfman bracket on E is expressed
as a derived bracket using the twisted de Rham differential (7.5). In the situa-
tion of our interest, assume that (T ⊕ T ∗ , h, i , [, ]H ) is equivariant over a T k -bundle
p : M → B. Consider the twisted de Rham complex of T k -invariant differential
k
forms (Γ(S)T , dH ) given by the Z2 -graded vector space
k k k
Γ(S)T = Γ(Λeven T ∗ M )T ⊕ Γ(Λodd T ∗ M )T
10.2. T-DUALITY AND COURANT ALGEBROIDS 225
with differential dH in (7.5). The following result is due to Bouwknegt, Evslin, and
Mathai [24].
Theorem 10.10 ([24]). Let (M, τ ) and (M̂ , τ̂ ) be T-dual. Let H ∈ τ and
Ĥ ∈ τ̂ be representatives as in Lemma 10.5 with q̂ ∗ Ĥ − q ∗ H = dB. Then there is
an isomorphism of differential complexes
Z
k k
τ : (Γ(S)T , dH ) → (Γ(Ŝ)T , dĤ ), τ (ρ) = eB ∧ q ∗ ρ,
Tk
where the integration is along the fibers of q̂ : M → M̂ .

Proof. First, let us make more explicit the definition of τ . Notice that eB ∧q ∗ ρ
is linear in ρ, and hence it suffices to define the fiber-wise integral above for a degree-
k
m element ρ ∈ Γ(Λm T ∗ M )T . For this choice of ρ, consider a connection θ on M
and using (10.4) we decompose
m
X m
M k
ρ= ρj ∈ Γ(Λj T ∗ M ⊗ ∧m−j t∗ )T .
j=0 j=0
Then, τ is defined taking the component of eB ∧ q ∗ ρ which has degree k along the
fiber of p̂ and integrating, that is,
Xm Z
1 k−m+j
τ (ρ) = B ∧ q ∗ ρj .
j=0
(k − m + j)! T k
To write this expression, we have used that H and Ĥ are as in Lemma 10.5, and
therefore B belongs to F 1 in the filtration (10.3) for p : M → B. We prove next
that τ is a morphism of differential complexes:
Z
dĤ (τ (ρ)) = dĤ (eB ∧ ρ)
Tk
Z
= (Ĥ − H) ∧ eB ∧ ρ + eB ∧ dρ − Ĥ ∧ eB ∧ ρ
k
ZT
= eB ∧ dρ − H ∧ eB ∧ ρ
Tk
= τ (dH ρ)
where, for simplicity, we have omitted pull-backs in the formula above. It remains
to show that τ so defined is an isomorphism. For this, one can work on a local
coordinate patch U on the base manifold B, so that M |U = U × T k × T̂ k , and take
a global frame for T ∗ T k and T ∗ T̂ k . We leave the details as an exercise.
We are ready for the main result of this section, which states that T-duality
induces an isomorphism of equivariant Courant algebroids, upon reduction to the
base. Let (E, M, T k ) be an equivariant Courant algebroid with base B. Consider
the vector bundle
E/T k → B,
whose sheaf of sections is given by the invariant sections of E, that is, Γ(E/T k ) =
k
Γ(E)T . We can endow E/T k with a natural Courant algebroid (E/T k , h, i , [, ], πE/T k )
as in Definition 2.8, with anchor πE/T k : E/T k → T B and pairing and Dorf-
man bracket given by the restriction of the neutral pairing and the Dorfman
226 10. T-DUALITY
k
bracket on E to Γ(E)T . Observe that E/T k has the same rank as E, and hence
E/T k is not an exact Courant algebroid over B (see Definition 2.9). We will call
(E/T k , h, i , [, ], πE/T k ) the simple reduction of E by T k .
Theorem 10.11 ([39]). Let (E, M, T k ) and (Ê, M̂ , T̂ k ) be equivariant exact

Courant algebroids over the same base M/T k = B = M̂ /T̂ k . Assume that (M, [E])
is T-dual to (M̂ , [Ê]). Then, there exists a Courant algebroid isomorphism
ψ : E/T k → Ê/T̂ k
which exchanges the twisted differentials in Lemma 7.37 and which is compatible
with Clifford multiplication, that is,
τ (d /ˆ0 (τ ρ),
/0 ρ) = d and τ (e · ρ) = ψ(e) · τ (ρ)
k k
for all e ∈ Γ(E)T and ρ ∈ Γ(S)T .
Proof. We choose equivariant isotropic splittings of E and Ê, and B as in

k
Definition 10.1. Given X + ξ ∈ Γ(T M ⊕ T ∗ M )T , choose the unique lift X of X to
k k
Γ(T M)T ×T̂ such that
(10.6) q ∗ ξ(Y ) − B(X, Y ) = 0 for all Y ∈ t.
Due to this condition the form q ∗ ξ − B(X, ·) is basic for the bundle determined by
q̂, and can therefore be pushed forward to M̂ . We define a map
ψ(X + ξ) = q̂∗ (X + q ∗ ξ − B(X, ·)).
Let us see next that ψ is independent of the choice of isotropic splittings. For a
different choice of isotropic splittings we have
H ′ = H + dB, Ĥ ′ = Ĥ + dB̂,
and hence
′
q̂ ∗ Ĥ ′ − q ∗ H ′ = d(B + q̂ ∗ B̂ − q ∗ B) = dB
′
for B = B − q ∗ B + q̂ ∗ B̂. Moreover, the action of q ∗ B and q̂ ∗ B̂ on t ⊗ t̂ is trivial.
′
Hence when lifting vectors to the configuration space, using either B or B yields
the same result, as claimed.
We check next that ψ preserves the Courant structure. First, ψ is an isometry
by
hψ(X + ξ), ψ(X + ξ)i = q̂∗ (X + q ∗ ξ − B(X, ·)), q̂∗ (X + q ∗ ξ − B(X, ·))
= q̂∗ (q ∗ ξ − B(X, ·))(q̂∗ (X))
= (q ∗ ξ − B(X, ·))(X)
= ξ(X).
To see that ψ preserves the Dorfman bracket, we prove first that it is compatible
with Clifford multiplication
τ (e · ρ) = ψ(e) · τ (ρ).
10.3. GEOMETRIC T-DUALITY 227
k
For e = X + ξ ∈ Γ(T M ⊕ T ∗ M )T , we have
Z
ψ(e) · τ (ρ) = ψ(e) · eB ∧ ρ
T k
Z
= (X + q ∗ ξ − B(X, ·)) · (eB ∧ ρ)
Tk
Z
= iX B ∧ eB ∧ ρ + eB ∧ iq∗ X ρ + (q ∗ ξ − iX B) ∧ eB ∧ ρ
Tk
Z
= eB ∧ (iX ρ + ξ ∧ ρ)
Tk
= τ (e · ρ).
k
Applying now Theorem 10.10 and Lemma 7.37, for any a, b ∈ Γ(T M ⊕ T ∗ M )T ,
we have
ψ([a, b]) · τ (ρ) = τ ([v1 , v2 ] · ρ)
/0 , a·], b·](ρ))
= τ ([[d
/̂0 , ψ(a)·], ψ(b)·](τ (ρ))
= [[d
= [ψ(a), ψ(b)] · τ (ρ).
The fact that ψ preserves the anchor map is left as an exercise.
Remark 10.12. Given T-dual pairs (M, [E]) and (M̂ , [Ê]), the isomorphism ψ
in Theorem 10.11 is not unique, as it depends on the choice of two-form B as in
Definition 10.1. For instance, there is the freedom to modify B by adding a closed
two-form on M ×B M̂ . The change of ψ can be seen explicitly in Example 10.31 by
introducing an extra parameter multiplying the second summand in the definition
of B.
10.3. Geometric T-duality

In this section we explain the fundamental interplay between T-duality, the
generalized Ricci tensor, and the generalized Ricci flow. For this, we need to in-
troduce a notion of duality for pairs (G, div), given by a generalized metric and
a divergence operator. Given an equivariant Courant algebroid (E, M, T k ) and a
T k -invariant generalized metric G on E, we can push-forward V+ along the bundle
projection p : M → B
p∗ V+ ⊂ E/T k
to obtain a generalized metric p∗ G on the simple reduction. This way we obtain
an identification between T k -invariant generalized metrics on E and generalized
metrics on E/T k . Similarly, given a T k -invariant divergence operator div : Γ(E) →
C ∞ (M ) it induces a first-order differential operator
k k
p∗ div : Γ(E)T = Γ(E/T k ) → C ∞ (M )T ∼= C ∞ (B).
The πE -Leibniz rule for div restricted to p∗ C ∞ (B) (see (2.26)) is precisely the
πE/T k -Leibniz rule for p∗ div, which thus defines a divergence operator on E/T k .
Definition 10.13. Let (E, M, T k ) and (Ê, M̂ , T̂ k ) be equivariant exact Courant

algebroids over the same base M/T k = B = M̂ /T̂ k . Assume that (M, [E]) is T-dual
to (M̂ , [Ê]) and fix an isomorphism ψ : E/T k → Ê/T̂ k as in Theorem 10.11. We
228 10. T-DUALITY
ˆ on Ê is the T-dual via ψ of a T k -invariant

will say that a T̂ k -invariant pair (Ĝ, div)
pair (G, div) on E if
ψ ◦ p∗ G = p̂∗ Ĝ ◦ ψ and ˆ
p∗ div = ψ ∗ p̂∗ div,
where p : M → B and p̂ : M̂ → B are the bundle projections.
The notion of T-dual pairs depends upon the choice of isomorphism ψ, which is
by no means unique (see Remark 10.12). Nonetheless, we will often abuse notation
ˆ is the T-dual of (G, div), assuming that some background
and say that (Ĝ, div)
isomorphism has been chosen.
In order to prove that the generalized Ricci tensor in Definition 3.31 is nat-
urally exchanged under T-duality, we need to understand the relation between
T k -invariant generalized connections on E and generalized connections on E/T k
Tk
(see §3.1). Denote by DE the space of generalized connections on E, by DE ⊂ DE
k
the space of T -invariant generalized connections, and by DE/T k the space of gen-
eralized connections on the simple reduction E/T k .
k
T
Lemma 10.14. There is a canonical identification DE = DE/T k .
Proof. The space DE/T k is non-empty, since we can construct a generalized
k
connection Ď out of an orthogonal connection ∇E/T on E/T k by
E/T k
Ďa b = ∇π (a) b.
E/T k
k
T
Similarly, the space DE is non-empty, since T k is compact and hence we can
k
T Tk
construct D ∈ DE by averaging an arbitrary element in DE . The affine spaces DE
k
and DE/T k are both modeled on Γ(E ∗ ⊗ o(E))T , via the identification Γ(E/T k ) =
k
Γ(E)T , and therefore it is enough to prove that a generalized connection on E/T k
induces an invariant connection on E. For this, we note that E is locally generated
k
by Γ(E)T , and therefore given Ď ∈ DE/T k we can define a generalized connection
k
D on E by extending its action from Γ(E)T to Γ(E) imposing the Leibniz rule
with respect to the anchor πE . The generalized connection D is compatible with
the pairing, as for f ∈ C ∞ (M ) and a, b, c ∈ Γ(E)T we have
πE (a)hf b, ci = df (πE (a)) hb, ci + f πE/T k (a)(hb, ci)
= hdf (πE (a)) + f Ďa b, ci + f hb, Ďa ci
= hDa (f b), ci + hb, Da ci.
For the first equality we have used that hb, ci is the pull-back of a function on B
and consequently
πE (a) hb, ci = πE/T k (a) hb, ci .

The next result shows that the generalized Ricci tensor in Definition 3.31 is
naturally exchanged under T-duality.
Proposition 10.15. Let (E, M, T k ) and (Ê, M̂ , T̂ k ) be equivariant exact Courant
10.3. GEOMETRIC T-DUALITY 229
ˆ on Ê be the T-dual of an invariant pair (G, div) on E.

to (M̂ , [Ê]) and let (Ĝ, div)
Then
(10.7) ˆ
p∗ Rc(G, div) ◦ ψ = ψ ◦ p̂∗ Rc(Ĝ, div),
where Rc is as in Definition 3.31. In particular, generalized Einstein pairs are
exchanged under T-duality (see Definition 3.52).
Proof. By definition of the generalized Ricci tensor (see Definition 3.31) it
suffices to prove that
ˆ
p∗ Rc± (G, div) = ψ ∗ p̂∗ Rc± (Ĝ, div),
where Rc± are as in Definition 3.28. Given a T k -invariant pair (G, div), let D ∈
Tk
D0 (G, div) (see Lemma 3.17). Averaging over T k if necessary, we obtain D ∈ DE .
Using the isomorphism ψ in Theorem 10.11 we have that ψ∗ p∗ D is a generalized
connection on Ê/T̂ k , which induces a generalized connection D̂ on Ê by Lemma
ˆ are T-dual, and hence
10.14. By hypothesis (G, div) and (Ĝ, div)
p̂∗ D̂Ĝ = (ψp∗ Dψ −1 )(ψp∗ Gψ −1 ) = ψp∗ DGψ −1 = 0,

and also
ˆ
p̂∗ divD̂ (a) = tr D̂a = tr ψp∗ Dψ −1 (a) = tr D(ψ −1 a) = (ψ −1 )∗ div(a) = p̂∗ div(a)
ˆ
for any a ∈ Γ(Ê/T̂ k ). Thus, it follows that D̂ ∈ D0 (Ĝ, div), which implies (see
Definition 3.28)
p∗ Rc± (G, div) = p∗ Rc± ∗ ± ∗ ± ˆ

D = ψ p̂∗ RcD̂ = ψ p̂∗ Rc (Ĝ, div).
Remark 10.16. A proof of the fact that generalized Ricci flat metrics are ex-
changed under T-duality has been obtained by several authors and with different
methods [12, 70, 162], some including an extension to more general Courant al-
gebroids. A more general result, for Poisson-Lie T-duality, in the sense of Klimčik
and Ševera [112], was proved in [151].
Our next goal is to understand the behavior of the scalar curvatures S ± in
(3.37) under T-duality. The reason why we do not consider here the generalized
scalar curvature S in Definition 3.41 is that, in principle, this would require showing
that T-duality preserves the notion of closed pair in Definition 3.40. This turns out
to be a delicate question, related to a phenomenon known in the literature as dilaton
shift, which will be discussed in the next section.

ˆ on Ê be the T-dual of an invariant pair (G, div) on E.
Then
(10.8) ˆ
S ± (G, div) = S ± (Ĝ, div),
where S ± is as in equation (3.37). In particular, solutions of the equation S ± = 0
are exchanged under T-duality.
230 10. T-DUALITY
Proof. Given a point on the base manifold B, we can consider a local spinor
bundle S± for the bundle V± /T k . Via the duality isomorphism ψ in Theorem 10.11
k
we can identify S± with a local spinor bundle for V̂± /T̂ k . Let D ∈ D0 (G, div)T
k
ˆ T̂ the induced generalized connection on Ê via ψ, as
and consider D̂ ∈ D0 (Ĝ, div)
in the proof of Proposition 10.15. Then, (10.8) follows by Definition 3.37, using D
and D̂ to evaluate the scalar curvatures S ± and applying Lemma 3.11 and Lemma
3.36.
Remark 10.18. A proof of the fact that solutions of the equations
Rc+ = 0, S+ = 0
are exchanged under T-duality was first given in [110, 152] for the case of exact
ˆ and
pairs. Notice that when (G, div) is a compatible pair, so is the T-dual (Ĝ, div),
in this case the equation Rc+ = 0 is equivalent to Rc = 0 (see Lemma 3.32).
We state next our main result of this section, which follows as a consequence
of the proof of Proposition 10.15.
Theorem 10.19. Let (E, M, T k ) and (Ê, M̂ , T̂ k ) be equivariant exact Courant

to (M̂ , [Ê]). Then, if (Gt , divt ) is a T k -invariant solution of the flow
∂
(10.9) G −1 G = −2Rc(Gt , divt )
∂t
ˆ t ) is the family of T̂ k -invariant T-dual pairs, then (Ĝt , div
and (Ĝt , div ˆ t ) is also a
solution of (10.9). In particular, solutions of gauge-fixed generalized Ricci flow,
and generalized Einstein pairs, are exchanged under T-duality.
Proof. The proof is a direct consequence of Definition 10.13 and equation

(10.7).
Remark 10.20. Preservation of the flow (10.9) under T-duality, as defined in

the present work, was proved in [70] following [162]. A more general result for
Poisson-Lie T-duality was proved in [151].
It is important to notice that, in the strictest sense, the generalized Ricci flow
(4.1) is not necessarily preserved under T-duality, since it is formulated in terms of
the generalized Ricci tensor for the special class of pairs (G, divG ). This is due to
the fact that the Riemannian divergence divG of a generalized metric is defined in
terms of the anchor map πE : E → T M (see Definition 2.46), which is not preserved
under T-duality. Instead, a solution of (4.1) is typically exchanged with a solution
of the gauge-fixed generalized Ricci flow, corresponding to an exact pair (G, div) (see
Definition 3.48). This is again related to the dilaton shift, which will be discussed
in the next section.
Remark 10.21. As observed in [39], pluriclosed metrics and generalized Kähler

structures are preserved under T-duality. Relying on Theorem 10.19, we expect that
flow lines for pluriclosed flow and generalized Kähler Ricci flow are also preserved
by the T-duality isomorphism.
10.4. BUSCHER RULES AND THE DILATON SHIFT 231
10.4. Buscher rules and the dilaton shift

In the present section we describe more explicitly the T-dual of a pair (G, div),
as considered in Definition 10.13. The specific local formulae for calculating the
T-dual are the famous “Buscher rules”, discovered in [32], which we will express
here in terms of global tensors. Our discussion will illustrate further the simplicity
of Definition 10.13, and the value of adopting the viewpoint of Courant algebroids.
For the sake of clarity, we will assume that the torus T k is one-dimensional.
The discussion can be easily generalized to the case of higher dimensional tori. We
set the notation T 1 = S 1 . Let (E, M, S 1 ) and (Ê, M̂ , Ŝ 1 ) be equivariant exact
Courant algebroids over the same base M/S 1 = B = M̂ /Ŝ 1 . Assume that (M, [E])
is T-dual to (M̂ , [Ê]). Let H ∈ [E] and Ĥ ∈ [Ê] be representatives as in Lemma
10.5 with q̂ ∗ Ĥ − q ∗ H = dB. Then, we can choose equivariant isotropic splittings
σ : T M → E, σ : T M̂ → Ê
such that E is given by the twisted Courant algebroid (T M ⊕ T ∗ M, h, i , [, ]H ) over
M , and similarly for Ê. Recall from Proposition 2.38 that an S 1 -invariant gener-
alized metric G on E is determined by an S 1 -invariant pair (g, b) of metric g and
two-form potential b on M . Upon a choice of connection θ on p : M → B, the pair
(g, b) can be expressed as
g = g0 θ ⊙ θ + g1 ⊙ θ + g2
(10.10)
b = b1 ∧ θ + b2
where bi are basic forms of degree i and gi are symmetric tensors of degree i. By
Lemma 10.5, we can assume that
B = −q ∗ θ ∧ q̂ ∗ θ̂
for a suitable connection θ̂ on p̂ : M̂ → B.
With this explicit decomposition of the data (g, b) in hand we can now give
the Buscher rules, which determine their relationship under T-duality. We give two
versions of this relationship, first expressed as an explicit formula for the associated
pair (ĝ, b̂), then expressed as a mapping of the invariant decomposition of (g, b).
Lemma 10.22. Given G an S 1 -invariant generalized metric on (T M ⊕T ∗M, h, i , [, ]H ),
consider the T-dual generalized metric Ĝ on (T M̂ ⊕ T ∗ M̂ , h, i , [, ]Ĥ ), as in Defini-
tion 10.13. Then, if the pair (g, b) associated to G is given by (10.10), then the pair
(ĝ, b̂) determined by Ĝ takes the form
1 b1 b1 ⊙ b1 − g1 ⊙ g1
ĝ = θ̂ ⊙ θ̂ − ⊙ θ̂ + g2 +
g0 g0 g0
(10.11)
g1 g1 ∧ b1
b̂ = − ∧ θ̂ + b2 + .
g0 g0
Proof. We denote by eθ the canonical vector field of θ, given by the unique
1
S 1 -invariant vertical vector field such that θ(eθ ) = 1. Let X +ξ ∈ Γ(T M ⊕T ∗ M )S .
Then, we can write
X = θ ⊥ Y + fX eθ ,
ξ = p∗ ζ + fξ θ,
232 10. T-DUALITY
where Y + ζ ∈ Γ(T B ⊕ T ∗ B), and θ⊥ Y denotes the horizontal lift with respect to
θ, and fX , fξ ∈ C ∞ (B). Going back to the definition of ψ in the proof of Theorem
10.11, the vector X on M satisfying (10.6) is given by
X = θ̂⊥ X + fξ eθ̂ ,
and therefore
ˆ
ψ(X + ξ) = θ̂⊥ Y + fξ eθ̂ + p̂∗ ζ + fX θ̂ = X̂ + ξ.
Using now that
eb (X + g(X)) = θ⊥ Y + fX eθ + (fX g1 + iY g2 − fX b1 + iY b2 )
+ (fX g0 + g1 (Y ) + b1 (Y ))θ,
it is not difficult to see that
ψ(eb (X + g(X))) = eb̂ (X̂ + ĝ(X̂)),
where (ĝ, b̂) are as in (10.11). The explicit calculation is left as an exercise.
For the calculations to come later, it will be useful to give the T-duality rela-
tionship explicitly in terms of the canonical decomposition of an S 1 -invariant pair
(g, b) on a principal bundle which we recall next.
Lemma 10.23. Given M a principal S 1 bundle over B with canonical vector
field eθ , an S 1 -invariant metric on M is uniquely determined by a metric h on B, a
smooth function φ ∈ C ∞ (B), and a principal connection θ. More precisely, g may
be expressed
g = φθ ⊙ θ + p∗ h,
where
g(eθ , ·)
φ = g(eθ , eθ ), θ= , h(·, ·) = g(π θ ·, π θ ·),
g(eθ , eθ )
and π θ is the horizontal projection determined by θ, i.e.
π θ (X) = X − θ(X)eθ .
Also, an S 1 -invariant two-form b is uniquely determined by basic forms η, µ. More
precisely, b may be expressed
b = θ ∧ η + µ,
where
η = eθ yb, µ = b − θ ∧ η.
Proof. Exercise.
The following result provides a version of the T-duality relationship which is

adapted to the canonical connections determined by the S 1 -invariant metrics.
Proposition 10.24. Given G an S 1 -invariant generalized metric on (T M ⊕
T ∗ M, h, i , [, ]H ), consider the T-dual generalized metric Ĝ on (T M̂ ⊕T ∗ M̂ , h, i , [, ]Ĥ ),
as in Definition 10.13. Let (g, b) and (ĝ, b̂) be the pairs associated to G and Ĝ,
10.4. BUSCHER RULES AND THE DILATON SHIFT 233
respectively. Let θg , φg , hg , ηg , µg denote the decomposition determined by (g, b) via

Lemma 10.23, with θĝ , etc. associated to (ĝ, b̂). Then
1
φĝ =
φg
θĝ = θ̂ + ηg
hĝ = hg
ηĝ = θg − θ
µĝ = µg − ηg ∧ ηĝ .
Proof. First we compute
g1
θg = θ + .
g0
Then we obtain
ηg = eθ yb = − b1 .
Then we may express
µg = b − θg ∧ ηg

g1
= b1 ∧ θ − θ + ∧ (−b1 ) + b2
g0
g1
= b2 + ∧ b1 .
g0
Furthermore we obtain
θ̂ĝ = θ̂ − b1 = θ̂ + ηg
Then, according to the Buscher rules,
g1
ηĝ = eθ̂ yb̂ = = θg − θ.
g0
Then we obtain
µĝ = b̂ − θ̂ĝ ∧ ηĝ
g1 g1 ∧ b1
= − ∧ θ̂ + b2 + − θ̂ + ηg ∧ (θg − θ)
g0 g0
g1 g1
= (θ̂ − b1 ) ∧ + b2 − θ̂ĝ ∧ ( )
g0 g0
= b2
g1
= µg − ∧ b1
g0
= µg − ηg ∧ ηĝ .

Having described explicitly the change of a generalized metric in terms of the
Buscher rules in Lemma 10.22 and Proposition 10.24, we turn next to the interplay
between divergences and T-duality. Our discussion provides a theoretical framework
to understand the so-called dilaton shift in string theory. To provide a more invari-
ant description, we shall work in the generality of Definition 10.1 and Definition
10.13, for arbitrary torus bundles. Let (E, M, T k ) and (Ê, M̂ , T̂ k ) be equivariant
234 10. T-DUALITY
exact Courant algebroids over the same base M/T k = B = M̂ /T̂ k . Assume that
ˆ
(M, [E]) is T-dual to (M̂ , [Ê]) and consider T-dual pairs (G, div) and (Ĝ, div). As
in Proposition 10.24, the T k -invariant metric g on M induced by G is equivalent
to a triple (gV , θg , hg ), where hg is a metric on B, θ is a principal connection on
M → B, and gV is a family of metrics on the fibers of M , that is,
g = gV (θg , θg ) + p∗ hg .
Using the Riemannian divergence divG of G (see Definition 2.46), we can express
div(a) = divG (a) − he, ai
k
for a suitable e ∈ Γ(E)T . We denote by ν ∈ Γ(| det V M ∗ |) the density induced by
gV on the vertical bundle V M ⊂ T M . Identifying
VM ∼ = M × t,
we regard ν = | det gV | as an equivariant map ν : M → | det t∗ | and, using that T k
is unimodular, we obtain a well-defined map
ν : B → | det t∗ |.
Observe that the following expression defines a well-defined one-form on the base
manifold
d log ν := ν −1 dν ∈ Γ(T ∗ B).
Similarly, we denote by (ê, ĝ, hĝ , θĝ , ĝV , ν̂) the corresponding data associated to
ˆ
(Ĝ, div). By Proposition 10.24, we have an equality h = hg = hĝ for the induced
metric on the base manifold B.
Proposition 10.25 (Dilaton shift). Let (E, M, T k ) and (Ê, M̂ , T̂ k ) be equi-
variant exact Courant algebroids over the same base M/T k = B = M̂ /T̂ k . Assume
ˆ
that (M, [E]) is T-dual to (M̂ , [Ê]) and consider T-dual pairs (G, div) and (Ĝ, div).
Then,
(10.12) ê = ψ(e) + 2d log(ν̂/ν),
where d log(ν̂/ν) := d log ν̂ − d log ν is regarded as a section of Ê/T̂ k using the
anchor map.
Proof. We denote π = πE/T k . The statement follows from the formula
Lπ(a) µh
div(a) = + 2hν −1 dν, ai − he, ai,
µh
k
for any a ∈ Γ(E)T , since the metric h on B induced by g and ĝ coincide and ψ
is the identity along the kernel of the anchor map π. To prove this formula, we
calculate
LπE (a) µg Lπ(e′ ) µh
= + ν −1 dν(π(e′ )),
µg µh
using the natural decomposition µg = µh ⊗ ν.
Note that if e = ϕ and ê = ϕ̂ for invariant 1-forms ϕ on M and ϕ̂ on M̂ ,
respectively, then (10.12) is equivalent to
!
| det gV |
(10.13) ϕ̂ = ϕ − 2d log = ϕ − 4d log | det gV |.
| det ĝV |
10.5. EINSTEIN-HILBERT ACTION 235
where we have used Proposition 10.24 for

| det ĝV | = | det gV |−1 .
Assuming that ϕ and ϕ̂ are exact, with the normalization ϕ = 8dφ and ϕ̂ = 8dφ̂,
we obtain
1
φ̂ = φ − log | det gV |.
2
This equation was encountered by physicists in their computations of the dual Rie-
mannian metric and dilaton field for T-dual sigma-models [32] and carries the name
of dilaton shift. Formula (10.12) was first obtained in [70], and later generalized to
the context of Poisson-Lie T-duality in [110]. It would be interesting to compare
our formula (10.12) with [56, Eq. (3.16)], in the context of non-abelian duality in
physics.
10.5. Einstein-Hilbert action

In this section we derive the preservation of the Einstein-Hilbert action 3.49
under T-duality. The main point is to complete our analysis of the interplay between
T-duality and the generalized scalar curvature initiated in Proposition 10.17. We
will use in an essential way the dilaton shift formula (10.12).
ˆ on Ê be the T-dual of an invariant pair (G, div) on E,
in the sense of Definition 10.13. Then, (G, div) is exact in the sense of Definition
ˆ is exact. Furthermore, if this is the case, then
3.48 if and only if (Ĝ, div)
ˆ
S(G, div) = S(Ĝ, div),
where S is the generalized scalar curvature in Definition 3.37. In particular, gen-
eralized scalar flat metrics with exact divergence are exchanged under T-duality.
Proof. If (G, div) is exact, e = dφ for φ ∈ C ∞ (M ). But then φ must be
k
T -invariant and consequently e is basic. By the dilaton shift formula (10.12)
ê = ψ(e) = dφ.
ˆ is exact. The last part of the
Therefore, (G, div) is exact if and only if (Ĝ, div)
statement follows from Proposition 10.17.
Remark 10.27. In general, the T-dual of a closed pair, in the sense of Definition
3.40, may not be closed. A counterexample is given in Example 10.31. Nonetheless,
ˆ
if (G, div) is closed then it must be a compatible pair, and hence the T-dual (Ĝ, div)
is also compatible. This suggests an extension of the notion of generalized scalar
curvature in Definition 3.41 to the case of compatible pairs.
Observe that by combining Proposition 10.26, Proposition 10.15, and Theorem
3.50, it follows that critical points of the generalized Einstein-Hilbert functional
(3.41), given by the solutions of (3.43), are exchanged under T-duality. This in-
dicates that the generalized Einstein-Hilbert functional is itself preserved by T-
duality, and this is the content of our last result. Notice from §3.7 that the diver-
gence of a T k -invariant exact pair (G, div) on M is of the form
div = divµ
for some T k -invariant volume form µ on M .
236 10. T-DUALITY
Theorem 10.28. Let (E, M, T k ) and (Ê, M̂ , T̂ k ) be equivariant exact Courant

to (M̂ , [Ê]) and let (G, divµ ) and (Ĝ, divµ̂ ) be exact T-dual invariant pairs on E and
Ê, respectively, such that
Z Z
µ= µ̂.
M M̂
Then,
(10.14) EH(G, divµ ) = EH(Ĝ, divµ̂ ).
Proof. By hypothesis there exist smooth functions f, fˆ ∈ C ∞ (B) such that

ˆ
µ = e−f µg , µ̂ = e−f µĝ .
Using the decompositions µg = µhg ⊗ ν and µĝ = µhĝ ⊗ ν̂ as in the proof of
Proposition 10.25, it follows from (10.12) that
(10.15) dfˆ = df + d log(ν̂/ν).
Define functions t, t̂ ∈ C ∞ (B) by integration along the fibers of p and p̂, that is,
Z Z
t(b) = ν, t̂(b) = ν̂.
p−1 (b) p̂−1 (b)
Then, (10.15) now reads

ˆ
e−f tµh = Ce−f t̂µh
and hence
Z Z Z Z Z
ˆ ˆ
µ= e−f tµh = C e−f t̂µh = C e−f t̂µh = C µ̂.
M B B B M̂
R R
The normalization M µ= M̂
µ̂ implies C = 1, and applying Proposition 10.26 we
conclude
Z Z
µ ˆ ˆ
EH(G, div) = S(G, div )e −f
tµ =h
S(Ĝ, divµ̂ )e−f t̂µh = EH(Ĝ, div),
B B
as claimed.
The equality (10.14) gives geometric content to the dilaton shift equation
(10.12), and provides a conceptual explanation of why critical points of the gener-
alized Einstein-Hilbert functional (3.41), given by generalized Ricci flat and scalar
flat metrics with exact divergence (3.43), are exchanged under T-duality. From
the point of view of the analysis in Chapter 6, Theorem 10.28 implies that the
energy functional and the invariant of generalized metrics λ(G) in Definition 6.1
are compatible with topological T-duality.
Remark 10.29. In the context of type II string theory (see Remark 3.43),
Theorem 10.28 shows that the low-energy string effective action functional (3.42) is
invariant under T-duality, as observed originally in [15]. In this physics setup, T-
duality is regarded as a symmetry of the Σ-model action (2.25), while the functional
(3.42) is obtained from (2.25) via a complicated Feynman diagram expansion. Thus,
Theorem 10.28 can be regarded as a ‘quantum effect’ of T-duality.
10.6. EXAMPLES 237
10.6. Examples
We start with examples of T-dual generalized Ricci flat and scalar flat metrics.
The first example is classical, and shows that T-duality recovers ‘topological mirror
symmetry’ for semi-flat Calabi-Yau metrics.
Example 10.30. Consider the non-compact manifold M = D × T k , where
D ⊂ Rk is the unit disk. We wish to endow M with a T k -invariant Ricci flat
metric and apply T-duality to it. For this, we take coordinates (x1 , . . . xk ) on D
and (y 1 , · · · , y k ) on T k , so that {dy 1 , . . . , dy k } is a global frame for T ∗ T k . By a
theorem by Cheng and Yau [46], given a positive constant C there exists a unique
solution
φ: D → R
to the real Monge-Ampère equation
!
∂2φ
det = C,
∂xi ∂xj
satisfying the boundary condition and convexity property
!
∂2φ
φ|∂D = 0, > 0.
∂xi ∂xj
We define a T k -invariant metric on M by

X
g= φij (dxi ⊗ dxj + dy i ⊗ dy j ),
i,j
2 √
where φij = ∂x∂i ∂xφ
j . Taking complex coordinates z
j
= xj + −1y j it turns out
that g is Kähler, with Kähler form
√
−1 X
ω= φij dz i ∧ dz j ,
2 i,j
and hence its Chern Ricci form is

√
ρC = − −1∂∂ log det(φij ) = 0.
Therefore, g is Ricci flat. Observe that g is flat along the k-dimensional torus fibers,
and dim M = 2k, and hence g receives the name of semi-flat Calabi-Yau metric on
M [125].
Consider now the T k -invariant standard exact Courant algebroid T M ⊕ T ∗ M
over M with vanishing three-form H = 0. Take T̂ k another k-dimensional torus
with coordinates (ŷ 1 , . . . , ŷ k ) and define M̂ = D × T̂ k . Taking the closed two-form
X
B= dy j ∧ dŷ j
j
k k
on M = D × T × T̂ , it follows directly from Definition 10.1 that (M, 0) is T-dual
to (M̂ , 0). Arguing as in the proof of Lemma 10.22, the isomorphism ψ in the proof
of Theorem 10.11 is given by
! !
∂ ∂ j j ∂ ∂
ψ j
= j
, ψ(dx ) = dx , ψ j
= −dŷ j , ψ(dy j ) = − j .
∂x ∂x ∂y ∂ ŷ
238 10. T-DUALITY
Define the generalized metric G on T M ⊕ T ∗ M by

V+ = {X + g(X) | X ∈ T M } ⊂ T M ⊕ T ∗ M,
and notice that * +
∂ k ∂ k
V+ = + φjk dx , + φjk dy .
∂xj ∂yj
Therefore, the T-dual generalized metric Ĝ is determined by
* +
∂ ∂
V̂+ = ψ(V+ ) = + φjk dxk , + φjk dŷ k ⊂ T M̂ ⊕ T ∗ M̂
∂xj ∂ ŷj
where (φjk ) = (φjk )−1 . From this, V̂+ = {X + ĝ(X) | X ∈ T M̂ } where

X
ĝ = φij dxi ⊗ dxj + φij dŷ i ⊗ dŷ j .
i,j
Consider next the pair (G, divG ), where G is the generalized metric above and divG
is the associated Riemannian divergence (see Definition 2.46). By Propositions
3.30 and 3.39, this pair is trivially Ricci flat and scalar flat, and hence so is the
ˆ by Propositions 10.15 and 10.26. We already
corresponding T-dual pair (Ĝ, div)
ˆ applying the dilaton shift formula in
have a formula for Ĝ, so let us calculate div
Proposition 10.25. For this, notice that
ν = det(φij )1/2 dy1 ∧ . . . dyk = C 1/2 dy1 ∧ . . . ∧ dyk ,
ν̂ = det(φij )1/2 dŷ1 ∧ . . . ∧ dŷk = C −1/2 dŷ1 ∧ . . . ∧ dŷk ,
and therefore by (10.12)
ê = −2d log C = 0.
By Proposition 3.30, we conclude that ĝ is a Ricci flat metric in the standard sense.
To see this more explicitly, we apply the Legendre transform to the base coor-
dinates, defining new coordinates x̂j on the disk D by
∂ x̂i
= φij .
∂xj
In these new coordinates, the T-dual metric reads
X
ĝ = φij (dx̂i ⊗ dx̂j + dŷ i ⊗ dŷ j ),
i,j
√
and taking complex coordinates ẑ j = x̂j + −1ŷ j we conclude as before that ĝ is
a semi-flat Calabi-Yau metric on M̂ .
Our second example shows the effect of T-duality on the generalized Ricci flat
and scalar flat metrics on S 3 × S 1 considered in Example 3.51.
Example 10.31. Consider the compact four-dimensional manifold M = SU (2)×
U (1) in Example 3.51. We regard M as a principal T 2 -bundle over S 2 ∼
= U (1)\SU (2),
via the natural left action of the abelian subgroup
T 2 = U (1) × U (1) ⊂ SU (2) × U (1).
Consider the T 2 -equivariant Courant algebroid (T M ⊕ T ∗ M, h, i , [, ]H ) over M ,
where
H = −ke123
10.6. EXAMPLES 239
for k ∈ R.
We claim that (M, [H]) is self-T-dual. To see this, we regard the correspondence
space M in Definition 10.1 inside the Lie group
ι : M ֒→ M × M̂ ,
where M̂ denotes another copy of M endowed with the three-form Ĥ = −kê123 .
Then, define B as the pull-back to M of a left-invariant two-form
B = −ι∗ (ke1 ∧ ê1 + e4 ∧ ê4 ).
Then, we have
dB = −ι∗ (ke23 ∧ ê1 − ke1 ∧ ê23 ) = −ι∗ (kê123 − ke123 ) = p̂∗ Ĥ − p∗ H
where we used that ι∗ de1 = ι∗ e23 = ι∗ ê23 = ι∗ dê1 , since e23 and ê23 are basic.
Consider now the Ricci flat and scalar flat closed pair (G, div) defined in Exam-
ple 3.51. By Propositions 10.15 and 10.17 the corresponding T-dual pair (Ĝ, div) ˆ
±
is also Ricci flat, and has vanishing scalar curvatures S = 0. Let us calculate the
T-dual pair explicitly. A direct calculation as in the proof of Lemma 10.22 shows
that the isomorphism ψ in Theorem 10.11 satisfies
ψ(ej ) = êj , ψ(ej ) = êj , for j = 2, 3.
1
ψ(e1 ) = kê1 , ψ(e4 ) = ê4 , ψ(e1 ) = ê1 , ψ(e4 ) = ê4 .
k
By definition of G we have
V+ = he2 + ke2 , e3 + ke3 , e1 + ke1 , e4 + kx2 e4 i ⊂ T M ⊕ T ∗ M,
and therefore the T-dual generalized metric Ĝ is
V̂+ := ψ(V+ ) = hê2 + kê2 , ê3 + kê3 , ê1 + kê1 , kx2 ê4 + ê4 i ⊂ T M̂ ⊕ T ∗ M̂ .
Notice that
V̂+ := {X + ĝ(X) | X ∈ T M̂ },
where
ĝ = k(ê1 ⊗ ê1 + ê2 ⊗ ê2 + ê3 ⊗ ê3 + (kx)−2 e4 ⊗ e4 ).
Finally, we calculate the T-dual divergence
ˆ = divĜ − hê, ·i
div
applying the dilaton shift formula in Proposition 10.25. Note that
ν = kxe14 , ν̂ = x−1 ê14 ,
and therefore ν and ν̂ are constant on B = S 2 . By (10.12) we conclude
ê = ψ(ϕ) = ψ(−xe4 ) = −xê4 .
ˆ is a compatible pair, but it is no longer closed. Similarly as
Observe that (Ĝ, div)
ˆ is Ricci flat,
in Example 3.51, one can check now explicitly that the pair (Ĝ, div)
and has vanishing scalar curvatures S ± = 0. We leave this last part as an exercise
for the reader.
We finish this section with two explicit examples of T-dual solutions of the
generalized Ricci flow.
240 10. T-DUALITY
Example 10.32. Consider M ∼ = S 3 with the Hopf fibration S 1 → S 3 → S 2 ,

and let θ denote the connection one-form on S 3 satisfying dθ = ωS 2 , where ωS 2
denotes the standard area form on S 2 , and furthermore let H = 0. Next let
M̂ ∼= S 2 × S 1 , and consider the trivial fibration S 1 → S 1 × S 2 → S 2 . Let θ̂ denote
the pullback of the canonical line element on S 1 to M̂ , so that dθ̂ = 0. Furthermore
let Ĥ = −θ̂ ∧ ωS 2 , which is closed. As observed in Example 10.7,

p∗ H − p̂∗ Ĥ = p̂∗ ωS 2 ∧ θ̂ = dp∗ θ ∧ θ̂ = d p∗ θ ∧ p̂∗ θ̂ ,
hence (M, H, θ) and (M̂ , Ĥ, θ̂) are topologically T-dual. Let gS 2 denote the round
metric on S 2 and consider an S 1 -invariant metric on S 3 of the form
g = Kθ ⊗ θ + LgS 2 .
Observe that by applying Proposition 10.24 we obtain that (g, 0) is T-dual to (ĝ, b̂)
with
ĝ = K −1 θ̂ ⊗ θ̂ + LgS 2 ,
(10.16)
b̂ = 0.
The solution to generalized Ricci flow with initial condition (g, 0) on S 3 has b ≡ 0
for all time, and takes the form
K2 K
K̇ = − , L̇ = −2 + .
L2 L
Expressing the T-dual data as ĝ = K̂ θ̂ ⊗ θ̂ + L̂gS 2 and using (10.16) we obtain the
evolution equation for ĝ as
˙ 1 ˙ 1
K̂ = , L̂ = −2 + ,
L̂ 2 K̂ L̂
which one can check is a solution to generalized Ricci flow. Observe that S 3 shrinks
to a round point under the flow, whereas on S 2 × S 1 the S 2 shrinks to a point while
the S 1 fiber blows up.
Example 10.33. Let M ∼ = S 2n+1 and consider the Hopf fibration S 1 →
2n+1
S → CP , and let θ denote the connection one-form on S 2n+1 satisfying
n
dθ = ωF S , where ωF S is the Kähler form of the Fubini-Study metric on CPn ,

and furthermore let H = 0. Next let M̂ ∼ = CPn × S 1 , and consider the trivial
1 1 n n
fibration S → S × CP → CP . Let θ̂ denote the pullback of the canonical line
element on S 1 to M̂ , and let Ĥ = −θ̂ ∧ ωS 2 . As in the previous example one easily
checks that (M, H, θ) and (M̂ , Ĥ, θ̂) are topologically T-dual.
Now let g0 denote any metric on S 2n+1 with positive curvature operator.
Choose H0 = 0, and let (gt , bt ) denote the solution to generalized Ricci flow with
initial condition (g0 , 0). By Proposition 4.20 we know that Ht = 0 for all time and
(gt , bt ) = (gt , 0), where gt is the unique solution to Ricci flow with initial condition
g0 . By the theorem of Bohm-Wilking [20], we have that gt exists on some finite
time interval [0, T ), and converges to a round point as t → T . It follows from Theo-
rem 10.19 and Proposition 10.24 that the dual solution (ĝt , b̂t ) to generalized Ricci
flow also exists on a finite time interval, asymptotically converging to a solution
which homothetically shrinks the CPn base and expanding the S 1 fiber, analogously
to the previous example. In fact, any of the Ricci flow ‘sphere theorems’ in odd
10.6. EXAMPLES 241
dimension (for instance [27], [92]) can be applied to invariant metrics satisfying
curvature positivity conditions to generate similar examples.
Bibliography
[1] Ilka Agricola and Thomas Friedrich, On the holonomy of connections with skew-symmetric
torsion, Math. Ann. 328 (2004), no. 4, 711–748. MR 2047649 54
[2] , A note on flat metric connections with antisymmetric torsion, Differential Geom.
Appl. 28 (2010), no. 4, 480–487. MR 2651537 63
[3] A. Alekseev and P. Xu, Derived brackets and Courant algebroids, preprint. 33, 34, 56
[4] D. V. Alekseevskii and B. N. Kimelfeld, Structure of homogeneous Riemannian spaces with
zero Ricci curvature, Funkcional. Anal. i PriloŽen. 9 (1975), no. 2, 5–11. MR 0402650 65
[5] Ben Andrews and Christopher Hopper, The Ricci flow in Riemannian geometry, Lecture
Notes in Mathematics, vol. 2011, Springer, Heidelberg, 2011. MR 2760593 4
[6] V. Apostolov, P. Gauduchon, and G. Grantcharov, Bi-Hermitian structures on complex
surfaces, Proc. London Math. Soc. (3) 79 (1999), no. 2, 414–428. MR 1702248 154, 162
[7] Vestislav Apostolov and Marco Gualtieri, Generalized Kähler manifolds, commuting com-
plex structures, and split tangent bundles, Comm. Math. Phys. 271 (2007), no. 2, 561–575.
MR 2287917 157, 160
[8] Vestislav Apostolov and Jeffrey Streets, The nondegenerate generalized Kähler Calabi-Yau
problem, arXiv:1703.08650. 177, 182, 199
[9] Thierry Aubin, Équations du type Monge-Ampère sur les variétés Kählériennes compactes,
Bull. Sci. Math. (2) 102 (1978), no. 1, 63–95. MR 494932 1
[10] Michael Bailey, Gil R. Cavalcanti, and Joey van der Leer Durán, A neighbourhood theorem
for submanifolds in generalized complex geometry, arXiv:1906.12069. 136
[11] Michael Bailey and Marco Gualtieri, Local analytic geometry of generalized complex struc-
tures, Bulletin of the London Mathematical Society 49 (2017), no. 2, 307319. 136
[12] David Baraglia and Pedram Hekmati, Transitive Courant algebroids, string structures and
T -duality, Adv. Theor. Math. Phys. 19 (2015), no. 3, 613–672. MR 3418511 86, 229
[13] Arnaud Beauville, Variétés Kähleriennes dont la première classe de Chern est nulle, J.
Differential Geom. 18 (1983), no. 4, 755–782 (1984). MR 730926 182, 220
[14] Lionel Bérard-Bergery, Sur la courbure des métriques Riemanniennes invariantes des
groupes de Lie et des espaces homogènes, Ann. Sci. École Norm. Sup. (4) 11 (1978), no. 4,
543–576. MR 533067 65
[15] Eric Bergshoeff, Chris Hull, and Tomas Ortin, Duality in the type-II superstring effective
action, Nuclear Physics B 451 (1995), no. 3, 547575. 236
[16] Francis Bischoff, Morita equivalence and generalized Kähler geometry, Ph.D. thesis, Univer-
sity of Toronto, 2019. 162
[17] Francis Bischoff, Marco Gualtieri, and Maxim Zabzine, Morita equivalence and the gener-
alized Kähler potential, arXiv:1804.05412. 154, 161, 162
[18] Jean-Michel Bismut, A local index theorem for non-Kähler manifolds, Math. Ann. 284
(1989), no. 4, 681–699. MR 1006380 45, 54, 150
[19] F. A. Bogomolov, The decomposition of Kähler manifolds with a trivial canonical class,
Mat. Sb. (N.S.) 93(135) (1974), 573–575, 630. MR 0345969 182, 220
[20] Christoph Böhm and Burkhard Wilking, Manifolds with positive curvature operators are
space forms, Ann. of Math. (2) 167 (2008), no. 3, 1079–1097. MR 2415394 240
[21] Jess Boling, Homogeneous solutions of pluriclosed flow on closed complex surfaces, J. Geom.
Anal. 26 (2016), no. 3, 2130–2154. MR 3511471 75
[22] Sébastien Boucksom, Philippe Eyssidieux, and Vincent Guedj (eds.), An introduction to
the Kähler-Ricci flow, Lecture Notes in Mathematics, vol. 2086, Springer, Cham, 2013.
MR 3202578 185
[23] Laurence Boulanger, Toric generalized Kähler structures, arXiv:1509.06785. 49
243
244 BIBLIOGRAPHY
[24] Peter Bouwknegt, Jarah Evslin, and Varghese Mathai, T -duality: topology change from
H-flux, Comm. Math. Phys. 249 (2004), no. 2, 383–415. MR 2080959 221, 225
[25] Peter Bouwknegt, Keith Hannabuss, and Varghese Mathai, T-duality for principal torus
bundles, Journal of High Energy Physics 2004 (2004), no. 03, 018018. 223
[26] Simon Brendle, Ricci flow and the sphere theorem, Graduate Studies in Mathematics, vol.
111, American Mathematical Society, Providence, RI, 2010. MR 2583938 4
[27] Simon Brendle and Richard Schoen, Manifolds with 1/4-pinched curvature are space forms,
J. Amer. Math. Soc. 22 (2009), no. 1, 287–307. MR 2449060 241
[28] Robert Bryant, Ricci flow solitons in dimension three with SO(3) symmetries,
https://services.math.duke.edu/ bryant/3DRotSymRicciSolitons.pdf. 83
[29] Nicholas Buchdahl, On compact Kähler surfaces, Ann. Inst. Fourier (Grenoble) 49 (1999),
no. 1, vii, xi, 287–302. MR 1688136 192
[30] , A Nakai-Moishezon criterion for non-Kähler surfaces, Ann. Inst. Fourier (Greno-
ble) 50 (2000), no. 5, 1533–1538. MR 1800126 192
[31] Henrique Bursztyn, Gil R. Cavalcanti, and Marco Gualtieri, Reduction of Courant algebroids
and generalized complex structures, Adv. Math. 211 (2007), no. 2, 726–765. MR 2323543
20, 86
[32] T.H. Buscher, A symmetry of the string background field equations, Ann. Inst. Fourier
(Grenoble) 194 (1987), 59–62. 221, 231, 235
[33] Eugenio Calabi, Improper affine hyperspheres of convex type and a generalization of a the-
orem by K. Jörgens, Michigan Math. J. 5 (1958), 105–126. MR 0106487 204
[34] C. G. Callan, D. Friedan, E. J. Martinec, and M. J. Perry, Strings in background fields,
Nuclear Phys. B 262 (1985), no. 4, 593–609. MR 819433 53, 57, 60, 68
[35] Huai Dong Cao, Deformation of Kähler metrics to Kähler-Einstein metrics on compact
Kähler manifolds, Invent. Math. 81 (1985), no. 2, 359–372. MR 799272 1
[36] Élie Cartan and Jan Arnoldus Schouten, On Riemannian geometries admitting an absolute
parallelism, Proc. Amsterdam 29 (1926), 933–946. 42, 63
[37] , On the geometry of the group-manifold of simple and semi-simple groups, Proc.
Amsterdam 29 (1926), 803–815. 42, 63
[38] Gil R. Cavalcanti, Hodge theory and deformations of SKT manifolds, arXiv:1203.0493. 144,
146
[39] Gil R. Cavalcanti and Marco Gualtieri, Generalized complex geometry and T -duality, A
celebration of the mathematical legacy of Raoul Bott, CRM Proc. Lecture Notes, vol. 50,
Amer. Math. Soc., Providence, RI, 2010, pp. 341–365. MR 2648900 221, 224, 226, 230
[40] , Blowing up generalized Kähler 4-manifolds, Bull. Braz. Math. Soc. (N.S.) 42 (2011),
no. 4, 537–557. MR 2861778 198
[41] , Stable generalized complex structures, Proceedings of the London Mathematical
Society 116 (2017), no. 5, 10751111. 136
[42] Jeff Cheeger, Comparison and Finiteness Theorems for Riemannian Manifolds, ProQuest
LLC, Ann Arbor, MI, 1967, Thesis (Ph.D.)–Princeton University. MR 2616706 108, 109
[43] Jeff Cheeger and Detlef Gromoll, The splitting theorem for manifolds of nonnegative Ricci
curvature, J. Differential Geometry 6 (1971/72), 119–128. MR 0303460 65
[44] Bing-Long Chen and Xi-Ping Zhu, Uniqueness of the Ricci flow on complete noncompact
manifolds, J. Differential Geom. 74 (2006), no. 1, 119–154. MR 2260930 94
[45] BingLong Chen and HuiLing Gu, Canonical solitons associated with generalized Ricci flows,
Sci. China Math. 56 (2013), no. 10, 2007–2013. MR 3102622 127
[46] Shiu Yuen Cheng and Shing-Tung Yau, Complete affine hypersurfaces. I. The completeness
of affine metrics, Comm. Pure Appl. Math. 39 (1986), no. 6, 839–866. MR 859275 237
[47] Shiing-Shen Chern, Complex manifolds without potential theory (with an appendix on the
geometry of characteristic classes), second ed., Universitext, Springer-Verlag, New York,
1995. MR 2244174 4
[48] Claude Chevalley, The algebraic theory of spinors and Clifford algebras, Springer-Verlag,
Berlin, 1997, Collected works. Vol. 2, Edited and with a foreword by Pierre Cartier and
Catherine Chevalley, With a postface by J.-P. Bourguignon. MR 1636473 135
[49] Bennett Chow, The Ricci flow on the 2-sphere, J. Differential Geom. 33 (1991), no. 2,
325–334. MR 1094458 85
BIBLIOGRAPHY 245
[50] Bennett Chow and Dan Knopf, The Ricci flow: an introduction, Mathematical Surveys and
Monographs, vol. 110, American Mathematical Society, Providence, RI, 2004. MR 2061425
4
[51] , The Ricci flow: an introduction, Mathematical Surveys and Monographs, vol. 110,
American Mathematical Society, Providence, RI, 2004. MR 2061425 107, 110
[52] Bennett Chow, Peng Lu, and Lei Ni, Hamilton’s Ricci flow, Graduate Studies in Mathe-
matics, vol. 77, American Mathematical Society, Providence, RI; Science Press Beijing, New
York, 2006. MR 2274812 4, 94
[53] André Coimbra, Charles Strickland-Constable, and Daniel Waldram, Supergravity as gener-
alised geometry I: type II theories, J. High Energy Phys. (2011), no. 11, 091, 35. MR 2913236
49, 53, 58
[54] Vicente Cortés and Liana David, Twist, elementary deformation, and kk correspondence in
generalized complex geometry, Int. J. Math. (2020). 82
[55] Theodore James Courant, Dirac manifolds, Trans. Amer. Math. Soc. 319 (1990), no. 2,
631–661. MR 998124 7, 137
[56] Xenia C. de la Ossa and Fernando Quevedo, Duality symmetries from non-abelian isometries
in string theory, Nuclear Physics B 403 (1993), no. 1-2, 377394. 235
[57] S. K. Donaldson, Anti self-dual Yang-Mills connections over complex algebraic surfaces and
stable vector bundles, Proc. London Math. Soc. (3) 50 (1985), no. 1, 1–26. MR 765366 207
[58] I. Y. Dorfman, Dirac structures of integrable evolution equations, Physics Letters A 125
(1987), 240–246. 7
[59] Nicola Enrietti, Anna Fino, and Luigi Vezzoni, The pluriclosed flow on nilmanifolds and
tamed symplectic forms, J. Geom. Anal. 25 (2015), no. 2, 883–909. MR 3319954 75
[60] Lawrence C. Evans, Classical solutions of fully nonlinear, convex, second-order elliptic equa-
tions, Comm. Pure Appl. Math. 35 (1982), no. 3, 333–363. MR 649348 204
[61] , Partial differential equations, second ed., Graduate Studies in Mathematics, vol. 19,
American Mathematical Society, Providence, RI, 2010. MR 2597943 4, 68
[62] Teng Fei, Bin Guo, and Duong H. Phong, Parabolic dimensional reductions of 11d super-
gravity, arXiv:1806.00583. 86
[63] Teng Fei and Duong H. Phong, Unification of the Kähler-Ricci and anomaly flows,
arXiv:1905.02274. 189
[64] Michael Feldman, Tom Ilmanen, and Lei Ni, Entropy and reduced distance for Ricci ex-
panders, J. Geom. Anal. 15 (2005), no. 1, 49–62. MR 2132265 118
[65] Anna Fino, Maurizio Parton, and Simon Salamon, Families of strong KT structures in six
dimensions, Comment. Math. Helv. 79 (2004), no. 2, 317–340. MR 2059435 146
[66] Anna Fino and Adriano Tomassini, Blow-ups and resolutions of strong Kähler with torsion
metrics, Adv. Math. 221 (2009), no. 3, 914–935. MR 2511043 198
[67] G. Folland, Weyl manifolds, J. Diff. Geom. 4 (1970), 145–153. 29
[68] Daniel Harry Friedan, Nonlinear models in 2 + ε dimensions, Ann. Physics 163 (1985),
no. 2, 318–419. MR 811072 1
[69] M. Garcia-Fernandez, Torsion-free generalized connections and heterotic supergravity,
Comm. Math. Phys. 332 (2014), no. 1, 89–115. MR 3253700 33, 42, 46, 49
[70] , Ricci flow, Killing spinors, and T-duality in generalized geometry, Adv. Math. 350
(2019), 1059–1108. 28, 33, 37, 49, 53, 54, 56, 57, 58, 61, 70, 86, 229, 230, 235
[71] Mario Garcia-Fernandez, Roberto Rubio, C. S. Shahbazi, and Carl Tipler, Canonical metrics
on holomorphic Courant algebroids, arXiv:1803.01873. 144, 148, 149, 183, 200
[72] Mario Garcia-Fernandez, Roberto Rubio, and Carl Tipler, Gauge theory for string alge-
broids, arXiv:2004.11399. 144, 146, 147, 148, 149, 150
[73] , Holomorphic string algebroids, arXiv:1807.10329. 149
[74] S. J. Gates, Jr., C. M. Hull, and M. Roček, Twisted multiplets and new supersymmetric
nonlinear σ-models, Nuclear Phys. B 248 (1984), no. 1, 157–186. MR 776369 1, 129, 150
[75] P. Gauduchon and S. Ivanov, Einstein-Hermitian surfaces and Hermitian Einstein-Weyl
structures in dimension 4, Math. Z. 226 (1997), no. 2, 317–326. MR 1477631 178
[76] Paul Gauduchon, Fibrés hermitiens à endomorphisme de Ricci non négatif, Bull. Soc. Math.
France 105 (1977), no. 2, 113–140. MR 0486672 181
[77] , La 1-forme de torsion d’une variété hermitienne compacte, Math. Ann. 267 (1984),
no. 4, 495–518. MR 742896 146
246 BIBLIOGRAPHY
[78] , Structures de Weyl-Einstein, espaces de twisteurs et variétés de type S 1 × S 3 , J.

Reine Angew. Math. 469 (1995), 1–50. MR 1363825 178
[79] , Hermitian connections and Dirac operators, Boll. Un. Mat. Ital. B (7) 11 (1997),
no. 2, suppl., 257–288. MR 1456265 167
[80] Krzysztof Gawȩdzki, Conformal field theory: a case study, Conformal field theory (Istanbul,
1998), Front. Phys., vol. 102, Adv. Book Program, Perseus Publ., Cambridge, MA, 2000,
p. 55. MR 1881386 18, 27
[81] Matthew Gibson and Jeffrey Streets, Deformation classes in generalized Kähler geometry,
arXiv:2005.03062. 194
[82] David Gilbarg and Neil S. Trudinger, Elliptic partial differential equations of second or-
der, Classics in Mathematics, Springer-Verlag, Berlin, 2001, Reprint of the 1998 edition.
MR 1814364 68
[83] Matt Gill, Convergence of the parabolic complex Monge-Ampère equation on compact Her-
mitian manifolds, Comm. Anal. Geom. 19 (2011), no. 2, 277–303. MR 2835881 189
[84] Steven Gindi and Jeffrey Streets, Structure of collapsing solutions of generalized Ricci flow,
arXiv:1811.08877. 127
[85] Ryushi Goto, Deformations of generalized complex and generalized Kähler structures, J.
Differential Geom. 84 (2010), no. 3, 525–560. MR 2669364 154, 162
[86] , Scalar curvature as moment map in generalized Kähler geometry, Journal of Sym-
plectic Geometry 18 (2020), no. 1, 147190. 49
[87] Phillip Griffiths and Joseph Harris, Principles of algebraic geometry, Wiley-Interscience
[John Wiley & Sons], New York, 1978, Pure and Applied Mathematics. MR 507725 4
[88] Marco Gualtieri, Generalized complex geometry, Ph.D. thesis, St. John’s College, University
of Oxford, 11 2003. 1, 129, 137, 150, 151, 152, 174
[89] , Branes on Poisson varieties, The many facets of geometry, Oxford Univ. Press,
Oxford, 2010, pp. 368–394. MR 2681704 33, 34, 37, 38, 41, 144
[90] , Generalized Kähler geometry, Comm. Math. Phys. 331 (2014), no. 1, 297–331.
MR 3232003 144, 146, 148, 149, 154
[91] , Generalized Kähler metrics from Hamiltonian deformations, Geometry and physics.
Vol. II, Oxford Univ. Press, Oxford, 2018, pp. 551–579. MR 3931787 162
[92] Richard S. Hamilton, Three-manifolds with positive Ricci curvature, J. Differential Geom.
17 (1982), no. 2, 255–306. MR 664497 1, 241
[93] , The Ricci flow on surfaces, Mathematics and general relativity (Santa Cruz, CA,
1986), Contemp. Math., vol. 71, Amer. Math. Soc., Providence, RI, 1988, pp. 237–262.
MR 954419 85
[94] , A compactness property for solutions of the Ricci flow, Amer. J. Math. 117 (1995),
no. 3, 545–572. MR 1333936 108, 109, 110
[95] , The formation of singularities in the Ricci flow, Surveys in differential geometry,
Vol. II (Cambridge, MA, 1993), Int. Press, Cambridge, MA, 1995, pp. 7–136. MR 1375255
101
[96] Robert Haslhofer and Aaron Naber, Weak solutions for the Ricci flow I, arXiv:1504.00911.
69
[97] Chun-Lei He, Sen Hu, De-Xing Kong, and Kefeng Liu, Generalized Ricci flow. I. Local
existence and uniqueness, Topology and physics, Nankai Tracts Math., vol. 12, World Sci.
Publ., Hackensack, NJ, 2008, pp. 151–171. MR 2503395 86
[98] Nigel Hitchin, Compact four-dimensional Einstein manifolds, J. Differential Geometry 9
(1974), 435–441. MR 0350657 65
[99] , Generalized Calabi-Yau manifolds, Q. J. Math. 54 (2003), no. 3, 281–308.
MR 2013140 129, 135
[100] , Brackets, forms and invariant functionals, Asian J. Math. 10 (2006), no. 3, 541–
560. MR 2253158 37, 38
[101] , Instantons, Poisson structures and generalized Kähler geometry, Comm. Math.
Phys. 265 (2006), no. 1, 131–164. MR 2217300 154
[102] , Bihermitian metrics on del Pezzo surfaces, J. Symplectic Geom. 5 (2007), no. 1,
1–8. MR 2371181 154, 162, 165
[103] Chris Hull, Ulf Lindström, Martin Roček, Rikard von Unge, and Maxim Zabzine, Gener-
alized Kähler geometry in (2, 1) superspace, J. High Energy Phys. (2012), no. 6, 013, front
matter+19. MR 3006894 150
BIBLIOGRAPHY 247
[104] Chris M. Hull, Ulf Lindström, Martin Roček, Rikard von Unge, and Maxim Zabzine,
Generalized Kähler geometry and gerbes, J. High Energy Phys. (2009), no. 10, 062, 25.
MR 2607445 150
[105] , Generalized Calabi-Yau metric and generalized Monge-Ampère equation, J. High
Energy Phys. (2010), no. 8, 060, 23. MR 2756087 150
[106] Christopher Hull and Ulf Lindstróm, The Generalised Complex Geometry of (p, q) Hermit-
ian Geometries, Comm. Math. Phys. 375 (2020), no. 1, 479–494. MR 4082174 144
[107] S. Ivanov and G. Papadopoulos, Vanishing theorems and string backgrounds, Classical Quan-
tum Gravity 18 (2001), no. 6, 1089–1110. MR 1822270 170, 172
[108] Joshua Jordan and Jeffrey Streets, On a Calabi-type estimate for pluriclosed flow, Adv.
Math. 366 (2020), 107097, 18. MR 4073308 150, 204, 206, 209
[109] Branislav Jurčo and Jan Visoký, Courant algebroid connections and string effective actions,
Noncommutative geometry and physics. 4, World Sci. Publ., Hackensack, NJ, 2017, pp. 211–
265. MR 3674835 33, 49
[110] Branislav Jurčo and Jan Vysoký, Poisson-Lie T-duality of string effective actions: a new
approach to the dilaton puzzle, J. Geom. Phys. 130 (2018), 1–26. MR 3807115 49, 230, 235
[111] Keiji Kikkawa and Masami Yamasaki, Casimir effects in superstring theories, Physics Let-
ters B 149 (1984), no. 4, 357 – 360. 221
[112] C. Klimčk and P. Ševera, Dual non-abelian duality and the Drinfeld double, Physics Letters
B 351 (1995), no. 4, 455462. 229
[113] Shoshichi Kobayashi, First Chern class and holomorphic tensor fields, Nagoya Math. J. 77
(1980), 5–11. MR 556302 206
[114] Shoshichi Kobayashi and Katsumi Nomizu, Foundations of differential geometry. Vol. II,
Wiley Classics Library, John Wiley & Sons, Inc., New York, 1996, Reprint of the 1969
original, A Wiley-Interscience Publication. MR 1393941 44
[115] Eva Kopfer and Karl-Theodor Sturm, Heat flow on time-dependent metric measure spaces
and super-Ricci flows, Comm. Pure Appl. Math. 71 (2018), no. 12, 2500–2608. MR 3869036
69
[116] Brett Kotschwar, An energy approach to the problem of uniqueness for the Ricci flow,
Comm. Anal. Geom. 22 (2014), no. 1, 149–176. MR 3194377 94
[117] , Ricci flow and the holonomy group, J. Reine Angew. Math. 690 (2014), 133–161.
MR 3200338 101
[118] N. V. Krylov, Boundedly inhomogeneous elliptic and parabolic equations, Izv. Akad. Nauk
SSSR Ser. Mat. 46 (1982), no. 3, 487–523, 670. MR 661144 204
[119] Ahcène Lamari, Courants Kählériens et surfaces compactes, Ann. Inst. Fourier (Grenoble)
49 (1999), no. 1, vii, x, 263–285. MR 1688140 192
[120] H. Blaine Lawson, Jr. and Marie-Louise Michelsohn, Spin geometry, Princeton Mathematical
Series, vol. 38, Princeton University Press, Princeton, NJ, 1989. MR 1031992 53
[121] John M. Lee, Riemannian manifolds, Graduate Texts in Mathematics, vol. 176, Springer-
Verlag, New York, 1997, An introduction to curvature. MR 1468735 4
[122] , Introduction to smooth manifolds, second ed., Graduate Texts in Mathematics, vol.
218, Springer, New York, 2013. MR 2954043 4
[123] Man-Chun Lee, Compact Hermitian manifolds with quasi-negative curvature,
arXiv:1810.07325. 189
[124] Man-Chun Lee and Jeffrey Streets, Complex manifolds with negative curvature operator,
arXiv:1903.12645, to appear IMRN. 215
[125] Naichung Conan Leung, Mirror symmetry without corrections, Comm. Anal. Geom. 13
(2005), no. 2, 287–331. MR 2154821 237
[126] Gary M. Lieberman, Second order parabolic differential equations, World Scientific Publish-
ing Co., Inc., River Edge, NJ, 1996. MR 1465184 68
[127] Ulf Lindström, Martin Roček, Rikard von Unge, and Maxim Zabzine, Generalized Kähler
manifolds and off-shell supersymmetry, Comm. Math. Phys. 269 (2007), no. 3, 833–849.
MR 2276362 150
[128] , Linearizing generalized Kähler geometry, J. High Energy Phys. (2007), no. 4, 061,
28. MR 2318766 150
[129] , A potential for generalized Kähler geometry, Handbook of pseudo-Riemannian ge-
ometry and supersymmetry, IRMA Lect. Math. Theor. Phys., vol. 16, Eur. Math. Soc.,
Zürich, 2010, pp. 263–273. MR 2681593 150, 161
248 BIBLIOGRAPHY
[130] Zhang-Ju Liu, Alan Weinstein, and Ping Xu, Manin triples for Lie bialgebroids, J. Differ-
ential Geom. 45 (1997), no. 3, 547–574. MR 1472888 7, 10
[131] John Lott, On the long-time behavior of type-III Ricci flow solutions, Math. Ann. 339
(2007), no. 3, 627–666. MR 2336062 127
[132] , Dimensional reduction and the long-time behavior of Ricci flow, Comment. Math.
Helv. 85 (2010), no. 3, 485–534. MR 2653690 127
[133] Robert J. McCann and Peter M. Topping, Ricci flow, entropy and optimal transportation,
Amer. J. Math. 132 (2010), no. 3, 711–730. MR 2666905 69
[134] John Milnor, Curvatures of left invariant metrics on Lie groups, Advances in Math. 21
(1976), no. 3, 293–329. MR 425012 74
[135] John Morgan and Gang Tian, Ricci flow and the Poincaré conjecture, Clay Mathematics
Monographs, vol. 3, American Mathematical Society, Providence, RI; Clay Mathematics
Institute, Cambridge, MA, 2007. MR 2334563 4
[136] James Morrow and Kunihiko Kodaira, Complex manifolds, AMS Chelsea Publishing, Prov-
idence, RI, 2006, Reprint of the 1971 edition with errata. MR 2214741 4
[137] A. Newlander and L. Nirenberg, Complex analytic coordinates in almost complex manifolds,
Ann. of Math. (2) 65 (1957), 391–404. MR 88770 129
[138] T. Oliynyk, V. Suneeta, and E. Woolgar, A gradient flow for worldsheet nonlinear sigma
models, Nuclear Phys. B 739 (2006), no. 3, 441–458. MR 2214659 53, 68, 111, 112
[139] Fabio Paradiso, Generalized Ricci flow on nilpotent Lie groups, arXiv:2002.01514. 75
[140] Grisha Perelman, The entropy formula for the Ricci flow and its geometric applications,
arXiv:0211159. 1, 111
[141] Grisha Perelman, Finite extinction time for the solutions to the Ricci flow on certain three-
manifolds, arXiv:0307245. 1
[142] Grisha Perelman, Ricci flow with surgery on three-manifolds, arXiv:0303109. 1
[143] Peter Petersen, Riemannian geometry, third ed., Graduate Texts in Mathematics, vol. 171,
Springer, Cham, 2016. MR 3469435 4, 68, 108
[144] Massimiliano Pontecorvo, Complex structures on Riemannian four-manifolds, Math. Ann.
309 (1997), no. 1, 159–177. MR 1467652 154
[145] Michael Reed and Barry Simon, Methods of modern mathematical physics. IV. Analysis
of operators, Academic Press [Harcourt Brace Jovanovich, Publishers], New York-London,
1978. MR 0493421 112
[146] O. S. Rothaus, Logarithmic Sobolev inequalities and the spectrum of Schrödinger operators,
J. Functional Analysis 42 (1981), no. 1, 110–120. MR 620582 121
[147] Martin Roček and Erik Verlinde, Duality, quotients, and currents, Nuclear Physics B 373
(1992), no. 3, 630646. 221
[148] Roberto Rubio and Carl Tipler, The Lie group of automorphisms of a Courant algebroid
and the moduli space of generalized metrics, Rev. Mat. Iberoam. 36 (2020), no. 2, 485–536.
MR 4082916 109
[149] H. Samelson, A class of complex-analytic manifolds, Portugaliae Math. 12 (1953), 129–132.
MR 0059287 145, 183
[150] Pavol Ševera, Letters to Alan Weinstein about Courant algebroids, arXiv:1707.00265. 12,
15, 17, 56
[151] Pavol Ševera and Fridrich Valach, Ricci flow, Courant algebroids, and renormalization of
Poisson-Lie T-duality, Lett. Math. Phys. 107 (2017), no. 10, 1823–1835. MR 3690034 49,
229, 230
[152] , Courant algebroids, Poisson-Lie T-duality, and type II supergravities, Comm.
Math. Phys. 375 (2020), 307–344. 49, 53, 112, 230
[153] Wan-Xiong Shi, Deforming the metric on complete Riemannian manifolds, J. Differential
Geom. 30 (1989), no. 1, 223–301. MR 1001277 94
[154] Jian Song and Gang Tian, The Kähler-Ricci flow through singularities, Invent. Math. 207
(2017), no. 2, 519–595. MR 3595934 1
[155] Jian Song and Ben Weinkove, An introduction to the Kähler-Ricci flow, An introduction to
the Kähler-Ricci flow, Lecture Notes in Math., vol. 2086, Springer, Cham, 2013, pp. 89–188.
MR 3185333 185
[156] Jeffrey Streets, Pluriclosed flow and the geometrization of complex surfaces,
arXiv:1808.09490. 220
BIBLIOGRAPHY 249
[157] , Regularity and expanding entropy for connection Ricci flow, J. Geom. Phys. 58
(2008), no. 7, 900–912. MR 2426247 68, 87, 118
[158] , Singularities of renormalization group flows, J. Geom. Phys. 59 (2009), no. 1, 8–16.
MR 2479257 86
[159] , Ricci Yang-Mills flow on surfaces, Adv. Math. 223 (2010), no. 2, 454–475.
MR 2565538 86
[160] , Pluriclosed flow, Born-Infeld geometry, and rigidity results for generalized Kähler
manifolds, Comm. Partial Differential Equations 41 (2016), no. 2, 318–374. MR 3462132
148, 150, 199, 203, 204, 213
[161] , Pluriclosed flow on manifolds with globally generated bundles, Complex Manifolds
3 (2016), no. 1, 222–230. MR 3550300 215
[162] , Generalized geometry, T -duality, and renormalization group flow, J. Geom. Phys.
114 (2017), 506–522. MR 3610057 70, 229, 230
[163] , Generalized Kähler-Ricci flow and the classification of nondegenerate generalized
Kähler surfaces, Adv. Math. 316 (2017), 187–215. MR 3672905 199
[164] , Global viscosity solutions of generalized Kähler-Ricci flow, Proc. Amer. Math. Soc.
146 (2018), no. 2, 747–757. MR 3731708 160, 219
[165] , Pluriclosed flow on generalized Kähler manifolds with split tangent bundle, J. Reine
Angew. Math. 739 (2018), 241–276. MR 3808262 176, 199, 204, 217
[166] , Classification of solitons for pluriclosed flow on complex surfaces, Math. Ann. 375
(2019), no. 3-4, 1555–1595. MR 4023384 172, 181
[167] Jeffrey Streets and Gang Tian, A parabolic flow of pluriclosed metrics, Int. Math. Res. Not.
IMRN (2010), no. 16, 3101–3133. MR 2673720 187, 220
[168] , Hermitian curvature flow, J. Eur. Math. Soc. (JEMS) 13 (2011), no. 3, 601–634.
MR 2781927 189, 217
[169] , Generalized Kähler geometry and the pluriclosed flow, Nuclear Phys. B 858 (2012),
no. 2, 366–376. MR 2881439 189, 193
[170] , Regularity results for pluriclosed flow, Geom. Topol. 17 (2013), no. 4, 2389–2429.
MR 3110582 191, 192, 198, 199, 220
[171] Jeffrey Streets and Yury Ustinovskiy, Classification of generalized Kähler-Ricci solitons on
complex surfaces, arXiv:1907.03819, to appear in Comm. Pure. Appl. Math. 181
[172] Jeffrey Streets and Micah Warren, Evans-Krylov estimates for a nonconvex Monge-Ampère
equation, Math. Ann. 365 (2016), no. 1-2, 805–834. MR 3498927 204
[173] Jeffrey D. Streets, Ricci Yang-Mills flow, ProQuest LLC, Ann Arbor, MI, 2007, Thesis
(Ph.D.)–Duke University. MR 2709943 86
[174] Michael Struwe, Curvature flows on surfaces, Ann. Sc. Norm. Super. Pisa Cl. Sci. (5) 1
(2002), no. 2, 247–274. MR 1991140 85
[175] Karl-Theodor Sturm, Super-Ricci flows for metric measure spaces, J. Funct. Anal. 275
(2018), no. 12, 3504–3569. MR 3864508 69
[176] Andrei Teleman, The pseudo-effective cone of a non-Kählerian surface and applications,
Math. Ann. 335 (2006), no. 4, 965–989. MR 2232025 192
[177] John A. Thorpe, Some remarks on the Gauss-Bonnet integral, J. Math. Mech. 18 (1969),
779–786. MR 0256307 65
[178] Gang Tian and Zhou Zhang, On the Kähler-Ricci flow on projective manifolds of general
type, Chinese Ann. Math. Ser. B 27 (2006), no. 2, 179–192. MR 2243679 198, 214
[179] Valentino Tosatti and Ben Weinkove, On the evolution of a Hermitian metric by its Chern-
Ricci form, J. Differential Geom. 99 (2015), no. 1, 125–163. MR 3299824 189
[180] K. Uhlenbeck and S.-T. Yau, On the existence of Hermitian-Yang-Mills connections in
stable vector bundles, vol. 39, 1986, Frontiers of the mathematical sciences: 1985 (New
York, 1985), pp. S257–S293. MR 861491 207
[181] Yury Ustinovskiy, The Hermitian curvature flow on manifolds with non-negative Griffiths
curvature, Amer. J. Math. 141 (2019), no. 6, 1751–1775. MR 4030526 189
[182] J. L. van der Leer Durán, Blow-ups in generalized Kähler geometry, Comm. Math. Phys.
357 (2018), no. 3, 1133–1156. MR 3769747 198
[183] Qingsong Wang, Bo Yang, and Fangyang Zheng, On Bismut flat manifolds, Trans. Amer.
Math. Soc. 373 (2020), no. 8, 5747–5772. MR 4127891 183
[184] Frank W. Warner, Foundations of differentiable manifolds and Lie groups, Scott, Foresman
and Co., Glenview, Ill.-London, 1971. MR 0295244 4
250 BIBLIOGRAPHY
[185] Kentaro Yano, Affine connexions in an almost product space, Kōdai Math. Sem. Rep. 11
(1959), 1–24. MR 0107275 42
[186] Shing Tung Yau, On the Ricci curvature of a compact Kähler manifold and the complex
Monge-Ampère equation. I, Comm. Pure Appl. Math. 31 (1978), no. 3, 339–411. MR 480350
1, 182, 204, 220
[187] Andrea Nicole Young, Modified Ricci flow on a principal bundle, ProQuest LLC, Ann Arbor,
MI, 2008, Thesis (Ph.D.)–The University of Texas at Austin. MR 2712036 86

Generalized Ricci Flow

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Generalized Ricci Flow

Uploaded by

Copyright:

Available Formats

Generalized Ricci Flow

arXiv:2008.07004v2 [math.DG] 28 Aug 2020

Mario Garcia-Fernandez and Jeffrey Streets

University of California, Irvine, CA 92617

Chapter 2. Generalized Riemannian Geometry 7

Chapter 3. Generalized Connections and Curvature 33

Chapter 4. Fundamentals of Generalized Ricci Flow 67

Chapter 5. Local Existence and Regularity 87

Chapter 6. Energy and Entropy Functionals 111

Chapter 7. Generalized Complex Geometry 129

The primary purpose of this book is to provide an intro-

The secondary purpose of this book is to formulate ques-

These questions and conjectures will be interspersed througout the text.

We prove the fundamental analytic properties of generalized Ricci flow in Chap-

Generalized Riemannian Geometry

On the road to understanding Riemannian geometry from the modern point of

2.1. Courant algebroids

As there is the natural determinant pairing between Λn V and Λn V ∗ , this induces a

Proof. Setting a = X + ξ, b = Y + η, we directly compute using the properties

The differential operator [X + ξ, ·] is furthermore compatible with the inner

Proof. Setting a = X + ξ, b = Y + η, b = Z + ζ and using the identity

Proof. Setting a = X + ξ, b = Y + η, c = Z + ζ, one has, using the identities

Proof. We set a = X + ξ, b = Y + η, c = Z + ζ. To prove (1), we directly

Definition 2.9. A Courant algebroid is exact if it fits into an exact sequence

2.2. Symmetries of the Dorfman bracket

Lemma 2.13. Given M a smooth manifold and B ∈ Λ2 T ∗ , the map eB is

Proof. Given X + ξ, Y + η ∈ T ⊕ T ∗ we directly compute

Proposition 2.14. Given M a smooth manifold and B ∈ Λ2 T ∗ , we have

for any X + ξ, Y + η ∈ Γ(T ⊕ T ∗ ). Consequently, the map eB is an automorphism

Proof. Fix X + ξ, Y + η ∈ Γ(T ⊕ T ∗ ) and B ∈ Γ(Λ2 T ∗ ), and then compute

Having exhibited that closed B-field transformations are Courant automor-

Proof. Since the given f is a diffeomorphism, we can define a bundle auto-

Proposition 2.17. Given M a smooth manifold and H ∈ Λ3 T ∗ such that

As an immediate consequence of the proof of the previous theorem we obtain the

Another direct consequence of the proof of Theorem 2.19 is an explicit char-

space of smooth maps from Σ to M , denoted C ∞ (Σ, M ). Given a closed three-form

In terms of the corresponding functionals, setting H ′ = f∗ (H − dB) we have

2.3. Generalized metrics

(1) a choice of maximal compact subgroup O(n) × O(n) ≤ O(n, n),

We refer to the generalized metric determined by (g, σ) as G(g, σ).

Proof. Let Gt be a one-parameter family of generalized metrics on W , with

By the second condition in (2.21), a tangent vector Φ at G induces well-defined

In particular, the associated tangent vector to the space of generalized metrics in

Proof. Using the identification W = V ⊕ V ∗ induced by σ, following the

Remark 2.34. Given Φ ∈ End(W ) satisfying (2.21) it is an exercise to verify

(2.24) π ◦ Φ|V− ◦ σ− (X) = g −1 (ġ(X) − ḃ(X)).

Proof. Using the identification W = V ⊕ V ∗ induced by σ we have σ− (X) =

2.3.2. Generalized metrics on manifolds. We next turn to the geometry,

Definition 2.39. Given M a smooth manifold and an exact Courant algebroid

Remark 2.42. With the notation in Proposition 2.38, the action of F = f ◦ eB

Here, |dϕ|2 denotes the norm squared of dϕ ∈ Γ(Σ, T ∗ Σ ⊗ ϕ∗ T ) with respect to

where ∇∗ ∇ is the rough Laplacian for the product of Levi-Civita connections of η

Proof. A calculation in coordinates shows that the functional SG can be al-

2.4. Divergence operators

Proof. This follows from a simple calculation:

More explicitly, consider an isotropic splitting σ : T → E and the induced

Proof. Consider the identification E = T ⊕ T ∗ given by G = G(g, σ). Then,

An interesting class of infinitesimal isometries, and hence of compatible pairs,

Generalized Connections and Curvature

Having established the fundamental properties of generalized metrics, we turn

3.1. Generalized connections

TD (a, b, c) = TD0 (a, b, c) + hχa b − χb a, ci + hχc a, bi

where we have used that hχb a, ci = −hχb c, ai. Setting now a = X + ξ, b = Y + η,

= hN, ci + 12 η(∇Z X) + 12 iY (∇∗Z ξ)