Exact flow equation for the divergence functional

arXiv:2303.04082v1 [hep-th] 7 Mar 2023

Stefan Floerchinger
Theoretisch-Physikalisches Institut, Max-Wien Platz 1, 07743 Jena, Germany

E-mail: stefan.floerchinger@uni-jena.de

Abstract: An exact functional renormalization group flow equation is derived for the
divergence functional which is a generalization of the Kullback-Leibler divergence to quan-
tum field theories in the Euclidean domain. It compares distributions with different sources
and field expectation values. The renormalization group flow for a regularized version of
this functional connects two limits: one where the functional is known in terms of the
microscopic action or probability distribution, and the other where all fluctuations are
taken into account. In the latter limit one can obtain full correlation functions from func-
tional derivatives of the divergence functional. The flow equation provides a possiblity to
determine this functional non-perturbatively.
Contents

1 Introduction 1

2 Modified exponential class of probability densities 2

3 Polchinskis and Wetterichs equations 2

4 Flow of the divergence functional 3

5 Limits of large and vanishing regulator 5

6 Conclusions 6

1 Introduction

The functional renormalization group is a method widely employed in quantum field theory
and statistical physics, see refs. [1–15] for reviews. It allows to connect the microscopic
description of a quantum field theory in terms of its microscopic action to the macroscopic
regime where the theory is described by a generating functional for correlation functions
such as the Schwinger functional or its Legendre transform, the quantum effective ac-
tion. The method is based on exact renormalization group flow equations, the Polchinski
equation for a modified version of the Schinger functional [16], or the Wetterich equa-
tion for a modified version of the quantum effective action [17]. For an extension with a
scale-dependent Hubbard-Stratonovich transformation see ref. [18]. Predecessors of these
equations are the equation of Wegner and Houghton [19] or the Callan-Symmanzik equation
[20–22].
In the present article we derive a similar renormalization group flow equation for the
divergence functional, which is a functional version of the Kullback-Leibler divergence
[23, 24]. Formally it corresponds to the relative information entropy between probability
distributions characterized by two different values of the source or field expectation value,
respectively. Interestingly, this functional can also be seen as a generating functional for
correlation functions, similar as the quantum effective action [25].
The Kullback-Leibler divergence plays an important role in probability and information
theory [24, 26]. Essentially it describes how well one distribution can be differentiated
from another. It governs also the probability for distributions of field configuations after
asymptotically many drawings through Sanov’s theorem [26]. This object also plays an
important role in information geometry [27, 28]. Geometric quantities like the Fisher
information metric or the two dual Amari-Chentsov connections can be derived from it,
see also [25].

–1–
2 Modified exponential class of probability densities

We consider a class of probability densities with respect to a measure Dφ for random

variables or fields φα ,
 
1 αβ α
pk [φ, J] = exp −S[φ] − Rk φα φβ + J φα − Wk [J] , (2.1)
2

with the Schwinger functional Wk [J] defined such that the probability density is properly
normalized,  
1 αβ
Z
Wk [J] α
e = Dφ exp −S[φ] − Rk φα φβ + J φα . (2.2)
2
We use an abstract index notation where α combines continuous indices like spatial posi-
tion and abstract indices like field components, for axample α = (x, j) and φα = φj (x).
Einsteins summation convention implies a sum over discrete indices and an integral over
continuous indices. It is usually clear from the context how expressions in this shorthand
notation can be made more concrete.
The sources J α are also fields, while Rkαβ is considered as a matrix parameterized by
a flow parameter k. We are interested in deriving differential equations for the flow with
this parameter k. An example would be Rkαβ = k2 δαβ , and typically Rkαβ is large (has large
eigenvalues) for large k2 and vanishes for k2 → 0. In this case the functional probability
distribution (2.1) approaches a deformed Gaussian for large k2 while for k2 = 0 it is the
standard unmodified probability density of a Euclidean quantum field theory. However,
for many steps of the development one can leave the form of Rkαβ open, and could even
consider it as a bilocal source term.

3 Polchinskis and Wetterichs equations

As a first step we note the flow equation of the k-dependent Schwinger functional,
    
∂ 1 ∂ αβ 1 ∂ αβ δWk [J] δWk [J] δWk [J]
Wk [J] = − R hφα φβ i = − R + . (3.1)
∂k 2 ∂k k 2 ∂k k δJ α δJ β δJ α δJ β

This is the Polchinski equation [16]. It uses the connected correlation function

δ2 Wk [J]
Gkαβ [J] = hφα φβ i − hφα ihφβ i = , (3.2)
δJ α δJ β
and the expectation value
δWk [J]
Φα [J] = hφα i = . (3.3)
δJ α
Incidentally the connected correlation function (3.2) is also the Fisher information metric
corresponding to the class of probability densities (2.1) when sources J α are taken as
coordinates [25].
Let us also introduce the Legendre transform

Γ̃k [Φ] = sup (J α Φβ − Wk [J]) , (3.4)

J

–2–
for which the flow equation is
 
∂  ∂  1 ∂ αβ
Γ̃k [Φ]Φ = − Wk [J]J = R [Gkαβ [Φ] + Φα Φβ ] . (3.5)
∂k ∂k 2 ∂k k
On the right hand side one can write the conncted two-point function as
(2)
Gkαβ [Φ] = (Γ̃k [Φ])−1
αβ , (3.6)

which is the inverse of the second functional derivative of eq. (3.4),

(2) δ2
(Γ̃k [Φ])αβ = Γ̃k [Φ]. (3.7)
δΦα δΦβ
The inverse propagator in (3.7) is actually the Fisher information metric corresponding to
(2.1) in expectation value coordinates [25].
Eq. (3.5) becomes even nicer in terms of the flowing action
1
Γk [Φ] = Γ̃k [Φ] − Rkαβ Φα Φβ , (3.8)
2
which has the flow equation
 
∂ 1 ∂ αβ (2)
Γk [Φ] = R (Γk [Φ] + Rk )−1
αβ . (3.9)
∂k 2 ∂k k
This is Wetterichs equation [17].

4 Flow of the divergence functional

Consider now the functional Kullback-Leibler divergence between the distributions at

source fields J and J ′ , respectively,
Z
D̃k [JkJ ] = Dφ pk [φ, J] ln(pk [φ, J]/pk [φ, J ′ ])
′

(4.1)
α ′α δWk [J] ′
=(J − J ) − Wk [J] + Wk [J ],
δJ α
where the first line is the general definition of a Kullback-Leibler divergence and the last
line uses the concrete form in eq. (2.1). It has the form of a Bregman divergence with
reversed arguments.
Alternatively one may replace the sources J and J ′ by the corresponding expectation
values Φ and Φ′ as coordinates. One finds
δΓ̃k [Φ′ ]
D̃k [ΦkΦ′ ] = Γ̃k [Φ] − Γ̃k [Φ′ ] − (Φλ − Φ′λ ). (4.2)
δΦ′λ
It is convenient to also define a divergence functional with subtracted regulator terms
1
Dk [ΦkΦ′ ] =D̃k [ΦkΦ′ ] − Rkαβ (Φα − Φ′α )(Φβ − Φ′β )
2
δΓk [Φ′ ] (4.3)
=Γk [Φ] − Γk [Φ′ ] − (Φλ − Φ′λ ).
δΦ′λ

–3–
We will call this the flowing divergence. As we will see below it has particularly nice limits.
Note, however, that only D̃k [ΦkΦ′ ] has all the mathematical properties of a relative entropy
functional and it is the Bregman divergence associated to Γ̃k [Φ]. The flowing divergence
(4.3) also vanishes when Φ = Φ′ , but could also be negative for some choices of argument.
First derivatives are given by

δ δΓk [Φ] δΓk [Φ′ ]

Dk [ΦkΦ′ ] =J α − J ′α − Rkαβ (Φβ − Φ′β ) = − ,
δΦα δΦα δΦ′α
(4.4)
δ ′ δ2 Γk [Φ′ ]
Dk [ΦkΦ ] = − (Φλ − Φ′λ ).
δΦ′α δΦ′α δΦ′λ

One may easily work out the second functional derivatives

(2,0) δ2 (2)
(D̃k [ΦkΦ′ ])αβ = D̃k [ΦkΦ′ ] =(Γ̃k [Φ])αβ ,
δΦα δΦβ
(1,1) δ2
(D̃k [ΦkΦ′ ])αβ = D̃k [ΦkΦ′ ] = − (Γ̃k [Φ′ ])αβ ,
δΦα δΦ′β
(0,2) δ2 δ3
(D̃k [ΦkΦ′ ])αβ = D̃k [ΦkΦ ′
] = − (Φ λ − Φ ′
λ ) Γ̃k [Φ′ ] + (Γ̃k [Φ′ ])αβ .
δΦ′α δΦ′β δΦ′α δΦ′β δΦ′λ
(4.5)

It is interesting to note that the right hand side of the first line depends actually only
on Φ. Higher order functional derivatives give one-particle irreducible vertex functions in
the presence of the regulator Rk . In a similar way, the right hand side of the second line
depends only on Φ′ . Finally, the term on the right hand side of the third line is linear in
Φ but depends non-linearly on Φ′ .
From eq. (3.5), together with its derivative with respect to the field expectation value
argument, we find a flow equation for the divergence functional
 
∂ ′ 1 ∂ αβ (2) (2) ′ −1
D̃k [ΦkΦ ] = R (Φα − Φ′α )(Φβ − Φ′β ) + (Γ̃k [Φ])−1 αβ − (Γ̃k [Φ ])αβ
∂k 2 ∂k k
(4.6)
δ3 Γ̃k [Φ′ ]

(2) (2) ′ −1
+ (Γ̃k [Φ′ ])−1 (
αµ kΓ̃ [Φ ]) (Φ λ − Φ ′
λ ) .
βν δΦ′µ δΦ′ν δΦ′λ

Here one may use (4.5) which yields a closed flow equation for the modified divergence
functional,
 
∂ 1 ∂ αβ (2,0)
′
D̃k [ΦkΦ ] = R (Φα − Φ′α )(Φβ − Φ′β ) + (D̃k [ΦkΦ′ ])−1
∂k 2 ∂k k αβ

(1,1) ′ −1 (0,2) ′ λκ (1,1) ′ −1
− (D̃k [ΦkΦ ])αλ (D̃k [ΦkΦ ]) (D̃k [ΦkΦ ])κβ .
(4.7)

On the right hand side we have greyed out those arguments on which the correspondonding
functionals do not depend any more after the functional derivative have been taken. Finally,

–4–
let us write this in terms of the flowing divergence,
 
∂ 1 ∂ αβ (2,0)
′
Dk [ΦkΦ ] = Rk (Dk [ΦkΦ′ ] + Rk )−1αβ
∂k 2 ∂k

(1,1) (0,2) (1,1)
− (Dk [ΦkΦ′ ] + Rk )−1
αλ (Dk [ΦkΦ′ ] + λκ
Rk ) (Dk [ΦkΦ′ ] + Rk )−1
κβ .

(4.8)

This exact flow equation is our main result.

5 Limits of large and vanishing regulator

From the definitions in (4.3), (3.4) and (2.2) one can obtain the functional integral relation
for the flowing divergence,
 
1 αβ δ
R
′
Dφ exp −S[φ] − 2 Rk (φα − Φα )(φβ − Φβ ) + δΦα Γk [Φ](φα − Φα )
e−Dk [ΦkΦ ] = R  .
D φ̃ exp −S[φ̃] − 21 Rkαβ (φ̃α − Φ′α )(φ̃β − Φ′β ) + δΦδ ′ Γk [Φ′ ](φ̃α − Φα )
α
(5.1)
One could further replace here
δ δ
Γk [Φ] = Dk [ΦkΦeq
k ],
δΦα δΦα
(5.2)
δ δ
Γk [Φ′ ] = − Dk [ΦkΦ′ ]Φ=Φeq ,

δΦα ′ δΦα k

where Φeq
k is the expectation value configuration solving δΓk [Φ]/δΦα = 0.
For large regulator scale k, where Rkαβ ∼ k2 δαβ is assumed, fluctuations around the
expectation values are suppressed and a saddle point approximation becomes valid
δ
lim Dk [ΦkΦ′ ] = S[Φ] − S[Φ′ ] − S[Φ′ ](Φα − Φ′α ). (5.3)
k→∞ δΦ′α

This can be supplemented by the next-to-leading order which is a one-loop term in the
presence of the regulator Rk . The crucial feature of eq. (5.3) is that the right hand side is
known! It is fully specified by the microscopic action S[φ] entering the probability density
(2.1).
In the opposite limit the flowing divergence approaches the full Kullback-Leibler di-
vergence or relative entropy functional,

lim Dk [ΦkΦ′ ] = D[ΦkΦ′ ]. (5.4)

k→0

The physical significance of D[ΦkΦ′ ] is discussed in ref. [25].

We find thus that D[ΦkΦ′ ] is fully determined by the starting point (5.3) and the
solution to the flow equation (4.8).

–5–
6 Conclusions

We have derived here a renormalization group flow equation for the flowing divergence
functional, which is a generalization of the Kullback-Leibler divergence to the setup of
Euclidean quantum field theories or statistical field theories. The new flow equation is a
close relative of Wetterich’s flow equation for the flowing action [17].
The flowing action is defined such that it approaches for large regulator an expression
fully defined by the microscopic action S[φ] or, in other words, the probability density at
vanishing source J. In the opposite limit of a vanishing regulator the flowing divergence
equals the full functional Kullback-Leibler divergence. The latter can be seen as a gener-
ating functional for correlation functions and it has an information theoretic significance
expressed for example through Sanov’s theorem [? ].
The flow equation we derived here is a functional differential equation for the steps
between the two limiting cases. It can therefore be used to determine the divergence
functional through the solution of a renormalization group flow instead of a perturbative
or Monte-Carlo calculation, for example. This makes new approximation schemed possible,
very much as in other applications of the functional renormalization group in quantum or
statistical field theory.
Future generalizations of the setup discussed here concern in particular quantum rela-
tive entropies and dynamical situations in quantum field theories. Relative entropies have
the advantage that they are well defined in the context of local quantum field theories [29],
in contrast to von-Neumann entanglement entropies which suffer from ultraviolet diver-
gences for reduced states corresponding to bounded regions of space.
Our new flow equation contributes also to the growing field of information geometry
and could find applications beyond physics.

References

[1] T.R. Morris, Elements of the Continuous Renormalization Group,

Progress of Theoretical Physics Supplement 131 (1998) 395.
[2] C. Bagnuls and C. Bervillier, Exact renormalization group equations: An introductory review,
Physics Reports 348 (2001) 91.
[3] M. Salmhofer and C. Honerkamp, Fermionic Renormalization Group Flows: Technique and
Theory, Progress of Theoretical Physics 105 (2001) 1.
[4] J. Berges, N. Tetradis and C. Wetterich, Non-Perturbative Renormalization Flow in Quantum
Field Theory and Statistical Physics, Physics Reports 363 (2002) 223 [hep-ph/0005122].
[5] J. Polonyi, Lectures on the functional renormalization group method,
Open Physics 1 (2003) 1.
[6] B. Delamotte, An Introduction to the Nonperturbative Renormalization Group, in
Renormalization Group and Effective Field Theory Approaches to Many-Body Systems,
A. Schwenk and J. Polonyi, eds., vol. 852, (Berlin, Heidelberg), pp. 49–132, Springer Berlin
Heidelberg (2012), DOI.

–6–
 [7] H. Gies, Introduction to the Functional RG and Applications to Gauge Theories, in
Renormalization Group and Effective Field Theory Approaches to Many-Body Systems,
A. Schwenk and J. Polonyi, eds., vol. 852, (Berlin, Heidelberg), pp. 287–348, Springer Berlin
Heidelberg (2012), DOI.
[8] J.M. Pawlowski, Aspects of the functional renormalisation group,
Annals of Physics 322 (2007) 2831.
[9] M. Reuter and F. Saueressig, Functional Renormalization Group Equations, Asymptotic
Safety, and Quantum Einstein Gravity, .
[10] W. Metzner, M. Salmhofer, C. Honerkamp, V. Meden and K. Schönhammer, Functional
renormalization group approach to correlated fermion systems,
Reviews of Modern Physics 84 (2012) 299.
[11] O.J. Rosten, Fundamentals of the exact renormalization group,
Physics Reports 511 (2012) 177.
[12] G. Camacho, C. Klöckner, D.M. Kennes and C. Karrasch, Review of recent developments of
the functional renormalization group for systems out of equilibrium, .
[13] P. Kopietz, L. Bartosch and F. Schütz, Introduction to the Functional Renormalization
Group, vol. 798 of Lecture Notes in Physics, Springer Berlin Heidelberg, Berlin, Heidelberg
(2010), 10.1007/978-3-642-05094-7.
[14] N. Dupuis, L. Canet, A. Eichhorn, W. Metzner, J. Pawlowski, M. Tissier et al., The
nonperturbative functional renormalization group and its applications,
Physics Reports 910 (2021) 1.
[15] A. Wipf, Statistical Approach to Quantum Field Theory: An Introduction, no. volume 992 in
Lecture Notes in Physics, Springer, Cham, second edition ed. (2021).
[16] J. Polchinski, Renormalization and effective lagrangians, Nuclear Physics B 231 (1984) 269.
[17] C. Wetterich, Exact evolution equation for the effective potential,
Physics Letters B 301 (1993) 90.
[18] S. Floerchinger and C. Wetterich, Exact flow equation for composite operators,
Physics Letters B 680 (2009) 371.
[19] F.J. Wegner and A. Houghton, Renormalization Group Equation for Critical Phenomena,
Physical Review A 8 (1973) 401.
[20] C.G. Callan, Broken Scale Invariance in Scalar Field Theory,
Physical Review D 2 (1970) 1541.
[21] K. Symanzik, Small distance behaviour in field theory and power counting,
Communications in Mathematical Physics 18 (1970) 227.
[22] K. Symanzik, Small-distance-behaviour analysis and Wilson expansions,
Communications in Mathematical Physics 23 (1971) 49.
[23] S. Kullback and R.A. Leibler, On Information and Sufficiency,
The Annals of Mathematical Statistics 22 (1951) 79.
[24] S. Kullback, Information Theory and Statistics, Dover Books on Mathematics, Dover
Publications, Mineola, N.Y (1997).
[25] S. Floerchinger, Information geometry of Euclidean quantum fields, to appear (2023) .

–7–
[26] T.M. Cover and J.A. Thomas, Elements of Information Theory, Wiley-Interscience,
Hoboken, N.J, 2nd ed ed. (2006).
[27] S.-i. Amari, Information Geometry and Its Applications, Springer Publishing Company,
Incorporated, 1st ed. (2016).
[28] N. Ay, J. Jost, H.V. Lê and L. Schwachhöfer, Information Geometry, vol. 64 of Ergebnisse
Der Mathematik Und Ihrer Grenzgebiete 34, Springer International Publishing, Cham (2017),
10.1007/978-3-319-56478-4.
[29] H. Araki, Relative entropy for states of von Neumann algebras. II,
Publications of the Research Institute for Mathematical Sciences 13 (1977) 173.

–8–