Prepared for submission to JHEP

The ABJM Amplituhedron

Song Hea,b,c Yu-tin Huangd,e Chia-Kai Kuod,e

a
CAS Key Laboratory of Theoretical Physics, Institute of Theoretical Physics, Chinese Academy
of Sciences, Beijing 100190, China
b
School of Fundamental Physics and Mathematical Sciences, Hangzhou Institute for Advanced
arXiv:2306.00951v2 [hep-th] 25 Sep 2023

Study; ICTP-AP International Centre for Theoretical Physics Asia-Pacific, UCAS, Hangzhou
310024, China
c
Peng Huanwu Center for Fundamental Theory, Hefei, Anhui 230026, P. R. China
d
Department of Physics and Center for Theoretical Physics, National Taiwan University, Taipei
10617, Taiwan
e
Physics Division, National Center for Theoretical Sciences, Taipei 10617, Taiwan

E-mail: songhe@[Link], yutinyt@[Link], chiakaikuo@[Link]

Abstract: In this paper, we take a major step towards the construction and applications
of an all-loop, all-multiplicity amplituhedron for three-dimensional planar N = 6 Chern-
Simons matter theory, or the ABJM amplituhedron. We show that by simply changing
the overall sign of the positive region of the original amplituhedron for four-dimensional
planar N = 4 super-Yang-Mills (sYM) and performing a symplectic reduction, only three-
dimensional kinematics in the middle sector of even-multiplicity survive. The resulting form
of the geometry, combined with its parity images, gives the full loop integrand. This simple
modification geometrically enforces the vanishing of odd-multiplicity cuts, and manifests
the correct soft cuts as well as two-particle unitarity cuts. Furthermore, the so-called
“bipartite structures” of four-point all-loop negative geometries also directly generalize to
all multiplicities. We introduce a novel approach for triangulating loop amplituhedra based
on the kinematics of the tree region, resulting in local integrands tailored to “prescriptive
unitarity”. This construction sheds fascinating new light on the interplay between loop
and tree amplituhedra for both ABJM and N = 4 sYM: the loop geometry demands that
the tree region must be dissected into chambers, defined by the simultaneous positivity
of maximal cuts. The loop geometry is then the “fibration” of the tree region. Using
the new construction, we give explicit results of one-loop integrands up to ten points and
two-loop integrands up to eight points by computing the canonical form of ABJM loop
amplituhedron.
Contents

1 Introduction and summary of results 1

2 Tree geometries 5

3 ABJM integrands from loop geometries 10

3.1 Definition and some all-loop cuts of the loop amplituhedron 11
3.2 Triangulating loop amplituhedron and one-loop integrands 15
3.3 Higher loops and bipartite negative geometries 27

4 Beyond positive solutions and parity action 33

4.1 The complete six-point amplitude 34
4.2 The complete eight-point amplitude 35

5 Conclusion and outlook 39

A Momentum-twistor Grassmannian formula for ABJM 43

B Sectors of Tn for even n 46

C Details on two-particle cuts 50

D Local integrand basis 54

1 Introduction and summary of results

The amplituhedron of planar N = 4 super Yang-Mills theory (sYM) [1–3] is a collection

of positive geometries whose canonical forms [4] encode all-loop, all-multiplicity scattering
amplitudes of the theory, where unitarity and causality etc. emerge from the geometries.
In the past decade there has been a lot of progress in the mathematics and physics of
the amplituhedron (c.f. [5–23]), as well as in search for such positive geometries in other
theories and contexts (c.f. [24–36]).
It has been long known that the scattering amplitude of three-dimensional N = 6
Chern-Simons-matter theory, or ABJM theory [37, 38], share many properties with N = 4
sYM. This includes the presence of Yangian (dual superconformal) symmetry for the leading
singularities [39, 40], as well as its correspondence with cells of the positive (orthogonal)
Grassmannian [41–43]. This leads to the natural question: is there a positive geometry
associated with ABJM amplitudes, and if so, is the new geometry related to that of N = 4
sYM in a natural way?

–1–
 Evidence for the existence of a geometry has been realized in [33, 34] where for par-
ticular component amplitudes, in the context of reduced supersymmetry, the answer can
be identified as a momentum amplituhedron in a construction that is similar to the N = 4
sYM [31]. However, the component nature of the construction becomes an obstruction
for generalizing to the loop level, since all components eventually appear in the cut of
the amplitude. More recently, a new positive geometry was found in [44]: remarkably the
canonical form of the geometry gives the four-point planar integrand of ABJM theory to
all loops. Not only has the geometry manifests various all-loop cuts such as soft cuts,
unitarity cuts, and vanishing cuts due to odd-particle amplitudes, but one can also exploit
it and compute four-point L-loop integrands up to L = 5 [44] without much work. More
recently, new integrated results, which contain e.g. Γcusp of the theory, have been obtained
by integrating these four-point loop integrands up to L = 3 [45, 46] (see [47, 48] for results
in sYM).
The new-found geometry has a very simple relation with that of N = 4 sYM: one
imposes a symplectic condition on all adjacent momentum twistors [49], and an additional
overall sign for all momentum-twistor four-brackets. The motivation for the former is
straightforward: the symplectic condition reduces the original four-dimensional conformal
group SU(4) to its three-dimensional counterpart Sp(4). The reason for the additional sign,
however, is less clear. Furthermore, as the four-point tree geometry is trivial, it is unclear
if the construction can be generalized to all multiplicity, where the subtle issue of branches,
inherent in the definition of orthogonal Grassmannians, arises.
In this paper, we demonstrate that the answer is affirmative by introducing a novel
interpretation of the amplituhedron. We begin by defining the “tree-region” in the space of
momentum twistors. This is the region for which the n set of momentum twistors satisfying:

⟨ii+1jj+1⟩ < 0, {⟨123i⟩} has k sign flips,

Tn : , (1.1)
⟨⟨ii+1⟩⟩ = 0 for i, j = 1, 2, . . . , n.

where the four-bracket represents the 4 × 4 determinant of Zi s and the two-bracket repre-
sents contraction with Sp(4) metric. If we neglect the symplectic condition and reverse the
overall sign of the four-bracket, we obtain the image of the tree N = 4 sYM amplituhedron
in momentum twistor space. The change of sign leads to an immediate consequence: the
region is non-empty only when n = even and k = n2 −2. This is exactly the configura-
tion where one has non-trivial amplitudes in ABJM. Thus the symplectic reduction and
inversion of sign reduce the four-dimensional kinematics to a subspace of three-dimensional
kinematics where ABJM has support!
Next, for each point within Tn we can define its associated loop-region via:
n
⟨(AB)a ii+1⟩ < 0, {⟨(AB)a 1i⟩} has k + 2 = sign flip,
2
Ln : ⟨(AB)a (AB)b ⟩ < 0, for any two loop a, b ∈ 1, 2, . . . , L, (1.2)
⟨⟨Aa Ba ⟩⟩ = 0 for a ∈ 1, . . . , L.

We claim that the planar loop-integrand of ABJM theory is given by the canonical form
of Ln . Note such geometric definition immediately manifests properties associated with

–2–
the integrand of ABJM theory, such as reduction under soft-cuts or vanishing cuts with
odd-multiplicity on one side.

3
1

2 2

1 3

4
4

Figure 1: An illustration of the chambers for the tree-region at eight points T8 . Each
chambers is characterized by the set of maximal cuts that are “positive”, i.e. that lives in
the amplituhedron Ln .

Since we are considering canonical forms of a space (Ln ) that is dependent on where
we are in Tn , the form maybe distinct from point to point. This leads us to define a new
structure within Tn , which we termed “chambers”. As we traverse Tn , a given chamber
refers to regions whose loop-form are invariant. Thus Tn is dissected into several chambers.
For example T6 has 1 chamber, T8 has 4 and T10 has 50. It turns out, this remarkable
new feature of Tn has an invariant definition independent of loops. For each point in Tn ,
consider the solution to
n
X
(DΓi )α,j Zj = 0, α = 1, · · · , k , (1.3)
j=1

where (DΓi )α,j are 4k-dimensional cells in the momentum twistor orthogonal Grassman-
nian [50]. The collection of cells for which the solution is positive then defines the chambers.
Thus the chambers can be considered as the “overlap” of the cells. For example at eight-
points, each of the four chambers are the overlap of two cells. As the 4k-dimensional cells
can also be identified as the kinematic solution for a k-loop maximal cut, the chambers can
be said to be defined through the collection of maximal cut that resides in Ln , see fig. 1.
Thus as one traverses Tn , one sees that the loop-geometry is demanding that the tree-
region is further dissected into chambers. Note that sometimes distinct chambers might be
degenerate at low loops, and are only fully resolved at higher loops: indeed the chambers
for T10 are only fully resolved at two-loops. As we will see, this picture leads to a fibration
of loop amplituhedron over the tree one:
X
Bi × Ωi (1.4)
i

–3–
where i labels the chambers of Tn , Tn = ∪i Bi and Bi , Ωi are the tree and loop-form
associated with the chamber. We provide an illustration of this structure for eight-points:

Note that we are essentially triangulating the loop-region through Tn , we do not in-
troduce new boundaries in loop space. Thus the resulting forms are naturally local! As we
will see that already for one loop ABJM, there is a unique loop form associated with each
chamber, which turns out to be the sum of local integrals (boxes and triangles), matching
the correct one-loop integrands up to n = 10 as we have checked. For two loops, similarly
we have obtained a unique two-loop form for each chamber as a linear combination of two-
loop local integrals, which indeed produces correct results up to n = 8 integrands [51, 52].
There is a single chamber/cell for n = 6 (so any two points in the positive region give
you the same loop form), and we have 4 and 50 such chambers by intersecting all possible
BCFW chambers for n = 8, 10 respectively. As the distinct forms are associated with
distinct maximal cut solutions, the representation is precisely that of “prescriptive unitar-
ity” [53]. That is, the integrand consists of terms that evaluate to 1 on particular maximal
cut solutions, and 0 on others.
Note that our new way of triangulation is applicable to both ABJM and N = 4 sYM.
Indeed it is an interesting mathematical problem of classifying such chambers both in
ABJM and sYM, and we will also apply this new way of triangulating loop amplituhedron
for sYM in [54]. However, there is a subtlety that is special to ABJM. Just as the orthogonal
group is a direct product of SO(n) and parity reflection {I, −I}, the amplituhedron for
ABJM also consists of the positive sector, and its parity conjugate image. For k = n2 −2,
we have 2k −1 such images. This counting can be naturally understood from the number
of maximal cut solutions for k-loops, which as discussed previously, are what characterizes
the chambers in Tn . Thus the full amplitude can be represented as:
X
AL−loop
n = Πn,α [Bi × Ωi ] (1.5)
i,α

where Πn,α with α = 0, 1, · · · , 2k −1 are the n-point parity operators. We have identified
the resulting operators for n = 6, 8 (with 2k = 2, 4).
Independent of the interpretation as ABJM amplituhedron, already for n = 4 the
reduced geometry provides a nice simplified model for the original (all-loop) amplituhedron

–4–
of N = 4 sYM. This becomes particularly clear when the amplituhedron is decomposed
into the so-called negative geometries [47], which can be viewed as natural building blocks
for multi-loop amplitudes in N = 4 sYM. As shown in [44], the reduction to D = 3 has
simplified negative geometries enormously: only those negative geometries corresponding
to bipartite graphs survive the reduction, which not only drastically reduces the number
of possible topologies but also put very strong constraints on their pole structure. We
will see that essentially all these multi-loop structures are untouched when generalized to
n > 4, and we find that negative geometries again simplify the computation of loop forms
for higher points. We will explicitly construct the L = 2 forms of negative geometries of
each chamber of n = 6, 8 cases, and explore consequences for some all-loop cuts.
The rest of the paper is organized as follows. In section 2 we motivate the tree geometry
for ABJM theory in momentum twistor space by reviewing the momentum-space geometry
implied by the Grassmannian integral and momentum-space amplituhedron. Equipped
with the tree region, we proceed to define the loop amplituhedron for ABJM theory in
section 3, where we demonstrate how the geometry manifests various consistency conditions
unique to ABJM as well as the most important unitarity cuts; we then proceed to derive
the loop-form for the positive sector for n = 6, 8, 10 at one loop and n = 6, 8 at two loops.
The full amplitude is then derived by introducing parity action on the positive sector in
section 4.

2 Tree geometries

It is instructive to begin with the tree-level amplitude, where its building blocks are related
to (n−3)-dimensional cells of the orthogonal Grassmannian OG>0 (k ′ , 2k ′ ) where n = 2k ′ .
More precisely, the BCFW building blocks in the tree-level recursion can be written as
residues of the integral formula [41]:
Z k′ ×2k′
d C δ(C · C T ) 2k′ |3k′
Qk ′ δ (C · Λ) , (2.1)
GL(k ′ ) i=1 Mi
where C denotes a k ′ ×2k ′ matrix that parameterizes k ′ -null planes in 2k ′ dimensional space;
all inner products “ · ” are with respect to diagonal n × n metric with alternating signs
and Mi is the i-th k ′ × k ′ minor of C. The integral is localized to the n−3-dimensional
subspace by evaluating the residues on the loci of vanishing minors Mi . Thus the sub-
space corresponds to the co-dimension (k ′ −2)(k ′ −3)/2 cells of the positive Grassmannian
OG>0 (k ′ , 2k ′ ), e.g. for n = 4, 6 the amplitude is given by the top-cell while for n = 8
by co-dimension one cells. The remaining n−3 degrees of freedom are then fixed by the
following constraints (2k ′ − 3 of them are independent)
n
X
Cα,i λai = 0, α = 1, · · · , k ′ , (2.2)
i=1

which imply momentum-conservation conditions ni=1 (−1)i−1 λai λbi = 0. Note importantly,
P

the orthogonal condition C · C T = 0 implies that the minors satisfy

Mi+k′
= ±1 . (2.3)
Mi

–5–
We will refer to the ratio being +1 (−1) as the positive (negative) branch.
It will be useful to utilize a graphical representation of these cells using on-shell dia-
grams, comprised of connected planar graphs of quartic vertices with n external legs. For
example for n = 6, where the cell is top-dimensional, the representative graph is

3 5

1 .

The n−3 = 3 degrees of freedom of this cell are associated with the three vertices. Each
graph gives rise to a permutation pattern that encodes the linear dependency of the columns
of Cα,i , i.e. the pattern of vanishing minors. For a more detailed discussion, see [42, 43].
For our purpose, it is sufficient to keep in mind that for a given on-shell diagram, solutions
to eq. (2.2) represents a configuration where all internal lines are on-shell for a given set of
external kinematics. Said in another way, each vertex in the graph represents a four-point
on-shell kinematics.

The tree region The combination of cells in OG(k ′ , n) that “tile” the tree amplitude
can be mapped to a positive geometry through the moment map introduced in [33, 34]
(see also [55]), following that of N = 4 sYM [31, 56]. In particular, a collection of ordered
2k ′ points on an k+4-dimensional moment can be represented as ΛA i with A = 1, · · · , 2+k
′

and i = 1, · · · , 2k ′ ; then the momentum amplituhedron YαA is the image of the top-cell in
OG(k ′ , 2k ′ ) through the following map:
n
X
YαA = Cα,i ΛA
i . (2.4)
i=1

This space is equivalent to requiring the Gr(k ′ , k ′ +2) Grassmannian YαA satisfying

⟨Y, i, i+1⟩ > 0,

n
X
(−1)i ⟨Y, i, a⟩⟨Y, i, b⟩ = 0, ∀a, b = 1, · · · , n
i=1
{⟨Y, 1, i⟩}, has k ′ sign flips , (2.5)
A ′ A ′ A ′
where ⟨Y, i, j⟩ ≡ ϵA1 A2 ···Ak′ +2 YαA11 · · · Yαkk′ Λi k +1 Λj k +2 .
The constraints in eq. (2.5) that defines the momentum amplituhedron can be directly
framed in kinematic space λi . To see this, note that constraints on ⟨Y, i, j⟩ are really
statements on the components of Λi,j that are perpendicular to Y . Thus it is natural to
identify
λi = Y ⊥ (ΛA T
i ) . (2.6)

–6–
Then we can simply identify ⟨Y, i, j⟩ = ⟨i, j⟩, where ⟨i, j⟩ = ϵαβ λαi λβj , with λi defined
through eq. (2.6). Thus the tree amplituhedron is now defined through

⟨i, i+1⟩ > 0, {⟨1, i⟩}, has k ′ sign flips

n
Kn : X (2.7)
(−1)i−1 ⟨i, a⟩⟨i, b⟩ = 0, ∀a, b = 1, · · · , n .
i=1

We will use Kn to denote any set of n real spinors that satisfy the above sign constraints.
Now, given a set of {λi }s, one can straightforwardly construct an associate set of momentum
twistors defined as !
i−1
λi 
Zi = (−1)  Pi−1 . (2.8)
j=1 pj λi

Then the sign patterns of eq. (2.7) translates to:1

⟨ii+1jj+1⟩ < 0, {⟨123i⟩} has k sign flips,

Tn : . (2.9)
⟨⟨ii+1⟩⟩ = 0 for i, j = 1, 2, . . . , n.

Here, k = k ′ −2 and the brackets are defined as ⟨ijkl⟩ := ϵIJKL ZiI ZjJ ZkK ZlL and ⟨⟨ij⟩⟩ :=
ΩIJ ZiI ZjJ . The Sp(4) symplectic metric ΩIJ is given as
! !
0 ϵ2×2 0 1
ΩIJ = , where ϵ2×2 = . (2.10)
ϵ2×2 0 −1 0

From now on we will refer to this as the tree (positive) region Tn . Note that under the
constraint ⟨ii+1jj+1⟩ < 0, one sequence ⟨aa+1a+2i⟩ having the correct number of sign
flips is sufficient to guarantee the same for others, and we have chosen ⟨123i⟩ as a repre-
sentative. There is a similar T-dual relation for the cells of the orthogonal Grassmannian.
Starting with a cell CΓ in OG(k ′ , 2k ′ ), on the support of eq. (2.2), one can map to a cell
DΓ in OG(k,
g 2k+4), where the e indicates the orthogonal condition is defined with respect
to a kinematic dependent n × n metric. In particular, the co-dimension (k ′ −2)(k ′ −3)/2
cell CΓ will be mapped to a co-dimension k(k−1)/2 DΓ and the k ′ × k ′ minor MiC maps
to k × k minor Mi+1 D , where the superscript labels the Grassmannian. This relation can be

derived on the level of the integral formula in eq. (2.1) which we review in appendix A.

Absence of non-middle sectors and chambers At this point, we have motivated

the definition of the tree-region Tn as an image from the T-dualization of Kn . As it
turns out, Tn in itself is a natural starting point for the study of ABJM amplitudes in
momentum twistor space, without ever referring to Kn . To see this, first note that Tn
is very similar to the tree-region for N = 4 sYM, except for two aspects: the symplectic
constraint and the overall negative sign for ⟨ii+1jj+1⟩. The first is simply reflecting 3-
dimensional kinematics, but remarkably, the second, i.e. whether ⟨ii+1jj+1⟩ is strictly
1
Here ⟨ii+1jj+1⟩ is defined with an extra k-dependent minus sign when jj+1 wraps around the last
entry, i.e. we have ⟨ii+1n1⟩(−1)k−1 < 0.

–7–
positive or negative, determines which 3-dimensional theory we are talking about. Indeed
with ⟨ii+1jj+1⟩ > 0, the sign flipping and symplectic constraint, we are simply considering
the tree-level amplitude of N = 4 sYM, with the kinematics lying on a three-dimensional
subspace. If we instead choose ⟨ii+1jj+1⟩ < 0, Tn is actually empty whenever n is odd!
Indeed using the identity

⟨abcd⟩ = −⟨⟨ab⟩⟩⟨⟨cd⟩⟩+⟨⟨bc⟩⟩⟨⟨da⟩⟩−⟨⟨ac⟩⟩⟨⟨db⟩⟩ , (2.11)

we have ⟨ii+1i+2i+3⟩ = ⟨⟨ii+2⟩⟩⟨⟨i+1i+3⟩⟩ and thus

Qn−3 !
i=1 ⟨ii+1i+2i+3⟩
⟨n−2n−1n1̂⟩⟨n1̂23⟩ = ⟨n−1n12⟩ Qn−3 . (2.12)
2
i=2 ⟨⟨ii+2⟩⟩

For odd n and Z ∈ R, the parenthesis on the RHS is a definite positive sign. Thus
if ⟨ii+1jj+1⟩ < 0, the LHS is strictly positive while the RHS is negative, leading to
a contradiction. For even n, one can similarly show that contradiction will arise unless
k = n2 − 2. We demonstrate this for n = 6, 8 in appendix B. Thus, all non-middle sectors
are absent and Tn is non-trivial exactly when ABJM on-shell kinematics is defined! From
now on, we will only require that Zi lives in Tn , not necessarily derived from Kn using
eq. (2.8). So for example, the first two components of Zi do not necessarily correspond to
the two-component spinors that satisfy momentum conservation.
Now given a point in Tn , we say that it is associated with a cell Γi if solving
n
X
(DΓi )α,j Zj = 0, α = 1, · · · , k , (2.13)
j=1

DΓ
yields a positive DΓi , i.e. all the minors Mi i are non-negative. Note that there are in total
2k number of solutions for eq. (2.13), so here we are referring to one of the solutions being
positive. We will refer to this as the positive solution. For a given BCFW triangulation,
the point can be associated with only one cell since such triangulations yield non-overlap
tilling.2 However, it is perfectly acceptable for a point to simultaneously lie in multiple
cells, as long as the set of “overlapping” cells do not participate in a triangulation. Thus
the maximal number of BCFW cells that can be associated with a given point is the total
number of distinct BCFW triangulations. Since for n = 4, 6 the tree amplitude corresponds
to the top cell, we will use n = 8 to illustrate this feature. The BCFW cells of OG(2,
g 8)
can be categorized by the vanishing of its 2 × 2 ordered minors:

Γ1 : M2D = 0, Γ2 : M3D = 0, Γ3 : M4D = 0, Γ4 : M5D = 0 . (2.14)

The tree amplitude is then given by the sum of the Grassmannian integral localized on
cells Γ1 and Γ3 or Γ2 and Γ4 . The two are equivalent via Cauchy’s theorem. Now T8 can
be schematically viewed as in fig. 2.
2
Since we do not have the tree amplituhedron in momentum twistor space, at this point, the non-
overlapping feature can only be phrased in momentum amplituhedron.

–8–
 Γ1 Γ3

3
1 3∩2
Γ2
1∩2
2 2

1∩4 Γ4
1 3
3∩4

4
4

Figure 2: An illustration of the chambers for the tree-region at eight points T8 . Each
chamber is characterized by the set of BCFW cells (Γi ) that are positive under eq. (2.13)
(or eq. (2.2)). Equivalent, it is identified by the set of maximal cuts that are “positive”,
i.e. there is one maximal cut solution that lives in the amplituhedron.

Points that lie in cell Γ1 cannot lie in Γ3 , since the two combine to give the tree amplitude.
On the other hand, it can simultaneously lie in Γ2 or Γ4 , although not both due to the
same reason. This suggests that we can use the intersection of BCFW cells to triangulate
T8 into four chambers:
1 ∩ 2, 1 ∩ 4, 3 ∩ 2, 3 ∩ 4 . (2.15)
These four chambers are related via cyclic rotation by one site. Importantly, this is not
merely a choice of being more fine-grained in partitioning the tree-region, or being demo-
cratic to all BCFW triangulations. As we will see in the next section, such partitioning
is demanded by the loop geometry. For 10-points, as we demonstrate in sec. 3.2, we will
have 50 chambers. To summarize, we determine the chamber associated with the external
kinematic point by solving eq. (2.13) for all possible 4k-cells Γi . The set of cells that are
positive, i.e. the Γi s such that the solution for DΓi have positive semi-definite minors, then
defines the chamber. Thus the tree-region can be schematically represented as
[
Tn = Bi (2.16)
i

where Bi represents the chambers. Note that here, we are considering BCFW cells. At
n = 6, 8, 10, it appears that all 4k-dimensional cells are BCFW cells. It will be interesting
to see if this is true for generic k. We will come back to this in the outlook.
Before moving on to loops, we briefly discuss other solutions to eq. (2.13). Recall that
when a point in Tn lies in Γi , only one of the 2k solutions will lead to a positive DΓi . The
other solutions, as it turns out, can be obtained by parity transformation (Π) acting on
subsets of the twistor variables, Zi → −Zi on a subset of twistors. The corresponding

–9–
operation for 6, 8 points are:

n=6 Π6 : {2, 4, 6}
n=8 Π8,1 : {2, 4, 6}, Π8,2 : {2, 6, 8}, Π8,3 ≡ Π8,2 Π8,1 : {4, 8} , (2.17)

where the curly bracket denotes the twistors on which the parity operator acts, and the
last entry in the second line represents the requisite parity flip can be viewed as consecutive
action of Π8,2 Π8,1 . With these parity operations, the remaining solutions can be converted
into the positive solution. The availability of such an operation will be crucial for us to
obtain the full amplitude.

3 ABJM integrands from loop geometries

We have established a natural region Tn in momentum twistor space which characterizes

the kinematics of tree-level scattering in ABJM. On the one hand it is the direct image of
momentum amplituhedron via T-duality, and on the other it is non-trivial precisely for the
kinematic configuration that ABJM amplitudes have support. Given this region, we would
now like to “attach” some loop geometries whose canonical forms give the loop integrand.
At four points, we only have four twistors and a single four-bracket, ⟨1234⟩. Thus up to
an Sp(4) rotation all points in T4 is projectively equivalent to
 
0 0 0 −1
 0 0 −1 0 
{Z} =  . (3.1)
 
−1 0 0 0 
0 −1 0 0

The geometry for four-point all loop amplitude was found in [44] where one considers the
canonical form for the loop-region

⟨(AB)a ii+1⟩ < 0,

⟨(AB)a (AB)b ⟩ < 0, for any two loop a, b ∈ 1, 2, . . . , L, (3.2)
⟨⟨Aa Ba ⟩⟩ = 0 for a ∈ 1, . . . , L.

where A, B denotes the loop twistors parameterized with respect to the columns of eq. (3.1).
Once again the four-point ABJM loop amplituhedron is defined in a fashion similar to
N = 4 super Yang-Mills, just with a definite negative sign for the four-bracket and a
symplectic constraint for the loop twistors. As we now show, by considering twistors living
in Tn , extending the definition to n-points with an appropriate assignment of sign flips, we
conjecture to have (positive branch of) the n-point ABJM amplituhedron which gives the
correct n-point loop integrand!

– 10 –
3.1 Definition and some all-loop cuts of the loop amplituhedron
Let us begin with the definition of the amplituhedron for n-points ABJM loop amplitude.
Starting with Zi ∈ Tn , and we define the loop-region Ln by requiring
n
⟨(AB)a ii+1⟩ < 0, {⟨(AB)a 1i⟩} has k + 2 = sign flip,
2
Ln : ⟨(AB)a (AB)b ⟩ < 0, for any two loop a, b ∈ 1, 2, . . . , L, (3.3)
⟨⟨Aa Ba ⟩⟩ = 0 for a ∈ 1, . . . , L.

Here once again, we can equivalently choose any other reference point for the sign flip
conditions, i.e. for the sequence {⟨(AB)a ji⟩|i ̸= j} with distinct js. First of all, the degree
of this space matches the degree of freedom for the loop integrand. Indeed for each loop
one has 4 × 2−4−1 = 3 degrees of freedom, where −4 and −1 corresponds to the GL(2)
redundancy and symplectic constraint. Parameterizing the loop twistors by

(ZA )a = Zi−1 + xa Zi − wa Zj
(3.4)
(ZB )a = ya Zi + Zj−1 + za Zj .

The symplectic condition ⟨⟨AB⟩⟩ = 0 can be used to solve xa . Thus for each “point” in Tn ,
eq. (3.3) defines a 3L-dimensional geometry in (wa , ya , za ) whose canonical form we will
identify as the L-loop integrand.
Soft-Cuts:
Before studying loop forms in detail, one can already see that the region defined in
eq. (3.3) geometrically manifests many properties of the loop integrand of ABJM theory.
For example it is known that due to the Chern-Simons interaction, the triple cut of the
L-loop integrand has milder behaviour and simply reduces to the (L−1)-loop integrand,

Aℓ−loop
n cut = (−1)i A(ℓ−1)−loop
n (3.5)
i−1, i, i+1

where we have used the dual regions to define the cut

Zi+1
Zi
i+1 ···
···

i L − 1-loop

···
i−1 .
Zi−1 Zi−2

Choosing the parameterization (i−2, i−1, i, i+1) for the loop twistors, the constraints

⟨(AB)a i−2 i−1⟩ = ⟨(AB)a i−1 i⟩ = ⟨(AB)a i i+1⟩ = 0 (3.6)

lead to (AB)a being identical to the line (i−1, i). On this solution, the mutual negativity
between the cut and any remaining loop variable in eq. (3.3) becomes:

– 11 –
 ⟨(AB)a (AB)b ⟩ = ⟨i−1i(AB)b ⟩ < 0 , (3.7)
which is nothing but the inequalities that define the (L−1)-loop amplituhedron.
Vanishing Cuts:
Slightly more non-trivial, are cuts which isolate odd-point kinematics, which should
vanish due to the absence of odd-point amplitudes. Note that since we’ve already shown
that Tn is empty for n = odd, all we need to show is that on any cut which isolates
odd-kinematics, eq. (3.3) completely factorizes into two parts which consist of odd regions.
Consider for example a two-particle cut ⟨ABi−1i⟩ = ⟨ABj−1j⟩ = 0
i
Zi−1 Zi
···

···
Zj Zj−1 .
j

As shown in [17], on such cut the inequalities for the amplituhedron of N = 4 sYM can be
recast into a union of “left” and “right” amplituhedron inequalities. Recall the difference
between the definition of ABJM amplituhedron in eq. (3.3) and the N = 4 counterpart is
the overall sign of the four-brackets and the symplectic constraint. The former does not
affect the analysis of [17], since it only changes the overall sign, while the latter immediately
sets odd-factorizations to zero.
For multi-loop cuts, we need to take into account the mutual negativity constraint
for the loop twistors. Using the familiar parametrization of n = 4, we expand each loop
variable using Z1 , Z2 , Z3 , Z4 :

(AB)a = (Z1 + xa Z2 − wa Z4 , ya Z2 + Z3 + za Z4 ), (3.8)

The condition ⟨⟨AB⟩⟩ = 0 becomes

⟨⟨14⟩⟩ ⟨⟨13⟩⟩
xa za + ya wa +za =− . (3.9)
⟨⟨24⟩⟩ ⟨⟨24⟩⟩

Since ⟨1234⟩ = ⟨⟨13⟩⟩⟨⟨24⟩⟩ < 0, we can rescale the RHS to 1. Note that here for n > 4 we
no longer have ⟨⟨14⟩⟩ = 0, however the mutual negativity condition

⟨(AB)a (AB)b ⟩
= (wa − wb )(ya − yb ) + (za − zb )(xa − xb ) < 0 (3.10)
⟨2341⟩

is unaffected by this extra term if we solve for x. Thus the implications of the mutual
negativity explored in [44] for n = 4 carries over to general multiplicity. In particular,
solving for xa and using the fact that za > 0 as implied from ⟨AB23⟩ < 0, we instead have

da,b ≡ (wa zb −wb za )(ya zb −yb za ) − (za,b )2 < 0. (3.11)

– 12 –
Now consider the following cut, which incises a three-point kinematics:
Zi−1 i Zi

···

···
··· .

With i = 2, we have ⟨(AB)a 12⟩ = ⟨(AB)b 12⟩ = 0 which implies wa = wb = 0. For such
case, da,b = −(za,b )2 < 0 which is always satisfied, and hence da,b = 0 is not a boundary.
Thus this cut does not appear as a boundary.
Finally, consider a three-particle cut with one of the mutual propagators put on-shell:

⟨(AB)a j−1j⟩ = ⟨(AB)b i−1i⟩ = ⟨(AB)a (AB)b ⟩ = 0. (3.12)

Schematically we have:
Zi−1 Zi
i
···

···

j .
Zj Zj−1

We would like to show that this boundary does not exist if one has an odd-multiplicity on
both sides of the cut. Using the parameterization in eq. (3.4) it is straight forward to see
that on the cut ⟨(AB)a j−1j⟩ = ⟨(AB)b i−1i⟩ = 0, the mutual negativity condition becomes
 
2 ⟨i−1 i j−1 j⟩
⟨(AB)a (AB)b ⟩ = ⟨i−1 i j−1 j⟩ wa yb − za,b
⟨⟨i+1Ba ⟩⟩⟨⟨i+1Bb ⟩⟩
!
zb 2 ⟨(AB)a j−2j−1⟩ ⟨⟨i−1i+1⟩⟩
 
= ⟨i−1 i j−1 j⟩ wa yb − 1 − < 0.
za ⟨(AB)b ii+1⟩ ⟨⟨j−2 j⟩⟩
(3.13)

The first term in the parenthesis is strictly positive since w, y > 0. The sign of the second
term will be determined by the ratio of ⟨⟨i−1i+1⟩⟩
⟨⟨j−2 j⟩⟩ .
To understand the sign of this ratio let’s digress and consider further implications of
the definite-sign condition for ⟨r r+1 s s+1⟩:

⟨r r+1 r+2 r+3⟩, (−1)k−1 ⟨n−2 n−1 n 1⟩, (−1)k−1 ⟨n123⟩ < 0. (3.14)

Note that when the entries are consecutive, the symplectic condition and (2.11) will lead to
complete factorization of the four-brackets: ⟨r r+1 r+2 r+3⟩ = ⟨⟨r r+2⟩⟩⟨⟨r+1 r+3⟩⟩. The
definite-sign condition will then lead to the following two sets of solutions for the two-
brackets:
⟨⟨13⟩⟩, ⟨⟨35⟩⟩, · · · , (−)k−1 ⟨⟨n−11⟩⟩ ≷ 0,
(3.15)
⟨⟨24⟩⟩, ⟨⟨46⟩⟩, · · · , (−)k−1 ⟨⟨n2⟩⟩ ≶ 0.

– 13 –
Thus we see that ⟨⟨rr+2⟩⟩ will have the same sign for all odd r and the opposites sign
for all even r. If one has an odd-multiplicity on both sides of the cut then ⟨⟨i−1i+1⟩⟩
⟨⟨j−2 j⟩⟩ will
be strictly negative and for even-multiplicity it will be strictly positive. Thus we see that
for an odd-multiplicity, all terms in the parenthesis of eq. (3.13) are strictly positive and
hence when combined with ⟨r r+1 s s+1⟩ < 0 trivializes the sign, and hence it is no longer
a boundary. It is amusing that properties of Tn are essential to ensure the absence of
all odd-multiplicity cuts. This phenomenon of Tn telling loops how to “behave”, will be
the central theme of this paper and we will see how it manifests when determining the
loop-form.
Two-particle cuts:
Finally, let us consider two-particle cuts, where the amplitude factorizes into “left”
and “right” counterparts. We will show that this simply follows from the geometrical
factorization of the amplituhedron. Before going into the detail, it is crucial to note that
the complete amplitude is the sum of various sectors obtained from parity operators Πs
(to be discussed in sec. 4). As the parity operators are defined from the interchanging of
maximal cut solutions, their presence under unitarity cuts factorizes. Thus for simplicity,
here we will simply focus on the positive sector.
In the context of N = 4 sYM, the unitarity of the amplitude can be translated into
geometrical factorization of the amplituhedron for all n and degrees k. More precisely,
under the cut ⟨ABi−1i⟩ = ⟨ABj−1j⟩ = 0, geometry factorization refers to the inequalities
of the amplituhedron being equivalent to summing over unions of inequalities associated
with the left and right amplituhedron, with the external legs suitably arranged. In the
sum, one includes left/right amplituhedron will run over all possible values of L1(2) , n1(2)
and k1(2) that satisfy

L = L1 + L2 + 1, n1 + n2 = n + 2, k = k1 + k2 , (3.16)

and the external legs of the left and right sets could be chosen as

L = {Z1 , · · · , Zi , A, B, Zj−1 , · · · , Zn }, R = {A, Zi+1 , · · · , Zj , B}. (3.17)

Note that since any rescaling of Zi s does not change the lines defined by pairs of momentum
twistors when considering the factorized geometry, one should allow for rescaling Zi ∈ L, R
by factors Zi → σL(R) (i)Zi . The central point of geometric factorization is that there
exists a set of rescaling such that on the cut, the inequalities of the amplituhedron can be
factorized into a union of inequalities which can be identified as the rescaled twistors living
inside left and right amplituhedron. This was beautifully established in [17].
As the ABJM amplituhedron can be viewed as a reduction of the N = 4 sYM am-
plituhedron, with an additional change of sign, geometrical factorization simply follows.
Indeed such a discussion for n = 4 was already presented in [45]. For n > 4, note that due
to the negative sign, the available geometry automatically satisfies ki = n2i − 2 with even
ni for i = 1, 2. For example, for n = 6, on the cut ⟨AB23⟩ = ⟨AB45⟩ = 0 for 6-4 channel,
(AB) can be parameterized by

A = Z2 + xZ3 , B = Z4 + zZ5 . (3.18)

– 14 –
Here, x = −(⟨⟨25⟩⟩z + ⟨⟨24⟩⟩)/⟨⟨35⟩⟩z from solving ⟨⟨AB⟩⟩ = 0. The inequalities associated
with {Zi , A, B} in T6 and L6 can be shown to be equivalent to the intersection of inequalities
involving external left and right sets L = {1, 2, A, B, 5, 6}, R = {A, 3, 4, B}:

Left: ⟨i i+1 j j+1⟩L < 0, {⟨123i⟩L } has 1 sign flip, ⟨⟨ii+1⟩⟩L = 0.

(3.19)
Right: ⟨1234⟩R < 0, ⟨⟨ii+1⟩⟩R = 0.

Here the labelling i in ⟨· · · ⟩L(R) denotes the i-th twistor in the set L(R). Furthermore,
the inequalities involving the remaining uncut loops (AB)a can be replaced by the union
of those defining the loop space with external sets L and R, according to the distribution
of L − 1 loop on the left and right sets:
! !
⟨(AB)a ii+1⟩L < 0, {⟨(AB)a 1i⟩}L has k1 + 2 sign flip,
∧ L, k1 , L1 → R, k2 , L2 ,
⟨(AB)a (AB)b ⟩L < 0, for any two loop a, b ∈ {L1 }.
(3.20)
satisfying ⟨⟨Aa Ba ⟩⟩ = 0 for a ∈ 1, . . . , L − 1 and with k1 = 1, k2 = 0. Here, {L1 } ⊂
{1, 2, · · · , L−1} has L1 elements and {L2 } is the complement with L2 = L−1−L1 elements.
For n = 8, the double cut for the 6-6 channel is ⟨AB23⟩ = ⟨AB67⟩ = 0, and the
double-cut geometry can be replaced by the inequalities involving external left and right
sets L = {1, 2, A, −B, −7, −8}, R = {A, 3, 4, 5, 6, B}, living in the region:

Left: ⟨i i+1 j j+1⟩L < 0, {⟨123i⟩L } has 1 sign flip, ⟨⟨ii+1⟩⟩L = 0.

(3.21)
Right: ⟨i i+1 j j+1⟩R < 0, {⟨123i⟩R } has 1 sign flip, ⟨⟨ii+1⟩⟩R = 0.

The remaining inequalities for the uncut loop variables can be replaced as (3.20) by ad-
justing k1 = 1, k2 = 1.
Not only does the region factorize, but the resulting form can also be matched with
the explicit computation in momentum space, which we present in detail in appendix C.

3.2 Triangulating loop amplituhedron and one-loop integrands

We now proceed to construct the loop integrand by deriving the canonical form for the
positive geometry defined in eq. (3.3), starting at one-loop. We will use a point in Tn as
input, i.e. a set of momentum twistors {Z} satisfying eq. (2.9). Then using this set of {Z}
for eq. (3.3) we can derive a canonical form.
Note that given two points in Tn , the canonical form derived from eq. (3.3) need not
be the same. Remarkably, points from the same chamber in Tn yields the same canonical
form! This is extremely non-trivial since while two points in the same chamber give
identical forms, their sign-flipping triangulation yields different blocks. Writing

Tn = ∪i Bi (3.22)

where Bi represents the chambers. For example at eight points, i = 1, 2, 3, 4 representing

the four distinct overlaps in eq. (2.15). Then the full loop amplitude should be written as
X
Bi × Ωi (3.23)
i

– 15 –
where Ωi is the canonical form from eq. (3.3) triangulated by the point in Bi , and Bi is the
associated tree-form for Bi . In other words, we have a product geometry, schematically
taking the form:

Note that we have not properly defined what the tree-form Bi is. As the chambers are
defined as the overlap of various (n−3)-dimensional cells, one expects that linear combina-
tions of Bi simply match to various leading singularities. For example at eight-point, where
we have four chambers 1 ∩ 2, 1 ∩ 4, 3 ∩ 2, 3 ∩ 4 , one expects the sum of tree-forms
B1∩2 , B1∩4 gives the leading singularity associated with the cell Γ1 . This is almost correct
except for one caveat: there are four leading singularities associated with cell Γ1 , which
one should be assigned to B1∩2 ⊕ B1∩4 ? As mentioned previously, for a given point in Bi
there are 2k solutions to eq. (2.2) of which only one is positive for cells that overlap in the
chosen chamber. The remaining becomes positive after parity transformation Π. Thus we
conclude that the combination of tree-forms Bi gives the positive leading singularity, while
further parity action yields the remaining 2k − 1 counterparts. The full integrand of ABJM
amplitude is then given as
X
AL−loop
n = Πn,α [Bi × Ωi ] (3.24)
i,α

where α = 0, 1, · · · 2k −1 labels the 2k −1 parity operators with Πn,0 being the identity.
Notice at loop level the parity operation acts on the tree-form and the loop-form: its
action on loop forms is simply equivalent to the canonical form of eq. (3.3) where Π acts
on the external momentum twistors. We will often refer to the “bare” geometry, i.e. acted
by Πn,0 , as positive sector. Note that the amplitude in eq. (3.24) is the momentum
twistor space integrand, which is equivalent to the momentum space integrand with an
k+3
overall ( ni=1 ⟨ii+1⟩)− 2 stripped off. This prefactor is identified through the “T-duality”
Q

transformation that relates the momentum space Grassmannian to the momentum twistor
space Grassmannian in appendix A.
To better illustrate the proposal, in this section we will study ABJM loop amplituhe-
dron and focus on the positive part first, i.e. Πn,0 part of eq. (3.24). We start with one-loop
case, and at this point we collect all possible sign patterns for ⟨AB1i⟩ that are compatible
with the sign flipping criteria in eq. (3.3). This gives a collection of regions defined by

– 16 –
inequalities of the remaining parameters (y, z, w). The sum of canonical forms (3-form) for
these regions gives the loop integrand in the parameterization of eq. (3.4). To reconstruct
the integrand, we introduce the following local integral basis:
Z
ϵ(a, i, j, k, X)
Ibox (i, j, k) :=
a (a · i)(a · j)(a · k)(a · X)
Z p (3.25)
(i · j)(j · k)(k · i)
Itri (i, j, k) := .
a (a · i)(a · j)(a · k)
The integrals are presented in their dual form, using the notation of [51, 52]. Although we
write the integrals in dual variables, as they take simpler form, their translation to twistor
variables are straight forward. For example, the epsilon numerators can be rewritten as:
1 
ϵ(i, j, k, l, m) = Q ⟨⟨i−1j⟩⟩⟨⟨j−1k⟩⟩⟨⟨k−1l⟩⟩⟨⟨l−1m⟩⟩⟨⟨m−1i⟩⟩
s=i,j,k,l,m ⟨s−1s⟩ (3.26)



−(i−1 ↔ i) − (k−1 ↔ k) · · · − (m−1 ↔ m) .
The coefficients of these integrals can now be fixed by using the explicit parameterization
for the loop variables in eq. (3.4) and matching to the 3-form derived from the geometry.
As it turns out, these coefficients are completely fixed already by matching to the co-
dimension one boundary of the geometry, i.e. single cuts. Such local representation allows
us to directly compare the result determined by the geometry with the known results in
the literature [51, 57].
The fact that the result obtained from the geometry can be directly matched with a
local integrand basis is not surprising. After all, we are “triangulating” the loop-form by
breaking Tn into chambers and don’t introduce new boundaries in the loop space (AB).
Thus our procedure gives rise to a representation satisfying prescriptive unitarity [53], i.e.
local integrands that contain only local singularities and match to only one maximal cut
solution while vanishing on others. Note that from this point of view, it is natural that
Tn is dissected into chambers that are identified from overlapping cells. This is the result
of the tension between 1. prescriptive unitarity, where all leading singularities (cells) are
realized once, and 2. the fact that Tn can be tiled by a distinct subset of cells. Thus to
realize prescriptive unitarity, Tn must be dissected into units that represent the overlap of
cells, i.e. chambers.

Six-point case Here we present the construction in detail for the six-point one-loop
amplitude. The eight- and ten-point amplitude can be derived following similar steps, and
we will only spell out the differences originating from the more intricate triangulation of
Tn .
We will use on-shell diagrams to represent BCFW cells. At six-point, the two available
on-shell diagrams are equivalent, reflecting the fact that the tree amplitude corresponds to
the top cell:
2

1 3 =
6 4
.
5

– 17 –
Let us begin with an arbitrary point in T6 . Since the tree amplitude is given by the top-
cell, there is only one chamber. This implies that all points in T6 should yield the same
loop-form through eq. (3.3). As we will see, while distinct points lead to the same form,
their sign-flipping triangulation yields distinct representations! We will now demonstrate
these properties using two different points:
Point 1.
 193133 
− 3848 − 501011
11544
14121
481
15545
481 − 665245
11544 − 236387
3848
 − 30635 − 27015 1975 2263
− 28521 − 30725 
 7696 7696 962 962 7696 7696 
{Z} =  1907188375 183507575 85910836 1640762167 . (3.27)
 0 11105328 2082249 − 2082249 − 11105328 0 
302520625 2794125
0 22210656 462722 − 763400 213262225
231361 − 22210656 0
One can check that these twistors indeed satisfy eq. (2.9) and are in T6 . For simplicity, let’s
focus on the two-form living on the co-dimension-one boundary ⟨AB34⟩ = 0. Choosing
i = 2 and j = 4 in eq. (3.4), this boundary translate to y = 0 and parameterized by w, z.
We can triangulate this space by considering all sign patterns consistent with three sign
flips. There are two consistent sign-flipping patterns:

{⟨(AB)1i⟩} = {−, +, −, −, +}, {−, +, −, +, +} . (3.28)

Other sign flipping such as {−, −, +, −, +} and {−, +, +, −, +} are not consistent with
eq. (3.27). For the first pattern, we have the following triangulation:
 p
⟨1235⟩ ⟨1235⟩ ⟨⟨13⟩⟩⟨1256⟩ −⟨1234⟩⟨3456⟩⟨1256⟩
{−, +, −, −, +} : − <z<− +
⟨1245⟩ ⟨1245⟩ ⟨1236⟩⟨1245⟩⟨1456⟩

2⟨1235⟩ ⟨1235⟩⟨1456⟩ ⟨1245⟩ ⟨1235⟩⟨1236⟩ 1
∧0<w < − + z+
⟨2345⟩ ⟨1345⟩⟨2456⟩ ⟨2345⟩ ⟨1246⟩⟨2345⟩ z
 p
⟨1235⟩ ⟨⟨13⟩⟩⟨1256⟩ −⟨1234⟩⟨3456⟩⟨1256⟩ 2⟨1235⟩ ⟨1245⟩
∨ − + <z∧ + z
⟨1245⟩ ⟨1236⟩⟨1245⟩⟨1456⟩ ⟨2345⟩ ⟨2345⟩

⟨1235⟩⟨2356⟩ 1 2⟨1235⟩ ⟨1235⟩⟨1456⟩ ⟨1245⟩ ⟨1235⟩⟨1236⟩ 1
+ <w< − + z+ .
⟨2345⟩⟨2456⟩ z ⟨2345⟩ ⟨1345⟩⟨2456⟩ ⟨2345⟩ ⟨1246⟩⟨2345⟩ z
(3.29)
It is straightforward to compute the associated 2-form:
!
dz dw dw
− √ ∧ −
z+ ⟨1235⟩
⟨⟨13⟩⟩⟨1256⟩ −⟨1234⟩⟨3456⟩⟨1256⟩ w w− ⟨2345⟩ − ⟨2345⟩ z− ⟨1235⟩⟨2356⟩
2⟨1235⟩ ⟨1245⟩ 1
⟨1245⟩ − ⟨1236⟩⟨1245⟩⟨1456⟩ ⟨2345⟩⟨2456⟩ z
!
dz dw dw
+ ⟨1235⟩
∧ − ⟨1235⟩⟨1456⟩ ⟨1245⟩ ⟨1235⟩⟨1236⟩ 1
.
z + ⟨1245⟩ w w − 2⟨1235⟩
⟨2345⟩ + ⟨1345⟩⟨2456⟩ − ⟨2345⟩ z − ⟨1246⟩⟨2345⟩ z
(3.30)
For the other sign flip pattern, we have
 ⟨1235⟩   ⟨1235⟩
{−, +, −, +, +} : 0<z<− ∧w >0 ∨ z >−
⟨1245⟩ ⟨1245⟩
(3.31)
2⟨1235⟩ ⟨1235⟩⟨1456⟩ ⟨1245⟩ ⟨1235⟩⟨1236⟩ 1 
∧ − + z+ <w ,
⟨2345⟩ ⟨1345⟩⟨2456⟩ ⟨2345⟩ ⟨1246⟩⟨2345⟩ z

– 18 –
with the associated 2-form
!
dz ∧ dw dz dw dw
− ∧ − .
zw z + ⟨1235⟩
⟨1245⟩
w w− 2⟨1235⟩
⟨2345⟩ + ⟨1235⟩⟨1456⟩
⟨1345⟩⟨2456⟩ − ⟨1245⟩
⟨2345⟩ z − ⟨1235⟩⟨1236⟩ 1
⟨1246⟩⟨2345⟩ z
(3.32)
By taking the sum of eqs. (3.30) and (3.32), the spurious poles in these two subregions are
cancelled, resulting in a complete 2-form on the boundary ⟨AB34⟩ = 0.
We can now find the combination of local integral basis that matches the above form.
Since we are considering the geometry for the boundary ⟨AB34⟩ = 0, not surprisingly only
integrals involving dual region 4 will participate. Substituting eq. (3.4) into the integrals
and taking into account the appropriate Jacobian factors we find that the result:
−Ibox (2, 3, 4) + Ibox (3, 4, 5) − Ibox (4, 5, 6) − Ibox (2, 4, 6) + Itri (2, 4, 6) . (3.33)
Point 2. Let us now consider another point,
 17457 
− 305 − 92931
2135
73236
2135
53184
2135 − 106989
2135 − 20343
305
 − 2703 − 14949 5022 4008
− 13611 − 2697 
 610 4270 2135 2135 4270 610
{Z} =  (3.34)

55932228 29095632 1232138304 388876788
− −

 0 130235 130235 4558225 651175 0 
4330206 1950264 90652608 25777926
0 130235 130235 − 4558225 − 651175 0
Now the allowed sign flipping patterns are {−, −, +, −, +}, {−, +, −, −, +}. For the first,
we have:
⟨1235⟩⟨3456⟩ 2⟨1235⟩ ⟨1245⟩ ⟨1235⟩⟨2356⟩ 1
{−, −, +, −, +} : z > − ∧w > + z+
⟨1345⟩⟨2456⟩ ⟨2345⟩ ⟨2345⟩ ⟨2345⟩⟨2456⟩ z
(3.35)
where the 2-form is
dz dw
∧ . (3.36)
⟨1235⟩⟨3456⟩
z + ⟨1345⟩⟨2456⟩ w − ⟨2345⟩ − ⟨2345⟩ z − ⟨1235⟩⟨2356⟩
2⟨1235⟩ ⟨1245⟩ 1
⟨2345⟩⟨2456⟩ z

For the other pattern:

 p 
⟨1235⟩ ⟨⟨13⟩⟩⟨1256⟩ −⟨1234⟩⟨3456⟩⟨1256⟩
0<z<− + ∧w >0
⟨1245⟩ ⟨1236⟩⟨1245⟩⟨1456⟩
 p
⟨1235⟩ ⟨⟨13⟩⟩⟨1256⟩ −⟨1234⟩⟨3456⟩⟨1256⟩ ⟨1235⟩⟨3456⟩
{−, +, −, −, +} : ∨ − + <z<−
⟨1245⟩ ⟨1236⟩⟨1245⟩⟨1456⟩ ⟨1345⟩⟨2456⟩

2⟨1235⟩ ⟨1245⟩ ⟨1235⟩⟨2356⟩ 1
∧w > + z+ .
⟨2345⟩ ⟨2345⟩ ⟨2345⟩⟨2456⟩ z
(3.37)
Its 2-form is
!
dz dw dw
− √ ∧ −
z+ ⟨1235⟩
⟨⟨13⟩⟩⟨1256⟩ −⟨1234⟩⟨3456⟩⟨1256⟩ w w− ⟨2345⟩ − ⟨2345⟩ z− ⟨1235⟩⟨2356⟩
2⟨1235⟩ ⟨1245⟩ 1
⟨1245⟩ − ⟨1236⟩⟨1245⟩⟨1456⟩ ⟨2345⟩⟨2456⟩ z

dz ∧ dw dz dw
+ − ⟨1235⟩⟨3456⟩
∧ 2⟨1235⟩ ⟨1245⟩ ⟨1235⟩⟨2356⟩ 1
.
zw z+ w− − −
⟨1345⟩⟨2456⟩ ⟨2345⟩ ⟨2345⟩ z ⟨2345⟩⟨2456⟩ z
(3.38)

– 19 –
We see that the triangulation arising from kinematic point eq. (3.34) is distinct from (3.27).
However, the sum of eqs. (3.36) and (3.38) gives the same as the sum of eqs. (3.30) and
(3.32).
The local form that contributes to the single cut ⟨AB34⟩ = 0 in eq. (3.33), can be
compared to the known complete result:
6
!
X
LS6,+ Ibox (1, 3, 5)+Itri (1, 3, 5)−Ibox (2, 4, 6)+Itri (2, 4, 6)+ (−)i Ibox (i−1, i, i+1)
i=1
6
!
X
i
+LS6,− Ibox (1, 3, 5)−Itri (1, 3, 5)−Ibox (2, 4, 6)−Itri (2, 4, 6)+ (−) Ibox (i−1, i, i+1) ,
i=1
(3.39)

where LS6,± correspond to the leading singularity associated with the positive and negative
branch of the orthogonal Grassmannian. For convenience, we give their explicit form in
momentum (spinor) space:
P 
δ 6 (Q) i,j,k=I ϵijk ⟨ij⟩ηk ± {1, 3, 5 → 2, 4, 6}
LS6,± = , I = {1, 3, 5} (3.40)
A± ± ±
52 A36 A14

where A±
P
a,b ≡ i∈I ⟨ia⟩⟨ib⟩ ± ⟨a+2a−2⟩⟨b−2b+2⟩. It is easy to see that the canonical form
in eq. (3.33) for the boundary ⟨AB34⟩ = 0 indeed matches the terms proportional to LS6,+
that contains such a single cut. To obtain the remaining part proportional to LS6,− requires
the action of parity operators, which we leave to section 4 for detail.

Eight-point case. We now move on to the eight-point one-loop amplitude. At eight-

points the BCFW representation for the tree-level amplitude consists of two terms, each
corresponding to an on-shell diagram that can be interpreted as a two-loop maximal cut.
There are in fact four in-equivalent diagrams:
i+4

Γi i+2 i+6

i ,

where i = 1, · · · , 4. Here the subscript for the cells Γi , also represents the vanishing 2 × 2
minor for OG(2,
g 8). The tree amplitude is given by the sum of the residues from the
Grassmannian integral in eq. (2.1) on these cells:
X X
Γ1 + Γ3 = Γ2 + Γ4 = Atree8 , (3.41)
sol. sol.

– 20 –
where sol. sums over the 22 = 4 solutions of eq. (2.2). As discussed previously, there are
P

four chambers for T8 :

Γ1

1∩2 Γ3
3∩2 Γ2
1∩4
3∩4

Γ4
Unlike the six-point case, where any point in T6 yields the same canonical form, here the
four chambers in T8 yield distinct forms! For example for points in the chamber 1 ∩ 2, the
corresponding canonical form can be matched to:
1
8
X
1∩2: (−1)i Ibox (i−1, i, i+1) + Itri (1, 3, 5) + Ibox (1, 3, 5) + Itri (5, 7, 1) + Ibox (5, 7, 1)
i=1
+Itri (2, 4, 6) − Ibox (2, 4, 6) + Itri (6, 8, 2) − Ibox (6, 8, 2).

2
(3.42)
Note that the last four integrands on the first line contain the one-loop maximal cut (1, 3, 5)
and (5, 7, 1), whose cut is precisely the leading singularity associated with cell Γ1

5
3
5
1
3 7
5
1
7
1
.
Similarly, the integrands on the second line are associated with that of cell Γ2 . This is
a general phenomenon: when a point is associated with a cell Γp the resulting loop-form
will include integrands that participate in the cuts contained by its on-shell diagram. Said
in another way, when we declare a point in Tn lies in Γp , this implies that cuts contained
in the on-shell diagram of Γp can be realized as the boundary of some positive geometry,
i.e. the loop amplituhedron. Thus the canonical form for eq. (3.3), from p ∩ q includes
integrands that participate in maximal cuts of Γp and Γq .
For convenience, let us consider the following combination of integrands as building

– 21 –
blocks:
8
X
I8L=1 (0) = (−1)i Ibox (i−1, i, i+1),
i=1
I8L=1 (1) = Ibox (1, 3, 5) + Itri (1, 3, 5) + Ibox (5, 7, 1) + Itri (5, 7, 1),
(3.43)
I8L=1 (2) = −Ibox (2, 4, 6) + Itri (2, 4, 6) − Ibox (6, 8, 2) + Itri (6, 8, 2),
I8L=1 (3) = Ibox (3, 5, 7) + Itri (3, 5, 7) + Ibox (7, 1, 3) + Itri (7, 1, 3),
I8L=1 (4) = −Ibox (4, 6, 8) + Itri (4, 6, 8) − Ibox (8, 2, 4) + Itri (8, 2, 4).

With these shorthand notations, the total canonical form is

X  
[p ∩ q]+ × I8L=1 (0)+I8L=1 (p)+I8L=1 (q) , {p, q} ∈ {1, 2}, {1, 4}, {3, 2}, {3, 4} .
p,q
(3.44)
Once again, the subscript [ ]+ indicates we are considering the positive sector. Note that
as the chambers are related by cyclic rotation, while the loop amplituhedron defined in
eq. (3.3) is cyclic invariant. Thus the combination I8L=1 (p)+I8L=1 (q) for each chamber
are related by cyclic rotation, as one can explicitly check. Since I8L=1 (p) appears in all
intersections involving p, collecting the terms proportional to I8L=1 (p) simply gives Γp .
Thus in total, we obtain
4
X
(Γ2,+ + Γ4,+ ) × I8L=1 (0) + Γi,+ × I8L=1 (i) , (3.45)
i=1

where Γi,+ represents the positive solution for the i-th cell. This yields the correct positive
sector of the one-loop integrand.

Ten-point case The anatomy of the tree region at n = 6, 8 is relatively simple: the
T6 consists of one chamber and T8 four. The four chambers are shown to be inherently
distinct by their extended one-loop form. As we will see, at ten-points we will need to go
to two-loops to fully distinguish all chambers.
At ten points, there exist three distinct topologies of on-shell diagrams, displayed below
along with their equivalent representation under the triangle move:

i+1 i+9
i+8 i i+2 i i
i+3
i+7
= =
i+8 i+2
i+6 i+4
i+6 i+4 i+6 i+4
ΓTi : (3.46)
i+9 i i+1
i
i+8 i+2
= = i+7
i+3

i+6 i+4 i+6 i+4

i+5

– 22 –
 i+9 i i+9 i

i+8 i
i+7 i+2
ΓPi : i+7 i+2 = = (3.47)
i+7 i+2
i+6 i+4

i+5 i+4 i+5 i+4

i i+1
i+2
i i+2 i i+3
i+9 i+9
ΓBi : =
i+4
= i+9 (3.48)
i+4 i+4
i+7 i+5
i+7
i+7 i+5
i+5

Including its cyclic images, there are a total of 10 + 10 + 5 = 25 in-equivalent BCFW

cells. From the recursion formula, the ten-point tree amplitude can be expressed as the
sum of three factorization channels: 8-4, 6-6, and 4-8. The residues on these cells from the
Grassmannian integral in eq. (2.1) have the following relations:
X
Ti + Ti+2 + Ti+4 + Ti+6 + Ti+8 = Atree
10 ,
sol.
X X X X (3.49)
Ti + Ti+4 = Bi + Pi , Ti+2 + Ti+8 = Bi+3 + Pi+8 ,
sol. sol. sol. sol.

where the summation is taken over the 23 = 8 solutions of eq. (2.2). The second line of
equality reflects the fact that in the 8-4 and 4-8 channels, the eight-point sub-amplitude
can be expressed in two ways

T1 B1 1 3
1 3
5
5

1 3 ⇒ =
+ +
T5 1 3 P1 1
5 3
5

.
Thus we have in total 22 inequivalent BCFW triangulations.
We will now use these cells to dissect T10 . Note that, unlike eight points, now cells
can appear in multiple triangulations. As the cells within a single BCFW triangulation do
not overlap, we can use this information to work out the overlapping regions throughout
T10 . For example consider a point residing in ΓT1 , the first line in eq. (3.49) tells us that it
cannot lie in ΓTi for i = 3, 5, 7, 9 at the same time. Now consider the BCFW triangulation
emerging from replacing T1 , T5 using the second line of eq. (3.49): {T3 , T7 , T9 , B1 , P1 }. Since
the point has been excluded from the images of T3 , T7 , and T9 , it must be located within

– 23 –
one of the images of either B1 or P1 . This tells us that T1 can be further separated into
T1 ∩ B1 and T1 ∩ P1 . Iteratively scanning through the 22 distinct BCFW triangulations,
one results in T10 being decomposed into fifty chambers! For example, the ten of chambers
involving T1 are:
T1,2 ∩ P1,2,7,8 , T1,2 ∩ P1,8 ∩ B2 , T1,4 ∩ P1,4,7,10 , T1,6 ∩ B1,2 , T1,6 ∩ P1,2,6,7 ,
T1,6 ∩ P1,6 ∩ B2 , T1,6 ∩ P2,7 ∩ B1 , T1,8 ∩ P1,4,7,8 , T1,10 ∩ P1,6,7,10 , (3.50)
and T1,10 ∩ P7,10 ∩ B1 ,
where T1,2 ∩ P1,8 ∩ B2 is a short hand notation for T1 ∩ T2 ∩ P1 ∩ P8 ∩ B2 . The remaining
40 corresponds to four cyclic shifts of the above by 2. The chambers in eq. (3.50) can be
represented diagrammatically as follows:

T10
T1
T2 T4 T6 T8 T10 T3
T5
T1,2 T1,6 T1,6 T1,10 T7
∩P1,2,7,8 T1,4 ∩B1,2 ∩P1,2,6,7 ∩P1,6,7,10
T1,8 T9
T1 ∩P1,4,7,10
∩P1,4,7,8
T1,6 T1,6 T1,10
T1,2 ∩ P1,8 ∩ B2
∩P1,6 ∩ B2 ∩P2,7 ∩ B1 ∩P7,10 ∩ B1

Figure 3: Here we present the chamber structure of T10 . Beginning with the BCFW
triangulation of T10 in terms of 5 pieces, as displayed in the upper right corner, each piece
is further dissected into smaller pieces via its overlap with other BCFW cells. These smaller
pieces constitute the chambers of T10 .

After listing the chambers of T10 we now can use it to geometrically obtain the one-
loop ten-point integrand. Unlike n = 8, we find that some chambers are degenerate at
the one-loop, i.e. they yield the same form. For example chambers T1,2 ∩ P1,2,7,8 and
T1,2 ∩ P1,8 ∩ B2 . In fact, this can be understood from the one-loop cuts visible form the
cells T1,2

9 1 1 3 2

7 9 1 3 5 8 6

7 5 10 2 4

T1 ∩ T2
1
8 6
10 2 2 4
7 5

8
. 6

We see that the cells T1 and T2 already contain the one-loop boundaries in Pi with i =
1, 2, 7, 8, as well as B2 . Thus further dissection of the region T1 ∩ T2 with Pi s or B2

– 24 –
is undetectable at one-loop. It turns out the intersection of two sets {Ti }i=1,3,5,7,9 and
{Ti }i=2,4,6,8,10 , are sufficient to characterize the one loop geometry of the ten-point. For
example T3 ∩ T6 covers another set of boundaries distinct from T1 ∩ T2 :

1 3 3 5 6

9 1 3 5 7 2 10

9 7 4 6 8

T3 ∩ T6
3
2 10
4 6 6 8
9 7

2 10
.
Thus from the one-loop geometry point of view, it is sufficient to write T10 as:
[
T10 = [2i−1 ∩ 2j] (3.51)
i,j=1,2,··· ,5

where 2i−1 ∩ 2j := T2i−1 ∩ T2j . These are the regions bounded by the black contours in
fig. 3 and its cyclic image. There are total of 25 sub-regions in this choice, and using them
we can obtain the one-loop ten-point integrand geometrically. In the regions [1 ∩ 2], the
corresponding loop canonical form is
1
10
X
1∩2: (−1)i Ibox (i−1, i, i+1)+Itri (1, 3, 5)+Ibox (1, 3, 5)+Itri (1, 5, 7)+Ibox (1, 5, 7)
i=1
+Itri (1, 7, 9)+Ibox (1, 7, 9)+Itri (2, 4, 6)−Ibox (2, 4, 6)+Itri (2, 6, 8)−Ibox (2, 6, 8)
+Itri (2, 8, 10)−Ibox (2, 8, 10).

2
(3.52)
Again, we can see that the loop-form is separated into three parts: the common soft part,
integrands related to T1 , and integrands related to T2 . We point out that Itri/box (1, 3, 5),
Itri/box (1, 5, 7), and Itri/box (1, 7, 9) produce the three one-loop maximal cuts in T1 while
Itri/box (2, 4, 6), Itri/box (2, 6, 8), and Itri/box (2, 8, 10) produce the three one-loop maximal
cuts in T2 :
Similarly, we can define the block functions to express the canonical form of other

– 25 –
regions:
10
X
L=1
I10 (0) := (−1)i Ibox (i−1, i, i+1),
i=1
L=1
I10 (2i−1) := Ibox (2i−1, 2i+1, 2i+3)+Itri (2i−1, 2i+1, 2i+3)+Ibox (2i−1, 2i+3, 2i+5)
+Itri (2i−1, 2i+3, 2i+5)+Ibox (2i−1, 2i+5, 2i+7)+Itri (2i−1, 2i+5, 2i+7),
L=1
I10 (2i) := −Ibox (2i, 2i+2, 2i+4)+Itri (2i, 2i+2, 2i+4)−Ibox (2i, 2i+4, 2i+6)
+Itri (2i, 2i+4, 2i+6)−Ibox (2i, 2i+6, 2i+8)+Itri (2i, 2i+6, 2i+8).
(3.53)

and the canonical form of the region 2i−1 ∩ 2j is

L=1 L=1 L=1
I10 (0) + I10 (2i−1) + I10 (2j). (3.54)

Summing over the 25 regions will give the one-loop ten-point amplitude in the positive
sector.
We see that the poles in the integrand, associated with an intersection region, exactly
match the local poles of the on-shell diagrams that define the intersection! As the number
of cells involved in an intersection increases, we will need higher loops to expose all possible
combinations of distinct boundaries. Indeed we have seen that for 10 points, some chambers
are degenerate at one loop. Such degeneracy is lifted as the loop order increases. In the
previous example of one-loop degeneracy for T1,2 ∩ P1,2,7,8 and T1,2 ∩ P1,8 ∩ B2 , will be
lifted at two loops. Indeed P2 contains the set of poles for the two-loop maximal cut
⟨AB12⟩ = ⟨AB34⟩ = ⟨AB56⟩ = ⟨ABCD⟩ = ⟨CD67⟩ = ⟨CD89⟩ = 0 that is not covered by
T1,2 ∩P1,8 ∩B2 , and thus we expect the degeneracy to be lifted. Indeed explicit computations
show that the two-loop forms for individual chambers are distinct. It will be interesting
to see if these chambers are sufficient to characterize all distinct two-loop forms. We leave
this to future work.
In summary, we find that at n = 6, 8, 10 there are one, four, and fifty chambers
respectively. The chambers yield distinct loop-forms at one-loop for n = 8 and at two-
loops for n = 10. This is illustrated below:

– 26 –
 Tree

L=1

L=2

n=6 n=8 n = 10

Figure 4: The loop-form for each distinct chamber might be degenerate at low loops. For
example at 10-points, the one-loop geometry can only distinguish between 25 subregions.
The degeneracy is lifted at two-loops, revealing the 50 chambers.

This analysis invites us to introduce a new characterization of the chambers. Recall that
each BCFW diagram also represents a set of maximal cuts. For n-points the diagram
represents an n/2−2 = k-loop maximal cut. This means that when solving for eq. (2.13)
(or equivalently (2.2)), we are also obtaining a maximal cut solution associated with the
given point in Tn . Thus asking for the solution to be positive is equivalent to asking
for the cut solution, which can be translated into a particular configuration of (A, B), to
reside on the boundary of the loop amplituhedron eq. (3.3). Said in another way, the
chambers of Tn are defined by the set of maximal cut that can simultaneously reside in the
loop amplituhedron! This also explains why we need to go to k-loops to fully resolve the
degeneracy of the chambers. Beyond k-loops, since the maximal cuts are associated with
on-shell diagrams that are reducible, the chambers will no longer bifurcate. At this point,
we have used the overlapping of BCFW cells to define the chambers. The question of
whether BCFW cells are sufficient simply translate to whether all of the n−3-dimensional
cells of OG>0 (k ′ , 2k ′ ) are BCFW cells. For k ′ = 3, the only 3-dimension cells is the top cell.
For k ′ = 4 we have four distinct co-dimension 1 cells that are 5-dimensional which exactly
correspond to our BCFW cells. At k ′ = 5 we have. It will be interesting to clarify whether
BCFW cells constitute all possible n−3-dimensional cells for general OG>0 (k ′ , 2k ′ ).

3.3 Higher loops and bipartite negative geometries

Now we move to higher-loops. We can proceed as before, simply triangulating the loop-
region using the chambers in Tn . As we will see for n = 6 and n = 8, there is no need for
further dissection, the loop-form continues to be invariant in each chamber. However, at
higher-loops, we can verify some important features of the multi-loop geometries defined
in eq. (3.3), which in turn provide some evidence that they indeed give all-loop ABJM
amplituhedron for all multiplicities. Recall that just as in the four-point case, the loop
geometry admits a natural decomposition by writing each “mutual positivity condition” as

– 27 –
“no condition” minus “negative condition”; in this way, our loop geometry becomes a sum
of “negative geometries” in D = 3: each negative geometry is represented by a labelled
graph with L nodes and E edges (edge (ij) for ⟨(AB)i (AB)j ⟩ > 0 since we reversed all
signs, and no condition otherwise), with an overall sign factor (−)E . We sum over all
graphs with L nodes without 2-cycles,
X
AL = (−)E(g) A(g) (3.55)
g

where A(g) denotes the (oriented) geometry for graph g. It suffices to consider all connected
graphs, whose (signed) sum gives the geometry for the logarithm of amplitudes [47]. Such a
decomposition is useful since each Ag is simpler, whose canonical form is easier to compute.
For example, the canonical forms for L = 2, 3 are

Ω2 = − + ,

Ω̃2

.
(3.56)
Ω3 = − + −

Ω̃3

where the connected part, or the integrand for logarithm of the amplitude, is denoted as
Ω̃L , e.g. Ω̃2 := Ω2 − 21 Ω21 . Similarly, the connected part of L = 4 is given by the sum of
graphs with 6 topologies (and so on for higher L),

+ + +

Ω̃4 = (3.57)

− − .

Now we show that exactly as in the n = 4 case, only those negative geometries with
bipartite graphs survive in D = 3, thus most negative geometries do not contribute. For
example, for Ω̃3 , the chain graph contributes but the triangle does not, i.e.
+ +
1 2 3 1 2 3 1 3 2
Ω̃3 = (3.58)
+ + + .
1 3 2 2 1 3 2 1 3

For Ω̃4 , only the two kinds of tree graphs and the box contribute. Note that the fraction
of bipartite graphs in all graphs tends to zero quickly as L increases: for L = 2, . . . , 7, the
number of topologies for connected graphs are 1, 2, 6, 21, 112, 853, while that of bipartite
topologies is 1, 1, 3, 5, 17, 44.
The reason that only bipartite negative geometries survive in D = 3 is exactly the same
as that for the n = 4 case, since only mutual conditions are needed in this reduction, which

– 28 –
 2 < 0 3 . Thus once again all mutual
we’ve proven in eq. (3.11) is still of the form wa,b ya,b −za,b
negativity conditions ⟨(AB)a (AB)b ⟩ > 0 translates to the inequality wa,b ya,b − za,b2 > 0 and

one concludes that (wa − wb ) and (ya − yb ) must have the same sign.
The remaining part of the proof goes exactly as for n = 4 (see appendix A of [52] for
details): for each edge of the graph we have an “arrow” which leads to “time ordering”
with no closed loop, and by the beautiful transitive reduction we find that except for
bipartite graphs where a black/white node denotes source/sink of the “ordering”, all other
graphs cancel in the sum [52]! Moreover, just as in n = 4 and as we will see for L = 2
n = 6, 8, the canonical form for these bipartite negative geometries has remarkably simple
pole structures. Since a sink (white node) can never have y, w → 0, it cannot have poles
⟨AB12⟩ = 0 or ⟨AB34⟩ = 0; similarly, a source (black node) cannot have poles ⟨AB23⟩ = 0
(note that ⟨AB14⟩ = 0 is not a physical pole for n > 4). One can immediately generalize
this by using other parametrizations, e.g. Z3 , Z4 , Z5 , Z6 , which means that each sink cannot
have poles in ⟨AB56⟩ = 0, and each source cannot have poles in ⟨AB34⟩ = 0. In this way,
we reach the conclusion, that each source/sink (black/white node) can only have poles
⟨ABi−1i⟩ = 0 for i odd/even, respectively (the latter includes the special case i = n,
which can only be a pole for white node).
Let us verify the above description by computing Ω̃α2 := Ωα2 − 21 (Ωα1 )2 for each chamber
labelled by α. The bipartite nature is reflected in that each black/white node for ℓ can
only have inverse propagators (ℓ · I) for I odd/even, which we will show matches nicely
with local integrands for Ωα2 [44].

Two-loop six-point integrand We start with n = 6 and first record the loop form Ωα2
for the top cell of the positive branch (LS6,+ ):
6
X 1 X X
− I critter (i) + I crab (i) + I+
2mh
(i) − I1bt (i) + I2bt (i) (3.59)
2
i=1 i=1,3,5 i=2,4,6

where the local (planar) integrands are given in the appendix D. Next, we reproduce this
result by computing canonical forms for the negative geometry, Ω̃α2 = Ωα2 −1/2(Ωα1 )2 , where
the one-loop form is for the same α, or LS+ . The result is naturally given in terms of some
integrands that look like ”bipartite graphs”.
The two-loop six-point form for negative geometry of the positive region is given by
a sum of three types of topologies, which are (non-planar) double-box, box-triangle, and
double-triangle integrals, and we denote them by Ia (B; W ), Ib (B; W ), Ic (B; W ), respec-
tively:

Ia (135, 246) + [Ib (135, 24) + (24 → 46, 26) + Ib (13, 246) + (13 → 35, 15)]
+[Ic (13, 24) + (24 → 26) + (13 → 35; 24 → 24, 46) + (13 → 15; 24 → 26, 46) (3.60)
+Ic (13, 46) + (13 → 35, 15; 46 → 26, 24)] + (ℓ1 ↔ ℓ2 ) ,
3
Here, we change the coordinate to the Poincare patch i.e. y ′ = y/z, w′ = w/z, z ′ = 1/z. Then eq. (3.11)
2
is proportional to wa,b ya,b − za,b < 0.

– 29 –
where B ⊂ {1, 3, 5} and W ⊂ {2, 4, 6} denote external points of the propagators involving
ℓ1 and ℓ2 , e.g.
Na (135, 246)
Ia (135, 246) = , (3.61)
(ℓ1 · 1)(ℓ1 · 3)(ℓ1 · 5)(ℓ1 · ℓ2 )(ℓ2 · 2)(ℓ2 · 4)(ℓ2 · 6)
and similarly for Ib , Ic ; note that we have 1 Ia integral, 2 × 3 = 6 Ib integrals, and 3 + 6 = 9
Ic integrals (6 cyclic images of Ic (13, 24) and 3 of Ic (13, 46), and in the end we symmetrize
over ℓ1 , ℓ2 . We give explicitly the numerator of these integrals
p p
:= Na (135, 246) = ϵ(ℓ1 , 1, 3, 5, µ )ϵ(ℓ2p, 2, 4, 6, µ ) − (ℓ1 · ℓ2 ) (1 · 3p · 5) (2 · 4 · 6)
1, 3, 5 2, 4, 6 −ϵ(ℓ1 , 1, 3, 5, ℓ2 ) (2 · 4 · 6) − ϵ(ℓ1 , 2, 4, 6, ℓ2 ) (1 · 3 · 5),
p
ϵ(ℓ1 , 2, 3, 4, 5) (1 · 3 · 5)
:= Nb (135, 24) = (ℓ1 · y4,(23) T(145) )(1 · 3) − ,
(3 · 5)
1, 3, 5 2, 4
p
ϵ(ℓ2 , 6, 1, 2, 3) (2 · 4 · 6)
:= Nb (13, 246) = (ℓ2 · y3,(12) T(634) )(2 · 6) − ,
(2 · 6)
1, 3 2, 4, 6
(1 · 3)(2 · 4)
:= Nc (13, 24) = − ,
2
1, 3 2, 4
(1 · 3)(4 · 6)
:= Nc (13, 46) = .
2
1, 3 4, 6
(3.62)
T
which are also denoted in terms of bipartite graphs. Here (i, j) (l, m, n) indicates the
intersection of a line (i, j) ≡ (Zi , Zj ) and a plane (l, m, n) ≡ (Zl , Zm , Zn ) is given by
\
(i, j) (l, m, n) = Zi ⟨j, l, m, n⟩ − Zj ⟨i, l, m, n⟩, (3.63)

and yp,q is a dual vector corresponding to the line (Zp , Zq ). In twistor space, the numerator
has very compact forms. For example, (ℓ1 · y4,(2,3) T(1,4,5) ) in twistor space (removing angle
brackets) is


(ℓ1 · y4,(2,3) T(1,4,5) ) → ⟨AB4 (2, 3) ∩ (1, 4, 5) ⟩ = ⟨AB24⟩⟨1345⟩−⟨AB34⟩⟨1245⟩. (3.64)

Two-loop eight-point integrand Next, we consider the two-loop eight-point case,

where the tree positive region has four chambers, labelled as (p∩q) for p = 1, 3 and q = 2, 4.
We focus on the first one, 1 ∩ 2, and we first give the expected loop form associated with
it:
1
8
X
db db 1 X X
1∩2: −IA (i)+IB (i)+ ICdb (i)+ db
ID,+ db
(i)+IE,+ (i)+ db
ID,+ db
(i)+IE,+ (i)
2
i=1 i=2,6 i=4,8
X
db 1 db X
bt bt
X
bt
X
db db
+ −IF,+,+ (i)+ IG,+,+ (i)− IA (i)+IA (i)− IB (i)+ ID,+ (i)+IE,+ (i)
2
i=3,7 i=1,5 i=3,7 i=3,7
X
db db
X
db 1 db X
bt bt
X
bt
+ ID,+ (i)+IE,+ (i)+ −IF,+,+ (i)+ IG,+,+ (i)+ IA (i)+IA (i)+ IB (i).
2
i=1,5 i=4,8 i=2,6 i=4,8
2
(3.65)

– 30 –
and for convenience let us consider the following combination of integrands as building
blocks:
8
X 1
I8L=2 (0) = db
−IA db
(p)+IB (p)+ ICdb (p),
2
p=1
X X X
I8L=2 (i) = db
ID,+ db
(i+p)+IE,+ (i+p) + db
ID,+ db
(i+p)+IE,+ (i+p) + db
−IF,+,+ (i+p)
p=1,5 p=3,7 p=2,6
1 db X X
+ IG,+,+ (i+p) + (−)i bt
IA bt
(i+p)+IA (i+p) + (−)i bt
IB (i+p).
2
p=4,8 i=2,6
(3.66)

These two-loop integrands are defined in appendix D. In complete analogy with the six-
point case above, we will reproduce these loop forms by computing canonical forms for the
negative geometry, Ω̃2 = Ω2 − 21 Ω21 for the chamber 1 ∩ 2. These results are again naturally
given in terms of loop forms for bipartite graphs, and we divide them into four groups
according to their pole structures, each of which resembles the six-point result. The first
group consists of those integrands with pole structures denoted by (135, 246):

p p
:= Na (135, 246) = ϵ(ℓ1 , 1, 3, 5, µ )ϵ(ℓ
p2 , 2, 4, 6, µ ) − (ℓ1 · ℓ2 ) (1 p· 3 · 5) (2 · 4 · 6)
1, 3, 5 2, 4, 6 −ϵ(ℓ1 , 1, 3, 5, ℓ2 ) (2 · 4 · 6) − ϵ(ℓ1 , 2, 4, 6, ℓ2 ) (1 · 3 · 5),


:= Nb (135, 24) = −(ℓ1 · 2)(1 · 4)(3 · 5)/2 + (ℓ1 · 3) (1 · p 4)(2 · 5)−(1 · 5)(2 · 4) /2
1, 3, 5 2, 4 −(ℓ1 · 4)(1 · 3)(2 · 5) − ϵ(ℓ1 , 2, 3, 4, 5) (1 · 3 · 5)/(3 · 5),



:= Nb (135, 46) = (ℓ1 · 4) (1 · 6)(3 p· 5)−(1 · 5)(3 · 6) /2 − (ℓ1 ↔ 1)
1, 3, 5 4, 6 +ϵ(ℓ1 , 3, 4, 5, 6) (1 · 3 · 5)/(3 · 5),

:= Nb (13, 246) = −(ℓ2 · 1)(2 · 4)(3 · 6)/2 + (ℓ2 · 2)(1 · 4)(3

p · 6)/2 + (ℓ2 · 3) (1 · 4)(2 · 6)
1, 3 2, 4, 6


−(1 · 6)(2 · 4) /2 + ϵ(ℓ2 , 1, 2, 3, 4) (2 · 4 · 6)/(2 · 4),


:= Nb (35, 246) = −(ℓ2 · 3)(2 · 5)(4 · 6)/2 + (ℓ2 · 4) (2 · p 5)(3 · 6)−(2 · 6)(3 · 5) /2
3, 5 2, 4, 6 −(ℓ2 · 5)(2 · 4)(3 · 6) + ϵ(ℓ2 , 3, 4, 5, 6) (2 · 4 · 6)/(4 · 6),

:= Nc (13, 24) = −(1 · 3)(2 · 4)/2.

1, 3 2, 4
(3.67)
We denote the sum of these integrands for bipartite graphs in the first group as
I(135, 246), which is essentially the same as the six-point case:

I(135, 246) =Ia (135, 246) + [Ib (135, 24) + Ib [135, 46] + Ib (13, 246) + Ib (35, 246)]
(3.68)
+ [Ic (13, 24) + (13 → 35) + (24 → 46) + (13 → 35; 24 → 46)].

where e.g. Ia (135, 246) = Na (135, 246)/((ℓ1 · 1)(ℓ1 · 3)(ℓ1 · 5)(ℓ2 · 2)(ℓ2 · 4)(ℓ2 · 6). Similarly,
there is a second group of bipartite graphs, related to the first ones by cyclic shifts, and
we denote the sum of these integrands as

I(157, 268) = I(135, 246) (3.69)

(135,246)→(571,682)

– 31 –
 The third group of integrands consists of the following bipartite graphs with pole
structures denoted by (135, 268):
p p
:= Na (135, 268) = ϵ(ℓ1 , 1, 3, 5, µ )ϵ(ℓp2 , 2, 6, 8 µ ) − (ℓ1 · ℓ2 ) (1 · p3 · 5) (2 · 6 · 8)
1, 3, 5 2, 6, 8 −ϵ(ℓ1 , 1, 3, 5, ℓ2 ) (2 · 6 · 8) − ϵ(ℓ1 , 2, 6, 8, ℓ2 ) (1 · 3 · 5),


:= Nb (135, 28) = (ℓ1 · 1)(2 · 5)(3 · 8)/2 + (ℓ1 · 2) (1 · 3)(5 p · 8)−(1 · 5)(3 · 8) /2
1, 3, 5 2, 8 −(ℓ1 · 8)(1 · 3)(2 · 5) − ϵ(ℓ1 , 8, 1, 2, 3) (1 · 3 · 5)/(1 · 3),
p
(6 · 8) (1 · 5 · 7)  
:= Nb (135, 68) = ϵ(ℓ1 , 1, 3, 6, 8)(3 · 5) − ϵ(ℓ1 , 1, 3, 5, 8)(3 · 6)
1, 3, 5 6, 8 (1 · 7)(5 · 7)(3
p · y5,8 )
+(ℓ2 · y5,8 ) (1 · 3 · 5 · 7)



:= Nb (13, 268) = (ℓ2 · 1) (2 · 8)(3 p· 6)−(2 · 6)(3 · 8) /2 − (ℓ2 ↔ 6)
1, 3 2, 6, 8 −ϵ(ℓ2 , 8, 1, 2, 3) (2 · 6 · 8)/(2 · 8),
p
(3 · 5) (2 · 4 · 6)  
:= Nb (135, 68) = − ϵ(ℓ2 , 6, 8, 3, 5)(2 · 8) − ϵ(ℓ2 , 6, 8, 2, 5)(3 · 8)
3, 5 2, 6, 8 (2 · 4)(4 ·p 6)(8 · y2,5 )
+(ℓ2 · y2,5 ) (2 · 4 · 6 · 8),
:= Nc (13, 28) = −(1 · 3)(2 · 8)/2, := Nc (35, 68) = (3 · 5)(6 · 8)/2.
1, 3 2, 8 3, 5 6, 8
(3.70)
The full integrand is given by the sum of these building blocks:

I(135, 268) =Ia (135, 268) + [Ib (135, 28) + Ib (135, 68) + Ib (13, 268) + Ib (35, 268)]
(3.71)
+ [Ic (13, 28) + Ic (35, 68)]
and finally the integrand for the last group is related by cyclic shift:

I(157, 246) = I(135, 268) (3.72)

(135,268)→(571,624)

The complete loop form associated with the chamber 1 ∩ 2 is given by the sum of these
four groups: I(135, 246) + I(157, 268) + I(135, 268) + I(157, 246) + (ℓ1 ↔ ℓ2 ).

Comments on all loops The bipartite structures of negative geometries, together with
the even/odd pole structures, have important implications for all loops. As we have seen
for L = 2 they greatly simplify canonical forms of these loop geometries, and they impose
stronger and stronger constraints for higher L. Exactly as in the n = 4 case, if we take the
log of the loop form for each chamber Ω̃α as the sum of (connected) negative geometries,
our prediction is that for L = 3 only chain topologies are needed for all multiplicities since
triangle (or any odd-gon) cannot be made bipartite. Similarly, we have chain, star, and
box topologies for L = 4 and 5 out of 21 topologies for L = 5 [52]. Moreover, in each case,
the black/white node only has odd/even poles as we have seen for L = 2, which reduces
the space of possible integrands enormously. For example, we expect that the L = 3, n = 6
integrand can be determined very similar to what we have done for L = 2. Since only
chain graphs are needed for the negative geometries, we can write down an ansatz with (a
subset of) 1, 3, 5 assigned for any black node, and (a subset of) 2, 4, 6 for any white node;
such constraining mutual conditions and pole structures make it very plausible that the
full L = 3 integrand can be computed from geometries without the need of other input.
We leave this computation (and possible ones for higher n and L) to future works.

– 32 –
 Last but not least, as we have discussed for n = 4 case in [45], these bipartite structures
are clearly related to the special feature that no odd-point amplitudes exist in ABJM
theory. As we have seen earlier, the fact that three-particle and five-particle cuts for all-
loop, all-multiplicity amplitudes vanish clearly follow from the geometry. More generally,
we find that arbitrarily complicated all-loop cuts of n-point geometry must vanish as long
as it isolates any (purely internal or partly external) odd-point amplitude. For purely
internal case with only Dij cut, this is simply because any bipartite graphs (which applies
to negative geometries for all n) cannot contain an odd cycle; for general cases involving
certain x, y, z, w cut, one can also generalize the argument given in [45] to all n.

4 Beyond positive solutions and parity action

One of the central themes of this paper is the dissection of Tn into chambers by analyzing
the positivity of the solution to,
X
(DΓ )α,i Zi = 0 . (4.1)
i

As mentioned previously, there are in general more than one solution and the remaining
are crucial in obtaining the full amplitude. For example at six-points, where the BCFW
cell is the top cell and there’s no dissection, we have two solutions corresponding to the
positive and negative branch of the Grassmannian. Denoting the number of solutions as
Γ[DΓ ], for BCFW cells in ABJM theory we have Γ[DΓ ] = 2k . Note that this is distinct
from N = 4 sYM, as one finds Γ[D] = 1.4 This can be attributed to the fact that BCFW
shifts in four-dimensions are linear in the deformation variable, while quadratic for three-
dimensions:
i i+1
i i+1

pi −zq pi+1 +zq pi,i+1 −zq−z −1 q̃ pi,i+1 +zq+z −1 q̃

There is another way to understand the 2k counting for ABJM. Each tree-level BCFW
term can be represented as an on-shell diagram with k-loops. Recall that each on-shell di-
agram can be understood as representing multi-loop maximal cut for a set of fixed external
kinematics. As there are two solutions for each loop we have 2k solutions.
Since for each loop the cut solutions are that to a quadratic equation, they are simply
related by taking different branches of the square root of the discriminant. For example
4
There are certainly non-BCFW cells for N = 4 that has Γ[DΓ ] > 1. We will discuss their relevance for
the dissection of Tn in the outlook.

– 33 –
for the following maximal cut
3 4

2 5

1 6
the square root is simply
p
−(p1 +p2 )2 (p3 +p4 )2 (p5 +p6 )2 = ⟨12⟩⟨34⟩⟨56⟩ (4.2)

thus to change between the two solutions will require us to change a sign in front of the
square root, which in this case can be achieved by simply applying a parity operation Π6
Zi → −Zi on legs {2, 4, 6} (equivalently {1, 3, 5}). In the following, we will use the six-
and eight-point examples to demonstrate how combining eq. (3.3) and parity operations
generate the full amplitude.

4.1 The complete six-point amplitude

Previously we’ve motivated the parity operation from interchanging between solutions of
the maximal cut. Here we will directly see how the action acts on the analytic solution
of the BCFW cell, which is straightforward at six-point since the cell is a top cell. At
six-point Γ[C6 ] = 2, and the positive branch is given as:
 
λ11 −λ12 λ13 −λ14 λ15 −λ16
C6,+ =  λ21 −λ22 λ23 −λ24 λ25 −λ26  . (4.3)
 
⟨35⟩ ⟨46⟩ ⟨51⟩ ⟨62⟩ ⟨13⟩ ⟨24⟩

Indeed one can check that with the two brackets living in Kn , i.e. eq. (2.7), C6,+ ∈ OG>0 .
The other solution is given as:
 
λ11 −λ12 λ13 −λ14 λ15 −λ16
 2
C6,− =  λ1 −λ22 λ23 −λ24 λ25 −λ26  . (4.4)

⟨35⟩ −⟨46⟩ ⟨51⟩ −⟨62⟩ ⟨13⟩ −⟨24⟩

Indeed the two cells can be interchanged by the action of Π6 (with overall scale a sign
on the columns 2, 4, 6). Similarly the corresponding leading singularities in eq. (3.40)
interchanges under Π6
{2,4,6}
LS6,+ ←−−−→ LS6,− . (4.5)
The fact that the two solutions are interchanged under Π6 , also tells us that the negative
solution actually becomes positive once Π6 acts on the external twistors. Since the solutions
also correspond to the solution of maximal cuts, the parity operation changes the solution
of the maximal cut. Said in another way, the negative cut solution lives on the boundary
of loop amplituhedron, i.e. eq. (3.3), if the external twistors are acted upon by Π6 .
Let us see this play out explicitly. The maximal cut requires:

⟨AB12⟩ = ⟨AB34⟩ = ⟨AB56⟩ = 0. (4.6)

– 34 –
The solution is given by

ZA = Z1 + αZ2 , ZB = βZ3 + Z4 (4.7)

with
p
⟨3561⟩ −⟨1234⟩⟨3456⟩⟨5612⟩ ⟨4561⟩ + α⟨2456⟩
α= ± , β=− . (4.8)
⟨2356⟩ ⟨⟨24⟩⟩⟨2356⟩ ⟨3561⟩ + α⟨2356⟩
We can see that the parity operator Π6 changes the relative sign in α, resulting in two so-
lutions interchanging. The positive cut solution will locate inside the amplituhedron (3.3),
while the negative one sits in the counterpart where the parity operator acts on the am-
plituhedron, explicitly the region defined by (with Ẑi := −Zi for i = 2, 4, 6) ⟨⟨AB⟩⟩ = 0,
⟨(AB)a (AB)b ⟩ < 0 and
⟨(AB)a 12̂⟩ < 0, ⟨(AB)a 2̂3⟩ < 0, ⟨(AB)a 34̂⟩ < 0,
⟨(AB)a 4̂5⟩ < 0, ⟨(AB)a 56̂⟩ < 0, ⟨(AB)a 6̂1⟩ < 0,
( ) (4.9)
⟨(AB)a 12̂⟩, ⟨(AB)a 13⟩, ⟨(AB)a 14̂⟩,
has 3 sign flip .
⟨(AB)a 15⟩, ⟨(AB)a 16̂⟩
Remarkably the associated canonical form from the above is the same as the unprojected
one except for a change of sign for the triangle integrals at one-loop in eq. (3.39) and
I 2mh , Ijbt at two-loops:
 
6
1
A2−loop
X X X
6 = LS6,+  − I critter (i) + I crab (i) + I+
2mh
(i) − I1bt (i) + I2bt (i)
2
i=1 i=1,3,5 i=2,4,6
 
6
X 1 X X
+LS6,−  − I critter (i) + I crab (i) + I−2mh
(i)+ I1bt (i)− I2bt (i)
2
i=1 i=1,3,5 i=2,4,6
(4.10)
Let’s understand this change of sign from two aspects. First, the momentum space image
of the Π6 is λi → (−)i−1 λi . Thus only spinor brackets ⟨ij⟩ are affected and not vector inner
products. As the only integrand in our local integrand basis that carries spinor brackets
are precisely the one-loop triangles and the two-loop box-triangles and (parity odd part
of) two-mass hard box (see eq. (3.25) and appendix D).
Before moving on to the eight-points, we note that we could have defined the parity
operation to act on legs {1, 3, 5} instead. As one can straightforwardly check that C6,±
interchanges amongst itself.

4.2 The complete eight-point amplitude

At eight points, Γ[C8 ] = 4 we have four solutions for any given cell. The relevant on-shell
diagrams are of the form
i+4

Γi i+2 i+6

i .

– 35 –
We would like to define a parity operation that interchanges the solutions of all four BCFW
cells Γi . Since the square root for each (two-loop) maximal cuts
q q
⟨ii+1⟩⟨i+2i+3⟩ −p2i+4,i+5,i−2,i−1 , ⟨i+4i+5⟩⟨i−2i−1⟩ −p2i,i+1,i+2,i+3 (4.11)
P
(pI := i∈I pi ), it is easy to see that parity flips {i, i+2, i+4}, {i, i+6, i+4} introduces
the requisite sign. We will choose Π8,1 : {2, 4, 6} and Π8,2 : {2, 6, 8}. To keep track of
how parity acts on the solutions for each cell, we denote the solutions as LS8,±,± [i], where
LS8,+,+ [i] is the positive solution for cell Γi . The remaining three are then defined through
the following sequence of projections:

LS8,+,+ [i]
Π8,1 Π8,2

LS8,+,− [i] Π8,3 Π8,1 Π8,2 LS8,−,+ [i] (4.12)

Π8,2 Π8,1
LS8,−,− [i]

where Π8,1 Π8,2 is equivalent to just flipping {4, 8}, which we will simply denote as Π8,3 .
Note that we will use the same operation for all four-cells.
At the eight-point, we have four chambers. The loop-form associated with the positive
sector is given for chamber 1 ∩ 2 in eq. (3.52) and (3.65). The form for chambers 3 ∩ 2,
3 ∩ 4, and 1 ∩ 4 are related by cyclic rotation. The full amplitude is then given by

A2−loop
X
8 = Bi × Ωi +Π8,1 [Bi × Ωi ] +Π8,2 [Bi × Ωi ] +Π8,3 [Bi × Ωi ] (4.13)
i

where i = 1 ∩ 2, 1 ∩ 4, 3 ∩ 2, 3 ∩ 4. In the positive sector, we identify

B1∩2 +B1∩4 = LS8,+,+ [1], B3∩2 +B3∩4 = LS8,+,+ [3],

B1∩2 +B3∩2 = LS8,+,+ [2], B1∩4 +B3∩4 = LS8,+,+ [4], (4.14)

Thus the action of parity operators on Bi can be read off from their action on LSs. The
action of Π8,i on the loop-forms is again obtained by computing the canonical form for the
region in eq. (3.3) with a subset of twistors replaced by Zi → Ẑi = −Zi . In the following
we give the loop-form associated with chamber 1 ∩ 2, the remaining are obtained by cyclic
shifts.
One-loop forms: The one-loop forms for chamber 1 ∩ 2 under parity operation Π8,1 are:
8
X
I8L=1 (0) = (−1)i Ibox (i−1, i, i+1) ,
i=1
(4.15)
I8L=1 (1) = Ibox (1, 3, 5) + Itri (1, 3, 5) + Ibox (5, 7, 1) − Itri (5, 7, 1) ,
I8L=1 (2) = −Ibox (2, 4, 6) + Itri (2, 4, 6) − Ibox (6, 8, 2) − Itri (6, 8, 2) ,

– 36 –
and Π8,2 ,
8
X
I8L=1 (0) = (−1)i Ibox (i−1, i, i+1) ,
i=1
(4.16)
I8L=1 (1) : Ibox (1, 3, 5) − Itri (1, 3, 5) + Ibox (5, 7, 1) + Itri (5, 7, 1) ,
I8L=1 (2) = −Ibox (2, 4, 6) − Itri (2, 4, 6) − Ibox (6, 8, 2) + Itri (6, 8, 2) ,

and Π8,3 ,
8
X
I8L=1 (0) = (−1)i Ibox (i−1, i, i+1) ,
i=1
(4.17)
I8L=1 (1) = Ibox (1, 3, 5) − Itri (1, 3, 5) + Ibox (5, 7, 1) − Itri (5, 7, 1) ,
I8L=1 (2) = −Ibox (2, 4, 6) − Itri (2, 4, 6) − Ibox (6, 8, 2) − Itri (6, 8, 2) .

Two-loop forms: The two-loop forms for chamber 1 ∩ 2 under parity operation Π8,1 is
given as,
8
X 1
I8L=2 (0) = db
−IA db
(i) + IB (i) + ICdb (i) ,
2
i=1
I8L=2 (1) = ID,+
db db
(2) + ID,− db
(6) + ID,+ db
(4) + ID,− db
(8) + IE,− db
(2) + IE,+ (6)
db db db db 1 db 1 db
+IE,− (4)+IE,+ (8)−IF,−,+ (3)−IF,+,− (7)+ IG,−,+ (3)+ IG,+,− (7)
2 2 (4.18)
bt bt bt bt bt bt
+ IA (1) − IA (5) − IA (1) + IA (5) − IB (3) + IB (7) ,
I8L=2 (2) = ID,+
db db
(3) + ID,− db
(7) + ID,− db
(1) + ID,+ db
(5) + IE,− db
(3) + IE,+ (7)
db db db db 1 db 1 db
+IE,+ (1)+IE,− (5)−IF,−,+ (4)−IF,+,− (8)+ IG,−,+ (4)+ IG,+,− (8)
2 2
bt bt bt bt bt bt
− IA (2) + IA (6) + IA (2) − IA (6) + IB (4) − IB (8) ,

and Π8,2 ,
8
X 1
I8L=2 (0) = db
−IA db
(i) + IB (i) + ICdb (i) ,
2
i=1
I8L=2 (1) = db
ID,− (2) db
+ ID,+ db
(6) + ID,− db
(4) + ID,+ db
(8) + IE,+ db
(2) + IE,− (6)
db db db db 1 db 1 db
+IE,+ (4)+IE,− (8)−IF,+,− (3)−IF,−,+ (7)+ IG,+,− (3)+ IG,−,+ (7)
2 2 (4.19)
bt bt bt bt bt bt
− IA (1) + IA (5) + IA (1) − IA (5) + IB (3) − IB (7) ,
I8L=2 (2) = ID,−
db db
(3) + ID,+ db
(7) + ID,+ db
(1) + ID,− db
(5) + IE,+ db
(3) + IE,− (7)
db db db db 1 db 1 db
+IE,− (1)+IE,+ (5)−IF,+,− (4)−IF,−,+ (8)+ IG,+,− (4)+ IG,−,+ (8)
2 2
bt bt bt bt bt bt
+ IA (2) − IA (6) − IA (2) + IA (6) − IB (4) + IB (8) ,

– 37 –
and Π8,3 ,
8
X 1
I8L=2 (0) = −IAdb db
(i)+IB (i)+ ICdb (i) ,
2
i=1
X X X
I8L=2 (1) = db
ID,− db
(i)+IE,− (i)+ db
ID,− db
(i)+IE,− (i)+ db
−IF,−,− (i)
i=2,6 i=4,8 i=3,7
X 1 X X
db bt bt bt
+ IG,−,− (i)+ IA (i)+IA (i)+ IB (i) , (4.20)
2
i=3,7 i=1,5 i=3,7
X X X
I8L=2 (2) = db
ID,− db
(i)+IE,− (i)+ db
ID,− db
(i)+IE,− (i)+ db
−IF,−,− (i)
i=3,7 i=1,5 i=4,8
X 1 X X
db bt bt bt
+ IG,−,− (i)− IA (i)+IA (i)− IB (i) .
2
i=4,8 i=2,6 i=4,8

The complete eight-point integrand: The complete one/two-loop integrand is

8
!
A1−loop
X X
8 = (LS8,a,b [1] + LS8,a,b [3]) × (−)i Ibox (i−1, i, i+1)
a,b=±,± i=1
(4.21)
X 4
X  
+ LS8,a,b [i] × Ia (i, i+2, i+4) + Ib (i+4, i+6, i)
a,b=±,± i=1

with I± (i, i+2, i+4) := (−)i−1 Ibox (i, i+2, i+4) ± Itri (i, i+2, i+4), and
8  
1 db
A82−loop
X X
db db
= (LS8,a,b [1]+LS8,a,b [3]) × −IA (i)+IB (i)+ IC (i)
2
a,b=±,± i=1
4
X 
db db db db db db
+ LS8,a,b [i] × ID,a (i+1)+ID,b (i+5)+ID,a (i+3)+ID,b (i+7)+IE,b (i+1)+IE,a (i+5)
i=1
db db db db 1 db 1 db
+IE,b (i+3)+IE,a (i+7)−IF,b,a (i+2)−IF,a,b (i+6)+ IG,b,a (i+2)+ IG,a,b (i+6)+(−)i bIA
bt
(i)
2 2 !

+(−)i aIAbt
(i+4)+(−)i aIA
bt
(i)+(−)i bIA
bt
(i+4)+(−)i aIB
bt
(i+2)+(−)i bIB
bt
(i+6) .

(4.22)

Prescriptive unitarity
The full eight-point amplitude provides a non-trivial example of how triangulating
through Tn gives a representation that satisfies prescriptive unitarity. More precisely, the
resulting form consists of integrands that evaluate to 1 on one of the maximal cut solutions,
and zero on the others. Let’s take the two-loop maximal cut as an example, which involves
4 solutions. Consider the cut

⟨AB12⟩ = ⟨AB34⟩ = ⟨AB56⟩ = ⟨CD56⟩ = ⟨CD78⟩ = ⟨CD12⟩ = 0. (4.23)

– 38 –
 db
The contributing integrand consist of IG,±,± (4) which appears in the loop-form associated
with chambers [1 ∩ 2] and [3 ∩ 2]. The cut solutions are

ZA = Z1 + αZ2 , ZB = βZ3 + Z4 ,
(4.24)
ZC = Z5 + γZ6 , ZD = δZ7 + Z8 ,

with
p
⟨1356⟩ −⟨1234⟩⟨3456⟩⟨5612⟩ ⟨1456⟩ + α⟨2456⟩
α± = − ± , β=− ,
⟨2356⟩ ⟨⟨24⟩⟩⟨2356⟩ ⟨1356⟩ + α⟨2356⟩
p (4.25)
⟨1257⟩ −⟨1256⟩⟨5678⟩⟨7812⟩ ⟨1258⟩ + γ⟨1268⟩
γ± = − ± , δ=− .
⟨1267⟩ ⟨⟨68⟩⟩⟨1267⟩ ⟨1257⟩ + γ⟨1267⟩

We denote the cut solution by cut±,± according to the sign of α± , γ± . One can straight-
forwardly check that
db
IG,±,± (4) = δ±,a δ±,b , (4.26)
cuta,b

as required.
Equipped with the explicit cut solution, one can further verify that the leading singu-
larity from the cut, matches associated tree-form B. Using the map in eq. (4.14), we see
that the tree form associated with B1∩2 +B3∩2 is LS8,+,+ [2]. The action of Π8,1 , Π8,2 and
Π8,3 then leads to LS8,−,+ , LS8,+,− and LS8,−,− respectively. Now let’s consider the gluing
of the five tree amplitudes in the cut:
5 6
4 7
k3 k4

k2 k5

k1 k6
3 8
2 1

Substituting the four cut-solutions one indeed finds:

δ (6) (Q1 ) δ (6) (Q2 ) δ (6) (Q3 ) δ (6) (Q4 ) δ (6) (Q5 )
Z Y
1
dηki
J ⟨23⟩⟨3k2 ⟩ ⟨45⟩⟨5k3 ⟩ ⟨67⟩⟨7k5 ⟩ ⟨81⟩⟨1k6 ⟩ ⟨k3 k4 ⟩⟨k4 k6 ⟩ (4.27)
ki
cut±,±

= LS8,±,± [2] .
q q
with Jacobian factor J = −p22,3 p24,5 p26,7,8,1 −p26,7 p28,1 p22,3,4,5 . Therefore, the tree form
exactly matches the leading singularity.

5 Conclusion and outlook

In this paper, we have proposed certain new positive geometries relevant for the n-point
all-loop planar integrand of ABJM theory. The canonical form of the all-loop ABJM
amplituhedron can be obtained in three steps:

– 39 –
 • With an overall sign flip, we obtain the tree positive region Tn by taking the di-
mensional reduction to D = 3 of that of n-point sYM amplituhedron, which turns
out to be non-empty only for middle sector k = n2 − 2; restricted on a suitable 3k-
dim subspace, we conjecture that it gives the tree ABJM amplituhedron, related via
T-duality to the ABJM momentum amplituhedron of [33, 34].

• By reducing the loop part to D = 3 in exactly the same way, we conjecture that the
geometry is the all-loop ABJM amplituhedron; concretely we use intersections of all
possible BCFW cells (chambers) to triangulate the 3L-dim loop geometry and obtain
a unique loop form for any points in a given chamber.

• By wedging the form of each tree chamber with the L-loop form, which turns out to
be given by local integrals, and summing over all chambers, we obtain the integrand
of the positive sector; we further include all the 2k sectors (k = 1, 2 for n = 6, 8)
related to the positive one by parity operations, and that gives the complete results
of ABJM L-loop integrands at n = 6, 8.

We have provided strong evidence for these claims, e.g. by showing how geometries guar-
antee all-loop vanishing cuts, soft cuts, and most importantly unitarity cuts. We have
triangulated loop geometry based on tree chambers and computed canonical forms for
L = 1, n = 6, 8, 10 (which involves summing over 1, 4, 25 chambers, respectively), and for
L = 2, n = 6, 8; in all cases, we obtain correct loop integrands, and the method is rather
efficient. We have also shown that the same bipartite structures of multi-loop negative
geometries seen earlier for n = 4 [44] extend nicely to all multiplicities and plays an im-
portant role in our two-loop computations. Perhaps the most immediate next step is to go
to higher multiplicities and loops: we expect to triangulate L = 2 geometries for n ≥ 10
once tree chambers are systematically understood; going to L = 3 for at least n = 6 poses
no more difficulties with bipartite structures the key in simplifying the computations.
There are many open questions to be clarified:

• Chambers: Currently, we have used BCFW cells and their overlap to triangulate
Tn in ABJM. Such triangulation is in a sense “hand-picked” by the loop geometry,
where each region yields a unique loop-form at given loop order. An obvious question
is then whether such BCFW triangulation is sufficient for all loops. The question
can be rephrased as whether all n−3 dimensional cells in the momentum orthogonal
Grassmannian corresponds to BCFW cells. Note that we have a negative answer
for the corresponding question for N = 4 sYM: indeed the four-mass box leading
singularity at N2 MHV [58]

– 40 –
 ,
is not a BCFW cell.

• Prescriptive unitarity for N = 4 sYM: Note that triangulations of the loop

amplituhedron using chambers of Tn,k can be straightforwardly applied to N =
4 sYM. We will use the one-loop 6-point NMHV amplitude as an example. It is
known that all 2n−4 cells for this configuration are BCFW cells, and hence we expect
that the chambers of NMHV T6,1 are comprised of the overlap of the BCFW cells.
Denoting the BCFW cells as [ i ], the tree amplitude is given by

[1] ∪ [3] ∪ [5] = [2] ∪ [4] ∪ [6]. (5.1)

Therefore, NMHV T6,1 can be partitioned into nine chambers:

1
3
1∩6
3∩6
1∩4
5
1∩2
5∩6
3∩4
6
3∩2
5∩4
5∩2
4
2
The canonical form of the overall geometry is equal to the sum of the canonical forms
of these nine chambers, each with its own loop geometry. The canonical form for the
chamber 1 ∩ 2 can be presented in the following local form,
 
 6 
 X 1mb 2mh
X
tri

[p ∩ q] × 
 Ij (i) + Ij (i) + I (i)
 . (5.2)
 i = {p, q} i=1 
j = {1, 2}

Here Ij1mb (i) and Ij2mh (i) refer to the one-mass box and two-mass-hard integrals,
which we list in appendix D for completeness. Using that [1] = [1∩2]+[1∩4]+[1∩6],

– 41 –
 and [2] = [1 ∩ 2] + [3 ∩ 2] + [5 ∩ 2], summing with the other eight chambers, we arrive
at the following NMHV6 one-loop integrands
 
6 6
1
A1−loop
X X X
6 = [i] ×  Ij1mb (i) + Ij2mh (i)+ I tri (k) . (5.3)
2
i=1 j=1,2 k=1

Note that all leading singularities are present, reflecting the fact such triangulations
leads to representation satisfying prescriptive unitarity.
What is more interesting is the eight-point N2 MHV case, where the four-mass-box
leading singularity has Γ[C] = 2. Since we only expect a single positive branch
for N = 4, the two solutions must participate in the dissection of T8,2 . It will
be interesting to see how prescriptive unitarity is realized through proper chamber
analysis for T8,2 [54].

• Soft-collinear singularities: The geometric definition of the loop integrand renders

many physical properties manifest, including the recursive nature of soft-cuts, the
correct factorization of maximal cuts as well as vanishing cuts. However, there is
a glaring exception, the absence of non-factorizable soft-collinear singularities. In
fact, it is this property, combined with the previously mentioned cuts, that fixes the
integrand at two-loops up to eight point [51, 52]. It will be interesting to understand
this further.

• Parity operation for general n: As we have discussed for k = n2 −2, we need k

parity operators. We have found the correct operators for k = 1, 2. However, when
generalizing to the higher multiplicities, some situations where changing the sign on
Zi s does not interchange the solutions. For example, at n = 12 the following maximal
cut
5 8

4 9

1 12
its square root is simply q
−p21,2,3,4 p25,6,7,8 p29,10,11,12 (5.4)
which could not be written as products of spinor brackets. It is important to figure
out how to modify and identify the parity operators in general and how their inclusion
can be understood as a necessity from a geometric point of view.

• The ABJM tree amplituhedron: Finally, the existence of loop amplituhedron

for ABJM suggests that there is a well-defined subspace for Tn whose canonical form
gives the tree amplitude. It would be nice to have this construction.
This would also help address our observation that the loop canonical form constructed
from a given block in Tn only exhibits local poles associated with on-shell diagrams

– 42 –
 that defined the block. For N = 4 sYM, such phenomenon can be understood from
the fact that the amplituhedron subspace is defined as:
 
C
 Dℓ1 
 
Y = C · Z, C =  D 
 (5.5)
 ℓ2 

..
.

where (C, Dℓi ) are the tree and loop parameterization respectively. Since all ordered
(k+2ℓ) × (k+2ℓ) minors of C must be positive, it is natural that the positive cells of
C dictates the allowed boundaries for the associated form in D, as the former must
appear as the boundary of the form in D.

• BDS geometry: As observed for two-loops n = 4 [59], n = 6 [51], and n = 8 [52],

the infrared divergence of the two-loop amplitude is fully captured by the BDS ansatz
of N = 4 sYM [60]. Furthermore, the BDS piece is fully captured by a set of local
integrands that are characterized by the presence of IR divergence and free of all
unphysical cuts. Given that our triangulation exposes the possibility of local trian-
gulation of the amplituhedron, it would be interesting to see if one can characterize
a subregion of Ln that gives loop integrands for the BDS piece.

Acknowledgement

It is our great pleasure to thank Nima Arkani-Hamed, Jacob Bourjaily, Tomasz Lukowski,
Matteo Parisi, Justinas Rumbutis, Jonah Stalknecht, Jaroslav Trnka, Congkao Wen, Ak-
shay Yelleshpur, Shun-Qing Zhang, and Yaoqi Zhang for stimulating discussions. The
authors would also like to thank DIAS for hosting the conference “Amplituhedron at 10”
where the results of this paper were discussed. SH thanks IAS (Princeton) for hospitality
during the completion of the work; his research is supported in part by National Natural
Science Foundation of China under Grant No. 11935013, 12047502, 12047503, 12247103,
12225510. Y-t H and C-k Kuo are funded by the National Science and Technology Council
of Taiwan, Grant No. 111-2811-M-002 -125.

A Momentum-twistor Grassmannian formula for ABJM

This appendix provides a comprehensive review of the Grassmannian formula for ABJM
in momentum twistor space [50]. As discussed in section 3 and the outlook, the complete
tree geometry has not yet been established, and this formula may be crucial for such a
definition. It is possible that uplifting the formula to the product map with Z could yield
a subspace of the tree region, akin to the original amplituhedron in N = 4 sYM. Below we
have gathered some essential properties of this formula.
The momentum-twistor space Grassmannian formula is subject to the delta functions:
n
X n
X
Dᾱ,j Zj = 0, g̃i,j Dᾱ,i Dβ̄,j = 0, for ᾱ, β̄ = 1, · · · , k . (A.1)
j=1 i,j=1

– 43 –
The second constraint is the orthogonal condition defining OG(k, g 2k+4). Here we use f
to indicate the orthogonal condition for the Grassmannian is defined with respect to a kine-
matic dependent n×n metric, i.e. g̃ = [g̃]n×n is defined by g̃i,j ≡ gi,j /((−)k det [g]n×n )1/(k+1)
with gi,j = sgn[j − i]⟨⟨ij⟩⟩, det [g]n×n = 22k ni=1 ⟨⟨i i+2⟩⟩.
Q

The R-invariant-like ABJM amplitude (in which a prefactor has been extracted from
the amplitude) can be defined as the contour integral over the Grassmannian G(k, n):

dk×n Dᾱ,j
Z
1
δ k(k+1)/2 D · g̃ · DT δ 4k (D · Z)δ 3k (D · χ),


Rn = (A.2)
γ Vol(GL(k)) m1 m2 · · · mk

where ml represent the l-th consecutive minors:

i1 i2 ik
ml ≡ ϵi1 ,i2 ,··· ,ik Dl+1 Dl+2 · · · Dl+k . (A.3)

To localize the integral (A.2), the matrix D needs to be restricted to 4k cells (after solving
the orthogonal condition).
Compared to the original formula in the momentum space, the matrix D is projected
out λ-plane. In momentum space, We fix the first two rows of C matrix to be (−1)i−1 λ
 
λ11 −λ12 · · · λ1n−1 −λ1n
 λ2 −λ2 · · · λ2 2
 1 2 n−1 −λn 
· · · c1,n−1 c1,n  ,
 
C= c c (A.4)
 1,1 1,2
 .. .. . . .. .. 

 . . . . . 
ck,1 ck,2 · · · ck,n−1 ck,n

and it is subject to the conditions:

n
X n
X
Cα,j λj = 0, (−1)j−1 Cα,j Cβ,j = 0, for α, β = 1, · · · , k + 2 . (A.5)
j=1 j=1

For α and β equal to 1 or 2, the delta function conditions are automatically satisfied as a
consequence of momentum conservation, i (−1)i−1 λai λbi = 0. The integral representation
P

of amplitude in λ-space is

d(k+2)×n Cα,i
Z
1
Ln = δ (k+2)(k+3)/2 (C · [(−1)a δa,b ]n×n · C T )
γ Vol(GL(k+2)) M1 M2 · · · Mk (A.6)
2(k+2) 3(k+2)
δ (C · λ)δ (C · η).

where Ml represent the l-th consecutive minors:

i
Ml ≡ ϵi1 ,i2 ,··· ,ik+2 Cli1 Cl+1
i2 k+2
· · · Cl+k+1 . (A.7)

The C matrix and D matrix are related by the projection matrix Qa,b , given by

⟨b−1 b+1⟩δa,b + ⟨b b−1⟩δa,b+1 + ⟨b+1 b⟩δa,b−1

Qa,b = (A.8)
⟨b+1 b⟩⟨b b−1⟩

– 44 –
through X
Dᾱ,j = (−1)i+j−1 Cᾱ,i Qi,j . (A.9)
i=1,··· ,n
It’s worth noting that the projection through Qa,b preserves the cell structure and trans-
forms the positive matrix C into a positive matrix D. Additionally, there is a relationship
between the minors of C and D, and we define the ratio of them to be
Ml
Jl ≡ = ⟨l, l+1⟩⟨l+1, l+2⟩ · · · ⟨l+k, l+k+1⟩. (A.10)
ml
The two integral formulas, Rn and Ln , are related up to a Jacobian factor and delta
functions that enforce momentum conservation and super momentum conservation,
1
Ln = δ 3 (p)δ (6) (Q) Q Rn . (A.11)
i=1,··· ,k+2 Ji
Specifically, for n = 6
δ 3 (p)δ (6) (Q) d6 Di
Z
1
δ D · g̃ · DT δ 4 (D · Z)δ 3 (D · χ), (A.12)


L6 =
⟨12⟩⟨23⟩2 ⟨34⟩2 ⟨45⟩ Vol(GL(1)) D2 D3 D4
In momentum space, since the minors and their conjugates are equal, the minors of the
positive matrix D and their conjugates are closely related. This allows us to choose al-
ternative sets of minors in the denominator of the Grassmannian formula to simplify the
Jacobian factor. Specifically, we can write
δ 3 (p)δ (6) (Q) d6 Di
Z
1
δ D · g̃ · DT δ 4 (D · Z)δ 3 (D · χ).


L6 = Q (A.13)
i ⟨i i+1⟩ Vol(GL(1)) D2 D4 D6
To facilitate the comparison between the two formulas, we consider the component
amplitudes. For n = 6, we have
1 M13 1 J3 1
= Q 1 ; (A.14)
2 M1 M 2 M3 C=C ∗ 2 i=1,2,3 Ji m1 m2 m3
D=D∗
and for n = 8 in the cell M1 = 0 (m1 = 0) for C (D),
1 M33 J1 1 m33
q = q Q . (A.15)
4 −p21,2,3,4 M2 M3 M4 C=CM ∗
1 =0
4 −p21,2,3,4 i=1,2,3,4 Ji m2 m3 m4 D=Dm∗
1 =0
 q −1  q −1
In the equation above, the terms 4 −p21,2,3,4 and J1 4 −p21,2,3,4 represent the
Jacobian factors resulting from localizing the delta function in eqs. (A.5) and (A.1), re-
spectively.
We can see that the selection of minors from the D matrix has a significant impact
on the overall pulling-out factor. Opting for a more symmetric approach, such as choosing
each minor to be half the size, can yield favorable results:
dk×n Dᾱ,j
Z
1
δ k(k+1)/2 D · g̃ · DT δ 4k (D · Z)δ 3k (D · χ), (A.16)


Rn = √
γ Vol(GL(k)) m1 m2 · · · mn

with prefactor 1/(⟨12⟩⟨23⟩ · · · ⟨n1⟩)(k+1)/2 such that

 (k+1)/2
3 (6) 1
Ln = δ (p)δ (Q) Rn . (A.17)
⟨12⟩⟨23⟩ · · · ⟨n1⟩

– 45 –
B Sectors of Tn for even n

In Section 2 we have shown that Tn is an empty set when n is odd; here we would like to
demonstrate that for even n Tn only exists for the middle sector with k = n2 − 2.
Recall that in D = 3 we can factor out a common factor ⟨⟨a−1 a+1⟩⟩ for each sign-

flip sequence ⟨a−1 a a+1 j⟩=⟨⟨a−1 a+1⟩⟩⟨⟨a j⟩⟩ , without affecting the sign-flip number.
Therefore, we can translate the sign-flipping condition on the sequences of double two-
brackets by using twisted cyclic symmetry with Zn+i ≡ (−1)k−1 Zi ≡ Ẑi :
 a+n−2
⟨⟨aj⟩⟩ j=a+2
for a = 1, · · · , n. (B.1)

It is useful to arrange the sign-flip sequence into a zigzag shape, as shown in eq. (B.4).
The definite sign of the boundaries ensures that the minor of every 2 × 2 adjacent block is
strictly negative:

⟨⟨a i⟩⟩ ⟨⟨a i+1⟩⟩ ⟨⟨a i⟩⟩⟨⟨a+1 i+1⟩⟩−⟨⟨a i+1⟩⟩⟨⟨a+1 i⟩⟩

··· ··· ⇒ (B.2)
⟨⟨a+1 i⟩⟩ ⟨⟨a+1 i+1⟩⟩ = ⟨a a+1 i i+1⟩ < 0.

This also implies that 2 × 2 adjacent blocks with the following signs are forbidden:
! ! ! !
+ − + + − − − +
, , , . (B.3)
+ + − + + − − −

We will see that for even n, only the middle k is a nonempty set in Tn , which be easily
shown in the case of n = 6:

⟨⟨13⟩⟩ ⟨⟨14⟩⟩ ⟨⟨15⟩⟩

+ (−)k
⟨⟨24⟩⟩ ⟨⟨25⟩⟩ ⟨⟨26⟩⟩
(B.4)
− ? (−)k+1
⟨⟨35⟩⟩ ⟨⟨36⟩⟩ ⟨⟨31̂⟩⟩
+ (−)k

The sequences {⟨⟨24⟩⟩, ⟨⟨25⟩⟩, ⟨⟨26⟩⟩} and {⟨⟨51⟩⟩, ⟨⟨52⟩⟩, ⟨⟨53⟩⟩} always have a conjugate sign
flip (k1 + k2 = n − 4), regardless of the sign of ⟨⟨25⟩⟩. Therefore, the middle sector
k = n/2 − 2 = 1 is the only allowed one.

Moving to the case n = 8, we can always show that if a sign-flip sequence is given
for the rows, a column sequence with the conjugate sign-flip number exists in this zigzag

– 46 –
arrangement, for any non-middle k. Specifically, when k = 0:

⟨⟨8̂6⟩⟩
−
⟨⟨13⟩⟩ ⟨⟨14⟩⟩ ⟨⟨15⟩⟩ ⟨⟨16⟩⟩ ⟨⟨17⟩⟩
+ + + + +
⟨⟨24⟩⟩ ⟨⟨25⟩⟩ ⟨⟨26⟩⟩ ⟨⟨27⟩⟩ ⟨⟨28⟩⟩
(B.5)
− − − − −
⟨⟨35⟩⟩ ⟨⟨36⟩⟩ ⟨⟨37⟩⟩ ⟨⟨38⟩⟩ ⟨⟨31̂⟩⟩
+ + + + +
⟨⟨46⟩⟩
−

By requiring that the sequences {⟨⟨1i⟩⟩}, {⟨⟨2i⟩⟩}, and {⟨⟨3i⟩⟩} have zero sign-flips, each
sequence becomes monotonic in sign. This, in turn, forces the sequence {⟨⟨6i⟩⟩} to be
alternating, with the maximum number of sign-flips.
When k = 4:

⟨⟨8̂6⟩⟩
−
⟨⟨13⟩⟩ ⟨⟨14⟩⟩ ⟨⟨15⟩⟩ ⟨⟨16⟩⟩ ⟨⟨17⟩⟩
+ − + − +
⟨⟨24⟩⟩ ⟨⟨25⟩⟩ ⟨⟨26⟩⟩ ⟨⟨27⟩⟩ ⟨⟨28⟩⟩
(B.6)
− + − + −
⟨⟨35⟩⟩ ⟨⟨36⟩⟩ ⟨⟨37⟩⟩ ⟨⟨38⟩⟩ ⟨⟨31̂⟩⟩
+ − + − +
⟨⟨46⟩⟩
−

Similar to its conjugate case k = 0, imposing that the three sequences {⟨⟨1i⟩⟩}, {⟨⟨2i⟩⟩},
and {⟨⟨3i⟩⟩} have the maximum number of sign-flips results in each sequence becoming
alternating in sign. As a consequence, the sign-flip number of the sequence {⟨⟨6i⟩⟩} is
guaranteed to be zero.
When k = 1, it is more convenient to divide it into two cases according to the sign of
⟨⟨26⟩⟩:

– 47 –
Case 1: ⟨⟨26⟩⟩ > 0

⟨⟨8̂6⟩⟩
+
⟨⟨13⟩⟩ ⟨⟨14⟩⟩ ⟨⟨15⟩⟩ ⟨⟨16⟩⟩ ⟨⟨17⟩⟩
+ − −
⟨⟨24⟩⟩ ⟨⟨25⟩⟩ ⟨⟨26⟩⟩ ⟨⟨27⟩⟩ ⟨⟨28⟩⟩
(B.7)
− ? + + +
⟨⟨35⟩⟩ ⟨⟨36⟩⟩ ⟨⟨37⟩⟩ ⟨⟨38⟩⟩ ⟨⟨31̂⟩⟩
+ ? −
⟨⟨46⟩⟩
−

Requiring the sequence {⟨⟨2i⟩⟩} has one sign-flip, ⟨⟨27⟩⟩ is determined to be negative. Using
the constraint that any 2 × 2 block’s minor is negative, the sign of ⟨⟨16⟩⟩ is then determined
to be positive. Consequently, the sequence {⟨⟨6i⟩⟩} has 3 sign-flips.

Case 2: ⟨⟨26⟩⟩ < 0

⟨⟨8̂6⟩⟩
+
⟨⟨13⟩⟩ ⟨⟨14⟩⟩ ⟨⟨15⟩⟩ ⟨⟨16⟩⟩ ⟨⟨17⟩⟩
+ ? −
⟨⟨24⟩⟩ ⟨⟨25⟩⟩ ⟨⟨26⟩⟩ ⟨⟨27⟩⟩ ⟨⟨28⟩⟩
(B.8)
− − − ? +
⟨⟨35⟩⟩ ⟨⟨36⟩⟩ ⟨⟨37⟩⟩ ⟨⟨38⟩⟩ ⟨⟨31̂⟩⟩
+ + −
⟨⟨46⟩⟩
−
This case ⟨⟨25⟩⟩ is determined to be negative and ⟨⟨36⟩⟩ to be positive, which is deduced
from the sequence {⟨⟨2i⟩⟩} has one sign flips and any minor of 2 × 2 block is negative. We
conclude that the sign-flip number of sequence {⟨⟨6i⟩⟩} is 3.
When k = 3, we can perform a similar analysis as for k = 1, and conclude that the
column sequence {⟨⟨6i⟩⟩} has only one sign flip. The corresponding zigzag diagrams and
their sign are displayed below:

– 48 –
Case 1: ⟨⟨26⟩⟩ > 0
⟨⟨8̂6⟩⟩
+
⟨⟨13⟩⟩ ⟨⟨14⟩⟩ ⟨⟨15⟩⟩ ⟨⟨16⟩⟩ ⟨⟨17⟩⟩
+ + −
⟨⟨24⟩⟩ ⟨⟨25⟩⟩ ⟨⟨26⟩⟩ ⟨⟨27⟩⟩ ⟨⟨28⟩⟩
(B.9)
− ? + − +
⟨⟨35⟩⟩ ⟨⟨36⟩⟩ ⟨⟨37⟩⟩ ⟨⟨38⟩⟩ ⟨⟨31̂⟩⟩
+ ? −
⟨⟨46⟩⟩
−
Case 2: ⟨⟨26⟩⟩ < 0
⟨⟨8̂6⟩⟩
+
⟨⟨13⟩⟩ ⟨⟨14⟩⟩ ⟨⟨15⟩⟩ ⟨⟨16⟩⟩ ⟨⟨17⟩⟩
+ ? −
⟨⟨24⟩⟩ ⟨⟨25⟩⟩ ⟨⟨26⟩⟩ ⟨⟨27⟩⟩ ⟨⟨28⟩⟩
(B.10)
− + − ? +
⟨⟨35⟩⟩ ⟨⟨36⟩⟩ ⟨⟨37⟩⟩ ⟨⟨38⟩⟩ ⟨⟨31̂⟩⟩
+ − −
⟨⟨46⟩⟩
−
We have also performed similar analysis for the sequences {⟨⟨ai⟩⟩} at n = 10 and
n = 12, and found that given a sequence has k sign-flips it is impossible to require all 2 × 2
block minors to be positive for any k ̸= n/2 − 2. This result strongly suggests that only
the middle sector with k = n/2 − 2 has a non-empty positive tree region Tn for any even
n.

Claim The positive tree region Tn exists only for even n and is located at the middle
sector k = n2 − 2.

Positive kinematics for 3d sYM As emphasized, the negative sign of ⟨ii+1jj+1⟩ < 0
plays a crucial role in excluding non-physical kinematic configurations in ABJM. If we
were to stay with positive ⟨ii+1jj+1⟩ > 0, one could find solutions for all n and k. This
is not surprising, since one would be just considering the geometry in a three-dimensional
subspace, i.e. 3d sYM.
It is easy to see that solutions exist for all n and k. If we revert back to ⟨ii+1jj+1⟩ > 0,
then the conditions (3.14) will be modified to consecutive even and odd brackets have a
uniform sign:
⟨⟨13⟩⟩, ⟨⟨35⟩⟩, · · · , (−)k−1 ⟨⟨n−11⟩⟩ ≷ 0,
(B.11)
⟨⟨24⟩⟩, ⟨⟨46⟩⟩, · · · , (−)k−1 ⟨⟨n2⟩⟩ ≷ 0.

– 49 –
Thus we no longer have those contradictions that lead to the absence of kinematic config-
urations for odd n or k ̸= n2 − 2.
Flipping the sign of ⟨ii+1jj+1⟩ results in a consistent way of satisfying symplectic
conditions for all n and k. Nonetheless, it seems that such a modification makes it difficult
to extend the kinematic region to loops. One possible reason for this is that 3d sYM is no
longer dual conformal. Also, it remains an open question whether a subspace exists within
this positive 3d kinematic region that is directly related to the dimensional reduction sYM
amplitude. This interesting problem is left for future investigation.

C Details on two-particle cuts

In this appendix, we will give details about two-particle cuts at the form level based on the
discussion on regions, which will allow us to verify unitarity explicitly for generic n. The
form will factorize as region and should satisfy
L−1
dz Y
Ba Ω(L)
a = ⟨(AB)l d2 Al ⟩⟨(AB)l d2 Bl ⟩δ(ΩIJ AIl BlJ )
g(z)
l=1 (C.1)
(L1 )
X
Bb Ωb (Z1 , · · ·, Zi , A, B, Zj−1 , · · ·, Zn ) × Bc Ω(L
c
2)
(A, Zi+1 , · · · , Zj , B).
allowed sets

The factorization of the loop form follows trivially from eq. (3.20) since the uncut loop
attaching to the left and right regions are disjoint ({L1 } ∩ {L2 } = ∅). So, the inequalities
of the uncut loop on the left and right sets are unaffected by each other, and their form
just trivially product together i.e. Ωb × Ωc . However, the external left and right sets share
the same twistor variables {Zi , A, B}, thus their tree forms are not trivially factorized.
Besides, the lack of a subspace prevents us from directly obtaining the tree form, which
makes identifying the tree form of the two-cut as the product of amplitudes more involved.
Here we will finish the computation of the forms for n = 6, 8 cases discussed in section 3
and focus on the form of tree region and loop (AB). We compare the result with gluing
super-amplitudes associated with the positive solution in the momentum space.

Unitarity at n = 6. The inequalities involving the twistors {Zi , A, B} in the double-

cut geometry ⟨AB23⟩=⟨AB45⟩=0 can be replaced by the external left and right sets L =
{1, 2, A, B, 5, 6}, R = {A, 3, 4, B}, living in the region of amplituhedron:
Left: ⟨i i+1 j j+1⟩L < 0, {⟨123i⟩L } has 1 sign flip, ⟨⟨ii+1⟩⟩L = 0.
(C.2)
Right: ⟨1234⟩R < 0, ⟨⟨ii+1⟩⟩R = 0.
We will denote the resulting one-form for this 6-4 channel as ω6 .
From the generalized unitarity, the double cut of loop integrand equals the left and
right amplitude integrate out the cut momentum and Grassmann variables
δ (6) (QL )
P P

p̄,q̄,r̄=,2,4,6 ⟨p̄q̄⟩L ηr̄,L + p,q,r=,1,3,5 ⟨pq⟩L ηr,L
Z
3 3
LS6,+ ω6 = d ηk1 d ηk2
cL L L
2,5 c4,1 c6,3
(C.3)
δ (6) (QR ) 2 2 3
· δ(k1 )δ(k2 ) d ℓ
⟨12⟩R ⟨23⟩R

– 50 –
 P
where cr̄s = ⟨r̄−2 r̄+2⟩ − i=1,3,,5 ⟨r̄i⟩⟨si⟩, and the labelling L, R imply that the spinor
variables is taking from λL = {λ1 , λ2 , k1 , k2 , λ5 , λ6 }, λR = {−k1 , λ3 , λ4 , −k2 }. On the RHS
the Jacobian factor from solving the cut condition can be expressed as

1 ⟨23⟩⟨45⟩
δ(k12 )δ(k22 ) d3 ℓ = dz (C.4)
⟨⟨35⟩⟩z ⟨ABI⟩

where we’ve used the cut solution in eq. (3.18).

To convert the generalized unitarity result into factorization form in twistor space, we
pull out the prefactor of LS6,+ and two tree amplitudes in the RHS, which is defined in
appendix A:
!
δ 3 (p)δ (6) (Q) δ(D · χ)
LS6,+ = Q6 √ , (C.5)
i=1 ⟨i i+1⟩
D1 D2 D3 D4 D5 D6 [D]1×6 Z=0, D·g̃·DT =0

δ (6) (QL )
P P

p̄,q̄,r̄=,2,4,6 ⟨p̄q̄⟩L ηr̄,L + p,q,r=,1,3,5 ⟨pq⟩L ηr,L
cL L L
2,5 c4,1 c6,3
! (C.6)
δ 3 (pL )δ (6) (QL ) δ(L · χL )
= Q6 √ ,
i=1 ⟨i i+1⟩
L1 L2 L3 L4 L5 L6 [L]1×6 Z=0, L·g̃L ·LT =0

and the four-point amplitude becomes unity after stripping off the prefactor.
Intermixing fermionic variables η and χ in the equation requires carefully dealing
transformation factor of them when integrating ηk1 , ηk2 . To simplify the calculation, we
choose the component amplitude A6 (ϕ̄, ϕ, ψ̄, ψ, ψ̄, ϕ) by integrating out η1 , η2 , and η6 . The
contribution of this component can be easily related to each other in terms of the minors
of Grassmannian formula in momentum and that in momentum-twistor space:

(612) = ⟨61⟩⟨23⟩D1 , (612)L = ⟨61⟩L ⟨23⟩L L1 . (C.7)

By collecting a bunch of prefactors and Jocabian factors into the measure function
g(z), the remaining part becomes

D13 dz L31
√ [D]1×6 Z = 0
ω6 = · √ [L]1×6 ZL = 0 (C.8)
D1 D2 D3 D4 D5 D6 g(z) L1 L2 L3 L4 L5 L6
T
D·g̃·D = 0 L·g̃·LT = 0
q
1 ⟨23⟩
here the measure function g(z) = x(z)z 3/2 ⟨45⟩ is coming from

3 6
!−1 6
!
⟨12⟩3R

⟨61⟩L ⟨12⟩L ⟨23⟩L Y 1 1 Y
⟨61⟩⟨12⟩⟨23⟩ ⟨i i+1⟩ ⟨i i+1⟩L ⟨12⟩R ⟨23⟩R
i=1 i=1 (C.9)
 −1/2
1 ⟨23⟩⟨45⟩ ⟨ABI⟩
× .
⟨⟨35⟩⟩z ⟨ABI⟩ ⟨k1 k2 ⟩

We note that the last term in the contribution of g(z) is the scaling factor of A, B to the
cut momentum k1 , k2 such that the first component of A, B can be identified as spinors.

– 51 –
Unitarity at n = 8. We can use a similar strategy for n = 8. The difference is that the
eight-point has four chambers. Each chamber will give a partial contribution in eq. (C.1) of
the full amplitude. For example, the 8-4 channel from the double cut ⟨AB34⟩ = ⟨AB56⟩ = 0
requires to sit the kinematic on (1 ∩ 2), (3 ∩ 2), (1 ∩ 4), and (3 ∩ 4) and the inequalities are
replaced by
Left: ⟨i i+1 j j+1⟩L < 0, {⟨123i⟩L } has 2 sign flip, ⟨⟨ii+1⟩⟩L = 0.
(C.10)
Right: ⟨1234⟩R < 0, ⟨⟨ii+1⟩⟩R = 0 .
with external left and right sets L = {1, 2, 3, A, B, 6, 7, 8}, R = {A, 4, 5, B}. The first two
chambers will give the same loop form, denoted by ω8 (2), and the last two chambers will
give the loop form, denoted by ω8 (4).
In momentum space, the sum of the form (1 ∩ 2) and (3 ∩ 2) follows the unitarity and
gives

δ (12) (CL ηL )
Z
1
LS8,+,+ [2] ω8 (2) = d3 ηk1 d3 ηk2 q
p22,3,4,5 L (1234)L (3456)L (4567)L [CL ]+,+


4×8 λL =0, (2345)L =0
CLT ·CL =0.

δ (6) (QR )
· δ(k12 )δ(k22 ) d3 ℓ
⟨12⟩R ⟨23⟩R
(C.11)
Similarly, the sum of form (1 ∩ 4) and (3 ∩ 4) is

δ (12) (CL ηL )
Z
1
LS8,+,+ [4] ω8 (4) = d3 ηk1 d3 ηk2 q
(1234) (2345) (3456)
p24,5,6,7 L
L L L [CL ]+,+
4×8 λL =0, (2345)L =0
CLT ·CL =0.

δ (6) (QR )
· δ(k12 )δ(k22 ) d3 ℓ .
⟨12⟩R ⟨23⟩R
(C.12)
Here, the spinor variables λL = {λ1 , λ2 , λ3 , k1 , k2 , λ6 , λ7 , λ8 }, λR = {−k1 , λ3 , λ4 , k2 }. The
Jocobian factor from solving the cut condition is
1 ⟨34⟩⟨56⟩
δ(k12 )δ(k22 ) d3 ℓ = dz. (C.13)
⟨⟨46⟩⟩z ⟨ABI⟩
Taking the component by integrating out η6 , η7 , η8 , η1 , stripping off the prefactor and
collecting them into g(z) give

3
J2,D m37 dz J2,L mL 7
√ +,+ ω8 (2) = ·q
m1 m3 m4 m5 m7 m8 [D]2×8 Z = 0, g(z) mL mL mL mL mL mL [L]+,+ Z = 0,
1 3 4 5 7 8 L
D·g̃·DT =0, m2 =0. 2×8
L·g̃L ·LT =0, mL
2 =0.

3
J2,D m37 dz J2,L mL
7
√ [D]+,+
ω8 (4) = ·q
m1 m2 m3 m5 m6 m7 2×8 Z = 0, g(z) mL L L L L L
1 m2 m3 m5 m6 m7 [L]+,+
2×8 ZL = 0,
D·g̃·DT =0, m4 =0.
L·g̃L ·LT =0, mL
4 =0.

(C.14)

– 52 –
where D and L are chosen to be localized on the positive solutions. The Jocabian factors
are
 −1/2
⟨1256⟩ Y 1
J2,D =   ,
⟨12⟩⟨56⟩ ⟨i i+1⟩⟨i+1 i+2⟩⟨i+2 i+3⟩
i=2,6
 −1/2 !1/2
⟨1256⟩L
Y 1 ⟨ABI⟩
J2,L =  ,
⟨12⟩L ⟨56⟩L ⟨i i+1⟩L ⟨i+1 i+2⟩L ⟨i+2 i+3⟩L ⟨k1 k2 ⟩
i=2,6
 −1/2 (C.15)
⟨3478⟩ Y 1
J4,D =   ,
⟨34⟩⟨78⟩ ⟨i i+1⟩⟨i+1 i+2⟩⟨i+2 i+3⟩
i=4,8
 −1/2 !1/2
⟨3478⟩L
Y 1 ⟨ABI⟩
J4,L =  ,
⟨34⟩L ⟨78⟩L ⟨i i+1⟩L ⟨i+1 i+2⟩L ⟨i+2 i+3⟩L ⟨k1 k2 ⟩
i=4,8
q
1 ⟨34⟩
and the measure function g(z) = x3/2 z 3/2 ⟨56⟩ is coming from

3 8
!1/2
⟨14⟩3R

⟨78⟩L ⟨81⟩L ⟨12⟩L Y ⟨i i+1⟩⟨i+1 i+2⟩⟨i+2 i+3⟩
⟨78⟩⟨81⟩⟨12⟩ ⟨i i+1⟩L ⟨i+1 i+2⟩L ⟨i+2 i+3⟩L ⟨12⟩R ⟨23⟩R
i=1 (C.16)
1 ⟨34⟩⟨56⟩ ⟨ABI⟩ −1
 
× .
⟨⟨46⟩⟩z ⟨ABI⟩ ⟨k1 k2 ⟩
Now, we go beyond the n-4 factorization channel and consider 6-6 channel at n=8,
choosing the double cut to be ⟨AB23⟩ = ⟨AB67⟩ = 0. The double-cut geometry can be
replaced by the inequalities involving external left and right sets L = {1, 2, A, −B, −7, −8},
R = {A, 3, 4, 5, 6, B}, living in the region of amplituhedron:

Left: ⟨i i+1 j j+1⟩L < 0, {⟨123i⟩L } has 1 sign flip, ⟨⟨ii+1⟩⟩L = 0.

(C.17)
Right: ⟨i i+1 j j+1⟩R < 0, {⟨123i⟩R } has 1 sign flip, ⟨⟨ii+1⟩⟩R = 0.
The form on this channel only has support on (3 ∩ 2) and (3 ∩ 4). These two chambers
have the same loop form, denoted by ω8 (3).
We sum the form of double-cut geometry of the chambers (3 ∩ 2) and (3 ∩ 4), and it
satisfies
δ (6) (QL )
P P

p̄,q̄,r̄=,2,4,6 ⟨p̄q̄⟩L ηr̄,L + p,q,r=,1,3,5 ⟨pq⟩L ηr,L
Z
3 3
LS8,+,+ [3] ω8 (3) = d ηk1 d ηk2
cL L L
2,5 c4,1 c6,3
δ (6) (QR )
P P

p̄,q̄,r̄=,2,4,6 ⟨p̄q̄⟩R ηr̄,R + p,q,r=,1,3,5 ⟨pq⟩R ηr,R
· δ(k12 )δ(k22 ) d3 ℓ .
cR cR cR
2,5 4,1 6,3
(C.18)
Here, the spinor variables λL = {λ1 , λ2 , k1 , −k2 , −λ7 , −λ8 }, λR = {−k1 , λ3 , λ4 , λ5 , λ6 , −k2 },
and the Jacobian factor resulting from solving the cut condition can be written as
1 ⟨23⟩⟨67⟩
δ(k12 )δ(k22 ) d3 ℓ = dz . (C.19)
⟨⟨37⟩⟩z+⟨⟨36⟩⟩ ⟨ABI⟩

– 53 –
Notably, the term 1/(⟨⟨37⟩⟩z+⟨⟨36⟩⟩) is also the same as the Jocabian factor of the symplectic
condition δ(ΩIJ AI B J ).
Here we integrate out η1 , η2 , η3 , η8 , which corresponds to integrating out η1,L(R) , η2,L(R) ,
and η6,L(R) in the left and right sets,. Pulling out the prefactor from the amplitude and
collecting them together, we arrive
J3,D m38
√ +,+
[D]2×8 Z = 0,
ω8 (3)
m1 m2 m4 m5 m6 m8
D·g̃·DT =0, m3 =0.
(C.20)
dz L31 R13
= · √ [L]+
× √ [R]+
.
g(z) L1 L2 L3 L4 L5 L6 1×6 ZL =0 R1 R2 R3 R4 R5 R6 1×6 ZR = 0
T T
L·g̃L ·L =0 R·g̃R ·R =0

The Jocabian factors are

 −1/2
⟨2367⟩ Y 1
J3,D =  , (C.21)
⟨23⟩⟨67⟩ ⟨i i+1⟩⟨i+1 i+2⟩⟨i+2 i+3⟩
i=3,7
q
1 ⟨23⟩⟨8123⟩⟨2345⟩⟨5678⟩
and g(z) = x(z)z(⟨⟨37⟩⟩z+⟨⟨36⟩⟩) ⟨67⟩⟨1234⟩ is coming from
 3 Q8 !1/2
⟨61⟩L ⟨12⟩L ⟨61⟩R ⟨12⟩R i=1 ⟨i i+1⟩⟨i+1 i+2⟩⟨i+2 i+3⟩
Q6
⟨81⟩⟨12⟩⟨23⟩ i=1 ⟨i i+1⟩L ⟨i+1 i+2⟩L ⟨i i+1⟩R ⟨i+1 i+2⟩R (C.22)
⟨23⟩⟨67⟩ ⟨AZref I⟩ 2 ⟨BZref I⟩ −1
   
1
× .
⟨⟨37⟩⟩z+⟨⟨36⟩⟩ ⟨ABI⟩ ⟨k1 λref ⟩ ⟨k2 λref ⟩
Here, Zref is the reference twistor with the first two components λref .

D Local integrand basis

This appendix defines the two-loop local integrals used for n = 6, 8 ABJM theory and
one-loop integrals for n = 6 N = 4 sYM. For simplicity, we only explicitly present the
numerators of these integrals, and the denominators can be easily inferred from the corre-
sponding diagram. For instance, the integral I critter (i) at n = 6 (shown below) is
ncritter (i)
Z
I critter (i) := + (ℓ1 ↔ ℓ2 )
ℓ1 ,ℓ2 (ℓ1 · i−1)(ℓ1 · i)(ℓ1 · i+1)(ℓ1 · ℓ2 )(ℓ2 · i+2)(ℓ2 · i+3)(ℓ2 · i+4)
(D.1)
The integrals are written in their dual form, and readers can refer to the notation introduced
in [51, 52].

Six points: The local integrals contributing to the two-loop six-point are

i+1 i+2 i+1 i+3 i+1 i+1

i+2
i i+3 i i+4 i i+3 i
i+4
i−1 i+4 i−1 i−1 i−1
bt
I critter (i) I crab (i) I±2mh (i) I1(2) (i)

– 54 –
 ncritter (i) := ϵ(ℓ1 , i−1, i, i+1, µ
)ϵ(ℓ2 , i+2, i+3, i+4, µ )+(ℓ1 · i)(ℓ2 · i+3)


(i−1 · i+1)(i+2 · i+4) /2 ,
ncrab (i) := ϵ(ℓ1 , i−1, i, i+1, µ
)ϵ(ℓ2 , i+3, i+4, i−1, µ )+(ℓ1 · i)(ℓ2 · i+4)


(i−1 · i+1)(i+3 · i−1) /2 ,

n2mh
± (i) := ϵ(ℓ1 , i−1, i, i+1, µ
)ϵ(ℓ2 , i+1, i+3, i−1, µ )−(ℓ1 · i)(i−1 · i+1) (D.2)


(ℓ2 · i−1)(i+1 · i+3)+(ℓ2 · i+1)(i−1 · i+3)−(ℓ2 · i+3)(i−1 · i+1) /2
p √
± (−1)i ϵ(ℓ1 , i−1, i, i+1, ℓ2 ) (i+1 · i+3 · i−1)/ 2 ,
p √
nbt
1 (i) :=ϵ(ℓ1 , i−1, i, i+1, i+2) (i · i+2 · i+4)/ 2(i · i+2) ,
p √
nbt
2 (i) :=ϵ(ℓ1 , i−1, i, i+1, i+4) (i · i+2 · i+4)/ 2(i · i+4) .
p p
with the shorthand notation (i · j · k) := (i · j)(j · k)(k · i).

Eight points: For two-loop eight points, the local integrals are

i+1 i+2 i+1 i+1 i+3

i i+3 i i+2 i i+4

i−1 i+4 i−1 i+3 i−1 i+5

IAdb (i) IBdb (i) ICdb (i)

ndb
A (i) := ϵ(ℓ1 , i−1, i, i+1,
µ
)ϵ(ℓ2 , i+2, i+3, i+4, µ )+(ℓ1 · i)(ℓ2 · i+3)


(i−1 · i+1)(i+2 · i+4) /2 ,
ndb
B (i) := ϵ(ℓ1 , i−1, i, i+1,
µ
)ϵ(ℓ2 , i+1, i+2, i+3, µ )+(ℓ1 · i)(ℓ2 · i+2) (D.3)


(i−1 · i+1)(i+1 · i+3) /2 ,
ndb
C (i) :=ϵ(ℓ1 , i−1, i, i+1,
µ
)ϵ(ℓ2 , i+3, i+4, i+5, µ )/2 .

i+1 i+1 i+1 i+3 i+1

i i+3 i i+5 i i+5 i i+3

i−1 i−1 i−1 i−1 i+5

db db db db
ID,± (i) ID,± (i) IE,± (i) IE,± (i)

ndb
D,±
(i):= ϵ(ℓ1 , i−1, i, i+1, µ
)ϵ(ℓ2 , i+1, i+3, i−1, µ )+(ℓ1 · i)(ℓ2 · i+3)(i−1 · i+1)2
−(ℓ1 · i)(ℓ2 · i+1)(i−1 · i+1)(i−1 · i+3)−(ℓ1 · i)(ℓ2 · i−1)(i−1 · i+1)
 p √
(i+1 · i+3) /2±(−)i ϵ(ℓ1 , i−1, i, i+1, ℓ2 ) (i+1, i+3, i−1)/ 2 ,
 (D.4)
ndb
D,± (i):= ϵ(ℓ1 , i−1, i, i+1, µ
)ϵ(ℓ2 , i+1, i+5, i−1, µ )+(ℓ 1 · i)(ℓ 2 · i+5)(i−1 · i+1)2

−(ℓ1 · i)(ℓ2 · i+1)(i−1 · i+1)(i−1 · i+5)−(ℓ1 · i)(ℓ2 · i−1)(i−1 · i+1)

 p √
(i+1 · i+5) /2±(−)i ϵ(ℓ1 , i−1, i, i+1, ℓ2 ) (i+1, i+5, i−1)/ 2 .

– 55 –
 
ndb
E,±
(i) := ϵ(ℓ1 , i−1, i, i+1, µ
)ϵ(ℓ2 , i+3, i+5, i−1, µ) + (ℓ1 · i)(ℓ2 · i+5)

(i−1 · i+1)(i−1 · i+3)−(ℓ1 · i)(ℓ2 · i+3)(i−1 · i+1)(i−1 · i+5) /2
p √
± (−)i ϵ(ℓ1 , i−1, i, i+1, ℓ2 ) (i+3, i+5, i−1)/ 2 ,
 (D.5)
ndb
E,± (i) := ϵ(ℓ1 , i−1, i, i+1, µ )ϵ(ℓ2 , i+1, i+3, i+5, µ ) − (ℓ1 · i)(ℓ2 · i+5)

(i−1 · i+1)(i+1 · i+3)+(i−1 · i+1)(i+1 · i+5)(ℓ1 · i)(ℓ2 · i+3) /2
p √
± (−)i ϵ(ℓ1 , i−1, i, i+1, ℓ2 ) (i+1, i+3, i+5)/ 2 .

i+1 i+2 i+2

i i+4 i i+4

i−1 i+6 i−2

db db
IF,±,± ′ (i) IG,±,±′ (i)


ndb
F,±,± ′ (i):= ϵ(ℓ1 , i−1, i, i+1, µ )ϵ(ℓ2 , i+2, i+4, i+6, µ )−(ℓ1 · i)(i−1·i+1) (ℓ2 · i+2)(i+4·i+6)

−(ℓ2 · i+4)(i+2 · i+6)+(ℓ2 · i+6)(i+2 · i+4) /2−±±′ (ℓ1 · i)(ℓ2 · yi+1,i+6 ) (i·i+2·i+4·i+6)
p

√ ′
p
i (ℓ1 · i)(i−1 · i+1) (i·i+2·i+6)
i−1
p
±(−) ϵ(ℓ1 , i−1, i, i+1, ℓ2 ) (i+2·i+4·i+6)/ 2± (−) √
2(i · i+2)(i · i+6)(i+4 · yi+1,i+6 )
 
ϵ(ℓ2 , i−1, i+1, i+2, i+4)(i+4 · i+6)−ϵ(ℓ2 , i+6, i+1, i+2, i+4)(i+4 · i−1) ,
(D.6)
p p
where we define (i · j · k · l) := (i · j)(j · k)(k · l)(l · i). The double labeling ± in the
subscript of ndb
F,±,±′ (i) corresponds to the sign of the first (resp. second) term in the
definition. The same convention applies to the double subscripts in ndb G,±,±′ (i) below.


ndb
G,±,± ′ (i):= ϵ(ℓ1 , i−2, i, i+2, µ )ϵ(ℓ2 , i+2, i+4, i−2, µ )−(ℓ1 · i)(ℓ2 · i−2)(i−2·i+2)(i+2·i+4)


+(ℓ1 · i−2)(i · i+2) (ℓ2 · i−2)(i+2 · i+4)−(ℓ2 · i+4)(i−2 · i+2) + (ℓ1 · i)(i−2 · i+2)


−(ℓ1 · i+2)(i−2 · i) (ℓ2 · i+4)(i−2 · i+2)−(ℓ2 · i+2)(i−2 · i+4) /2 + ±±′ (ℓ1 · ℓ2 )


p p p √
(i−2 · i · i+2) (i+2 · i+4 · i−2) ± (−)i ϵ(ℓ1 , i−2, i, i+2, ℓ2 ) (i+2 · i+4 · i−2)/ 2
√
±′ (−)i−1 ϵ(ℓ1 , i+2, i+4, i−2, ℓ2 ) (i−2 · i · i+2)/ 2 .
p

(D.7)

i+1 i+1 i+1

i+4 i+2 i+2
i i i
i+6 i+4 i+6
i−1 i−1 i−1
IAbt (i) IAbt (i) IBbt (i)

– 56 –
 p √
nbt
A
(i) := ϵ(ℓ1 , i−1, i, i+1, i+6) (i · i+4 · i+6)/ 2(i · i+6) ,
p √
nbt
A (i) := ϵ(ℓ1 , i−1, i, i+1, i+2) (i · i+2 · i+4)/ 2(i · i+2) , (D.8)
p √
nbt
B (i) := ϵ(ℓ1 , i−1, i, i+1, i+2) (i · i+2 · i+6)/ 2(i · i+2) .

Finally, we list the local integrands used for the NMHV N = 4 sYM,

⟨AB(i−2 i−1 i) ∩ (i i+1 i+2)⟩⟨X i+1 i−1⟩

Z
I11mb (i) :=
A,B ⟨AB i−2 i−1⟩⟨AB i−1 i⟩⟨AB i i+1⟩⟨AB i+1 i+2⟩⟨ABX⟩
⟨AB i+4 i+2⟩⟨X (i+1 i+2 i+3) ∩ (i+3 i+4 i+5)⟩
Z
I21mb (i) :=
A,B ⟨AB i+1 i+2⟩⟨AB i+2 i+3⟩⟨AB i+3 i+4⟩⟨AB i+4 i+5⟩⟨ABX⟩


⟨AB i+2 (i i+1) ∩ i+2 i+3 (i+4 i+5 i+1) ∩ X − (i+2, i+3) ↔ (i+4, i+5) ⟩
Z
2mh
I1 (i) :=
A,B ⟨AB i i+1⟩⟨AB i+1 i+2⟩⟨AB i+2 i+3⟩⟨AB i+4 i+5⟩⟨ABX⟩


⟨AB i−2 (i−1 i) ∩ i−3 i−2 (i+1 i+2 i−1) ∩ X − (i−3, i−2) ↔ (i+1, i+2) ⟩
Z
2mh
I2 (i) := .
A,B ⟨AB i−3 i−2⟩⟨AB i−2 i−1⟩⟨AB i−1 i⟩⟨AB i+1 i+2⟩⟨ABX⟩
(D.9)

i+1 i+4 i+3 i

i i+2 i+3 i+5 i+2 i+5 i−1 i+2

i−1 i+2 i+1 i−2

I11mb (i) I21mb (i) I12mh (i) I22mh (i)

and the triangle integral I tri (i) is defined by

⟨i−1 i i+1 i+2⟩⟨X i+1 i⟩

Z
I tri (i) := . (D.10)
A,B ⟨AB i−1 i⟩⟨AB i i+1⟩⟨AB i+1 i+2⟩⟨ABX⟩

References
[1] N. Arkani-Hamed and J. Trnka, “The Amplituhedron,” JHEP 10 (2014) 030,
arXiv:1312.2007 [hep-th].
[2] N. Arkani-Hamed and J. Trnka, “Into the Amplituhedron,” JHEP 12 (2014) 182,
arXiv:1312.7878 [hep-th].
[3] N. Arkani-Hamed, H. Thomas, and J. Trnka, “Unwinding the Amplituhedron in Binary,”
JHEP 01 (2018) 016, arXiv:1704.05069 [hep-th].
[4] N. Arkani-Hamed, Y. Bai, and T. Lam, “Positive Geometries and Canonical Forms,” JHEP
11 (2017) 039, arXiv:1703.04541 [hep-th].
[5] S. Franco, D. Galloni, A. Mariotti, and J. Trnka, “Anatomy of the Amplituhedron,” JHEP
03 (2015) 128, arXiv:1408.3410 [hep-th].
[6] L. Ferro, T. Lukowski, A. Orta, and M. Parisi, “Towards the Amplituhedron Volume,” JHEP
03 (2016) 014, arXiv:1512.04954 [hep-th].
[7] Y. Bai, S. He, and T. Lam, “The Amplituhedron and the One-loop Grassmannian Measure,”
JHEP 01 (2016) 112, arXiv:1510.03553 [hep-th].

– 57 –
 [8] L. J. Dixon, M. von Hippel, A. J. McLeod, and J. Trnka, “Multi-loop positivity of the planar
N = 4 SYM six-point amplitude,” JHEP 02 (2017) 112, arXiv:1611.08325 [hep-th].
[9] S. N. Karp and L. K. Williams, “The m=1 amplituhedron and cyclic hyperplane
arrangements,” arXiv:1608.08288 [[Link]].
[10] S. N. Karp, L. K. Williams, and Y. X. Zhang, “Decompositions of amplituhedra,”
arXiv:1708.09525 [[Link]].
[11] L. Ferro, T. Lukowski, and M. Parisi, “Amplituhedron meets Jeffrey-Kirwan Residue,”
arXiv:1805.01301 [hep-th].
[12] P. Galashin and T. Lam, “Parity duality for the amplituhedron,” arXiv:1805.00600
[[Link]].
[13] N. Arkani-Hamed, C. Langer, A. Yelleshpur Srikant, and J. Trnka, “Deep Into the
Amplituhedron: Amplitude Singularities at All Loops and Legs,” Phys. Rev. Lett. 122 no. 5,
(2019) 051601, arXiv:1810.08208 [hep-th].
[14] G. Salvatori and S. L. Cacciatori, “Hyperbolic Geometry and Amplituhedra in 1+2
dimensions,” arXiv:1803.05809 [hep-th].
[15] R. Kojima, “Triangulation of 2-loop MHV Amplituhedron from Sign Flips,” JHEP 04 (2019)
085, arXiv:1812.01822 [hep-th].
[16] J. Rao, “4-particle amplituhedronics for 3-5 loops,” Nucl. Phys. B 943 (2019) 114625,
arXiv:1806.01765 [hep-th].
[17] A. Yelleshpur Srikant, “Emergent unitarity from the amplituhedron,” JHEP 01 (2020) 069,
arXiv:1906.10700 [hep-th].
[18] C. Langer and A. Yelleshpur Srikant, “All-loop cuts from the Amplituhedron,” JHEP 04
(2019) 105, arXiv:1902.05951 [hep-th].
[19] T. Lukowski, “On the Boundaries of the m=2 Amplituhedron,” arXiv:1908.00386
[hep-th].
[20] E. Herrmann, C. Langer, J. Trnka, and M. Zheng, “Positive geometry, local triangulations,
and the dual of the Amplituhedron,” JHEP 01 (2021) 035, arXiv:2009.05607 [hep-th].
[21] R. Kojima and J. Rao, “Triangulation-free Trivialization of 2-loop MHV Amplituhedron,”
JHEP 10 (2020) 140, arXiv:2007.15650 [hep-th].
[22] T. Lukowski, M. Parisi, and L. K. Williams, “The positive tropical Grassmannian, the
hypersimplex, and the m=2 amplituhedron,” arXiv:2002.06164 [[Link]].
[23] M. Parisi, M. Sherman-Bennett, and L. Williams, “The m=2 amplituhedron and the
hypersimplex: signs, clusters, triangulations, Eulerian numbers,” arXiv:2104.08254
[[Link]].
[24] N. Arkani-Hamed, P. Benincasa, and A. Postnikov, “Cosmological Polytopes and the
Wavefunction of the Universe,” arXiv:1709.02813 [hep-th].
[25] N. Arkani-Hamed, Y. Bai, S. He, and G. Yan, “Scattering Forms and the Positive Geometry
of Kinematics, Color and the Worldsheet,” arXiv:1711.09102 [hep-th].
[26] B. Eden, P. Heslop, and L. Mason, “The Correlahedron,” JHEP 09 (2017) 156,
arXiv:1701.00453 [hep-th].

– 58 –
[27] N. Arkani-Hamed, S. He, and T. Lam, “Stringy canonical forms,” JHEP 02 (2021) 069,
arXiv:1912.08707 [hep-th].
[28] N. Arkani-Hamed, S. He, G. Salvatori, and H. Thomas, “Causal Diamonds, Cluster
Polytopes and Scattering Amplitudes,” arXiv:1912.12948 [hep-th].
[29] N. Arkani-Hamed, S. He, T. Lam, and H. Thomas, “Binary Geometries, Generalized
Particles and Strings, and Cluster Algebras,” arXiv:1912.11764 [hep-th].
[30] S. He and C. Zhang, “Notes on Scattering Amplitudes as Differential Forms,” JHEP 10
(2018) 054, arXiv:1807.11051 [hep-th].
[31] D. Damgaard, L. Ferro, T. Lukowski, and M. Parisi, “The Momentum Amplituhedron,”
JHEP 08 (2019) 042, arXiv:1905.04216 [hep-th].
[32] L. Ferro and T. Lukowski, “The Loop Momentum Amplituhedron,” JHEP 05 (2023) 183,
arXiv:2210.01127 [hep-th].
[33] Y.-t. Huang, R. Kojima, C. Wen, and S.-Q. Zhang, “The orthogonal momentum
amplituhedron and ABJM amplitudes,” JHEP 01 (2022) 141, arXiv:2111.03037 [hep-th].
[34] S. He, C.-K. Kuo, and Y.-Q. Zhang, “The momentum amplituhedron of SYM and ABJM
from twistor-string maps,” JHEP 02 (2022) 148, arXiv:2111.02576 [hep-th].
[35] Y.-T. Huang, C.-K. Kuo, and C. Wen, “Dualities for Ising networks,” Phys. Rev. Lett. 121
no. 25, (2018) 251604, arXiv:1809.01231 [hep-th].
[36] N. Arkani-Hamed, Y.-T. Huang, and S.-H. Shao, “On the Positive Geometry of Conformal
Field Theory,” JHEP 06 (2019) 124, arXiv:1812.07739 [hep-th].
[37] K. Hosomichi, K.-M. Lee, S. Lee, S. Lee, and J. Park, “N=5,6 Superconformal Chern-Simons
Theories and M2-branes on Orbifolds,” JHEP 09 (2008) 002, arXiv:0806.4977 [hep-th].
[38] O. Aharony, O. Bergman, D. L. Jafferis, and J. Maldacena, “N=6 superconformal
Chern-Simons-matter theories, M2-branes and their gravity duals,” JHEP 10 (2008) 091,
arXiv:0806.1218 [hep-th].
[39] T. Bargheer, F. Loebbert, and C. Meneghelli, “Symmetries of Tree-level Scattering
Amplitudes in N=6 Superconformal Chern-Simons Theory,” Phys. Rev. D 82 (2010) 045016,
arXiv:1003.6120 [hep-th].
[40] Y.-t. Huang and A. E. Lipstein, “Dual Superconformal Symmetry of N=6 Chern-Simons
Theory,” JHEP 11 (2010) 076, arXiv:1008.0041 [hep-th].
[41] S. Lee, “Yangian Invariant Scattering Amplitudes in Supersymmetric Chern-Simons
Theory,” Phys. Rev. Lett. 105 (2010) 151603, arXiv:1007.4772 [hep-th].
[42] Y.-T. Huang and C. Wen, “ABJM amplitudes and the positive orthogonal grassmannian,”
JHEP 02 (2014) 104, arXiv:1309.3252 [hep-th].
[43] Y.-t. Huang, C. Wen, and D. Xie, “The Positive orthogonal Grassmannian and loop
amplitudes of ABJM,” J. Phys. A47 no. 47, (2014) 474008, arXiv:1402.1479 [hep-th].
[44] S. He, C.-K. Kuo, Z. Li, and Y.-Q. Zhang, “All-Loop Four-Point
Aharony-Bergman-Jafferis-Maldacena Amplitudes from Dimensional Reduction of the
Amplituhedron,” Phys. Rev. Lett. 129 no. 22, (2022) 221604, arXiv:2204.08297 [hep-th].
[45] S. He, C.-K. Kuo, Z. Li, and Y.-Q. Zhang, “Emergent unitarity, all-loop cuts and integrations
from the ABJM amplituhedron,” JHEP 07 (2023) 212, arXiv:2303.03035 [hep-th].

– 59 –
[46] J. M. Henn, M. Lagares, and S.-Q. Zhang, “Integrated negative geometries in ABJM,” JHEP
05 (2023) 112, arXiv:2303.02996 [hep-th].
[47] N. Arkani-Hamed, J. Henn, and J. Trnka, “Nonperturbative negative geometries: amplitudes
at strong coupling and the amplituhedron,” JHEP 03 (2022) 108, arXiv:2112.06956
[hep-th].
[48] D. Chicherin and J. M. Henn, “Symmetry properties of Wilson loops with a Lagrangian
insertion,” JHEP 07 (2022) 057, arXiv:2202.05596 [hep-th].
[49] A. Hodges, “Eliminating spurious poles from gauge-theoretic amplitudes,” JHEP 05 (2013)
135, arXiv:0905.1473 [hep-th].
[50] H. Elvang, Y.-t. Huang, C. Keeler, T. Lam, T. M. Olson, S. B. Roland, and D. E. Speyer,
“Grassmannians for scattering amplitudes in 4d N = 4 SYM and 3d ABJM,” JHEP 12
(2014) 181, arXiv:1410.0621 [hep-th].
[51] S. Caron-Huot and Y.-t. Huang, “The two-loop six-point amplitude in ABJM theory,” JHEP
03 (2013) 075, arXiv:1210.4226 [hep-th].
[52] S. He, Y.-t. Huang, C.-K. Kuo, and Z. Li, “The two-loop eight-point amplitude in ABJM
theory,” JHEP 02 (2023) 065, arXiv:2211.01792 [hep-th].
[53] J. L. Bourjaily, E. Herrmann, and J. Trnka, “Prescriptive Unitarity,” JHEP 06 (2017) 059,
arXiv:1704.05460 [hep-th].
[54] S. He, Y.-t. Huang, C.-K. Kuo, and M. Parisi, “Chambers unlocked: from elliptic leading
singularities to local triangulations,” in progress .
[55] T. Lukowski, R. Moerman, and J. Stalknecht, “On the geometry of the orthogonal
momentum amplituhedron,” JHEP 12 (2022) 006, arXiv:2112.03294 [hep-th].
[56] L. Ferro, T. Lukowski, and R. Moerman, “From momentum amplituhedron boundaries to
amplitude singularities and back,” JHEP 07 no. 07, (2020) 201, arXiv:2003.13704
[hep-th].
[57] A. Brandhuber, G. Travaglini, and C. Wen, “A note on amplitudes in N=6 superconformal
Chern-Simons theory,” JHEP 07 (2012) 160, arXiv:1205.6705 [hep-th].
[58] N. Arkani-Hamed, J. L. Bourjaily, F. Cachazo, A. B. Goncharov, A. Postnikov, and J. Trnka,
Grassmannian Geometry of Scattering Amplitudes. Cambridge University Press, 4, 2016.
arXiv:1212.5605 [hep-th].
[59] W.-M. Chen and Y.-t. Huang, “Dualities for Loop Amplitudes of N=6 Chern-Simons Matter
Theory,” JHEP 11 (2011) 057, arXiv:1107.2710 [hep-th].
[60] Z. Bern, L. J. Dixon, and V. A. Smirnov, “Iteration of planar amplitudes in maximally
supersymmetric Yang-Mills theory at three loops and beyond,” Phys. Rev. D 72 (2005)
085001, arXiv:hep-th/0505205.

– 60 –