Math. Ann.

(2013) 357:1005–1047
DOI 10.1007/s00208-013-0922-5 Mathematische Annalen

Integrability of higher pentagram maps

Boris Khesin · Fedor Soloviev

Received: 4 September 2012 / Revised: 26 February 2013 / Published online: 29 March 2013
© Springer-Verlag Berlin Heidelberg 2013

Abstract We define higher pentagram maps on polygons in Pd for any dimension

d, which extend R. Schwartz’s definition of the 2D pentagram map. We prove their
integrability by presenting Lax representations with a spectral parameter for scale
invariant maps. The corresponding continuous limit of the pentagram map in dimen-
sion d is shown to be the (2, d + 1)-equation of the KdV hierarchy, generalizing the
Boussinesq equation in 2D. We also study in detail the 3D case, where we prove inte-
grability for both closed and twisted polygons and describe the spectral curve, first
integrals, the corresponding tori and the motion along them, as well as an invariant
symplectic structure.

Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1006
2 Review of the 2D pentagram map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1010
3 Geometric definition of higher pentagram maps . . . . . . . . . . . . . . . . . . . . . . . . . . 1012
3.1 Pentagram map in 3D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1012
3.2 Pentagram map in any dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1013
3.3 General pentagram maps and duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1015
4 Continuous limit of the higher pentagram maps . . . . . . . . . . . . . . . . . . . . . . . . . . 1016
4.1 Definition of the continuous limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1016
4.2 Envelopes and the KdV hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1018

B. Khesin
School of Mathematics, Institute for Advanced Study, Princeton, NJ 08540, USA

B. Khesin (B) · F. Soloviev

Department of Mathematics, University of Toronto, Toronto, ON M5S 2E4, Canada
e-mail: [email protected]
F. Soloviev
e-mail: [email protected]

123
1006 B. Khesin, F. Soloviev

5 Explicit formulas for the 3D pentagram map . . . . . . . . . . . . . . . . . . . . . . . . . . . 1021

5.1 Two involutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1021
5.2 Cross-ratio type coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1022
5.3 Coordinates on twisted polygons: odd versus even n . . . . . . . . . . . . . . . . . . . . . 1024
6 Algebraic-geometric integrability of the 3D pentagram map . . . . . . . . . . . . . . . . . . . 1026
6.1 Scaling transformations and a Lax function in 3D . . . . . . . . . . . . . . . . . . . . . . 1027
6.2 Properties of the spectral curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1029
6.3 The spectral curve and invariant tori . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1033
7 Ramifications: closed polygons and symplectic leaves . . . . . . . . . . . . . . . . . . . . . . 1036
7.1 Closed polygons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1036
7.2 Invariant symplectic structure and symplectic leaves . . . . . . . . . . . . . . . . . . . . . 1037
8 A Lax representation in higher dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1040
Appendix A: Continuous limit in the 3D case . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1042
Appendix B: Higher terms of the continuous limit . . . . . . . . . . . . . . . . . . . . . . . . . . 1043
Appendix C: Bijection of the spectral map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1044

1 Introduction

The pentagram map was defined by Schwartz in [13] on plane convex polygons consid-
ered modulo projective equivalence. Figure 1 explains the definition: for a polygon P
the image under the pentagram map is a new polygon T (P) spanned by the “shortest”
diagonals of P. Iterations of this map on classes of projectively equivalent polygons
manifest quasiperiodic behaviour, which indicates hidden integrability [14].
The integrability was proved in [11] for the pentagram map on a larger class of
the so called twisted polygons in 2D, which are piecewise linear curves with a fixed
monodromy relating their ends. Closed polygons correspond to the monodromy given
by the identity transformation. It turned out that there is an invariant Poisson structure
for the pentagram map and it has sufficiently many invariant quantities. Moreover,
this map turned out to be related to a variety of mathematical domains, including
cluster algebras [2,4], frieze patterns, and integrable systems of mathematical physics:
in particular, its continuous limit in 2D is the classical Boussinesq equation [11].
Integrability of the pentagram map for 2D closed polygons was established in [12,15],
while a more general framework related to surface networks was presented in [3].
In this paper we extend the definition of the pentagram map to closed and twisted
polygons in spaces of any dimension d and prove its various integrability properties.
It is worth mentioning that the problem of finding integrable higher-dimensional gen-
eralizations for the pentagram map attracted much attention after the 2D case was

Fig. 1 The image T (P) of a

hexagon P under the 2D
pentagram map

123
Integrability of higher pentagram maps 1007

treated in [11].1 The main difficulty in higher dimensions is that diagonals of a poly-
gon are generically skew and do not intersect. One can either confine oneself to special
polygons (e.g., corrugated ones, [3]) to retain the intersection property or one has too
many possible choices for using hyperplanes as diagonals, where it is difficult to find
integrable ones, cf. [9].
Below, as an analog of the 2D shortest diagonals for a generic polygon in a projective
space RPd we propose to consider a “short-diagonal hyperplane” passing through d
vertices where every other vertex is taken starting with a given one. Then a new vertex
is constructed as the intersection of d consecutive diagonal hyperplanes. We repeat
this procedure starting with the next vertex of the initial polygon. The higher (or d-
dimensional) pentagram map T takes the initial polygon to the one defined by this
set of new vertices. As before, the obtained polygon is considered modulo projective
equivalence in RPd .
We also describe general pentagram maps T p,r in RPd enumerated by two integral
parameters p and r by considering p-diagonals (i.e., hyperplanes passing through
every pth vertex of the polygon) and by taking the intersections of every r th hyperplane
like that. There is a curious duality between them: the map T p,r is equal to Tr,−1
p modulo
a shift in vertex indices. However, we are mostly interested in the higher pentagram
maps, which correspond to T := T2,1 in RPd .
We start by describing the continuous limit of the higher pentagram map as the
evolution in the direction of the “envelope” for such a sequence of short-diagonal
planes as the number of vertices of the polygon tends to infinity. (More precisely, the
envelope here is the curve whose osculating planes are limits of the short-diagonal
planes.)
Theorem A (=Theorems 4.3, 4.5) The continuous limit of the higher pentagram map
in RPd is the (2, d +1)-equation in the KdV hierarchy, which is an infinite-dimensional
completely integrable system.
This generalizes the Boussinesq equation as a limit of the pentagram map in RP2
and this limit seems to be very robust. Indeed, the same equation appears for an almost
arbitrary choice of diagonal planes. It also arises when instead of osculating planes
one considers other possible definitions of higher pentagram maps (cf. e.g. [9]).
However, the pentagram map in the above definition with short-diagonal hyper-
planes exhibits integrability properties not only in the continuous limit, but as a discrete
system as well. To study them, we define two coordinate systems for twisted polygons
in 3D (somewhat similar to the ones used in 2D, cf. [11]), and present explicit formulas
for the 3D pentagram map using these coordinates (see Theorem 5.6).
Then we describe the pentagram map as a completely integrable discrete dynamical
system by presenting its Lax form in any dimension and studying in detail the 3D case
(see Sect. 6). For algebraic–geometric integrability we complexify the pentagram map.
The corresponding 2D case was investigated in [15].

1 There seem to be no natural generalization of the pentagram map to polytopes in higher dimension d ≥ 3.
Indeed, the initial polytope should be simple for its diagonal hyperplanes to be well defined. In order to
iterate the pentagram map the dual polytope has to be simple as well. Thus iterations could be defined only
for d-simplices, which are all projectively equivalent.

123
1008 B. Khesin, F. Soloviev

The key ingredient of the algebraic-geometric integrability for a discrete dynamical

system is a discrete Lax (or zero curvature) equation with a spectral parameter, which
in our case assumes the following form:

−1
L i,t+1 (λ) = Pi+1,t (λ)L i,t (λ)Pi,t (λ).

Here the index t represents the discrete time variable, the index i refers to the vertex
of an n-gon, and λ is a complex spectral parameter. (For the pentagram map in CPd
the functions L i,t (λ) and Pi,t (λ) are matrix-valued of size (d + 1) × (d + 1).) The
discrete Lax equation arises as a compatibility condition of an over-determined system
of equations:

L i,t (λ)i,t (λ) = i+1,t (λ)
Pi,t (λ)i,t (λ) = i,t+1 (λ),

for an auxiliary function i,t (λ).

Remark 1.1 Recall that for a smooth dynamical system the Lax form is a differential
equation of type ∂t L = [P, L] on a matrix L. Such a form of the equation implies that
the evolution of L changes it to a similar matrix, thus preserving its eigenvalues. If
the matrix L depends on a parameter, L = L(λ), then the corresponding eigenvalues
as functions of parameter do not change and in many cases provide sufficiently many
first integrals for complete integrability of such a system.
Similarly, an analogue of the Lax form for differential operators of type ∂x − L is a
zero curvature equation ∂t L − ∂x P = [P, L] . This is a compatibility condition which
provides the existence of an auxiliary function ψ = ψ(t, x) satisfying a system of
equations ∂x ψ = Lψ and ∂t ψ = Pψ . The above Lax form and auxiliary system are
discrete versions of the latter.

In our case, the equivalence of formulas for the pentagram map to the dynamics
defined by the Lax equation implies complete algebraic-geometric integrability of the
system. More precisely, the following theorem summarizes several main results on the
3D pentagram map, which are obtained by studying its Lax equation. The dynamics
is (generically) defined on the space Pn of projectively equivalent twisted n-gons in
3D, which we describe below, and has dimension 3n, while closed n-gons form a
submanifold of codimension 15 in it.
Later on we will introduce the notion of spectral data which consists of a Riemann
surface, called a spectral curve, and a point in the Jacobian (i.e., the complex torus) of
this curve, as well as a notion of a spectral map between the space Pn and the spectral
data.

Theorem B (=Theorems 6.15, 6.19, 7.1) A Zariski open subset of the complexified
space Pn of twisted n-gons in 3D is a fibration whose fibres are Zariski open subsets of
tori. These tori are Jacobians of the corresponding spectral curves and are invariant
with respect to the space pentagram map. Their dimension is 3n/2 for odd n and
3(n/2) − 3 for even n, where n/2 is the integer part of n/2.

123
Integrability of higher pentagram maps 1009

The pentagram dynamics on the Jacobians goes along a straight line for odd n and
along a staircase for even n (i.e., the discrete evolution is either a constant shift on a
torus, or its square is a constant shift).
For closed n-gons the tori have dimensions 3n/2 − 6 for odd n and 3(n/2) − 9
for even n.

Remark 1.2 One also has an explicit description of the the above fibration in terms of
coordinates on the space of n-gons. We note that the pentagram dynamics understood as
a shift on complex tori does not prevent the corresponding orbits on the space Pn from
being unbounded. The dynamics described above takes place for generic initial data,
i.e., for points on the Jacobians whose orbits do not intersect certain divisors. Points
of generic orbits with irrational shifts can return arbitrarily close to such divisors. On
the other hand, the inverse spectral map is defined outside of these special divisors
and may have poles there. Therefore the sequences in the space Pn corresponding to
such orbits may escape to infinity.

It is known that the pentagram map in 2D possesses an invariant Poisson structure

[11], which can also be described by using the Krichever–Phong universal formula
[15]. Although we do not present an invariant Poisson structure for the pentagram map
in 3D, we describe its symplectic leaves, as well as the action-angle coordinates. More
precisely, we present an invariant symplectic structure (i.e., a closed non-degenerate
2-form), and submanifolds where it is defined (Theorem 7.5). By analogy with the 2D
case, it is natural to suggest that these submanifolds are symplectic leaves of an invari-
ant Poisson structure, and that the inverse of our symplectic structure coincides with
the Poisson structure on the leaves. An explicit description of this Poisson structure
in 3D is still an open problem.
Note that the algebraic-geometric integrability of the pentagram map implies its
Arnold–Liouville complete integrability on generic symplectic leaves (in the real case).
Namely, the existence of a (pre)symplectic structure coming from the Lax form of
the pentagram map (see Sect. 7.2), together with the generic set of first integrals,
appearing as coefficients of the corresponding spectral curve, provides sufficiently
many integrals in involution. (Note that proving independence of first integrals while
remaining within the real setting is often more difficult than first proving the algebraic-
geometric integrability, which in turn implies their independence in the real case.)
Finally, in Sect. 8 we present a Lax form for the pentagram maps in arbitrary dimen-
sion (which implies their complete integrability) assuming their scaling invariance:

Theorem C (=Theorem 8.3) The scale-invariant pentagram map in CPd admits a

Lax representation with a spectral parameter.

The scaling invariance of the pentagram maps is proved for all d ≤ 6, with some
numerical evidence for higher values of d > 6 as well. It would be interesting to
establish it in full generality. There is a considerable difference between the cases of
even and odd dimension d, which can be already seen in the analysis of the 2D and
3D cases.

123
1010 B. Khesin, F. Soloviev

2 Review of the 2D pentagram map

In this section we recall the main definitions and results in 2D (see [11]), which will be
important for higher-dimensional generalizations below. We formulate the geometric
results in the real setting, while the algebraic-geometric ones are presented for the
corresponding complexification.
First note that the pentagram map can be extended from closed to twisted polygons.

Definition 2.1 Given a projective transformation M ∈ P S L(3, R) of the plane RP2 ,

a twisted n-gon in RP2 is a map φ : Z → RP2 , such that φ(k + n) = M ◦ φ(k) for
any k. M is called the monodromy of φ. Two twisted n-gons are equivalent if there is
a transformation g ∈ P S L(3, R) such that g ◦ φ1 = φ2 .
Consider generic n-gons, i.e., those that do not have any three consecutive vertices
lying on the same line. Denote by Pn the space of generic twisted n-gons consid-
ered up to P S L(3, R) transformations. The dimension of Pn is 2n. Indeed, a twisted
n-gon depends on 2n variables representing coordinates of vertices vk := φ(k)
for k = 1, . . . , n and on 8 parameters of the monodromy matrix M, while the
P S L(3, R)-equivalence reduces the dimension by 8. The pentagram map T is gener-
ically defined on the space Pn . Namely, for a twisted n-gon vertices of its image are
the intersections of pairs of consecutive shortest diagonals: T vk := (vk−1 , vk+1 ) ∩
(vk , vk+2 ). Such intersections are well defined for a generic point in Pn .
a) Results on integrability (in the twisted and closed cases). There is a Poisson
structure on Pn invariant with respect to the pentagram map. There are 2n/2 + 2
integrals in involution, which provide integrability of the pentagram map on Pn . Its
symplectic leaves have codimensions 2 or 4 in Pn depending on whether n is odd or
even, and the invariant tori have dimensions n − 1 or n − 2, respectively [11].
Moreover, when restricted to the space Cn of closed polygons (dim Cn = 2n − 8 =
2(n − 4)), the map is still integrable and has invariant tori of dimension n − 4 for odd
n and n − 5 for even n. Note that the space Cn of closed polygons is not a Poisson
submanifold in the space Pn of twisted n-gons, so the corresponding Poisson structure
on Pn cannot be restricted to Cn .
There is a Lax representation for the pentagram map. Coefficients of the corre-
sponding spectral curve are the first integrals of the dynamics. The pentagram map
defines a discrete motion on the Jacobian of the spectral curve. This motion is linear
or staircase-like depending on the parity of n, see [15].
b) Coordinates on Pn . The following two systems of coordinates on Pn are partic-
ularly convenient to work with, see [11]. Assume that n is not divisible by 3. Then
there exists a unique lift of points vk = φ(k) ∈ RP2 to the vectors Vk ∈ R3 satisfying
the condition det |V j , V j+1 , V j+2 | = 1 for each j. Associate a difference equation to
a sequence of vectors Vk ∈ R3 by setting

V j+3 = a j V j+2 + b j V j+1 + V j

for all j ∈ Z. The sequences (a j ) and (b j ) turn out to be n-periodic, which is a

manifestation of the fact that the lifts satisfy the relations V j+n = M V j , j ∈ Z, for a

123
Integrability of higher pentagram maps 1011

certain monodromy matrix M ∈ S L(3, R). The variables a j , b j , 0 ≤ j ≤ n − 1 are

coordinates on the space Pn .
There exists another coordinate system on the space Pn , which is more geometric.
Recall that the cross-ratio of 4 points in P1 is given by

(t1 − t2 )(t3 − t4 )
[t1 , t2 , t3 , t4 ] = ,
(t1 − t3 )(t2 − t4 )

where t is any affine parameter. Now associate to each vertex vi the following two num-
bers, which are the cross-ratios of two 4-tuples of points lying on the lines (vi−2 , vi−1 )
and (vi+1 , vi+2 ) respectively:
 
xi = vi−2 , vi−1 , ((vi−2 , vi−1 ) ∩ (vi , vi+1 )), ((vi−2 , vi−1 ) ∩ (vi+2 , vi+2 ))
 
yi = ((vi−2 , vi−1 ) ∩ (vi+1 , vi+2 )), ((vi−1 , vi ) ∩ (vi+1 , vi+2 )), vi+1 , vi+2

In these coordinates the pentagram map has the form

1 − xi−1 yi−1 1 − xi+2 yi+2

T ∗ xi = xi T ∗ yi = yi+1 .
1 − xi+1 yi+1 1 − xi yi

One can see that the pentagram map commutes with the scaling transformation [11,14]:
 
Rs : (x1 , y1 , . . . , xn , yn ) → sx1 , s −1 y1 , . . . , sxn , s −1 yn .

In these coordinates the invariant Poisson structure has a particularly simple form, see
[11].
c) Continuous limit: the Boussinesq equation. The n → ∞ continuous limit of a
twisted n-gon with a fixed monodromy M ∈ P S L(3, R) can be viewed as a smooth
parameterized curve γ : R → RP2 satisfying γ (x + 2π ) = Mγ (x) for all x ∈ R. The
genericity assumption that every three consecutive points of an n-gon are in general
position corresponds to the assumption that γ is a non-degenerate curve in RP2 , i.e.,
the vectors γ (x) and γ (x) are linearly independent for all x ∈ R.
The space of such projectively equivalent curves is in one-to-one correspondence
with linear differential operators of the third order: L = ∂ 3 + u 1 (x)∂ + u 0 (x), where
the coefficients u 0 and u 1 are periodic in x. Namely, a curve γ (x) in RP2 can be lifted
to a quasi-periodic curve G = {G(x)} in R3 satisfying det |G(x), G (x), G (x)| = 1
for all x ∈ R. The components of the vector function G(x) = (G 1 (x), G 2 (x), G 3 (x))
are homogenous coordinates of γ (x) in RP2 : γ (x) = (G 1 : G 2 : G 3 )(x) ∈ RP2 . The
vector function G(x) can be identified with a solution of the unique linear differential
operator L, i.e., the components of G(x) are identified with three linearly independent
solutions of the differential equation L y = 0.
A continuous analog of the pentagram map is obtained by the following construc-
tion. Given a non-degenerate curve γ (x), we draw the chord (γ (x − ), γ (x + )) at
each point x. Consider the envelope (x) of these chords. (Fig. 2 shows their lifts:
chords (G(x − ), G(x + )) and their envelope L (x).) Let u 1, and u 0, be the peri-
odic coefficients of the corresponding differential operator. Their expansions in have

123
1012 B. Khesin, F. Soloviev

Fig. 2 Constructing the

envelope L (x) in 2D

the form u i, = u i + 2 wi +O( 3 ) and allow one to define the evolution du i /dt := wi ,
i = 0, 1. After getting rid of u 0 this becomes the classical Boussinesq equation on the
periodic function u = u 1 , which is the (2, 3)-flow in the KdV hierarchy of integrable
equations on the circle: u tt + 2(u 2 )x x + u x x x x = 0.
Below we generalize these results to higher dimensions.

3 Geometric definition of higher pentagram maps

3.1 Pentagram map in 3D

First we extend the notion of a closed polygon to a twisted one, similar to the 2D
case. We present the 3D case first, before giving the definition of the pentagram map
in arbitrary dimension, since it is used in many formulas below.

Definition 3.1 A twisted n-gon in RP3 with a monodromy M ∈ S L(4, R) is a map

φ : Z → RP3 , such that φ(k + n) = M ◦ φ(k) for each k ∈ Z. (Here we consider the
natural action of S L(4, R) on the corresponding projective space RP3 .) Two twisted
n-gons are (projectively) equivalent if there is a transformation g ∈ S L(4, R), such
that g ◦ φ1 = φ2 .

Note that equivalent n-gons must have similar monodromies. Closed n-gons (space
polygons) correspond to the monodromies M = Id and −Id. Let us assume that
vertices of an n-gon are in general position, i.e., no four consecutive vertices belong
to one and the same plane in RP3 . Also, assume that n is odd. Then one can show
(see Sect. 5.3 below and cf. Proposition 4.1 in [11]) that there exists a unique lift of
the vertices vk := φ(k) ∈ RP3 to the vectors Vk ∈ R4 satisfying for all j ∈ Z the
identities det |V j , V j+1 , V j+2 , V j+3 | = 1 and V j+n = M V j , where M ∈ S L(4, R).2
These vectors satisfy difference equations

V j+4 = a j V j+3 + b j V j+2 + c j V j+1 − V j , j ∈ Z,

2 This explains our choice of the group S L(4, R) rather than the seemingly more natural group P S L(4, R):
since S L(4, R) is a twofold cover of P S L(4, R), a twisted n-gon would have two different lifts from RP3
to R4 corresponding to two different lifts of the monodromy M from the latter group.

123
Integrability of higher pentagram maps 1013

Fig. 3 In 3D the image T vk of

the vertex vk is the intersection
of three “short-diagonal” planes
Pk−1 , Pk , and Pk+1

with n-periodic coefficients (a j , b j , c j ) and we employ this equation to introduce the

(a, b, c)-coordinates on the space of twisted n-gons.
Define the pentagram map on the classes of equivalent n-gons in such a way.
Definition 3.2 Given an n-gon φ in RP3 , for each k ∈ Z consider the two-
dimensional “short-diagonal plane” Pk := (vk−2 , vk , vk+2 ) passing through 3 ver-
tices vk−2 , vk , vk+2 . Take the intersection point of the three consecutive planes
Pk−1 , Pk , Pk+1 and call it the image of the vertex vk under the space pentagram
map T , see Fig. 3. (We assume the general position, so that every three consecutive
planes Pk for the given n-gon intersect at a point.)
By lifting a two-dimensional plane Pk from RP3 to the three-dimensional plane
through the origin in R4 (and slightly abusing notation) we have Pk = ∗(Vk−2 ∧ Vk ∧
Vk+2 ) in terms of the natural duality ∗ between R4 and R4∗ . The lift of T vk to R4 is
proportional to ∗[Pk−1 ∧ Pk ∧ Pk+1 ].
Below we describe the properties of this space pentagram map in detail.

3.2 Pentagram map in any dimension

Before defining the pentagram map in RPd , recall that S L(d +1, R) is a twofold cover
of P S L(d + 1, R) for odd d and coincides with the latter for even d.
Definition 3.3 A twisted n-gon in RPd with a monodromy M ∈ S L(d + 1, R) is a
map φ : Z → RPd , such that φ(k + n) = M ◦ φ(k) for each k ∈ Z, and where M
acts naturally on RPd .
We define the S L(d +1, R)-equivalence of n-gons as above, and assume the vertices
vk := φ(k) to be in general position, i.e., in particular, no d + 1 consecutive vertices
of an n-gon belong to one and the same (d − 1)-dimensional plane in RPd .
Remark 3.4 One can show that there exists a unique lift of the vertices vk = φ(k) ∈
RPd to the vectors Vk ∈ Rd+1 satisfying det |V j , V j+1 , . . . , V j+d | = 1 and V j+n =
M V j , j ∈ Z, where M ∈ S L(d +1, R), if and only if the condition gcd(n, d +1) = 1
holds. The corresponding difference equations have the form

V j+d+1 = a j,d V j+d + a j,d−1 V j+d−1 + · · · + a j,1 V j+1 + (−1)d V j , j ∈ Z, (1)

123
1014 B. Khesin, F. Soloviev

with n-periodic coefficients in the index j. This allows one to introduce coordinates
{a j,k , 0 ≤ j ≤ n − 1, 1 ≤ k ≤ d} on the space of twisted n-gons in RPd .
For a generic twisted n-gon in RPd one can define the “short-diagonal” (d − 1)-
dimensional plane Pk passing through d vertices of the n-gon by taking every other
vertex starting at the point vk , i.e., through the vertices vk , vk+2 , . . . , vk+2d−2 . For
calculations, however, it is convenient to have the set of vertices “centered” at vk , and
then the definition becomes slightly different in the odd and even dimensional cases.
Namely, for odd dimension d = 2 + 1 we consider the short-diagonal hyperplane
Pk through the d vertices

Pk := (vk−2 , vk−2 +2 , . . . , vk , . . . , vk+2 )

(thus including the vertex vk itself), while for even dimension d = 2 we take Pk
passing through the d vertices

Pk := (vk−2 +1 , vk−2 +3 , . . . , vk−1 , vk+1 , . . . , vk+2 −1 )

(thus excluding the vertex vk ).

Definition 3.5 The higher pentagram map T takes a vertex vk of a generic twisted
n-gon in RPd to the intersection point of the d consecutive short-diagonal planes Pi
around vk . Namely, for odd d = 2 + 1 one takes the intersection of the planes

T vk := Pk− ∩ Pk− +1 ∩ · · · ∩ Pk ∩ · · · ∩ Pk+ ,

while for even d = 2 one takes the intersection of the planes

T vk := Pk− +1 ∩ Pk− +2 ∩ · · · ∩ Pk ∩ · · · ∩ Pk+ .

The corresponding map T is well defined on the equivalence classes of n-gons in RPd .
As usual, we invoke the generality assumption to guarantee that every d consecutive
hyperplanes Pi intersect at one point in RPd . It turns out that the pentagram map defined
this way has a special scaling invariance, which allows one to prove its integrability:
Theorem C (= Theorem 8.3) The scale-invariant higher pentagram map is a discrete
completely integrable system on equivalence classes of n-gons in RPd . It has an
explicit Lax representation with a spectral parameter.
Remark 3.6 The pentagram map defined this way in 1D is the identity map. In the 2D
case this definition was given in [13] and its integrability for twisted polygons was
proved in [11], while for closed ones it was proved in [12,15].
Remark 3.7 One can also give an “asymmetric definition” for planes Pk , where more
general sequences of d vertices vk j are used, and then T vk is defined as the intersection
of d consecutive planes Pk . It turns out, however, that exactly this “uniform” definition
of diagonal planes Pk , where Pk passes through every other vertex, leads to integrability
of the pentagram map.

123
Integrability of higher pentagram maps 1015

One of possible definitions of the pentagram map discussed in [9] coincides with
ours in 3D. In that definition one takes the intersection of a (possibly asymmetric,
but containing the vertex vk ) plane Pk and the segment [vk−1 , vk+1 ]: for our centered
choice of Pk this segment belongs to both planes Pk−1 and Pk+1 . The definitions
become different in higher dimensions.

3.3 General pentagram maps and duality

We define more general pentagram maps T p,r depending on two integral parameters
in arbitrary dimension d. These parameters specify the diagonal planes and which of
them to intersect.
Definition 3.8 For a generic twisted n-gon in RPd one can define a p-diagonal hyper-
plane Pk as the one passing through d vertices of the n-gon by taking every pth vertex
starting at the point vk , i.e.,

Pk := (vk , vk+ p , . . . , vk+(d−1) p ).

The image of the vertex vk under the general pentagram map T p,r is defined by
intersecting every r th out of the p-diagonal hyperplanes starting with Pk :

T p,r vk := Pk ∩ Pk+r ∩ · · · ∩ Pk+(d−1)r .

The corresponding map T p,r is considered on the space of equivalence classes of

n-gons in RPd .
For the higher pentagram map T discussed in Sect. 3.2 one has p = 2, r = 1, and
the indices in the definition of Pk are “centered” at vk . In other words, T = T2,1 ◦ Sh,
where Sh stands for some shift in the vertex index. Below we denote by Sh any shift
in the index without specifying the shift parameter. Note that T p, p = Sh.
Theorem 3.9 There is a duality between the general pentagram maps T p,r and Tr, p :

T p,r = Tr,−1
p ◦ Sh.

For example, the map T1,2 in 2D is defined by extending the sides of a polygon and
intersecting them with the “second neighbouring” sides. This corresponds exactly to
passing from T (P) back to P in Fig. 1, i.e., it is the inverse of T modulo the numeration
of vertices.
Proof To prove this theorem we introduce the following “duality maps,” cf. [11]. 

Definition 3.10 Given a generic sequence of points φ( j) ∈ RPd , j ∈ Z, and a
nonzero integer p we define the sequence α p (φ( j)) ∈ (RPd )∗ as the plane

α p (φ( j)) := (φ( j), φ( j + p), . . . , φ( j + (d − 1) p)).

123
1016 B. Khesin, F. Soloviev

The following proposition is straightforward for our definition of the general pen-
tagram map.

Proposition 3.11 For every nonzero p the maps α p are involutions modulo index
shifts (i.e., α 2p = Sh), they commute with the shifts (i.e., α p ◦ Sh = Sh ◦ α p ), and the
general pentagram map is the following composition: T p,r = αr ◦ α p ◦ Sh.

To complete the proof of the theorem we note that

Tr,−1
p = (α p ◦ αr ◦ Sh)
−1
= Sh −1 ◦ αr−1 ◦ α −1
p = Sh ◦ αr ◦ Sh ◦ α p ◦ Sh
= αr ◦ α p ◦ Sh = T p,r ◦ Sh,

since α −1
p = α p ◦ Sh from the Proposition above, while Sh
−1 = Sh and Sh ◦ Sh = Sh.

Below we construct a Lax form for the higher pentagram maps, i.e., for the maps
T2,1 (and hence for T1,2 as well) for any d. Integrability of the pentagram map on a
special class of the so-called corrugated twisted polygons in RPd was proved in [3],
which should imply the integrability of the pentagram map T p,1 in 2D. Then the above
duality would also give integrability of T1, p in RP2 , defined as the intersection of a
pair of polygon edges whose numbers differ by p. (One should also mention that for p
and r mutually prime with n one can get rid of one of the parameters by appropriately
renumbering vertices, at least for the closed n-gon case, cf. [12] in 2D. This reduces
the study to the map T p,1 for some values of n.) Complete integrability of general
pentagram maps for other pairs ( p, r ) in RPd is a wide open problem.

Problem 3.12 Which of the general pentagram maps T p,r in RPd are completely
integrable systems?

4 Continuous limit of the higher pentagram maps

4.1 Definition of the continuous limit

In this section we consider the continuous limit of polygons and the pentagram map on
them. In the limit n → ∞ a twisted n-gon becomes a smooth quasi-periodic curve γ (x)
in RPd . Its lift G(x) to Rd+1 is defined by the conditions: i) the components of the
vector function G(x) := (G 1 (x), . . . , G d+1 (x)) provide homogeneous coordinates
for γ (x) = (G 1 : . . . : G d+1 )(x) in RPd , ii) det |G, G , . . . , G (d) |(x) = 1 for all
x ∈ R, and iii) G(x + 2π ) = M G(x) for a given M ∈ S L(d + 1, R). Then G satisfies
a linear differential equation of order d + 1:

G (d+1) + u d−1 (x)G (d−1) + · · · + u 1 (x)G + u 0 (x)G = 0 (2)

with periodic coefficients u i (x). Here and below stands for d/d x.
Let us consider the case of odd d = 2 + 1. Fix a small > 0. A continu-
ous analog of the hyperplane Pk is the hyperplane P (x) passing through d points

123
Integrability of higher pentagram maps 1017

Fig. 4 The envelope L (x) in

3D. The point L (x) and the
vectors L (x) and L (x) belong
to the plane
(G(x − ), G(x), G(x + ))

γ (x − ), . . . , γ (x), . . . , γ (x + ) of the curve γ .3 Note that as → 0 the hyper-

planes P (x) tend to the osculating hyperplane of the curve γ spanned by the vectors
γ (x), γ (x), . . . , γ (d−1) (x) at the point γ (x).
Let (x) be the envelope curve for the family of hyperplanes P (x) for a fixed
. The envelope condition means that for each x the point (x) and the derivative
(d−1)
vectors (x), . . . , (x) belong to the plane P (x). This means that the lift of
(x) to L (x) in R d+1 satisfies the system of d = 2 + 1 equations (see Fig. 4):

det |G(x − ), G(x − ( − 1) ), . . . , G(x), . . . ,

G(x + ), L ( j) (x)| = 0, j = 0, . . . , d − 1. (3)

Similarly, for even d = 2 the lift L (x) satisfies the system of d equations:

det |G(x − (2 − 1) ), G(x − (2 − 3) ), . . . , G(x − ), G(x + ), . . . ,

G(x + (2 − 1) ), L ( j) (x)| = 0, j = 0, . . . , d − 1. (4)

The evolution of the curve γ in the direction of the envelope , as changes,

defines a continuous limit of the pentagram map T . Namely, below we show that the
expansion of L (x) has the form

L (x) = G(x) + 2
B(x) + O( 4 ).

The family of functions L (x) satisfies a family of differential equations:

L (d+1) + u d−1, (x)L (d−1) + · · · + u 1, (x)L + u 0, (x)L = 0,

where u j,0 (x) = u j (x).

Expanding the coefficients u j, (x) as u j, (x) = u j (x) + 2 w j (x) + O( 4 ), we
define the continuous limit of the pentagram map T as the system of evolution differ-
ential equations du j (x)/dt = w j (x) for j = 0, . . . , d − 1, i.e., 2 plays the role of
time.
Theorem A The continuous limit du j (x)/dt = w j (x), j = 0, . . . , d − 1 for x ∈ S 1
of the pentagram map is the (2, d + 1)-KdV equation of the Adler-Gelfand-Dickey
hierarchy on the circle.

3 For a complete analogy with the discrete case, one could take the points γ (x − 2 ),
γ (x − 2( − 1) ), . . . , γ (x), . . . , γ (x + 2 ). However, one can absorb the factor 2 by rescaling → 2 .

123
1018 B. Khesin, F. Soloviev

This theorem is proved as a combination of Theorems 4.3 and 4.5 below.

Remark 4.1 Recall the definition of the KdV hierarchy (after Adler et al. [1]). Let L
be a linear differential operator of order d + 1:

L = ∂ d+1 + u d−1 (x)∂ d−1 + u d−2 (x)∂ d−2 + · · · + u 1 (x)∂ + u 0 (x) (5)

with periodic coefficients u j (x), where ∂ k stands for d k /d x k . One can define its
fractional power L m/d+1 as a pseudo-differential operator for any positive integer m
and take its purely differential part Q m := (L m/d+1 )+ . In particular, for m = 2 one
has Q 2 = ∂ 2 + d+1 2
u d−1 (x). Then the (m, d + 1)-KdV equation is the following
evolution equation on (the coefficients of) L:

d
L = [Q m , L].
dt

Remark 4.2 a) For d = 1 the discrete pentagram map is the identity map, hence the
continuous limit is trivial, which is consistent with vanishing of the (2,2)-KdV
equation.
b) For d = 2 the (2,3)-KdV equation is the classical Boussinesq equation, found in
[11].
c) Apparently, the (2, d + 1)-KdV equation is a very robust continuous limit. One
obtains it not only for the pentagram map defined by taking every other vertex,
but also for a non-symmetric choice of vertices for the plane Pk , see Remark 4.4.
Also, the limit remains the same if instead of taking the envelopes one starts with
a map defined by taking intersections of various planes [9].

4.2 Envelopes and the KdV hierarchy

Theorem 4.3 For any dimension d, the envelope L (x) has the expansion
 
2
L (x) = G(x) + 2
Cd G (x) + u d−1 G(x) + O( 4 )
d +1

for a certain constant Cd , as → 0.

 
The 2 -term of this expansion can be rewritten as Cd ∂ 2 + d+1 2
u d−1 (x) G(x).
Consequently, it defines the following evolution of the curve G(x):
 
d 2
G = ∂2 + u d−1 G.
dt d +1

Proof Since L approaches G as → 0, one has the expansion L = G + A +

2 B + 3 C + O( 4 ). First we note that the expansion of L in has only even powers

of , since the Eqs. (3) and (4) defining L have the symmetry → − . Therefore,
we have A = C = 0 and L = G + 2 B + O( 4 ).

123
Integrability of higher pentagram maps 1019

Notice that G(x) with its first d derivatives form a basis in Rd+1 for each x. We
express the vector coefficient B in this basis: B = b0 (x)G+b1 (x)G +· · ·+bd (x)G (d) .
Recall that, e.g., for odd d = 2 + 1 the lift L (x) satisfies the system of d equations:

det G(x − ), G(x − ( − 1) ), . . . , G(x), . . . , G(x + ), L ( j) (x) = 0,

j = 0, . . . , d − 1.
2
Fix x and expand all terms in : e.g., G(x + ) = G(x) + G (x) + 2 G + · · · , etc. In
each equation consider the coefficients at the lowest power of , being 2 + d(d − 1)/2
here.
The equation with j = 0 gives Cd det |G, G , . . . , G (d−1) , B| = 0 for some
nonzero Cd , which implies that there is no G (d) -term in the expansion of B. Sim-
ilarly, for j = 1 we obtain Cd det |G, G , . . . , G (d−1) , B | = 0, which means that
there is no G (d) -term in the expansion of B , or, equivalently, there is no G (d−1) -term
in the expansion of B. Using this argument for j = 0, 1, . . . , d − 3, and d − 1, we
deduce that B contains no terms with G (d) , . . . , G and G .
The equation with j = d − 2 results in a different term in the expansion and gives

det G, G , . . . , G (d−1) , B (d−2) = Cd det G, G , . . . , G (d−1) , G (d) ,

which implies that B = Cd G + b(x)G for some function b(x).

(d)
Finally, the normalization det | L , L , . . . , L | = 1 allows one to find b(x) by
plugging in it L = G + (Cd G + b(x)G) + O( 4 ). For the 2 -terms one obtains
2

Cd det G, G , . . . , G (d−2) , G (d+1) , G (d) + Cd det G, G , . . . , G (d−2) , G (d−1) ,

G (d+2) + (d + 1)b(x) det G, G , . . . , G (d−2) , G (d−1) , G (d) = 0.

By using the linear differential equation G (d+1) +u d−1 (x)G (d−1) +· · ·+u 0 (x)G = 0
to express G (d+1) and G (d+2) via lower derivatives we see that the first and the second
determinants are equal to −u d−1 , while the last one is equal to 1. Thus one has
(d + 1)b(x) − 2Cd u d−1 (x) = 0, which gives b(x) = Cd d+1 2
u d−1 (x).
Hence L = G + Cd (G + d+1 u d−1 G) + O( ), as required.
2 2 4 

Remark 4.4 One can see that the only condition on vertices defining the hyperplane Pk
required for the proof above is that they are distinct. A different choice of vertices for
the hyperplane Pk changes the constant Cd , but does not affect the evolution equation
for G. The above theorem for an envelope L (x) is similar to an expansion in [9] for
a curve defined via certain plane intersections.
Theorem 4.5 In any dimension d the continuous limit of the pentagram map defined
by the evolution
 
d 2
G = ∂2 + u d−1 G
dt d +1

123
1020 B. Khesin, F. Soloviev

of the curve G coincides with the (2, d + 1)-KdV equation. Consequently, it is an

infinite-dimensional completely integrable system.
Proof Recall that the (2, d + 1)-KdV equation is defined as the evolution

L̇ = [Q 2 , L] := Q 2 L − L Q 2 ,

where the linear differential operator L of order d + 1 is given by formula (5) and
Q 2 = ∂ 2 + d+1
2
u d−1 . Here L̇ stands for d L/dt.
By assumption, the evolution of the curve G is described by the differential equation
Ġ = Q 2 G. We would like to find the evolution of the operator L tracing it. For any
t, the curve G and the operator L are related by the differential equation LG = 0 of
order d + 1. Consequently,

L̇G + L Ġ = 0.

Note that if the operator L satisfies the (2, d + 1)-KdV equation and G satisfies
Ġ = Q 2 G, we have the identity:

L̇G + L Ġ = (Q 2 L − L Q 2 )G + L Q 2 G = Q 2 LG = 0.

In virtue of the uniqueness of the linear differential operator of the form (5) for a given
fundamental set of solutions, we obtain that indeed the evolution of L is described by
the (2, d + 1)-KdV equation. 

Remark 4.6 The proof above is reminiscent of the one used in [10] to study symplectic
leaves of the Gelfand-Dickey brackets. Note that the absence of the term linear in is
related to the symmetric choice of vertices for the hyperplane Pk . For a non-symmetric
choice the evolution would be defined by the linear term in and given by the equation
Ġ = G . This is the initial, (1, d + 1)-equation of the corresponding KdV hierarchy,
manifesting the fact that the space x-variable can be regarded as the “first time”
variable. A natural question arises whether the whole KdV-hierarchy is hidden as an
appropriate limit of the pentagram map. An evidence to this is given by noticing that
the terms with higher powers in lead to equations similar to the higher equations in
the KdV hierarchy, see Appendix B.
Remark 4.7 One can see that the continuous limit of the general pentagram maps T p,r
for various p = r in RPd defined via envelopes for a centered choice of vertices is
the same (2, d + 1)-KdV flow, i.e., the limit depends only on the dimension.
Indeed, an analog of the p-diagonal is the plane P (x) passing through the points
G(x), G(x + p), . . . , G(x + (d − 1) p). Rescaling , we can assume the points to
be G(x), G(x + ), . . . , G(x + (d − 1)), which leads to the planes P (x) defined
in Sect. 4.1 after a shift in x. Then the definition of L (x) via the envelope of such
planes will give the same (2, d + 1)-KdV equation.
It would be interesting to define the limit of the intersections of every r th plane via
some higher-order terms of the envelope, as mentioned in the above remark, so that it
could lead to other (m, d + 1)-equations in the KdV hierarchy.

123
Integrability of higher pentagram maps 1021

5 Explicit formulas for the 3D pentagram map

5.1 Two involutions

Now we return to the 3D case. In this section we assume that n is odd and consider n-
gons in RP3 . Recall that in this case an n-gon with a given monodromy M ∈ S L(4, R)
lifts uniquely to R4 and is described by difference equations

V j+4 = a j V j+3 + b j V j+2 + c j V j+1 − V j , j ∈ Z, (6)

with n-periodic coefficients (a j , b j , c j ). In other words, for odd n the variables

(a j , b j , c j ), 0 ≤ j ≤ n − 1, provide coordinates on the space Pn of twisted
n-gons in RP3 considered up to projective equivalence (see Proposition 5.9).
In order to find explicit formulas for the pentagram map, we present it as a compo-
sition of two involutions α and β, cf. Sect. 3.3, and then find the formulas for each of
them separately. The same approach was used in [11] in 2D, although the formulas in
3D are more complicated.
Definition 5.1 Given a sequence of points φ( j) ∈ RP3 , j ∈ Z, define two sequences
α(φ( j)) ∈ (P3 )∗ and β(φ( j)) ∈ (RP3 )∗ , where
a) α(φ( j)) is the plane (φ( j − 1), φ( j), φ( j + 1));
b) β(φ( j)) is the plane (φ(( j − 2), φ( j), φ( j + 2)).
Proposition 5.2 The maps α and β are involutions, i.e., α 2 = β 2 = Id, while the
pentagram map is their composition: T = α ◦ β.
Note that the indices which define the planes are symmetric with respect to j. As
a result, we do not have an extra shift of indices, cf. Proposition 3.11 (unlike the 2D
case and the general map α p ).
Lemma 5.3 The involution α : Vi → Wi = ∗(Vi ∧ Vi−1 ∧ Vi+1 ) maps Eq. (6) to the
following difference equation:

Wi+4 = ci+1 Wi+3 + bi Wi+2 + ai−1 Wi+1 − Wi .

Lemma 5.4 The involution β : Vi → Wi = λi ∗ (Vi ∧ Vi−2 ∧ Vi+2 ) maps Eq. (6) to
the difference equation

Wi+4 = Ai Wi+3 + Bi Wi+2 + Ci Wi+1 − Wi ,

where the coefficients Ai , Bi , Ci are defined as follows:

Ai = ci−1 (ai ai+2 + ai+2 bi+1 ci + ci ci+2 )2 λi+1 λi+2 λi+4

2
,
Bi = ((ai−2 + bi−1 ci−2 )(ci+2 + ai+2 bi+1 ) − ai+2 ci−2 )
×(ai−1 ai+1 + ai+1 bi ci−1 + ci−1 ci+1 )λi λi+1 λi+3 λi+4 ,
Ci = ai+1 (ai ai+2 + ai+2 bi+1 ci + ci ci+2 )2 λi+2 λi+3 λi+4
2
.

123
1022 B. Khesin, F. Soloviev

The sequence λi , i ∈ Z, is n-periodic and is uniquely determined by the condition

λi λi+1 λi+2 λi+3

1
= .
(ai−2 ai + ai bi−1 ci−2 + ci−2 ci )(ai−1 ai+1 + ai+1 bi ci−1 + ci−1 ci+1 )

The proofs of these lemmas are straightforward computations, which we omit.

Combined together, these lemmas provide formulas for the pentagram map. They
have, however, one drawback: one needs to solve a system of equations in λi , i ∈ Z,
which results in the non-local character of the formulas in (a, b, c)-coordinates and
their extreme complexity.

5.2 Cross-ratio type coordinates

Similarly to the 2D case, there exist alternative coordinates on the space Pn . They are
defined for any n, and the formulas for the pentagram map become local, i.e., involving
the vertex φ( j) itself and several neighboring ones.
Definition 5.5 For odd n the variables
b j+1 aj c j+1
xj = , yj = , zj = , 0 ≤ j ≤ n − 1,
a j a j+1 b j+1 c j a j+1 b j

provide coordinates on the space Pn , where the n-periodic variables (a j , b j , c j ), j ∈

Z, are defined by the difference equation (6).
It turns out that the variables (x j , y j , z j ), 0 ≤ j ≤ n − 1 are well defined and
independent for any n, even or odd. Below we provide two (equivalent) ways to define
them for even n: a pure geometric (local) definition of these variables (see Proposition
5.7) and the above definition extended to quasi-periodic sequences (a j , b j , c j ) (see
Sect. 5.3).
Theorem 5.6 In the coordinates xi , yi , z i the pentagram map for any (either odd or
even) n is given by the formulas:

1 + yi−1 + z i+2 + yi−1 z i+2 − yi+1 z i

T ∗ (xi ) = xi+1 ,
1 + yi−1 + z i
xi−1 yi−1 z i (1 + yi+1 + z i+2 )(1 + yi−2 + z i−1 )
T ∗ (yi ) = ,
xi z i−1 (1 + yi + z i+1 )(1 + yi−1 + z i+2 + yi−1 z i+2 − yi+1 z i )
xi+1 z i (1 + yi+1 + z i+2 )(1 + yi−2 + z i−1 )
T ∗ (z i ) = .
xi (1 + yi−1 + z i )(1 + yi−2 + z i+1 − yi z i−1 + yi−2 z i+1 )

Before proving this theorem we describe the (xi , yi , z i ) coordinates in greater detail.
It turns out that they may be defined completely independently of (ai , bi , ci ) in the
following geometric way.
Recall that the x, y coordinates for the 2D pentagram map are defined as cross-
ratios for quadruples of points on the line (Vi , Vi+1 ), where two points are these

123
Integrability of higher pentagram maps 1023

vertices themselves, and two others are intersections of this line with extensions of
the neighbouring edges. Similarly, the next proposition describes the new coordinates
via cross-ratios of quadruples of points, 2 of which are the vertices of an edge, and
2 others are the intersection of the edge extension with two planes. For instance, the
variable yi is the cross-ratio of 4 points on the line (Vi , Vi+1 ), two of which are Vi
and Vi+1 , while two more are constructed as intersections of this line with the planes
via the triple (Vi+2 , Vi+3 , Vi+4 ) and with the plane via the triple (Vi+2 , Vi+4 , Vi+5 ).
More precisely, the following proposition holds.
Proposition 5.7 The coordinates xi , yi , z i are given by the cross-ratios:

012 , 123 ,
xi = − Vi+4 , Vi+5 , 45 45

234 , 245 ,
yi = − Vi , Vi+1 , 01 01

013 , 123 ,
z i = − Vi+4 , Vi+5 , 45 45

j ,j
where the point m11 ,m2 2 ,m 3 for a given i is the intersection of the line (Vi+ j1 , Vi+ j2 )
with the plane (Vi+m 1 , Vi+m 2 , Vi+m 3 ).
By the very definition these coordinates are projectively invariant.
Proof of proposition If ∗ is the Hodge star operator with respect to the Euclidean
metric in R4 (or C4 ), then
j ,j
m11 ,m2 2 ,m 3 = ∗ ∗(Vi+ j1 ∧ Vi+ j2 ) ∧ ∗(Vi+m 1 ∧ Vi+m 2 ∧ Vi+m 3 ) .

It suffices to prove the proposition in the case of an odd n, because the formulas are
local, and we can always add a vertex to change the parity of n. Therefore, we may
assume that (a j , b j , c j ) are global coordinates and use them for the proof.
A simple computation shows that

45
012 = −(bi+1 + ai ai+1 )Vi+4 + ai Vi+5 ,
01
234 = Vi − ci Vi+1 ,
01
245 = −bi+1 Vi + (ai + ci bi+1 )Vi+1 ,

013 = (ci+1 + bi ai+1 )Vi+4 − bi Vi+5 ,

45
45
123 = −ai+1 Vi+4 + Vi+5 .

Recall (see Lemma 4.5 in [11]) that if 4 vectors a, b, c, d ∈ R4 (or C4 ) lie in the same
2-dimensional plane and are such that

c = λ1 a + λ2 b, d = μ1 a + μ2 b,

then the cross-ratio of the lines spanned by these vectors in the plane is given by

λ2 μ1 − λ1 μ2
[a, b, c, d] = .
λ2 μ1

123
1024 B. Khesin, F. Soloviev

Comparing the cross-ratios with the original definition of the variables xi , yi , z i con-
cludes the proof. 

Now we are in a position to prove the explicit local formulas in Theorem 5.6. The
proof is similar, but more involved than that of Proposition 4.11 in [11].
Proof of theorem Due to the local character of the formulas for the pentagram map T
in (x j , y j , z j )-coordinates, we may always add an extra vertex to make the number n
of vertices odd, and then use coordinates (a j , b j , c j ) and Lemmas 5.3 and 5.4 for the
proof.
The pentagram map is a composition T = α ◦ β : Vi → Ui . Namely,
 
Ui = μi ∗ ∗(Vi−3 ∧ Vi−1 ∧ Vi+1 )∧∗(Vi−2 ∧Vi ∧ Vi+2 ) ∧ ∗(Vi−1 ∧ Vi+1 ∧ Vi+3 ) ,

where the constants μi are chosen so that det |U j , U j+1 , U j+2 , U j+3 | = 1 for all j.
At the level of the coordinates (a j , b j , c j ), we obtain:

Bi+1 Ci+1 Ai
T ∗ (xi ) = , T ∗ (yi ) = , T ∗ (z i ) = ,
Ci+1 Ci+2 Bi+1 Ai−1 Ci+2 Bi

where Ai , Bi , Ci are defined in Lemma 5.4. Eliminating the variables λi with different
i by using the formula for the product λi λi+1 λi+2 λi+3 concludes the proof. 


5.3 Coordinates on twisted polygons: odd versus even n

In this section we compare how one introduces the coordinates on the space Pn of
twisted n-gons for odd or even n, and how this changes the statements above.
Definition 5.8 Call a sequence (a j , b j , c j ), j ∈ Z, n-quasiperiodic if there is a
sequence t j , j ∈ Z, satisfying t j t j+1 t j+2 t j+3 = 1 and such that

tj tj tj
a j+n = a j , b j+n = b j , c j+n = c j (7)
t j+3 t j+2 t j+1

for each j ∈ Z.
Note that a sequence t j , j ∈ Z, must be 4-periodic, and it is defined by three
parameters, e.g., by α := t0 /t3 , β := t0 /t2 , and γ := t0 /t1 with αβγ > 0, and hence
t0 = (αβγ )1/4 . Thus the space QS n of n-quasiperiodic sequences has dimension
3n + 3, and {(a j , b j , c j ), j = 0, . . . , n − 1} × (α, β, γ ) are coordinates on it.
Now we associate a sequence of vectors V j ∈ R4 , j ∈ Z, and difference equations

V j+4 = a j V j+3 + b j V j+2 + c j V j+1 − V j , j ∈ Z, (8)

to each twisted n-gon v j := φ( j) ∈ RP3 , j ∈ Z, with a monodromy M ∈ S L(4, R).

This gives a correspondence between sequences (a j , b j , c j ), j ∈ Z, and twisted
n-gons.

123
Integrability of higher pentagram maps 1025

Proposition 5.9 There is a one-to-one correspondence between twisted n-gons

(defined up to projective equivalence) and three-parameter equivalence classes in the
space QS n of n-quasiperiodic sequences {(a j , b j , c j ), j = 0, . . . , n −1}×(α, β, γ ).
If n is odd, there exists a unique n-periodic sequence (a j , b j , c j ) in each class.
If n = 4 p, then the numbers α, β, γ are projective invariants of a twisted n-gon.
If n = 4 p + 2, then there is one projective invariant: αγ /β.
In other words, for odd n the equivalence classes are “directed along” the parameters
(α, β, γ ) and one can chose a representative with α = β = γ = 1 in each class. For
n = 4 p the classes are “directed across” these parameters, and hence the latter are
fixed for any given class. The case n = 4 p + 2 is in between: in a sense, two of the
(α, β, γ )-parameters and one of the (a, b, c)-coordinates can serve as coordinates on
each equivalence class.
This proposition can be regarded as an analogue of Proposition 4.1 and Remark 4.4
in [11] for d = 2.
Proof First, we construct the correspondence, and then consider what happens for
different arithmetics of n. For a given n-gon vk = φ(k) we construct a sequence of
vertices V j ∈ R4 , j ∈ Z, in the following way: choose the lifts φ(0) → V0 , φ(1) →
V1 , φ(2) → V2 arbitrarily, and then determine the vectors V j , j > 2, and V j , j < 0,
recursively using the condition det(V j , V j+1 , V j+2 , V j+3 ) = 1, which follows from
Eq. (8).
By definition of a twisted n-gon, we have φ( j + n) = M ◦ φ( j) for each
j ∈ Z, where M ∈ S L(4, R). Consequently, for each j ∈ Z there exists a num-
ber t j , such that V j+n = t j M V j , and the matrix M is independent of j. The equa-
tion det(V j+n , V j+1+n , V j+2+n , V j+3+n ) = 1 implies that t j t j+1 t j+2 t j+3 = 1 and
t j+4 = t j for each j ∈ Z. In other words, the whole sequence {t j } is determined
by t1 , t2 , t3 , and then t0 = 1/t1 t2 t3 . Quasiperiodic conditions (7) follow from the
comparison of the equation

V j+4+n = a j+n V j+3+n + b j+n V j+2+n + c j+n V j+1+n − V j+n

with Eq. (8).

Now we rescale the initial three vectors: V0 → k0 V0 , V1 → k1 V1 , V2 → k2 V2 ,
where k0 k1 k2 = 0. A different lift of the three initial vectors corresponds to a different
sequence Ṽ j = k j V j , where the sequence k j , j ∈ Z, must also be 4-periodic and
satisfy k0 k1 k2 k3 = 1. This rescaling gives the action of (R∗ )3 on the space QS n of
n-quasiperiodic sequences. By construction, the corresponding orbits (or equivalence
classes) of sequences are in a bijection with twisted n-gons. The group (R∗ )3 acts as
follows:

t j → t j (k j /k j+n ), a j → a j (k j /k j+3 ), b j → b j (k j /k j+2 ), c j → c j (k j /k j+1 ).

Now we have 3 cases:

• n is odd. Then the above (R∗ )3 -action on t0 , t1 , t2 allows one to make them all
equal to 1, which corresponds to the constant sequence {t j = 1, j ∈ Z} and a

123
1026 B. Khesin, F. Soloviev

periodic sequence {(a j , b j , c j ), j ∈ Z}. Indeed, e.g., for n = 4 p + 3 one has

the system of 3 equations: t0 (k0 /k3 ) = 1, t1 (k1 /k4 ) = 1, t2 (k2 /k5 ) = 1. Since
k4 = k0 , k5 = k1 , and k3 = 1/k0 k1 k2 , we obtain a system of 3 equations on the
unknowns k0 , k1 , k2 , which has the unique solution.
• n = 4 p + 2. One can check that the (R∗ )3 -action does not change the ratio
t0 t2 αγ
t1 t3 = β .
• n = 4 p. The (R∗ )3 -action does not change the three quantities t0 /t3 , t0 /t2 , t0 /t1 .


Now we can introduce coordinates on the space QS n of n-quasiperiodic sequences
using Definition 5.5 and quasiperiodic variables (a j , b j , c j ).
Proposition 5.10 For any n the variables (x j , y j , z j ), 0 ≤ j ≤ n − 1 are indepen-
dent and constant on the equivalence classes in QS n , i.e., they are well-defined local
coordinates on the space Pn = QS n / ∼.
Proof It is straightforward to check that the (R∗ )3 -action defined above is trivial on
the variables (x j , y j , z j ). For instance,

b j+1 b j+1 (k j+1 /k j+3 )

xj = → = xj.
a j a j+1 a j (k j /k j+3 ) a j+1 (k j+1 /k j+4 )

The independence of the new variables on QS n follows from that for the original ones.
Alternatively, it also follows from their local geometric definition (Proposition 5.7).


Remark 5.11 In the (a, b, c)-coordinates for even n some of the (α, β, γ )-parameters
were needed to describe the equivalence classes in QS n . In the (x, y, z)-coordinates
those parameters are functions of x j , y j , z j :
i) for n = 4 p + 2,


2p
x22 j y2 j z 2 j+1 αγ
= ;
x22 j+1 y2 j+1 z 2 j β
j=0

ii) for n = 4 p,


p−1
x4 j x4 j+2 y4 j+2 z 4 j+1 
p−1
y4 j+1 z 4 j 
p−1
y4 j z 4 j+3 γ
= α, = β, = .
x4 j+1 x4 j+3 y4 j+3 z 4 j+2 y4 j+3 z 4 j+2 y4 j+2 z 4 j+1 α
j=0 j=0 j=0

These identities follow from Definitions 5.5 and 5.8.

6 Algebraic-geometric integrability of the 3D pentagram map

In this section we complexify the pentagram map and assume that everything is defined
over C.

123
Integrability of higher pentagram maps 1027

6.1 Scaling transformations and a Lax function in 3D

Recall that a discrete Lax equation with a spectral parameter is a representation of a

dynamical system in the form
−1
L i,t+1 (λ) = Pi+1,t (λ)L i,t (λ)Pi,t (λ), (9)

where t stands for the discrete time variable, i refers to the vertex index, and λ is a
complex spectral parameter.
The pivotal property responsible for algebraic-geometric integrability of all penta-
gram maps considered in this paper is the presence of a scaling invariance. In the 2D
case, this means the invariance with respect to the transformations a j → a j s, b j →
b j /s, where s is an arbitrary number. In the 3D case, the pentagram map is invariant
with respect to the transformations a j → a j s, b j → b j , c j → c j s. In both cases the
invariance follows from the explicit formulas of the map. Note that formally one can
define other pentagram maps by choosing the intersection planes in many different
ways. However, only very few of these maps possess any scaling invariance. Below
we derive a Lax representation from the scaling invariance. First we do it for odd n,
when (a j , b j , c j ), 0 ≤ j ≤ n − 1, are coordinates on the space Pn .
Theorem 6.1 The 3D pentagram map on twisted n-gons with odd n admits a Lax
representation with the Lax function L j (λ) given by
⎛ ⎞ ⎛ ⎞−1
c j /λ 1/λ 0 0 0 0 0 −1
⎜ bj 0 ⎟ ⎜ cj ⎟
⎜
L j (λ) = ⎝
0 1 ⎟ = ⎜λ 0 0 ⎟
a j /λ 0 0 1/λ⎠ ⎝0 1 0 bj ⎠
−1 0 0 0 0 0 λ aj

in the coordinates a j , b j , c j . Its determinant is det L j ≡ 1/λ2 .

Note that we always consider a polygon and the corresponding Lax function at
a particular moment of time. Whenever necessary we indicate the moment of time
explicitly by adding the second index “t” to the Lax function (above L j := L j,t ),
while if there is no ambiguity we keep only one index. Before proving this theorem
we give the following
Definition 6.2 The monodromy operators T0,t , T1,t , . . . , Tn−1,t are defined as the fol-
lowing ordered products of the corresponding Lax functions:

T0,t = L n−1,t L n−2,t , . . . , L 0,t ,

T1,t = L 0,t L n−1,t L n−2,t , . . . , L 1,t ,
...
Ti,t = L i+n−1,t L i+n−2,t , . . . , L i+1,t L i,t ,
...
Tn−1,t = L n−2,t L n−3,t , . . . , L 0,t L n−1,t ,

where the (integer) index t represents the moment of time.

123
1028 B. Khesin, F. Soloviev

Proof of theorem First observe that the Lax equation implies that the corresponding
monodromy operators satisfy

−1
Ti,t+1 (λ) = Pi,t (λ)Ti,t (λ)Pi,t (λ),

i.e., Ti,t (λ) changes to a similar matrix when t → t + 1, and hence the eigenvalues of
the matrices Ti,t (λ) as functions of λ are invariants of the map. Conversely, if some
function Ti,t (λ) has this property, then there must exist a matrix Pi,t (λ) (defined up to
a multiplication by a scalar function) satisfying the above equation.
How to define such a monodromy depending on a parameter? The monodromy
matrix associated with the difference equation

V j+4 = a j V j+3 + b j V j+2 + c j V j+1 − V j

is M = N0 N1 N2 , . . . , Nn−1 , where
⎛ ⎞
0 0 0 −1
⎜1 0 0 cj ⎟
Nj = ⎜
⎝0
⎟.
1 0 bj ⎠
0 0 1 aj

For odd n the variables (a j , b j , c j ), 0 ≤ j ≤ n − 1, are well-defined coordinates on

the space of twisted n-gons. These variables are periodic: for any j we have a j+n =
a j , b j+n = b j , c j+n = c j . The vectors V j are quasi-periodic: V j+n = M V j , and
depend on the lift of the points from the projective space. This means that the pentagram
map preserves the eigenvalues of the matrix M, but not the matrix M itself.
Lemmas 5.3 and 5.4 imply that the pentagram map is invariant with respect to the
scaling transformations: a j → sa j , c j → sc j . Therefore, the pentagram map also
preserves the eigenvalues of the monodromy matrix M(s) corresponding to the n-gons
scaled by s. Namely, we have
⎛ ⎞
0 0 0 −1
⎜1 0 0 sc j ⎟
M(s) = N0 (s)N1 (s)N2 (s), . . . , Nn−1 (s), where N j (s) = ⎜
⎝0
⎟.
1 0 bj ⎠
0 0 1 sa j

Now one can see that the matrix N j (s) can be chosen as a Lax function. For tech-
nical reasons (which will be clear later), we define the Lax matrix as L −1
j (λ) :=
−1
(g N j (s)g)/s, where g := diag(1, s, 1, s), and λ := 1/s . This gives the required
2

matrix L j (λ). 

As we mentioned before, the formulas for the pentagram map are non-local in the
variables (a j , b j , c j ). As a result, an explicit expression for the matrix Pi,t (λ) becomes
non-local as well. On the other hand, one can use the variables (x j , y j , z j ) (given by
Definition 5.5) to describe a Lax representation. Their advantage is that all formulas
become local and are valid for any n, both even and odd.

123
Integrability of higher pentagram maps 1029

Theorem 6.3 For any n the equations for the 3D pentagram map are equivalent to
the Lax equation
−1
L̃ i,t+1 (λ) = P̃i+1,t (λ) L̃ i,t (λ) P̃i,t (λ),

where
⎛ ⎞−1
0 0 0 −1
⎜λxi yi 0 0 1⎟
⎜
L̃ i,t (λ) = ⎝ ⎟ ,
0 zi 0 1⎠
0 0 λxi 1
⎛ ⎞
0 ρi 0 −ρi
⎜λσi (1 + z i ) −ρi λσi ρi ⎟
⎜ ⎟
P̃i,t (λ) = ⎜ y θ z i−1
−θi 1+yi−2 ⎟ ,
⎝ i−1 i τi τi ⎠
− 1+yλyi−1
i−1
+z i 0 λ
1+yi−1 +z i 0

and the variables ρi , σi , τi , and θi stand for

1
ρi = ,
xi (1 + yi + z i+1 )
xi−1 yi−1 (1 + yi−2 + z i−1 )
σi = ,
xi z i−1 (1 + yi−1 + z i )(1 + yi + z i+1 )
τi = xi (1 + yi−2 − yi z i−1 + z i+1 + z i+1 yi−2 ),
1 + yi−2 + z i−1
θi = .
τi (1 + yi−1 + z i )

Proof The proof is a long but straightforward verification. 


Remark 6.4 The Lax functions L and L̃ in the (ai , bi , ci ) and (xi , yi , z i ) variables are
related to each other as follows:
 
−1
L̃ i,t = ai+1 h i+1 L i,t h i , where h i := diag(1, ci , bi , ai ).

6.2 Properties of the spectral curve

Recall that the monodromy operators Ti,t (λ) satisfy the equation
−1
Ti,t+1 (λ) = Pi,t (λ)Ti,t (λ)Pi,t (λ).

It implies that the function of two variables R(λ, k) = det (Ti,t (λ) − k Id) is indepen-
dent of i and t. Furthermore, R(λ, k) = 0 is a polynomial relation between λ and k:
R(λ, k) becomes a polynomial after a multiplication by a power of λ. Its coefficients
are integrals of motion for the pentagram map. The zero set of R(λ, k) = 0 is an
algebraic curve in C2 . A standard procedure (of adding the infinite points and normal-
ization with a few blow-ups) makes it into a compact Riemann surface, which we call

123
1030 B. Khesin, F. Soloviev

the spectral curve and denote by . In this section we explore some of the properties
of the spectral curve and, in particular, find its genus.

Definition 6.5 For an odd n define the spectral function R(λ, k) as

R(λ, k) := det (Ti,t (λ) − k Id),

i.e., using the Lax function in the (a, b, c)-coordinates from Theorem 6.1. The spectral
curve  is the normalization of the compactification of the curve R(λ, k) = 0.
We define the integrals of motion I j , J j , G j , 0 ≤ j ≤ q = n/2, as the coeffi-
cients of the expansion
⎛ ⎞ ⎛ ⎞
q 
q
R(λ, k) = k 4 − k 3 ⎝ G j λ j−n ⎠ + k 2 ⎝ J j λ j−q−n ⎠
j=0 j=0
⎛ ⎞

q
−k ⎝ I j λ j−2n ⎠ + λ−2n = 0.
j=0

When n is even, the sequence (a j , b j , c j ), j ∈ Z, is not n-periodic, and the mon-

odromy operator Ti,t (λ) cannot be defined. One should use the Lax function L̃ i,t (λ) in
the (x, y, z)-coordinates from Theorem 6.3 to define the monodromy operator T̃i,t (λ)
and the corresponding spectral curve.
Namely, first note that the integral of motion I0 has the following explicit expression:

n−1 −1/4

n−1 
I0 = ai = xi2 yi z i .
i=0 i=0

Definition 6.6 For any n (either even or odd), the spectral function is

R(λ, k) = R̃(λ, k I0 )/I04 , where R̃(λ, k) := det (T̃i,t (λ) − k Id),

and the monodromy operator T̃i,t (λ) is defined using the Lax function L̃ i,t (λ) from
Theorem 6.3.

The spectral function R(λ,

 k) defined  this way coincides with det (Ti,t (λ) − k Id)
−1
for odd n, since T̃i,t = I0 h i Ti,t h i , see Remark 6.4. It is convenient to have such
a unified definition for computations of integrals of motion.

Theorem 6.7 For generic values of the integrals of motion I j , J j , G j , the genus g of
the spectral curve  is g = 3q for odd n and g = 3q − 3 for even n, where q = n/2.

To prove it, we first describe the singularities of R(λ, k) = 0 by considering the

formal series solutions (the so-called Puiseux series).

123
Integrability of higher pentagram maps 1031

Lemma 6.8 If n is even, the equation R(λ, k) = 0 has 4 formal series solutions at
λ = 0:
1 I1
O1 : k1 = − 2 λ + O(λ2 ),
I0 I0
 
k∗ 1
O2,3 : k2,3 = q + O , where k∗ satisfies G 0 k∗2 − J0 k∗ + I0 = 0,
λ λq−1
G0 G1 G2
O4 : k4 = n + n−1 + n−2 + O(λ3−n ),
λ λ λ
and 4 solutions at λ = ∞:
 
k∞ 1
W1,2,3,4 : k1,2,3,4 = + O ,
λq λq+1
4
where k∞ − G q k∞
3
+ Jq k∞
2
− Iq k∞ + 1 = 0.

If n is odd, the equation R(λ, k) = 0 has 4 formal series solutions at λ = 0:

1 I1
O1 : k1 = − 2 λ + O(λ2 ),
I0 I
√ 0  
−I0 /G 0 J0 1
O2 : k2,3 = ± + + O ,
λn/2 2G 0 λ(n−1)/2 λ(n−2)/2
G0 G1 G2
O3 : k4 = n + n−1 + n−2 + O(λ3−n ),
λ λ λ
and 4 solutions at λ = ∞:
 
k∞ 1
W1,2 : k1,2,3,4 = n/2 + O , where k∞
4
+ Jq k∞
2
+ 1 = 0.
λ λ(n+1)/2

The remaining coefficients of the series are determined uniquely.

Proof of lemma One finds the series coefficients recursively by substituting the series
into the equation R(λ, k) = 0, which determines the spectral curve. 

Now we can complete the proof of Theorem 6.7.
Proof of theorem As follows from the definition of the spectral curve , it is a ramified
fourfold cover of CP1 , since the 4 × 4-matrix T̃i,t (or Ti,t ) has 4 eigenvalues. By the
Riemann–Hurwitz formula the Euler characteristic of  is χ () = 4χ (CP1 ) − ν =
8 − ν, where ν is the number of branch points. On the other hand, χ () = 2 − 2g,
and once we know ν it allows us to find the genus of the spectral curve  from the
formula 2 − 2g = 8 − ν.
The number ν of branch points of  on the λ-plane equals the number of zeroes
of the function ∂k R(λ, k) aside from the singular points. The function ∂k R(λ, k) is
meromorphic on , therefore the number of its zeroes equals the number of its poles.
One can see that for any n the function ∂k R(λ, k) has poles of total order 9n at z = 0,

123
1032 B. Khesin, F. Soloviev

and it has zeroes of total order 6n at z = ∞. Indeed, substitute the local series from
Lemma 6.8 to the expression for ∂k R(λ, k). (E.g., for n = 2q at O4 one has k ∼ λ−n .
The leading terms of ∂k R(λ, k) for the pole at λ = 0 are k 3 , k 2 λ−n , kλ−q−n , λ−2n .
The first two terms, being of order λ−3n , dominate and give the pole order of 3n = 6q.)
The corresponding orders of the poles and zeroes of ∂k R(λ, k) on  are summarized
as follows:

n = 2q Pole Zero n = 2q + 1 Pole Zero

O1 4q W1 3q O1 2n W1 3n
O2 4q W2 3q O2 4n W2 3n
O3 4q W3 3q O3 3n
O4 6q W4 3q

For instance, for n = 2q this gives the total order of poles: 4q + 4q + 4q + 6q =

18q = 9n, while the total order of zeroes is 4 × 3q = 12q = 6n.
Therefore, the number of zeroes of ∂k R(λ, k) at nonsingular points λ = {0, ∞} is
ν = 9n − 6n = 3n, and so is the total number of branch points of  in the finite part
of the (λ, k) plane. The difference between odd and even values of n arises because 
has 2 additional branch points at λ = ∞, and 1 branch point at λ = 0 for odd n, i.e.,
ν = 3n + 3.
Consequently, one has 2 − 2g = 8 − ν with ν = 3n for even n and ν = 3n + 3 for
odd n. The required expression for the genus g follows: g = 3q − 3 for n = 2q and
g = 3q for n = 2q + 1. 

Remark 6.9 Now we describe a few integrals of motion using the coordinates
(a j , b j , c j ), 0 ≤ j ≤ n − 1, when n is odd. The description is similar to that in
the 2D case (cf. Section 5.2 and Proposition 5.3 in [11]). Consider a code which is an
ordered sequence of digits from 1 to 4. The number of digits in a code is p, q, r, t,
respectively. The code is called “admissible” if p + 2q + 3r + 4t = n. The number
p + r is called its “weight.” Each code expands in a “word” of n characters in the
following way: 1, 2, 3, 4 are replaced by “a”,“*b”,“**c”,“****”, respectively. Now
we label the vertices of an n-gon by 0, 1, . . . , n − 1, and associate letters in a word
to them keeping the order. We obtain one monomial by taking the product of the vari-
ables ai ,bi , or ci that occur at the vertex i. The letter “*” corresponds to “1”. The
sign of the monomial is (−1)t . Next step is to permute the numbering of the vertices
cyclically and take the sum of the monomials. Note, however, that if, for example,
n = 9, then the code “333” corresponds to the sum c0 c3 c6 + c1 c4 c7 + c2 c5 c8 without
the coefficient 3. Finally, we repeat this procedure for all admissible codes of weight
p + r and denote the total sum by Iˆp+r . Additionally, we define the sum Ĝ p+r by
substituting ai → ci+1 , ci → ai−1 in Iˆp+r .
Consider, for example, the case n = 7. Then all admissible codes of weight 1 are
142, 124, 1222, 34, 223. The corresponding sum is

Iˆ1 = (−a1 b0 − a5 b0 + a5 b0 b2 b4 − c0 + c0 b2 b4 ).
cyclic

123
Integrability of higher pentagram maps 1033

Proposition 6.10 For odd n one has


n−1
I0 = a j , Iq = Iˆ1 , Ii = Iˆn−2i ,
j=0


n−1
G0 = c j , G q = Ĝ 1 , G i = Ĝ n−2i .
j=0

Proof The proof is analogous to the proof of Proposition 5.3 in [11]. 



6.3 The spectral curve and invariant tori

The spectral curve is a crucial component of algebraic-geometric integrability. Below

we always assume it to be generic. (As everywhere in this paper, “generic” means
values of parameters from some Zariski open subset in the space of parameters.) It
has a natural torus, its Jacobian, associated with it. It turns out that one can recover
a Lax function from the spectral curve and a point on the Jacobian, and vice versa:
in our situation this correspondence is locally one-to-one. The dynamics of the pen-
tagram map becomes very simple on the Jacobian. In this section, we construct this
correspondence and describe the dynamics of the pentagram map.

Definition 6.11 A Floquet-Bloch solution ψi,t of a difference equation ψi+1,t =

L̃ i,t ψi,t is an eigenvector of the monodromy operator:

T̃i,t ψi,t = kψi,t .


The normalization 4j=1 ψ0,0, j ≡ 1 (i.e., the sum of all components of the vector ψ0,0
is equal to 1) determines all vectors ψi,t with i, t ≥ 0uniquely. Denote the normalized
4
vectors ψi,t by ψ̄i,t , i.e., ψ̄i,t = ψi,t / j=1 ψi,t, j . (The vectors ψ0,0 and ψ̄0,0 are
identical in this notation.) We also denote by Di,t the pole divisor of ψ̄i,t on .

Remark 6.12 We use the Lax function L̃ i,t and the monodromy operator T̃i,t in the
above definition to allow for both even and odd n. In the case of odd n one can instead
employ the Lax function L i,t and the monodromy operator Ti,t , while all the statements
and proofs below remain valid.

Theorem 6.13 A Floquet-Bloch solution ψ̄i,t is a meromorphic vector function on .

Generically its pole divisor Di,t has degree g + 3.

Proof The proof of the fact that the function ψ̄i,t is meromorphic on the spectral curve
, as well as that its number of poles is deg Di,t = ν/2, is identical to the proof of
Proposition 3.4 in [15]. The number ν of the branch points of  is different: in Theorem
6.7 we found that 2 − 2g = 8 − ν, where g is the genus of the spectral curve. This
implies the required expression: deg Di,t = g + 3. 


123
1034 B. Khesin, F. Soloviev

Definition 6.14 Let J () be the Jacobian of the spectral curve , and [D0,0 ] is the
equivalence class of the divisor D0,0 , the pole divisor of ψ0,0 , under the Abel map.
The pair consisting of the spectral curve  (with marked points Oi and Wi ) and a point
[D0,0 ] ∈ J () is called the spectral data. The spectral map S associates to a given
generic twisted n-gon in CP3 its spectral data (, [D0,0 ]).

The algebraic–geometric integrability is based on the following theorem.

Theorem 6.15 For any n, the spectral map S defines a bijection between a Zariski
open subset of the space Pn = {(xi , yi , z i ), 0 ≤ i ≤ n − 1} and a Zariski open subset
of the spectral data.

Corollary 6.16 For odd n, the spectral map S defines a bijection between a Zariski
open subset of the space Pn  C3n = {(ai , bi , ci ), 0 ≤ i ≤ n − 1} and a Zariski
open subset of the spectral data.

Proof of Corollary 6.16 The statement follows from Theorem 6.15 and Definition 5.5
relating the coordinates (xi , yi , z i ) and (ai , bi , ci ) for odd n. 


The proof of Theorem 6.15 is based on Proposition 6.17 (which completes the con-
struction of the direct spectral map) and Proposition 6.18 (an independent construction
of the inverse spectral map), which are also used below to describe the corresponding
pentagram dynamics. It will be convenient to introduce the following notation for
divisors: O pq := O p + Oq and W pq := W p + Wq (e.g., O12 := O1 + O2 ).

Proposition 6.17 The divisors of the coordinate functions ψi,t,1 , . . . , ψi,t,4 for 0 ≤
i ≤ n − 1 and any integer t satisfy the following inequalities, provided that their
divisors remain non-special up to time t:
For odd n one has
• (ψi,t,1 ) ≥ −D + O2 − i O23 + (i + 1)W12 − t (W12 − O13 );
• (ψi,t,2 ) ≥ −D + (1 − i)O23 + i W12 − t (W12 − O13 );
• (ψi,t,3 ) ≥ −D − i O23 + (i + 1)W12 − t (W12 − O13 );
• (ψi,t,4 ) ≥ −D + O2 + (1 − i)O23 + i W12 − t (W12 − O13 );
For even n one has
• (ψi,t,1 ) ≥ −D + O2 +  i−t+2
2 W12 +  2 W34 −  2 O24 −  2 O34 + t O14 ;
i−t+1 i+1 i

• (ψi,t,2 ) ≥ −D +  2 W12 +  2 W34 −  2 O24 −  2 O34 + t O14 ;

i−t+1 i−t i−1 i

• (ψi,t,3 ) ≥ −D +  i−t+2
2 W12 +  2 W34 −  2 O34 −  2 O24 + t O14 ;
i−t+1 i+1 i

• (ψi,t,4 ) ≥ −D + O2 +  2 W12 +  2 W34 −  2 O34 −  2i O24 + t O14 ;

i−t+1 i−t i−1

where D = D0,0 corresponds to the divisor at t = 0 and is an effective divisor of

degree g + 3, while x is the floor (i.e., the greatest integer) function of x.

Proof The proof is a routine comparison of power expansions in λ at the points O p , Wq

for ki and L i,t and is very similar to the proof of Proposition 3.10 in the 2D case in
[15], although the 3D explicit expressions are more involved. See more details in
Appendix C. 


123
Integrability of higher pentagram maps 1035

Proposition 6.18 For any n, given a generic spectral curve with marked points and
a generic divisor D of degree g + 3 one can recover a sequence of matrices
⎛ ⎞−1
0 0 0 −1
⎜λxi yi 0 0 1⎟
⎜
L̃ i,t (λ) = ⎝ ⎟ ,
0 zi 0 1⎠
0 0 λxi 1

for 0 ≤ i ≤ n − 1 and any t.

We describe the reconstruction procedure and prove this proposition in Appendix C.
Proof of Theorem 6.15 The proof consists of constructions of the spectral map S and
its inverse. The spectral map was described in Definition 6.14 based on Theorem 6.13.
We comment on an independent construction of the inverse spectral map now.
Pick an arbitrary divisor D of degree g + 3 in the equivalence class [D0,0 ] ∈ J ()
and apply Proposition 6.18. “A Zariski open subset of the spectral data” is defined
by spectral functions which may be singular only at the points Oi , Wi and by such
divisors [D] ≡ [D0,0 ] ∈ J () that all divisors in Proposition 6.17 with 0 ≤ i ≤ n − 1
up to time t are non-special. 

The next theorem describes the time evolution of the pentagram map in the Jacobian
of . The difference between even and odd n is very similar to the 2-dimensional case.
Combined with Theorem 6.15, it proves the algebraic-geometric integrability of the
3D pentagram map. (It also implies that it is possible to obtain explicit formulas of the
coordinates of the pentagram map as functions of time using the Riemann θ -functions.)

Theorem 6.19 The equivalence class [Di,t ] ∈ J () of the pole divisor Di,t of ψ̄i,t
has the following time evolution:
• when n is odd,

[Di,t ] = [D0,0 − t O13 + i O23 + (t − i)W12 ],

• when n is even,
 
    
i +1 i i −t +1
[Di,t ] = D0,0 − t O14 + O3 + O2 + i O4 −
2 2 2
  
i −t
W12 − W34 .
2

where deg Di,t = g + 3 and x is the floor function of x, and provided that spectral
data remains generic up to time t.
For an odd n this discrete time evolution in J () takes place along a straight line,
whereas for an even n the evolution goes along a “staircase” (i.e., its square goes
along a straight line).

123
1036 B. Khesin, F. Soloviev

Proof The vector functions ψi,t with i, t = 0 are not normalized. The normalized
4
vectors are equal to ψ̄i,t = ψi,t / f i,t , where f i,t = j=1 ψi,t, j . Proposition 6.17
implies that the divisor of each function f i,t is:
• for odd n,

( f i,t ) = Di,t − D0,0 + t O13 − i O23 + (i − t)W12 ,

• for even n,
     
i +1 i i −t +1
( f i,t ) = Di,t − D0,0 + t O14 − O3 − O2 − i O4 +
2 2 2
 
i −t
W12 + W34 .
2

Since the divisor of any meromorphic function is equivalent to zero, the result of the
theorem follows. The staircase dynamics is related to alternating jumps in the terms
(i − t + 1)/2 and (i − t)/2 as t increases over integers. 

Note that although the pentagram map preserves the spectral curve, it exchanges
the marked points. The “staircase” dynamics on the Jacobian appears after the iden-
tification of curves with different marking. One cannot observe this dynamics in the
space of twisted polygons Pn , before the application of the spectral map.

7 Ramifications: closed polygons and symplectic leaves

7.1 Closed polygons

Closed polygons in CP3 correspond to the monodromies M = ±Id in S L(4, C).

They form a subspace of codimension 15 = dim S L(4, C) in the space of all twisted
polygons Pn . The pentagram map on closed polygons in 3D is defined for n ≥ 7.
Theorem 7.1 Closed polygons in CP3 are singled out by the condition that either
(λ, k) = (1, 1) or (λ, k) = (1, −1) is a quadruple point of . Both conditions are
equivalent to 9 independent linear constraints on I j , J j , G j . Generically, the genus
of  drops to g = 3q − 9 when n is even, and to g = 3q − 6 when n is odd, where
q = n/2. The dimension of the Jacobian J () drops by 6 for closed polygons for
any n. Theorem 6.13 holds with this genus adjustment, and Theorems 6.15 and 6.19
hold verbatim for closed polygons (i.e., on the subspace of closed polygons Cn ⊂ Pn ).
Proof For a twisted n-gon its monodromy matrix at a moment t is equal to T0,t (1)
in the (a, b, c)-coordinates or to T̃0,t (1) in the (x, y, z)-coordinates. An n-gon is
closed if and only if T0,t (1) = Id or T0,t (1) = −Id (respectively, T̃0,t (1) = I0 Id or
T̃0,t (1) = −I0 Id). For our definition of the spectral function, either of these conditions,
T0,t (1) = ±Id or T̃0,t (1) = ±I0 Id, implies that (λ, k) = (1, ±1) is a self-intersection
point for .

123
Integrability of higher pentagram maps 1037

The algebraic conditions implying that (1, ±1) is a quadruple point are:
• R(1, ±1) = 0,
• ∂k R(1, ±1) = ∂λ R(1, ±1) = 0,
• ∂k2 R(1, ±1) = ∂λ2 R(1, ±1) = ∂kλ
2 R(1, ±1) = 0,

• ∂k R(1, ±1) = ∂λ R(1, ±1) = ∂kkλ R(1, ±1) = ∂kλλ

3 3 3 3 R(1, ±1) = 0.

However, the function R(λ, k) is special at the points (1, ±1), because the following
relation holds:
1 1
R(1, ±1) = ±∂k R(1, ±1) − ∂k2 R(1, ±1) ± ∂k3 R(1, ±1).
2 6
Consequently, the above 10 conditions are equivalent to only 9 independent linear
equations on I j , J j , G j , 0 ≤ j ≤ q.
The proofs of Theorems 6.15 and 6.19 apply, mutatis mutandis, to the periodic
case. To define the Zariski open set of spectral data for closed polygons, we confine
to spectral functions that can be singular only at the point (λ, k) = (1, 1) or (1, −1)
in addition to singularities at Oi and Wi and use the same restrictions on divisors D
as in the proof of Theorem 6.15.
In the periodic case we also have to adjust the count of the number ν of branch
points of  and the corresponding calculation for the genus g of , cf. Theorem 6.13.
Namely, as before, the function ∂k R(λ, k) has poles of total order 9n over λ = 0,
and zeroes of total order 6n over λ = ∞. Now since R(λ, k) has a quadruple point
(1, ±1), ∂k R(λ, k) has a triple zero at (1, ±1). But λ = 1 is not a branch point of .
Consequently, ∂k R(λ, k) has triple zeroes on 4 sheets of  over λ = 1. The Riemann-
Hurwitz formula is 2 − 2g = 8 − ν, where the number of branch points for even n is
ν = 9n − 6n − 12 = 3n − 12, while for odd n it is ν = 9n − 6n − 12 + 3 = 3n − 9.
Therefore, we have g = 3q − 9 for even n, and g = 3q − 6 for odd n. 


7.2 Invariant symplectic structure and symplectic leaves

It was proved in [15] that in the 2D case an invariant symplectic structure on the space of
twisted polygons Pn provided by Krichever–Phong’s universal formula [5,6] coincides
with the inverse of the invariant Poisson structure found in [11] when restricted to the
symplectic leaves. We show that in 3D the same formula also provides an invariant
symplectic structure defined on leaves described below. While we do not compute
the symplectic structure explicitly in the coordinates (ai , bi , ci ) or (xi , yi , z i ) due to
complexity of the formulas, the proofs are universal and applicable in the higher-
dimensional case of CPd as well. Finding an explicit expression of the symplectic
structure or of the corresponding Poisson structure is still an open problem.
Definition 7.2 [5,6] Krichever–Phong’s universal formula defines a pre-symplectic
form on the space of Lax operators, i.e., on the space Pn . It is given by the expression:

1    dλ
ω=− res Tr 0−1 T̃0−1 δ T̃0 ∧ δ0 .
2 λ
λ=0,∞

123
1038 B. Khesin, F. Soloviev

The matrix 0 := 0,t (λ) is composed of the eigenvectors ψ0,t on different sheets of
 over the λ-plane, and it diagonalizes the monodromy matrix T̃0 := T̃0,t (λ). (In this
definition we drop the index t, because all variables correspond to the same moment
of time.)
The leaves of the 2-form ω are defined as submanifolds of Pn , where the expression
δ ln k (dλ/λ) is holomorphic. The latter expression is considered as a 1-form on the
spectral curve .
Proposition 7.3 For even n the leaves are singled out by 6 conditions:

δ I0 = δ Iq = δG 0 = δG q = δ J0 = δ Jq = 0;

For odd n the leaves are singled out by 3 conditions:

δG 0 = δ I0 = δ Jq = 0.

Proof These conditions follow immediately from the definition of the leaves and
Lemma 6.8. For example, at the point O1 we have
 
dλ 1 δ I0
δ ln k1 = + O(1) dλ.
λ λ I0

This 1-form is holomorphic in λ if and only if δ I0 = 0. Similarly, we obtain

δ(I0 /G 0 ) = 0 at the point O2 for odd n. (One has to keep in mind that the local
parameter around this point is λ1/2 .) 

Remark 7.4 The definition of a presymplectic structure ω on Pn uses 0 and T̃0 and
hence relies on the normalization of 0 . When restricted to the leaves from Proposi-
tion 7.3, the 2-form ω becomes independent of the normalization of the Floquet–Bloch
solutions. Additionally, the form ω becomes non-degenerate, i.e., symplectic, when
restricted to these leaves, as we prove below. The symplectic form is invariant with
respect to the evolution given by the Lax equation. The proof is very similar to that of
Corollary 4.2 in [8] (cf. [5,6] for other proofs).
Theorem 7.5 The rank of the invariant 2-form ω restricted to the leaves of Proposi-
tion 7.3 is equal to 2g.
Proof Since the 1-form δ ln k dλ/λ is holomorphic on , it can be represented as a
sum of the basis holomorphic differentials:

dλ 
g
δ ln k = δUi dωi , (10)
λ
i=1

where g is the genus of . The coefficients Ui can be found by integrating the last
expression over the basis cycles ai of H1 ():

dλ
Ui = ln k .
ai λ

123
Integrability of higher pentagram maps 1039

According to formula (5.7) in [7], we have:


g+3
ω= δ ln k( pi ) ∧ δ ln λ( pi ),
i=1

where the points pi ∈ , 1 ≤ i ≤ g + 3, constitute the pole divisor D0,0 of the

normalized Floquet-Bloch solution ψ0,0 .
After rearranging the terms, we obtain:
⎛ ⎞ ⎛ ⎞
g+3 
 ps  ps 
g
dλ ⎠
ω = δ⎝ δ ln k =δ ⎝ δUi dωi ⎠ = δUi ∧ δϕi ,
λ
s=1 s,i i=1

where
g+3 
 ps
ϕi = dωi
s=1

are coordinates on the Jacobian J (). The variables Ui and ϕi are natural Darboux
coordinates for ω, which also turn out to be action-angle coordinates for the pentagram
map. (The latter follows from the general properties of the Krichever–Phong universal
form for a given Lax representation, cf. [5,6].)
Let us show that the functions Ui are independent. Assume the contrary, then there
exists a vector v on the space Pn , such that δUi (v) = 0 for all i. Then it follows
from (10) that ∂v k ≡ 0. After applying the operator ∂v to R(λ, k), we conclude that k
satisfies an algebraic equation of degree 3, which is impossible, since  is a fourfold
cover of the λ-plane. 

Remark 7.6 In more details, there are the following two cases:
• even n = 2q. The dimension of the space Pn is 6q. The codimension of the leaves
is 6. Therefore, the dimension of the leaves matches the doubled dimension of the
tori: 2g = 6q − 6.
• odd n = 2q + 1. The dimension of the space Pn is 6q + 3. The codimension of
the leaves is 3. Again, the dimension of the leaves matches the doubled dimension
of the tori: 2g = 6q.
The algebraic-geometric integrability in the complex case implies Arnold-Liouville
integrability in the real one. Indeed, the pre-symplectic form depends on entries of
the monodromy matrix in a rational way, since it is independent of the permutation of
sheets of the spectral curve . Therefore, its restriction to the space of the real n-gons
provides a real pre-symplectic structure. One obtains invariant Poisson brackets on the
space of polygons Pn by inverting the real symplectic structure on the leaves, while
employing invariants of Proposition 7.3 as the corresponding Casimirs.
Problem 7.7 Find an explicit formula for an invariant Poisson structure with the
above symplectic leaves.

123
1040 B. Khesin, F. Soloviev

8 A Lax representation in higher dimensions

The origin of the integrability of the pentagram map is the presence of its scaling
invariance. Assume that gcd(n, d + 1) = 1. The difference equation (1)

V j+d+1 = a j,d V j+d + a j,d−1 V j+d−1 + · · · + a j,1 V j+1 + (−1)d V j

allows one to introduce coordinates a j,1 , a j,2 , . . . a j,d , 0 ≤ j ≤ n − 1, on the space

of twisted n-gons in any dimension d.
Proposition–conjecture 8.1 (The scaling invariance) The pentagram map on twisted
n-gons in CPd is invariant with respect to the following scaling transformations:
• for odd d = 2 + 1 the transformations are

a j,1 → sa j,1 , a j,3 → sa j,3 , a j,5 → sa j,5 , , . . . , , a j,d → sa j,d ,

while other coefficients a j,2l with l = 1, . . . , do not change;

• for even d = 2 the transformations are

a j,1 → s − a j,1 , a j,3 → s 1− a j,3 , . . . , a j,d−1 → s −1 a j,d−1 ,

a j,2 → sa j,2 , a j,4 → s 2 a j,4 , . . . , a j,d → s a j,d

for all s ∈ C∗4 .

Proof In any dimension d the pentagram map is a composition of involutions α and
β, see Sect. 5.1. (More precisely, α is not an involution for even d, but its square α 2 is
a shift in the vertex index, see [11] for the 2D case.) One can prove that the involution
α : Vi → Wi in any dimension has the form

W j+d+1 = (−1)d+1 (a,1 W j+d + a,2 W j+d−1 + · · · + a,d W j+1 − W j ),

where  stands for the first index, which is irrelevant for the scaling (Lemma 5.3 proves
the case d = 3).
We call this Proposition–conjecture because the proof of an analog of Lemma 5.4
(for the map β) in higher dimensions is computer assisted. One verifies that for a
given dimension d the coefficients consist of the terms that are consistent with the
scaling. 

We obtained explicit formulas, and hence a direct (theoretical) proof of the scaling
invariance for the pentagram maps up to dimension d ≤ 6. This bound is related to
computing powers to produce explicit formulas and might be extended. However, we
have no general purely theoretical proof valid for all d and it would be very interesting
to find it.

4 We thank G. Mari-Beffa for correcting an error in the scaling for even d in the first version of this
manuscript, as well as in the short version [16]. This error related to numerics with a different choice of
vertices for the diagonal planes leads to another system, different from T p,r , which also turns out to be
integrable and will be discussed elsewhere.

123
Integrability of higher pentagram maps 1041

Problem 8.2 Find a general proof of the scaling invariance of the pentagram map in
any dimension d.
Theorem 8.3 The scale-invariant pentagram map on twisted n-gons in any dimension
d is a completely integrable system. It is described by the Lax matrix
⎛ ⎞
0 0 · · · 0 (−1)d
⎜ a j,1 ⎟
⎜ ⎟
L −1 (λ) = ⎜ a ⎟
j,2 ⎟ ,
j ⎜ D(λ)
⎝ ··· ⎠
a j,d

where D(λ) is the following diagonal matrix of size d × d:

• for odd d = 2 + 1, one has D(λ) = diag(λ, 1, λ, 1, . . . , λ);
• for even d = 2 , one has D(λ) = diag(1, λ, 1, λ, . . . , 1, λ).
Proof sketch By using the scaling invariance of the pentagram map, one derives the
Lax matrix exactly in the same way as in 3D, see Sect 6.1. Namely, first construct
the (d + 1) × (d + 1)-matrix N j (s) depending on our scaling parameter s, and then
use the formula L −1 −1
j (λ) = g N j (s)g /s with a suitable choice of the diagonal
m

(d + 1) × (d + 1)-matrix g and an appropriate function of the parameter s.

For odd d = 2 + 1, we have g = diag(1, s, 1, s, . . . , 1, s), m = 1, and λ ≡ s −2 ,
whereas for even d = 2 , we have g = diag(1, s − , s, s 1− , s 2 , . . . , s −1 , s −1 , s ),
m = , and λ ≡ s −d−1 . The Lax representation with a spectral parameter is con-
structed as we described above.
Using the genericity assumptions similar to those used in the 2D and 3D cases,
one constructs the spectral map and its inverse, which is equivalent to algebraic-
geometric integrability of the pentagram map. Coefficients of the spectral curve form
a maximal family of first integrals. Along with a (pre)symplectic structure defined by
the Krichever–Phong formula, this provides the Arnold–Liouville integrability of the
system on the corresponding symplectic leaves in the real case. 

The scaling parameter has a clear meaning in the continuous limit:
Proposition 8.4 For any dimension d the continuous limit of the scaling transforma-
tions corresponds to the spectral shift L → L + λ of the differential operator L.
Proof In 2D this was proved in [11]. A continuous analog of the difference equation
(1) is

G(x + (d + 1) ) = ad (x, )G(x + d ) + · · · + a1 (x, )G(x + ) + (−1)d G(x),

where G(x) satisfies the differential equation (2) with a differential operator L of the
form (5). Using the Taylor expansion for G(x + j ) and the expansion a j (x, ) =
a 0j (x) + a 1j (x) + · · · , we obtain expressions of aki in terms of the coefficients of L,
i.e., in terms of functions u j (x) and their derivatives. We find that the terms ak0 are
constant, ak1 = 0 for all k, while aki for i ≥ 2 are linear in u d−i+1 and differential
polynomials in the preceding coefficients u d−1 , . . . , u d−i+2 .

123
1042 B. Khesin, F. Soloviev

The scaling parameter also has an expansion in : s = τ0 + τ1 + 2 τ2 +· · · . We apply

it to the coefficients ai (x, ) and impose the condition that ak0 and ak1 = 0 are fixed,
similarly to [11]. By term-wise calculations (different in the cases of even and odd d
and using the “triangular form” of the expressions for aki ), one successively obtains that
τ0 = 1, τ1 = · · · = τd = 0, i.e., s must have the form s = 1 + τd+1 d+1 + O( d+2 ).
Its action shifts only the last term of L: u 0 → u 0 +const·τd+1 d+1 , i.e., it is equivalent
to the spectral shift L → L + λ. 


Note that the spectral shift commutes with the KdV flows. Indeed, d/dt (L + λ) =
d/dt L = [Q 2 , L] = [Q 2 , L + λ], since Q 2 (L) := ∂ 2 + d+1
2
u d−1 = Q 2 (L + λ) for
operators L of degree d + 1 ≥ 3. Equivalently, the pentagram map commutes with
the scaling transformations in the continuous limit.

Acknowledgments We are grateful to M. Gekhtman and S. Tabachnikov for useful discussions. B.K. was
partially supported by the Simonyi Fund and an NSERC research grant.

Appendix A: Continuous limit in the 3D case

In this section we present explicit formulas manifesting Theorem A on the continuous

limit of the 3D pentagram map. Consider a curve G(x) in R4 given by the differential
equation

G + u(x)G + v(x)G + w(x)G = 0

with periodic coefficients u(x), v(x), w(x). To find the continuous limit, we fix and
consider a plane P (x) passing through the three points G(x − ), G(x), G(x + )
on this curve. We are looking for an equation of the envelope curve L (x) for these
planes.
This envelope curve L (x) satisfies the following system of equations:

det |G(x), G(x + ), G(x − ), L (x)| = 0

det |G(x), G(x + ), G(x − ), L (x)| = 0
det |G(x), G(x + ), G(x − ), L (x)| = 0.

By considering the Taylor expansion and using the normalizations det |L , L , L , L |

= 1 and det |G, G , G , G | = 1 we find that

2  u 
L (x) = G(x) + G (x) + G(x) + O( 4 ) (11)
6 2

123
Integrability of higher pentagram maps 1043

as → 0. Now, the equation L + u L + v L + w L = 0 implies that:

2
u =u+ (v − u ) + O( 4 ),
3
2
v =v+ (2w + v − uu − 2u ) + O( 4 ),
6
2
w =w+ (2w − uu − vu − u ) + O( 4 ).
12

These equations describe the (2, 4)-equation in the Gelfand-Dickey hierarchy:

⎧
⎨ u̇ = 2v − 2u ,
L̇ = [Q 2 , L] ⇔ v̇ = v + 2w − uu − 2u ,
⎩
ẇ = w − 21 vu − 21 uu − 21 u ,

where L = ∂ 4 + u∂ 2 + v∂ + w and Q 2 = (L 2/4 )+ = ∂ 2 + u2 .

Remark 9.1 A different choice of the points defining the plane P (x) on the original
curve leads to the same continuous limit. For instance, the choice of G(x −3 ), G(x +
), G(x + 2 ) results in the same expression for L (x), where in (11) instead of the
coefficient 2 /6 one has 7 2 /6. This leads to the same evolution of the curve G with
a different time parameterization, cf. Remark 4.4.

Appendix B: Higher terms of the continuous limit

Recall that in the continuous limit for the pentagram map in RPd the envelope for
osculating planes moves according to the (2, d + 1)-KdV equation (Theorem 4.5).
This evolution is defined by the 2 -term of the expansion of the function L (x).
The same proof works in the following more general setting. Let L be a differential
operator (5) of order d + 1 and G a non-degenerate curve defined by its solutions:
LG = 0.

Proposition 10.1 Assume that the curve G evolves according to the law Ġ = Q m G,
where Q m := (L m/(d+1) )+ is the differential part of the mth power of the operator
Q = L 1/(d+1) . Then this evolution defines the equation L̇ = [Q m , L], which is the
(m, d + 1)-equation in the corresponding KdV hierarchy of L.

Furthermore, one can define the simultaneous evolution of all terms in the
-expansion of L (x) using the following construction. For the pseudodifferential
2
operator Q := L 1/(d+1) consider the formal series exp( Q) := 1 + Q + 2 Q 2 + · · ·
and take its differential part:

123
1044 B. Khesin, F. Soloviev

 2  2 ∞
 m
(exp( Q))+ = 1 + Q + Q + ···
2
= 1 + Q1 + Q2 + · · · = Qm .
2 + 2 m!
0

For each power of this is a multiple of the differential operator Q m , which is the
differential part of the mth power Q m of the operator Q = L 1/(d+1) .
Corollary 10.2 The formal evolution equation Ġ = (exp( Q))+ G corresponds to
the full KdV hierarchy L̇ = [(exp( Q))+ , L], where the operator L is of order d + 1
and the (m, d + 1)-equation corresponds to the power m . 

A natural question is which equations of this hierarchy actually appear as the evo-
lution of the envelope L (x). Recall that only even powers of arise in the expansion
of the function L (x) for the continuous limit of the pentagram map. The 2 -term
gives the (2, d + 1)-KdV equation. It turns out that the 4 -term in the continuous limit
of the 2D pentagram map results in the equation very similar to the (4, 3)-equation
in the KdV hierarchy (which is a higher-order Boussinesq equation). Although the
numerical coefficients in these differential equations are different, one may hope to
obtain the exact equations of the KdV hierarchy for different m by using an appropriate
rescaling. This allows one to formulate
Problem 10.3 Do higher (m, d + 1)-KdV flows appear as the m -terms in the expan-
sion of the envelope L (x) for the continuous limit of the pentagram map for any even
m > 2?

Appendix C: Bijection of the spectral map

In this appendix we sketch the proof of Proposition 6.17 and prove Proposition 6.18,
which allows one to reconstruct the L-matrix from spectral data, and hence complete
the proof of Theorem 6.15 on the spectral map.
Proposition 11.1 (= Proposition 6.18) For any n, given a generic spectral curve with
marked points and a generic divisor D of degree g + 3 one can recover a sequence of
matrices
⎛ ⎞−1
0 0 0 −1
⎜λxi yi 0 0 1⎟
L̃ i,t (λ) = ⎜
⎝ 0
⎟ ,
zi 0 1⎠
0 0 λxi 1

for 0 ≤ i ≤ n − 1 and any t.

Proof Without loss of generality we describe the procedure to reconstruct the matrices
L i (λ) := L̃ i,0 (λ), for 0 ≤ i ≤ n − 1 and t = 0.
1. First, we pick functions ψi, j := ψi,0, j for 0 ≤ i ≤ n − 1, 1 ≤ j ≤ 4, and t = 0,
satisfying Proposition 6.17. Note that according to the Riemann-Roch theorem,
the functions ψi,1 and ψi,4 are defined up to a multiplication by constants, whereas
the functions ψi,2 and ψi,3 belong to 2-dimensional subspaces. The functions ψi,1

123
Integrability of higher pentagram maps 1045

and ψi,4 belong to the same subspaces. We pick the pairs of functions ψi,1 , ψi,3
and ψi,2 , ψi,4 to be linearly independent. Observe that any sets of functions
ψi,1 , . . . , ψi,4 satisfying Proposition 6.17 are related by gauge transformations
ψi → gi−1 ψi , where
⎛ ⎞
Ai 0 0 0
⎜0 Bi 0 Ei ⎟
gi = ⎜
⎝ Fi
⎟ , gi = gi+n ,
0 Ci 0⎠
0 0 0 Di

and ψi stands for ψi = (ψi,1 , . . . , ψi,4 )T . We also define ψn to be ψn = I0 kψ0

for any n in (x, y, z)-variables.
2. We find the unique matrix L i satisfying the equation ψi = (L i )−1 ψi+1 :

⎛ ⎞−1
0 0 0 ti,1
⎜λti,5 0 λti,6 ti,2 ⎟
L i (λ) = ⎜
⎝ 0
⎟ .
ti,7 0 ti,3 ⎠
λti,8 0 λti,9 ti,4

3. One can check that there exists the unique choice of the matrices gi , 0 ≤ i ≤ n−1,
such that the equality L i (λ) = gi+1 L i (λ)gi−1 is possible. The latter is equivalent
to the following system of equations (0 ≤ i ≤ n − 1):

Ai ti,1 Bi ti,2 + E i ti,4 Di ti,4

= −1; = 1; = 1;
Di+1 Di+1 Di+1
Bi+1 Fi ti,1 + Bi+1 Ci ti,3 − Ci E i+1 ti,7
= 1; Bi ti,6 + E i ti,9 = 0;
Bi+1 Di+1
Ci+1 ti,8 − Fi+1 ti,9 = 0.

These equations decouple and may be solved explicitly. One only needs to check
D ti,4
the solvability of n equations Dii+1 = 1, 0 ≤ i ≤ n − 1, for n variables Di , 0 ≤
"n−1
i ≤ n − 1. A non-trivial solution exists provided that i=0 ti,4 = 1. It depends
on an arbitrary constant, which corresponds to multiplication of all matrices gi by
the same number and does not affect the Lax matrices. One can check that

ψi+1,4 (O1 ) 
n−1
ψn,4 (O1 )
ti,4 = and ti,4 = = I0 k(O1 ).
ψi,4 (O1 ) ψ0,4 (O1 )
i=0

By using Lemma 6.8 we find the value k(O1 ) = 1/I0 as required. Now the
remaining variables Ai , Bi , Ci , E i , Fi , 0 ≤ i ≤ n − 1, are uniquely determined.




123
1046 B. Khesin, F. Soloviev

Corollary 11.2 For odd n, given a generic spectral curve with marked points and a
generic divisor D one can recover a sequence of matrices
⎛ ⎞−1
0 0 0 −1
⎜λ 0 0 cj ⎟
L i,t (λ) = ⎜
⎝0
⎟
1 0 bj ⎠
0 0 λ aj

with 0 ≤ i ≤ n − 1 and any t.

Proof The statement follows from Proposition 6.18 and the fact that (a j , b j , c j ), 0 ≤
i ≤ n − 1, are coordinates on the space Pn for odd n. 

We complete the exposition with a sketch of the proof for Proposition 6.17 for even
n (the case of odd n is similar).
Proposition 11.3 (= Proposition 6.17 ) For even n, the divisors of the coordinate
functions ψi,t,1 , . . . , ψi,t,4 for 0 ≤ i ≤ n − 1 and any integer t satisfy the following
inequalities, provided that their divisors remain non-special up to time t:
• (ψi,t,1 ) ≥ −D + O2 +  i−t+2
2 W12 +  2 W34 −  2 O24 −  2 O34 + t O14 ;
i−t+1 i+1 i

• (ψi,t,2 ) ≥ −D +  2 W12 +  2 W34 −  2 O24 −  2 O34 + t O14 ;

i−t+1 i−t i−1 i

• (ψi,t,3 ) ≥ −D +  i−t+2
2 W12 +  2 W34 −  2 O34 −  2 O24 + t O14 ;
i−t+1 i+1 i

• (ψi,t,4 ) ≥ −D + O2 +  2 W12 +  2 W34 −  2 O34 −  2i O24 + t O14 ;

i−t+1 i−t i−1

where D is an effective divisor of degree g + 3, and x is the floor function of x.

Proof First, we prove these inequalities for t = 0 and 0 ≤ i ≤ n − 1. For illustration
we find the multiplicities of the components of the vector ψi,0 at the point O2 , while
other points can be treated in a similar fashion. We employ the matrices L̃ i,t in the
coordinates xi , yi , z i .
Notice that a cyclic permutation of indices (n − 1, n − 2, . . . , 1, 0) changes the
monodromies Ti → Ti+1 and the Floquet-Bloch solutions ψ̄i → ψ̄i+1 . For even n,
it also permutes ψ̄i (O2 ) ↔ ψ̄i (O3 ) and W12 ↔ W34 , i.e., the corresponding pairs of
the vectors ψ̄i at the points (W1 , W2 ) and (W3 , W4 ) are swapped.
Using the asymptotic expansion of T̃0,t (λ) at λ = 0, the definition of the
Floquet-Bloch solution, and the normalization condition, one can show that ψ0,0 =
(O(λ), O(λ), 1 + O(λ), O(λ))T as λ → 0 at the point O2 . Since
⎛ ⎞
1 1 0 0  
⎜0 0 0 0⎟ 1 1
L 1,0 (λ)L 0,0 (λ) = ⎜
⎝ y1
⎟ +O as λ → 0,
y1 0 0⎠ x0 x1 y0 y1 λ2 λ
0 0 0 0

and ψ2,0 = L 1,0 L 0,0 ψ0,0 , generically one has ψ2,0 = (O(1), O(1), O(1/λ), O(1))T
at O2 .
By definition, the normalized vectors are ψ̄i,t = f i,t ψi,t . Using a cyclic permuta-
tion, we find that ψ̄2k,0 = (O(λ), O(λ), 1 + O(λ), O(λ))T and that f 2,0 (λ) = O(λ)

123
Integrability of higher pentagram maps 1047

at O2 . Using the permutation argument again, we derive that fi+2,0 (λ)/ f i,0 (λ) = O(λ)
at O2 for even i. Therefore, one has f 2k,0 (λ) = O(λk ) at O2 . Now the required
multiplicities for the vector ψ2k,0 at O2 follow. Furthermore, since ψ2k+1,0 =
L 2k ψ2k,0 , one can check that generically f 2k+1,0 (λ)/ f 2k,0 (λ) = O(1) and ψ̄2k+1,0 =
(O(1), O(λ), O(1), O(1))T at the point O2 . This establishes also the multiplicities
for the vector ψ2k+1,0 at O2 .
Having proved the proposition for t = 0, one can prove it for t > 0 by using the
formula ψi,t+1 = P̃i,t ψi,t . Note that it suffices to study the cases t = 0 and t = 1 only.
Consider, for example, the multiplicity of the function ψi,1,1 at the point O2 . Since
ψi,1,1 = (ψi,0,2 − ψi,0,4 )/(xi (1 + yi + z i+1 )), one can check that the multiplicity of
the right-hand side at O2 is 1 − k for i = 2k and it is equal to −k for i = 2k + 1,
i.e., ψi,1,1 and ψi,0,1 have the same multiplicities at O2 . Other cases are treated in a
similar way. 


References

1. Adler, M.: On a trace functional for formal pseudo differential operators and the symplectic structure
of the Korteweg-de Vries type equations. Invent. Math. 50(3), 219–248 (1978/1979)
2. Fomin, S., Zelevinsky, A.: Cluster algebras. IV. Coefficients. Compos. Math. 143, 112–164 (2007)
3. Gekhtman, M., Shapiro, M., Tabachnikov, S., Vainshtein, A.: Higher pentagram maps, weighted
directed networks, and cluster dynamics. Electron. Res. Announc. Math. Sci. 19, 1–17 (2012).
arXiv:1110.0472
4. Glick, M.: The pentagram map and Y-patterns. Adv. Math. 227, 1019–1045 (2011)
5. Krichever, I.M., Phong, D.H.: On the integrable geometry of soliton equations and N=2 supersymmetric
gauge theories. J. Differ. Geom. 45, 349–389 (1997)
6. Krichever, I.M., Phong, D.H.: Symplectic forms in the theory of solitons. Surv. Differ. Geom. IV,
239–313 (1998)
7. Krichever, I.M., Phong, D.H.: Spin chain models with spectral curves from M theory. Comm. Math.
Phys. 213, 539–574 (2000)
8. Krichever, I.M.: Vector bundles and Lax equations on algebraic curves. Comm. Math. Phys. 229(2),
229–269 (2002)
9. Marí-Beffa, G.: On generalizations of the pentagram map: discretizations of AGD flows. arXiv:1103.
5047
10. Ovsienko, V.Yu., Khesin, B.A.: Symplectic leaves of the Gelfand-Dikii brackets and homotopy classes
of nondegenerate curves. Funct. Anal. Appl. 24(1), 33–40 (1990)
11. Ovsienko, V., Schwartz, R., Tabachnikov, S.: The pentagram map: a discrete integrable system. Comm.
Math. Phys. 299, 409–446 (2010). arXiv:0810.5605
12. Ovsienko, V., Schwartz, R., Tabachnikov, S.: Liouville-Arnold integrability of the pentagram map on
closed polygons. Duke Math. J. arXiv:1107.3633 (to appear)
13. Schwartz, R.: The pentagram map. Exp. Math. 1, 71–81 (1992)
14. Schwartz, R.: Discrete monodromy, pentagrams, and the method of condensation. J. Fixed Point Theory
Appl. 3(2), 379–409 (2008)
15. Soloviev, F.: Integrability of the pentagram map. Duke Math. J. arXiv:1106.3950 (to appear)
16. Khesin, B., Soloviev, F.: The pentagram map in higher dimensions and KdV flows. Electron. Res.
Announc. Math. Sci. 19, 86–96 (2012)

123