A QUICK ESTIMATE FOR THE VOLUME OF A POLYHEDRON

Alexander Barvinok and Mark Rudelson

December 17, 2021

Abstract. Let P be a bounded polyhedron defined as the intersection of the non-

negative orthant Rn n
+ and an affine subspace of codimension m in R . We show that
a simple and computationally efficient formula approximates the volume of P within
a factor of γ m , where γ > 0 is an absolute constant. The formula provides the best
known estimate for the volume of transportation polytopes from a wide family.

1. Introduction
The problem of efficient computation (approximation) of the volume of a poly-
tope, and, more generally, of a given convex body has attracted a lot of attention,
see, for example, [GK18] for a survey. The most successful approach is via Markov
Chain Monte Carlo randomized algorithms, see [Ve05] for a survey. In particular,
randomized algorithms allow one to approximate the volume of a polytope in Rn
within relative error  > 0 in time polynomial in n and −1 . The polytope can be
defined as the convex hull of a finite set of points or as the intersection of halfspaces,
or by a membership oracle, in which case the algorithms extend to the class of all
“well-conditioned” convex bodies.
Deterministic algorithms enjoyed less success. For a general convex body B ⊂
n
R , the only available polynomial time algorithm approximates volume within a
factor of nO(n) , using an approximation of B by an ellipsoid, see [G+88]. For B
defined by a membership oracle, this approximation factor is basically the best
possible (up to some logarithmic terms) that can be achieved in deterministic poly-
nomial time [BF87]. If P is a polytope defined as the convex hull of a set of points or
as the intersections of halfspaces, deterministic algorithms in principle may turn out
to be as powerful as randomized ones, but so far the approximation ratio achieved

1991 Mathematics Subject Classification. 52B55, 52A38, 52A40, 52A41, 52B11, 68Q25,
68W25.
Key words and phrases. polytope, polyhedron, volume, deterministic algorithm, formula, log-
concave density.
Research of AB and MR is partially supported by NSF Grants DMS 1855428 and DMS 2054408
respectively.

Typeset by AMS-TEX
1
in deterministic polynomial time is the same as for general convex bodies. We re-
mark that if P ⊂ Rn is a polytope defined as the convex hull of n + O(1) points or
as the intersection of n + O(1) halfspaces, then vol P can be computed exactly in
polynomial time, in the former case by a triangulation into nO(1) simplices and in
the latter case by a dual procedure of expressing P as a signed linear combination
of nO(1) simplices, cf. [GK18]. Of course, for special classes of polytopes, such as
parallelepipeds, there can be computationally efficient explicit formulas.
In this paper, we consider the class of polyhedra P defined as the intersection
of the non-negative orthant Rn+ and an affine subspace in Rn . In coordinates, P is
defined by a system of linear equations Ax = b, where A is an m × n matrix, x is
an n-vector of variables and b is an m-vector, and inequalities x ≥ 0, meaning that
the coordinates of x are non-negative. We assume that m < n, that rank A = m
and that P has a non-empty relative interior, that is, contains a point x with all
coordinates positive. Hence dim P = n−m and we measure the (n−m)-dimensional
volume of P in its affine span with respect to the Euclidean structure inherited from
Rn . We also assume that P is bounded, that is, a polytope. Generally, any (n−m)-
dimensional polyhedron with n facets can be represented as the intersection of Rn+
and an affine subspace of codimension m. Furthermore, many interesting polyhedra,
such as transportation polytopes, see, for example [DK14], are naturally defined in
this way.
We present a deterministic polynomial time algorithm which approximates the
volume of such a polytope P within a factor of γ m , where γ > 0 is an absolute
constant (for m large enough, one can choose γ = 4.89). In fact, our algorithm
is basically a formula. The only “non-formulaic” part of our algorithm consists
of solving some standard convex optimization problem on P , namely finding its
“analytic center”, see [Re88]. After that, we only need to compute two m × m
determinants, which, as is well-known, can be accomplished in O(m3 ) time. While
the approximation factor γ m looks big compared to 1 +  achieved by randomized
algorithms, it appears to be the best achieved to date by a deterministic polynomial
time algorithm for many interesting classes of polytopes, such as transportation
polytopes. Since the algorithm is basically a formula, it allows one to analyze how
the volume changes as P evolves inside its class, which turns out to be important
for studying some statistical phenomena related to contingency tables, cf. [D+20].
The approximation factor looks more impressive when n  m, which is indeed
the case for many interesting classes of polytopes. Note that if we dilate a d-
dimensional polytope by a factor of (1 + ), its volume gets multiplied by (1 + )d .
Hence sometimes one considers the “volume ratio” (vol P )1/d , which scales more
manageably when the polytope P is scaled, cf. [Pi89]. Thus if n  m, our algorithm
(formula) approximates the volume ratio within a factor of 1 + o(1).

2. The main result and some applications

(2.1) The setup. Let A = (αij ) be an m × n matrix, let b = (β1 , . . . , βm ) be an
m-vector, and suppose that the polyhedron P ⊂ Rn is defined by the system of
2
equations
n
X
(2.1.1) αij ξj = βi for i = 1, . . . , m
j=1

and inequalities

(2.1.2) ξj ≥ 0 for j = 1, . . . , n.

We assume that m < n and that rank A = m, in which case the system (2.1.1)
defines an (n − m)-dimensional affine subspace.
Suppose that P has a non-empty relative interior, that is, contains a point x =
(ξ1 , . . . , ξn ) where ξj > 0 for j = 1, . . . , n, and is also bounded, that is, a polytope.
Thus P is an (n − m)-dimensional bounded polyhedron and our goal is to estimate
its volume vol P relative to the Euclidean structure inherited from Rn by the affine
subspace defined by (2.1.1).
We define a function f : Rn+ −→ R by
n
X
f (x) = n + ln ξj where
(2.1.3) j=1
x = (ξ1 , . . . , ξn ) and ξj > 0 for j = 1, . . . , n

and consider the following optimization problem

(2.1.4) Find z∈P such that f (z) = max f (x).

x∈P

The function f is strictly concave and hence the maximum point z can be found
efficiently (in polynomial time), see [NN94]. Also, the point z = (ζ1 , . . . , ζn ) is
unique and satisfies ζj > 0 for j = 1, . . . , n, see [BH10]. In fact, the point z was at
the very source of interior-point methods in optimization [Re88] under the name of
the analytic center of P . We note that if the symmetry group of P is sufficiently
rich, we can determine z without solving the optimization problem (2.1.3)–(2.1.4),
as z must be invariant under all permutations of the coordinates ξ1 , . . . , ξn that
map P onto itself.
Now we can state our result.
(2.2) Theorem. Let A be m × n matrix of rank A = m < n, let b be an m-vector
and suppose that the polyhedron P defined by the system (2.1.1)–(2.1.2) is bounded
and has a non-empty relative interior. Let z = (ζ1 , . . . , ζn ) be the analytic center of
P defined as the solution to the optimization problem (2.1.3)–(2.1.4). Let B be the
m × n matrix obtained by multiplying the j-th column of A by ζj for j = 1, . . . , n
and let √
f (z) det AAT
E(A, b) = e √ .
det BB T
3
 (1) Let α0 be the necessarily unique number in the interval (0, 1) satisfying
Z +∞
1
− 2α1
1 + s2 0 ds = 1, α0 ≈ 0.7148659168.
2π −∞

Then  m
1
vol P ≤ √ E(A, b) ≤ (1.19)m E(A, b);
α0
(2) We have
m+2


22Γ
vol P ≥ E(A, b);
π m/2 e(m+2)/2 (m + 2)m/2

(3) For any 0 <  < 1/2, there is γ() > 0 such that
n 1
o  1 m
+
vol P ≥ exp −γ()m 2 √ E(A, b)
2πe
n 1
o
≥ exp −γ()m 2 + (0.24)m E(A, b).

Some remarks are in order. Using the standard bound

√ 1
(2.2.1) Γ(t) ≥ 2πtt− 2 e−t for t ≥ 1,

we conclude that the right hand side of the formula in Part (2) decreases with m
roughly as  m
1
√ E(A, b) ≈ (0.14)m E(A, b).
e 2π
The lower bound in Part (3) is asymptotically stronger, although it contains a
constant which may render it weaker than the bound of Part (2) for smaller m. In
addition, the proof of Part (2) is rather elementary, whereas the proof of Part (3)
relies on the recent breakthrough by Chen
 in the KLS conjecture [Ch21], although
m
the same asymptotic behavior in the √1 term could be achieved by using
2πe
earlier and weaker thin shell estimates, see [LV17] for a survey.
(2.3) Gaussian approximation. It was proved in [BH10] that if some analytic
conditions on A and z are satisfied, we have asymptotically
√
ef (z) det AAT
(2.3.1) vol P ≈ √
(2π)m/2 det BB T

as m and n grow. The right hand side of (2.3.1) is called in [BH10] the maximum
entropy Gaussian approximation. Under typical
√ circumstances those analytic con-
ditions require, in particular, that m = O( n) and that the coordinates ζ1 , . . . , ζn
4
of the analytic center z of P are roughly of the same order. We explain the name
and the intuition behind this formula in Section 3.1. The estimate of Theorem 2.2
is much cruder, but its validity doesn’t depend on the particulars of A and b or the
relations between m and n. We note that to understand some statistical phenomena
related to contingency tables [D+20], it is important to understand the behavior of
the volume of P when the coordinates ζ1 , . . . , ζn have decidedly different orders of
magnitude.
(2.4) Example: 2-way transportation polytopes. Let us fix positive integers
k and l, a k-vector of positive real numbers r = (ρ1 , . . . , ρk ) and an l-vector of
positive real numbers c = (γ1 , . . . , γl ) such that
k
X l
X
(2.4.1) ρi = γj .
i=1 j=1

We consider the set T (r, c) of k × l non-negative real matrices with row sums r and
column sums c. In other words, T (r, c) is defined in the space Rk×l ∼ = Rkl of k × l
matrices x = (ξij ) by the equations
l
X k
X
(2.4.2) ξij = ρi for i = 1, . . . , k and ξij = γj for j = 1, . . . , l
j=1 i=1

and inequalities
ξij ≥ 0 for all i, j.
It is not hard to see that T (r, c) is a polyhedron of dimension (k −1)(l −1). Because
of the balance condition (2.4.1), which is necessary and sufficient for T (r, c) to be
non-empty, the equations (2.4.2) are not linearly independent, and to bring them
into the form required by Theorem 2.2 it suffices to drop precisely one equation
from the list. The polyhedron T (r, c) is called a 2-way transportation polytope with
margins r and c, see [DK14], and its volume was investigated, in particular, in
[CM09], [Ba09], [BH12] and [B+20].
Some margins are of a particular interest. If k = l and ρi = γj = 1 for all i and j,
which we write as r = c = 1, we get the polytope of k×k doubly stochastic matrices,
also known as the Birkhoff or Birkhoff - von Neumann polytope, cf. [DK14]. In this
case, Canfield and McKay [CM09] obtained an asymptotic formula for the volume
as k −→ ∞:
 
1 1 2
vol T (1, 1) = 1 exp + k (1 + o(1)) .
(2π)k− 2 k (k−1)2 3

By symmetry, the analytic center of T (1, 1) is the matrix Z = (ζij ) with

1
ζij = for all i, j.
k
5
Note that the formula differs from the Gaussian approximation (2.3.1) by a factor
of e1/3 . In [BH12], this factor was interpreted as the Edgeworth correction in
the Central Limit Theorem, and so corrected Gaussian approximation asymptotic
formula was extended for all “tame” margins, where k and l grow proportionately,
and all coordinates ζij of the analytic center are within a constant factor of each
other.
One can observe a curious phase transition destroying the tameness of margins
somewhat unexpectedly. Suppose that k = l and that

ρ1 = . . . = ρk−1 = γ1 = . . . = γk−1 = 1.

It is not hard to show that if we choose

ρk = γk = 2 −  for some small  > 0

then the entries of the analytic center satisfy

max ζij = O k −1


as k −→ ∞.
ij

However, if we choose

ρk = γk = 2 +  for some small  > 0,

then the ζkk entry becomes large:

ζkk > δ for some δ = δ() > 0.

This and similar phase transitions are investigated in [D+20]. Their existence
may serve as an indication that vol T (r, c) cannot be estimated too closely by a
smooth analytic expression as the margins r and c vary even mildly. The formula
of Theorem 2.2 approximates vol T (r, c) within a factor of exp {O(k + l)} and it
appears to be the only known formula where the bound on the approximation
factor does not depend on the numerics of the margins r and c. It also provides
the best known approximation for the volume of a generic 2-way transportation
polytope.
(2.5) Example: 3-way planar transportation polytopes. For an integer r >
0 we consider the polytope Pr of all r × r × r arrays (tensors) X = (ξijk ) satisfying
the equations
r
X r
X
ξijk = 1 for j, k = 1, . . . , r, ξijk = 1 for i, k = 1, . . . , r
i=1 j=1
(2.5.1) r
X
and ξijk = 1 for i, j = 1, . . . , r
k=1
6
and inequalities
ξijk ≥ 0 for all i, j, k.
The polytope Pr is known as a 3-way planar transportation polytope, see [DK14].
One can also consider 3-way axial transportation polytopes obtained by fixing sums
over 2-dimensional coordinate sections of the array; somehow, those turn out to
have a simpler structure than Pr . The linear equations (2.5.1) are not independent,
and it is not hard to check that dim Pr = (r − 1)3 .
As is well known, the vertices of the Birkhoff polytope of Section 2.4 are the
permutation matrices. The integer vertices of Pr correspond to Latin squares, but
there are plenty of non-integer vertices [LL14] and the arithmetic of their coordi-
nates can vary wildly [Gr92]. By symmetry, the analytic center Z = (ζijk ) of Pr
satisfies
1
ζijk = for all i, j, k.
r
2
Theorem 2.2 implies that up to a factor of γ r for some absolute constant γ > 0, the
3 3
volume of Pr is approximated by er r−(r−1) . Hence we obtain an asymptotically
exact estimate

(2.5.2) ln vol Pr = r3 − (r − 1)3 ln r + O(r2 ) as r −→ ∞.

It appears that (2.5.2) is the best estimate of the volume of Pr to date.

In the rest of the paper, we prove Theorem 2.2. In Section 3, we collect some
preliminaries. In Section 4, we prove the upper bound of Part (1). In Section 5, we
prove the lower bound of Part (2), and in Section 6, we prove the lower bound of
Part (3).

3. Preliminaries
(3.1) The maximum entropy density. Recall that a real-valued random vari-
able X has the standard exponential distribution if the density pX (t) of X satisfies
 −t
e if t ≥ 0
pX (t) =
0 if t < 0.
For the expectation and variance, we have

EX = 1 and var X = 1.

Let matrix A, vector b, function f , polyhedron P and point z = (ζ1 , . . . , ζn ) be

as in Theorem 2.2. Let a1 , . . . , an be the columns of matrix A, considered as m-
vectors. Suppose further that X1 , . . . , Xn are independent standard exponential
random variables and let us define a random vector Y with values in Rm by
n
X n
X
(3.1.1) Y = ζj Xj aj = Xj bj ,
j=1 j=1
7
where b1 , . . . , bn are the columns of matrix B, as defined in Theorem 2.2. It is
proved in [BH10] that the density pY at b can be expressed as
vol P
(3.1.2) pY (b) = √
ef (z) det AAT
and that for the expectation and the covariance matrix of Y , we have

(3.1.3) E Y = b and Cov Y = BB T .

We also need the characteristic function of Y . For t ∈ Rm , t = (τ1 , . . . , τm ), we

have
n√ o Yn n√ o
φY (t) =E exp −1hY, ti = E exp −1Xj hbj , ti
j=1
n
Y 1
= √ ,
j=1
1− −1hbj , ti

where h·, ·i is the standard inner product in Rm . Consequently, the density of Y

can be recovered as
Z n √ n
1 oY 1
(3.1.4) pY (y) = m
exp − −1hy, ti √ dt,
(2π) Rm j=1
1 − −1hb j , ti

see also [BH10]. We will be interested only in the situations when the integral
(3.1.4) converges absolutely.
Equations (3.1.1) – (3.1.4) are the only ones we need from this section for the
proof of Theorem 2.2. The rest contains some explanatory remarks.
Equations (3.1.3) and (3.1.4) are straightforward to check, while equation (3.1.2)
follows from the fact that the density of the random vector Z = (ζ1 X1 , . . . , ζn Xn ) is
constant on P and equal to e−f (z) , see Theorem 7 in [BH10]. The formal proof easily
follows from the Lagrange optimality condition for z. A more intuitive explanation
is that Z has the largest entropy among all random vectors supported on Rn+ and
with expectation in the affine subspace defined by the system Ax = b, just as the
standard exponential distribution has the largest entropy among all distributions
supported on R+ and with expectation 1.
Since Y is the sum of independent random variables and (3.1.3) holds, in view of
the (local) Central Limit Theorem it is not inconceivable that in the vicinity of b, the
distribution of Y can be close to a Gaussian distribution. By analyzing the integral
(3.1.4), it is shown in [BH10] that it is indeed the case under some conditions on
A and z, and hence we obtain the Gaussian approximation formula (2.3.1). It is
further shown in [BH12] that for transportation polytopes with “tame” margins,
the local Central Limit Theorem holds, albeit with the Edgeworth correction that
takes into account the 3rd and 4th moments of Y .
8
(3.2) Isotropic and log-concave densities. Recall that a non-negative measur-
able function f : Rm −→ R+ is called density if
Z
f (x) dx = 1.
Rm

A density f is called centered if

Z
xf (x) dx = 0,
Rm

where we assume that the integral converges absolutely. A density f is called

isotropic if it is centered and

1 if i = j
Z
ξi ξj f (x) dx = where x = (ξ1 , . . . , ξm ) .
Rm 0 if i 6= j,
A density f : Rm −→ R+ is called logarithmically concave or log-concave, if it can
be written as f (x) = eg(x) , where g : Rm −→ R ∪ {−∞} is a concave function.
We will use the following basic fact.
Let f : Rn −→ R+ be a log-concave density, let L ⊂ Rn be a subspace and let
n
R −→ L be the orthogonal projection. Then the push-forward density h on L
defined by Z
h(y) = f (x) dx,
y+L⊥

where L⊥ is the orthogonal complement of L, is also log-concave. This is a standard

corollary of the Prékopa - Leindler inequality, see, for example, Chapter I of [A+15].
(3.3) Transforming the density of Y . From (3.1.2), we express the volume of
P as
√
(3.3.1) vol P = ef (z) det AAT pY (b),
where pY is the density of the random variable Y defined by (3.1.1). The analytic
center z depends only on the affine subspace defined by the system Ax = b, but
not on a particular choice of a matrix A and vector b. If W is an invertible m × m
matrix, then the affine subspaces defined by systems Ax = b and A0 x = b0 where
A0 = W A and b0 = W b coincide. If we replace A by A0 = W A and b by b0 = W b
then the matrix B gets replaced by B 0 = W B, and we have
det AAT det A0 (A0 )T
= ,
det BB T det B 0 (B 0 )T
and hence the estimate E(A, b) of Theorem 2.2 does not change. Furthermore, we
have B 0 (B 0 )T = W (BB T )W T . Choosing an appropriate W if needed, without loss
of generality, we assume that B satisfies
(3.3.2) BB T = Im ,
9
where Im is the m × m identity matrix. If (3.3.2) holds, then in view of the formula
(3.1.3), the density pY −b of Y − b is isotropic. The crucial fact for us is that pY is
also log-concave. Because of (3.3.2), we can identify Rm isometrically with an m-
dimensional subspace L in Rn , so that B is the matrix of the orthogonal projection
Rn −→ L in some pair of orthonormal bases of Rn and L. Let
( n P o
n
exp − j=1 ξj if ξ1 , . . . , ξn ≥ 0
f (ξ1 , . . . , ξn ) =
0 otherwise.

be the standard exponential density on Rn . Obviously, f is log-concave. Then pY

is the push-forward of f and hence is also log-concave.

4. Proof of the upper bound

In this section, we prove Part (1) of Theorem 2.2. Our approach is inspired by
Ball’s work on the volume of a section of the cube [Ba89].
(4.1) More preliminaries. As is discussed in Section 3.3, we assume that matrix
B satisfies (3.3.2). We define the random variable Y by (3.1.1). In view of (3.3.1),
our goal is to bound the `∞ -norm kpY k∞ of the density pY of Y . From (3.1.4), we
have
Z n
1 Y
− 21
(4.1.1) kpY k∞ ≤ 1 + hbj , ti2 dt
(2π)m Rm j=1

For a vector a ∈ Rm , a = (α1 , . . . , αm ), by a ⊗ a we denote the m × m matrix

with the (i, j)-th entry equal αi αj . To bound the integral in the right hand side
of (4.1.1), we use the Brascamp - Lieb inequality in the form adapted by Ball, see
Theorem 2 in [Ba01].
(4.2) Lemma. Let u1 , . . . , un be unit vectors from Rm and let λ1 , . . . , λn be pos-
itive numbers such that
Xn
λj (uj ⊗ uj ) = Im ,
j=1

where Im is the m × m identity matrix. Then, for measurable functions f1 , . . . , fn :

R −→ R+ , we have

Z n n Z +∞ λj
λ
Y Y
fj j (huj , xi) dx ≤ fj (ξ) dξ .
Rm j=1 j=1 −∞


Next, we investigate 1-dimensional integrals.
10
(4.3) Lemma. For α ∈ (0, 1), let
Z +∞
1 1

− 2α
F (α) = 1 + ατ 2 dτ.
2π −∞

Then
(1) The function F (α) is increasing on the interval (0, 1);
(2) There is a unique α0 ∈ (0, 1) such that
1
F (α0 ) = √ .
α0

Numerically,
α0 ≈ 0.7148659168.

Proof. Let
 

1
2 − 2α 1 2


h(τ, α) = 1 + ατ = exp − ln 1 + ατ .
2α
Then
τ2
   
∂ 1 2 1
ln 1 + ατ 2



h(τ, α) = ln 1 + ατ − exp −
∂α 2α2 2α (1 + ατ 2 ) 2α
and

2 2
1 2

τ 2 (1 + ατ ) ln 1 + ατ − ατ 2
ln 1 + ατ − = .
2α2 2α (1 + ατ 2 ) 2α2 (1 + ατ 2 )
Finally, we observe that

g(σ) = (1 + σ) ln(1 + σ) − σ > 0 for σ > 0,

since g(0) = 0 and

g 0 (σ) = ln(1 + σ) > 0 for σ > 0.
Summarizing, the function α 7−→ h(τ, α) is increasing for all τ 6= 0 and constant
for τ = 0. The proof of Part (1) follows.
For 0 < α < 1, we have
√ Z +∞ Z +∞
√ α
1
2 − 2α 1
− 1
αF (α) = 1 + ατ dτ = 1 + σ 2 2α dσ.
2π −∞ 2π −∞
√
Hence as α changes from 0 to 1, the value of αF (α) increases from 0 to +∞.
√
We find α0 from the equation α0 F (α0 ) = 1, which we solve numerically. This
completes the proof of Part (2). 
Recall that b1 , . . . , bn are the columns of matrix B. By k · k we denote the
standard Euclidean norm in Rm .
11
(4.4) Corollary. Suppose that the columns b1 , . . . , bn of matrix B satisfy
n
X
bj ⊗ bj = Im
j=1

and that
√
kbj k ≤ α0 for j = 1, . . . , n,
where α0 is the constant of Lemma 4.3. Then
−m/2
kpY k∞ ≤ α0 .

Proof. We use (4.1.1). Without loss of generality, we assume that bj 6= 0 for

j = 1, . . . , n. Let
1
λj = kbj k2 and uj = p bj for j = 1, . . . , n.
λj
Hence uj are unit vectors,
n
X n
X
(4.4.1) λj (uj ⊗ uj ) = Im and λj = m,
j=1 j=1

where the second identity is obtained comparing the traces of matrices on both
sides of the first identity. Besides.
0 < λj ≤ α0 for j = 1, . . . , n.
By (4.1.1), we have
Z n
1 Y
−1/2
kpY k∞ ≤ 1 + λj huj , ti2 dt
(2π)m Rm j=1
Z Y n  λj
1
1
2 − 2λj
= 1 + λj huj , ti dt
(2π)m Rm j=1

n Z +∞ λj
1 Y
1
2 − 2λj
≤ 1 + λj τ dτ
(2π)m j=1 −∞
n Z +∞ λj
1 Y
1
2 − 2α0
≤ 1 + α0 τ dτ
(2π)m j=1 −∞
n  λj
1 Y 2π −m/2
= m
√ = α0 .
(2π) j=1 α0

We use Lemma 4.2 in the inequality of the third line, Part (1) of Lemma 4.3 in the
inequality of the fourth line and Part (2) of Lemma 4.3 and (4.4.1) in the last line.

To complete the proof, we need the following standard result.
12
(4.5) Lemma. Let f, g : R −→ R+ be densities and let h : R −→ R+ be its
convolution Z +∞
h(ξ) = f (ξ − τ )g(τ ) dτ.
−∞

Then h is a density and

khk∞ ≤ min {kf k∞ , kgk∞ } .


(4.6) Proof of Part (1) of Theorem 2.2. As before, we assume that the matrix
B satisfies (3.3.2), or equivalently, the columns b1 , . . . , bn of B satisfy
n
X
(4.6.1) bj ⊗ bj = Im ,
j=1

where Im is the m × m identity matrix. Another equivalent way to write (4.6.1) is

n
X
hbj , xi2 = kxk2 for all x ∈ Rm .
j=1

In view of (3.3.1), our goal is to extend the conclusion of Corollary 4.4, without
√
assuming that kbj k ≤ α0 for j = 1, . . . , n.
We proceed by induction on m. If m = 1 then
n
X n
X
Y = µj Xj where µ2j = 1.
j=1 j=1

If we have
√
|µj | ≤ α0 for j = 1, . . . , n,
√
the result follows from Corollary 4.4. If for at least one µj we have |µj | > α0 then
by Lemma 4.5, we have
1 1
kpY k∞ ≤ pµj Xj ∞ = < √ .
|µj | α0

Suppose now that m > 1. Let

λj = kbj k2 for j = 1, . . . , n.

If
λj ≤ α0 for j = 1, . . . , n,
13
the result follows by Corollary 4.4. Otherwise, we have λj > α0 for some j. Without
loss of generality, we assume that λn > α0 . From (4.6.1) it follows that λn ≤ 1.
Suppose first that λn = 1. Then from (4.6.1) we must have

(4.6.2) hbj , bn i = 0 for j = 1, . . . , n − 1.

We consider a decomposition Rm = Rm−1 ⊕ R, y = (y 0 , η), where we identify

span (b1 , . . . , bn−1 ) = Rm−1 and span (bn ) = R.

From (4.6.1) and (4.6.2), we have

n−1
X
bj ⊗ bj = Im−1 .
j=1

Pn−1
Let Y 0 = j=1 Xj bj be a random vector in Rm−1 and let Y 00 = Xn bn = ±Xn be
a random variable with values in R, so that Y = (Y 0 , Y 00 ). Then

pY (y 0 , η) = pY 0 (y 0 )pY 00 (η)

and applying the induction hypothesis, we obtain

−(m−1)/2 −m/2
kpY k∞ = kpY 0 k∞ kpY 00 k∞ = kpY 0 k∞ ≤ α0 < α0 .

It remains to consider the case where

(4.6.3) α0 < λn < 1.

We consider a decomposition Rm = Rm−1 ⊕ R, where R is identified with span (bn )

and Rm−1 is identified with the orthogonal complement b⊥ n . For j = 1, . . . , n − 1,
let b0j be the orthogonal projection of bj onto Rm−1 and let b00j be the orthogonal
projection of bj onto R. From (4.6.1) it follows that
n−1
X
b0j ⊗ b0j = Im−1 .
j=1

We introduce a random vector Y 0 with values in Rm−1 by

n−1
X
0
Y = Xj b0j
j=1

and a random variable Y 00 with values in R by

n−1
X
00
Y = Xn bn + Xj b00j ,
j=1
14
so that Y = (Y 0 , Y 00 ). By the induction hypothesis, we have
−(m−1)/2
(4.6.4) kpY 0 k∞ ≤ α0 .
Using conditional density, for y ∈ Rm , y = (y 0 , η) where y 0 ∈ Rm−1 and η ∈ R, we
write
if pY 0 (y 0 ) = 0

0 0
(4.6.5) pY (y , η) =
pY 0 (y 0 )pY 00 |Y 0 (η|y 0 ) if pY 0 (y 0 ) 6= 0.
Let
n−1
X
Z= Xj b00j .
j=1
Pn−1
Since (4.6.1) and (4.6.3) hold, the matrix i=1 bi ⊗ bi is invertible. Therefore, the
random vector
n−1
X
(Y 0 , Z) = Xj bj
j=1

has density and hence the conditional density pZ|Y 0 exists whenever pY 0 6= 0.
Now, we have Y 00 = Z + Xn bn and hence for the conditional densities we have
pY 00 |Y 0 = pZ|Y 0 ∗ pXn bn .
Therefore, by Lemma 4.5,

pY 00 |Y 0 ≤ kpX b k = 1 1
(4.6.6) ∞ n n ∞
≤ √ .
kbn k α0
Combining (4.6.4) – (4.6.6), we conclude that
−m/2
kpY k∞ ≤ α0 .
The proof now follows from (3.3.1). 

5. Proof of Part (2)

The bound of Part (2) of Theorem 2.2 will follow from some general estimate for
isotropic log-concave densities.
(5.1) Lemma. Let f : Rm −→ R+ be a centered log-concave density and let
H ⊂ Rn be a closed halfspace containing 0. Then
Z
1
f (x) dx ≥ .
H e

Proof. When f is the uniform density on a convex body, the result was proved by
Grünbaum [Gr60] (with a slightly better constant depending on n and decreasing
to 1/e as n grows). For an adaptation to general log-concave measures, see [LV07],
Proposition 1.5.16 in [A+15] or Lemma 2.2.6 in [B+14]. 
15
(5.2) Theorem. Let f : Rm −→ R+ be an isotropic log-concave density. Then

m+2


2Γ 2
f (0) ≥ m/2 (m+2)/2
.
π e (m + 2)m/2

Proof. Without loss of generality, we assume that f is not constant on open subsets
of Rn : a general f can be approximated bya sequence of strictly log-concave
isotropic densities fn obtained from f (x) exp −kxk2 /n by a scaling, shift and
linear transformation. Then the set
n o
K = x ∈ Rm : f (x) ≥ f (0)

is convex and such that 0 ∈ ∂K. There is a hyperplane supporting K at 0. Let H

be the open halfspace bounded by that hyperplane and disjoint from the interior
of K. We have

(5.2.1) f (x) ≤ f (0) for all x∈H

and by Lemma 5.1,

Z
1
(5.2.2) β := f (x) dx ≥ .
H e

It is clear now that f (0) > 0. From this point on, our proof mimics that of
Proposition 10.2.5 of [A+15] that provides a lower bound for the `∞ -norm of an
isotropic, but not necessarily log-concave density. Let

π m/2
κm = m+2


Γ 2

denote the volume of the unit ball in Rm and let

n o
Dτ = x ∈ Rm : kxk ≤ τ

denote the ball of radius τ > 0 in Rm . Since f is isotropic, we have

Z m Z
X
2
(5.2.3) kxk f (x) dx = ξi2 f (x) dx = m.
Rm i=1 Rm

Let ρ > 0 be a number, to be specified later.

16
 Taking into account (5.2.3) and then (5.2.1), we obtain

kxk2
Z Z Z Z !
m= kxk2 f (x) dx ≥ kxk2 f (x) dx = 1 dτ f (x) dx
Rm H H 0
! !
Z +∞ Z Z +∞ Z
= f (x) dx dτ = β− f (x) dx dτ
0 H\D√τ 0 H∩D√τ
!
Z ρ Z Z ρ  
1
≥ β− f (x) dx dτ ≥ β − κm f (0)τ m/2 dτ
0 H∩D√τ 0 2
κm f (0) m+2
= ρβ − ρ 2 .
m+2

Optimizing on ρ, we choose
 2/m
2β
ρ=
κm f (0)

and obtain
 2/m   m+2
2β κm f (0) 2β m
m ≥ β−
κm f (0) m+2 κm f (0)
22/m β (m+2)/m m
= 2/m
,
(κm f (0)) m+2

from which m+2 m+2

2β 2 Γ m+2


2β 2 2
f (0) ≥ = m/2 ,
κm (m + 2)m/2 π (m + 2)m/2
and the proof follows by (5.2.2). 
(5.3) Proof of Part (2). As before, without loss of generality, we assume that
matrix B satisfies (3.3.2). Then by (3.1.3), the density pY −b of Y − b is isotropic.
It is also, as we discussed in Section 3.3, log-concave. Hence by Theorem 5.2, we
have
2Γ m+2


pY (b) = pY −b (0) ≥ m/2 (m+2)/2 2 .
π e (m + 2)m/2
The proof now follows by (3.3.1). 

6. Proof of Part (3)

(6.1) More preliminaries. We start with recalling a well-known fact that a
bound on the Cheeger constant implies the thin-shell property, see Chapter 2 of
[A+21] or Chapters 13 and 14 of [B+14].
17
 Let µ be a Borel probability measure in Rm and let dist be the standard distance
in Rm . For a Borel set A ⊂ Rm and τ > 0, we define Aτ ⊂ Rm by
n o
m
Aτ = x ∈ R : dist(x, A) ≤ τ

and
µ(Aτ ) − µ(A)
µ(A+ ) = lim inf .
τ −→0+ τ
The Cheeger constant χµ is the largest χ ≥ 0 such that

µ(A+ ) ≥ χ min µ(A), 1 − µ(A)



for all Borel sets A ⊂ Rm . For a function F : Rm −→ R, we define

|F (x) − F (y)|
|∇F (x)| = lim sup .
dist(x,y)−→0+ dist(x, y)

We need the following result. For a locally Lipschitz integrable function F : Rm −→

R, we have
Z Z
2
(6.1.1) |F (x) − E F | dµ(x) ≤ |∇F (x)| dµ(x),
Rm χµ R m

see, for example, Proposition 1.4 in [BH97].

We will apply (6.1.1) with F (x) = kxk where we have |∇F (x)| = 1 for all x ∈ Rm .
Chen [Ch21] proved that if µ is a measure with a log-concave isotropic density
then
1 n √ o
(6.1.2) ≤ exp γ ln m ln ln m
χµ

for some absolute constant γ > 0 (we assume that m ≥ 3).

(6.2) Theorem. For any 0 <  < 1/2, there is γ() > 0 such that if f : Rm −→ R+
is an isotropic log-concave density, then
n 1
o  1 m
+
f (0) ≥ exp −γ()m 2 √ .
2πe

Proof. In view of Theorem 5.2, without loss of generality we may assume that
m ≥ 3. As in the proof of Theorem 5.2, we find a halfspace H ⊂ Rm such that
0 ∈ ∂H so that
Z
1
(6.2.1) f (x) dx ≥ and f (0) ≥ f (x) for all x ∈ H.
H e
18
Let Z
α= kxkf (x) dx.
Rm

Since f is isotropic, we have

1/2
√
Z
2
α ≤ kxk f (x) dx = m.
Rm

Next, we apply (6.1.1) and (6.1.2) for the probability measure µ with density f (x)
and the function F (x) = kxk. We obtain
Z n √ o
| kxk − α| f (x) dx ≤ 2 exp γ ln m ln ln m .
Rm

Then by the Markov inequality, for any 0 <  < 1/2, there is a constant γ1 () > 0
such that n √ o 1
µ x ∈ Rm : kxk ≥ m + γ1 ()m ≤ .
2e
Comparing this with (6.2.1), we conclude that
n √ o 1
µ x ∈ Rm : f (x) ≤ f (0) and kxk ≤ m + γ1 ()m ≥ .
2e
Hence
√
m 1
f (0)κm m + γ1 ()m ≥
2e
and
Γ m+2


1 √  −m

2
f (0) ≥ m + γ1 ()m = √ m.
2eκm 2eπ m/2 ( m + γ1 ()m )
The proof now follows by (2.2.1). 
 m
1
We remark that to obtain the asymptotic √2πe in the main term, we could
√
use in (6.1.2) any bound for χ−1
µ that is o( m), see survey [LV17] and references
therein.
(6.3) Proof of Part (3). The proof follows as in Section 5.3, except that we use
Theorem 6.2 instead of Theorem 5.2. 

References
[A+15] S. Artstein-Avidan, A. Giannopoulos, and V.D. Milman, Asymptotic Geometric Analy-
sis. Part I, Mathematical Surveys and Monographs, 202, American Mathematical Soci-
ety, Providence, RI, 2015.
[A+21] S. Artstein-Avidan, A. Giannopoulos, and V.D. Milman, Asymptotic Geometric Anal-
ysis. Part II, Mathematical Surveys and Monographs, 261, American Mathematical
Society, Providence, RI, 2021.
19
[Ba89] K. Ball, Volumes of sections of cubes and related problems, Geometric Aspects of Func-
tional Analysis (1987–88), Lecture Notes in Mathematics, vol. 1376, Springer, Berlin,
1989, pp. 251–260.
[Ba01] K. Ball, Convex geometry and functional analysis, Handbook of the Geometry of Banach
Spaces, Vol. I, North-Holland, Amsterdam, 2001, pp. 161–194.
[BF87] I. Bárány and Z. Füredi, Computing the volume is difficult, Discrete & Computational
Geometry 2 (1987), no. 4, 319–326.
[Ba09] A. Barvinok, Asymptotic estimates for the number of contingency tables, integer flows,
and volumes of transportation polytopes, International Mathematics Research Notices.
IMRN (2009), no. 2, 348–385.
[BH10] A. Barvinok and J.A. Hartigan, Maximum entropy Gaussian approximations for the
number of integer points and volumes of polytopes, Advances in Applied Mathematics
45 (2010), no. 2, 252–289.
[BH12] A. Barvinok and J.A. Hartigan, An asymptotic formula for the number of non-negative
integer matrices with prescribed row and column sums, Transactions of the American
Mathematical Society 64 (2012), no. 8, 4323–4368.
[BH97] S.G. Bobkov and C. Houdré, Some connections between isoperimetric and Sobolev-type
inequalities, Memoirs of the American Mathematical Society, vol. 129, no. 616, 1997.
[B+14] S. Brazitikos, A. Giannopoulos, P. Valettas and B-H. Vritsiou, Geometry of Isotropic
Convex Bodies, Mathematical Surveys and Monographs, 196, American Mathematical
Society, Providence, RI, 2014.
[B+20] P. Brändén, J. Leake and I. Pak, Lower bounds for contingency tables via Lorentzian
polynomials, preprint arXiv:2008.05907 (2020).
[CM09] E.R. Canfield and B.D. McKay, The asymptotic volume of the Birkhoff polytope, Online
Journal of Analytic Combinatorics (2009), No. 4, 4 pp.
[Ch21] Y. Chen, An almost constant lower bound of the isoperimetric coefficient in the KLS
conjecture, Geometric and Functional Analysis 31 (2021), no. 1, 34–61.
[DK14] J.A. De Loera and E.D. Kim, Combinatorics and geometry of transportation polytopes:
an update, Discrete Geometry and Algebraic Combinatorics, Contemporary Mathemat-
ics, 625, American Mathematical Society, Providence, RI, 2014, pp. 37–76.
[D+20] S. Dittmer, L. Hanbaek and I. Pak, Phase transition in random contingency tables with
non-uniform margins, Transactions of the American Mathematical Society 373 (2020),
no. 12, 8313–8338.
[GK18] P. Gritzmann and V. Klee, Computational convexity, Handbook of Discrete and Com-
putational Geometry, Third edition (J.E. Goodman, J. O’Rourke and C. Tóth eds.),
Discrete Mathematics and its Applications (Boca Raton), CRC Press, Boca Raton, FL,
2018, pp. 937–964.
[Gr92] M.B. Gromova, The Birkhoff-von Neumann theorem for polystochastic matrices, Selecta
Mathematica 11 (1992), 145–158.
[G+88] M. Grötschel, L. Lovász and A. Schrijver, Geometric Algorithms and Combinatorial
Optimization, Algorithms and Combinatorics: Study and Research Texts, 2, Springer-
Verlag, Berlin, 1988.
[Gr60] B. Grünbaum, Partitions of mass-distributions and of convex bodies by hyperplanes,
Pacific Journal of Mathematics 10 (1960), 1257–1261.
[LV17] Y.T. Lee and S.S. Vempala, The Kannan-Lovász-Simonovits conjecture, Current Devel-
opments in Mathematics 2017, Int. Press, Somerville, MA, 2019, pp. 1–36.
[LL14] N. Linial and Z. Luria, On the vertices of the d-dimensional Birkhoff polytope, Discrete
& Computational Geometry 51 (2014), no. 1, 161–170.
[LV07] L. Lovász and S. Vempala, The geometry of logconcave functions and sampling algo-
rithms, Random Structures & Algorithms 30 (2007), no. 3, 307–358.
[NN94] Yu. Nesterov and A. Nemirovskii, Interior-Point Polynomial Algorithms in Convex Pro-
gramming, SIAM Studies in Applied Mathematics, 13, Society for Industrial and Applied
20
 Mathematics (SIAM), Philadelphia, PA, 1994.
[Pi89] G. Pisier, The Volume of Convex Bodies and Banach Space Geometry, Cambridge Tracts
in Mathematics, 94, Cambridge University Press, Cambridge, 1989.
[Re88] J. Renegar, A polynomial-time algorithm, based on Newton’s method, for linear pro-
gramming, Mathematical Programming 40 (1988), no. 1, (Series A), 59–93.
[Ve05] S. Vempala, Geometric random walks: a survey, Combinatorial and Computational Ge-
ometry, Mathematical Sciences Research Institute Publications, 52, Cambridge Univer-
sity Press, Cambridge, 2005, pp. 577–616.

Department of Mathematics, University of Michigan, Ann Arbor, MI 48109-1043,

USA
E-mail address: [email protected], [email protected]

21