More on zeros and approximation of the Ising partition function

by   Alexander Barvinok, et al.

We consider the problem of computing ∑_x e^f(x), where f(x)=∑_ij a_ijξ_i ξ_j + ∑_i b_i ξ_i is a real-valued quadratic function and x=(ξ_1, ..., ξ_n) ranges over the Boolean cube {-1, 1}^n. We prove that for any δ >0, fixed in advance, the value of ∑_x e^f(x) can be approximated within relative error 0 < ϵ < 1 is quasi-polynomial n^O(ln n - lnϵ) time, as long as ∑_j |a_ij| ≤ 1-δ for all i. We apply the method of polynomial interpolation, for which we prove that ∑_x e^f(x) 0 for complex a_ij and b_i such that ∑_j | a_ij| ≤ 1-δ, ∑_j | a_ij| ≤δ^2/10 and | b_i| ≤δ^2/10 for all i, which is interpreted as the absence of a phase transition in the Lee - Yang sense in the corresponding Ising model. The bounds are asymptotically optimal. The novel feature of the bounds is that they control the total interaction of each vertex but not every pairwise interaction.


page 1

page 2

page 3

page 4


Integrating products of quadratic forms

We prove that if q_1, ..., q_m: R^n ⟶ R are quadratic forms in variable...

Smoothed counting of 0-1 points in polyhedra

Given a system of linear equations ℓ_i(x)=β_i in an n-vector x of 0-1 va...

Isoperimetric Inequalities for Real-Valued Functions with Applications to Monotonicity Testing

We generalize the celebrated isoperimetric inequality of Khot, Minzer, a...

Computing permanents of complex diagonally dominant matrices and tensors

We prove that for any λ > 1, fixed in advance, the permanent of an n × n...

Searching for dense subsets in a graph via the partition function

For a set S of vertices of a graph G, we define its density 0 ≤σ(S) ≤ 1 ...

Optimal Piecewise Linear Function Approximation for GPU-based Applications

Many computer vision and human-computer interaction applications develop...

Data-driven geometric scale detection via Delaunay interpolation

Accurate approximation of a real-valued function depends on two aspects ...

1. Introduction and main results

The Ising model is one of the oldest, most famous and most studied models in statistical physics, see [FV18] for a thorough introduction, description, results and references. In this paper, we look at the computational complexity and complex zeros of the partition function in the Ising model. This is a classical and also currently very active area of research, see [G+19], [J+19], [L+12], [L+19a], [L+19b], [PR18], [S+14] and [Z+11] for some recent results.

Formally, the partition function we work with is described as follows. Let be the -dimensional Boolean cube of all


, where for . We consider a quadratic polynomial ,

and define the partition function as

Often, there is an underlying with vertices numbered and the property that if and only if and span an edge of . In this case, the variable is interpreted as the spin of a vertex and as the energy of the configuration . The coefficients describe the interactions of vertices and : the interaction is ferromagnetic if and antiferromagnetic if . The coefficients describe the external field, see [FV18] for a thorough discussion.

In this paper, we discuss the computational complexity of computing (approximating) . Via the method of polynomial interpolation [Ba16], we relate the computational complexity to the absence of complex zeros of in the vicinity of real coefficients and . We prove the following main result.

(1.1) Theorem


Suppose that for some , we have

for . Then

Some remarks are in order.

The first remark is regarding notation. We treat indices in as an unordered pair, so is the coefficient of the monomial in the quadratic polynomial and the sum accounts for all coefficients of the monomials containing . For a complex number , we denote by and the real and imaginary parts of respectively.

The second remark concerns algorithmic consequences of Theorem 1.1. From Theorem 1.1, the by now standard polynomial interpolation argument (see [Ba16], [G+19], [L+19a], [L+19b], [PR17]), produces an algorithm for approximating the partition function when the coefficients and are real and satisfy the condition

where is fixed in advance. As there is no restriction on , the sum can be exponentially large in . To avoid dealing with exponentially large numbers, we assume that we are provided with numbers for . Then the complexity of the algorithm is quasi-polynomial: we approximate within relative error in time.

We describe the algorithm in Section 3 and prove Theorem 1.1 in Section 2. In the remainder of this section we relate Theorem 1.1 to what is known about the partition function in the Ising model.

(1.2) The bounds for the zero-free region are asymptotically optimal

Let be a graph with vertices . For a real number , let us choose

and let us choose for some for all . Let us fix some positive integer and choose either

(all interactions are ferromagnetic) or

(all interactions are antiferromagnetic). We consider the partition function as a function of a complex parameter . It is known that for a fixed , as grows and ranges over all graphs with the largest degree of a vertex, the zeros of the univariate function with either choice of can get arbitrarily close to , see [BG01], [PR18]. We have

The right hand side approaches as , which shows that “1” in the “” bound of Theorem 1.1 cannot be replaced by a larger number.

(1.3) The bounds for approximation are asymptotically optimal in the antiferromagnetic case

As in Section 1.2, let be a graph of the largest degree , let us choose all for all and let us define by (1.2.1). Suppose that

It is shown in [SS14] and also in [G+16] that the problem of approximating is NP-hard under randomized reduction. Hence unless the computational complexity hierarchy collapses, we cannot approximate in quasi-polynomial time in the class of problems where all and

for an arbitrarily small , fixed in advance.

(1.4) The ferromagnetic case is special

Suppose that for all and that for and some complex parameter . Lee and Yang [LY52] showed that the zeros of the univariate function lie on the line . If are allowed to vary, then as long as for . The ferromagnetic case is also special from the complexity point of view: Jerrum and Sinclair [JS93] constructed a randomized polynomial time algorithm approximating when for some real and . Deterministic approximation algorithms of quasi-polynomial (genuinely polynomial, if the largest degree of the underlying graph is fixed in advance) complexity, are constructed in [L+19b], assuming that for some constant , fixed in advance, and , see also Section 7.4 of [Ba16]. The complexity status of the approximation problem in the ferromagnetic case of and by a deterministic algorithm appears to be not known.

(1.5) The bounds for phase transition are asymptotically optimal

Theorem 1.1 can be interpreted as saying that there is no phase transition in the Lee - Yang sense [YL52] provided the parameters and are real and satisfy (1.1.1) for some , fixed in advance. There is a related, though not identical, concept of phase transition, based on the disappearance of correlation decay, see [FV18]. If we choose for all and define as in (1.2.1), then the correlation decay occurs precisely in the interval

where is the largest degree of a vertex of , see [Z+11], [S+14], and [L+19a]. In view of (1.2.2), the bound (1.1.1) is asymptotically optimal as .

(1.6) What’s new

The main novelty of our approach to approximation compared to those of [G+19], [L+12], [L+19a], [L+19b], [S+14] and [Z+11] is that we state our condition in terms of the mixed norm on the interactions , see (1.1.1), as opposed to the uniform bound

on the strength of individual interactions [Z+11], [S+14], where is the largest degree of the underlying graph. As we remarked above, our results are asymptotically optimal, when . Generally, the condition (1.1.1) appears to be more robust than (1.6.1), as (1.1.1) is independent on the degree and allows individual coefficients to be relatively large, as long as the sum for all interactions of any given vertex remains appropriately bounded. Of course, for any particular , the conditions (1.1.1) and (1.6.1) are in general position, as it is easy to construct examples where one holds and the other is violated.

Another novelty of our approach with respect to locating zero-free regions of , compared to those of [G+19], [L+19a], and [PR18], is that we allow all parameters and to vary: this concerns both Lee - Yang zeros [LY52], [PR18] of as a function of with fixed and the Fisher zeros [L+19a] of as a function of with fixed. It appears that Theorem 1.1 is the first result establishing an asymptotically optimal zero-free region when the interactions are allowed to differ for different pairs and even to be of different signs, so that we have a mixture of ferromagnetic and antiferromagnetic interactions.

Finally we note that when the energy is described by a higher degree polynomial

on the Boolean cube, some (apparently non-optimal) estimates can be found in


2. Proof of Theorem 1.1

(2.1) Definitions and notation

We consider the Boolean cube of vectors , where for . Let us choose a set and numbers for . The set

is called a face of . Any index is called a free index of the face and the number of free indices is called the dimension of and denoted . Indexes are called fixed. For example, if then the face is the whole cube and if then the face consists of a single point. Generally, a face of dimension consists of points.

Let be a function and let be a face. We define the partial sum of on by

We will use the following straightforward identity. Let be a face of of dimension at least . Let be a free index of and let and be the faces obtained by fixing the -th coordinate of vectors to and respectively. Then

Let us fix a real number and a vector such that for . We denote by the set of all polynomials ,


We can view as a convex subset of with a non-empty interior. In particular, for any and any , we have for .

In what follows, we view non-zero complex numbers as vectors in the plane and measure angles between them.

We start with a simple geometric lemma.

(2.2) Lemma

Let be non-zero numbers such that the angle between and does not exceed some and let .

Demonstration Proof

Let , and be the orthogonal projections of , and respectively onto the bisector of the angle between and . Then

and the proof of Part 1 follows.

To prove Part 2, let and . Then

is real and the angle between and does not exceed . Let

Without loss of generality, we may assume that and hence . Let

by Part 1. Since , we have

and hence

We have

Since the function is convex on the interval , the minimum value of

subject to the constraints and is attained at . Consequently,

which proves Part 2. ∎

Part 1 can be extended to the sum of more than two vectors, for which one should require , see Lemma 3.6.3 in [Ba16]. Part 2 with a stronger condition is extended to more than two vectors in [BD20], see Lemma 2.1 there.

(2.3) Lemma

Let be a face of . Suppose that for all we have and, moreover, the following condition is satisfied: if is a free index of and and are the faces obtained by setting the -th coordinate and respectively, then the angle between the numbers and does not exceed for some .

Since for all ,

and the set is simply connected, we can choose a branch of the function for .

Let us fix two indices , at most one of which is free for . Then

for all .

Demonstration Proof

Differentiating, we get

Suppose first that neither of the indices and is free. Then the value of is constant for all and hence

from which the proof follows.

Suppose now that only one of the indices and , say , is free. Let and be the faces obtained by setting the -th coordinate and respectively. Applying (2.1.1), we get

We apply Lemma 2.2 with

By (2.1.1), we have

and the angle between and does not exceed by the assumption of the lemma. Applying Part 1 of Lemma 2.2, we get

Applying Part 2 of Lemma 2.2, we conclude that

which completes the proof. ∎

(2.4) Corollary

Let be a face as in Lemma 2.3 and let be a polynomial. Suppose that is a fixed index of and let be a polynomial obtained by replacing the coefficient in for some by . Then the angle between and does not exceed

Demonstration Proof

For , let be the polynomial obtained by replacing with in , so and and for . Then


Applying Lemma 2.3, we complete the proof. ∎

(2.5) Proof of Theorem 1.1

First, we show that there is such that

Indeed, we can just choose . Using that

and that

we obtain

We prove by induction for the following statement.

Let be a face of dimension of . Then . Moreover, if and is a free index of then the following holds. Let and be the faces of obtained by fixing the -th variable to and respectively. Then the angle between and does not exceed .

The statement clearly holds for . Suppose that , let be a free index of and let be the corresponding faces, so . By the induction hypothesis, we have and . Moreover, , where is obtained from

by replacing the coefficients for by and the coefficient by . Since is a fixed index for , replacing by leads to multiplying by , which results in the rotation of by an angle of . By Corollary 2.4, replacing all with leads to a rotation of by at most an angle of

Hence the angle between and does not exceed

by (2.5.1). Since

applying Part 1 of Lemma 2.2 with

we conclude that

This concludes the proof of the induction step and hence of Theorem 1.1. ∎

3. Approximation


We suppose that the coefficients and are real and satisfy

and some , fixed in advance (there are no restrictions on ). Here we sketch an algorithm for approximating

by the method of polynomial interpolation. Without loss of generality, we assume that and that is large enough,

since for smaller the sum can be computed by brute force.

(3.1) Writing as for a univariate polynomial

First, we rewrite as a polynomial in some new variables. We have


For given , we consider a polynomial

of degree in the variables . Thus we have

and we want to approximate .

Next, for given and , we consider a univariate polynomial

of a complex variable . Our goal is to approximate .

(3.2) Showing that in a neighborhood of

Our next goal is to show that

Since and , from (3.1.1) we deduce that

In particular,

From (3.1.1), we obtain

Let us choose an arbitrary such that and let us define

where we choose the principal branch of the logarithm, so that . Let

From Section 3.1, we have

Combining (3.2.2)–(3.2.4), we obtain


In addition, if , then

Hence by Theorem 1.1 we have that and therefore by (3.2.5) we conclude that (3.2.1) holds.

(3.3) Computing

As discussed in Section 2.2 of [Ba16], see also [PR17] and [Ba19] for some enhancements, as soon as in some neighborhood of the interval , to approximate within relative error , it suffices to compute the derivatives for where . Since in our case , to approximate within relative error , it suffices to compute for , where the implied constant in the “” notation depends on alone.

From (3.1.2), we obtain