On Leonid Gurvits's Proof for Permanents

We give a concise exposition of the elegant proof given recently by Leonid Gurvits for several lower bounds on permanents. 1. PERMANENTS. The permanent of a square matrix A = (ai, j ) i, j=1 is defined by perA = ∑ π∈Sn n ∏ i=1 ai,π(i), (1) where Sn denotes the set of all permutations of {1, . . . , n}. (The name “permanent” has its root in Cauchy’s fonctions symétriques permanentes [2], as a counterpart to fonctions symétriques alternées—the determinants.) Despite its appearance as the simpler twin-brother of the determinant, the permanent has turned out to be much less tractable. Whereas the determinant can be calculated quickly (in polynomial time, with Gaussian elimination), determining the permanent is difficult (“number-P-complete”). As yet, the algebraic behaviour of the permanent function has appeared to a large extent unmanageable, and its algebraic relevance moderate. Most fruitful research on permanents concerns lower and upper bounds for the permanent (see the book of Minc [12]). In this paper we will consider only lower bounds. Indeed, most interest in the permanent function came from the famous van der Waerden conjecture [16] (in fact formulated as a question), stating that the permanent of any n × n doubly stochastic matrix is at least n!/nn, the minimum being attained only by the matrix with all entries equal to 1/n. (A matrix is doubly stochastic if it is nonnegative and each row and column sum is equal to 1.) This conjecture was unsolved for over fifty years, which, when contrasted with its simple form, also contributed to the reputation of intractability of permanents. Finally, Falikman [6] and Egorychev [4] were able to prove this conjecture, using a classical inequality of Alexandroff and Fenchel. The proof with eigenvalue techniques also revealed some unexpected nice algebraic behaviour of the permanent function (see, also for background, Knuth [9] and van Lint [10, 11]). Before the proof of the van der Waerden conjecture was found, a weaker conjecture was formulated by Erdős and Rényi [5]. It claims the existence of a real number α3 > 1 such that, for each nonnegative integer-valued n × n matrix A with all row and column sums equal to 3, the permanent of A is at least α 3 . This would follow from the van der Waerden conjecture, since 13 A is doubly stochastic, and hence perA = 3per ( 1 3 A ) ≥ 3 n! nn ≥ ( 3 e )n . (2) Erdős and Rényi also asked for the largest value of α3 one can take in this bound. More generally, for any natural number k, they asked for the largest real number αk such that each nonnegative integer-valued n × n matrix A = (ai, j ) with all row doi:10.4169/000298910X523380 December 2010] ON LEONID GURVITS’S PROOF FOR PERMANENTS 903 Integre Technical Publishing Co., Inc. American Mathematical Monthly 117:10 July 23, 2010 10:00 a.m. schrijver.tex page 904 and column sums equal to k has permanent at least α k . Note that this permanent is equal to the number of perfect matchings in the k-regular bipartite graph with vertices u1, . . . , un, v1, . . . , vn, where ui and v j are connected by ai, j edges. In 1979, before the van der Waerden conjecture was settled, the first conjecture of Erdős and Rényi was proved, by Bang [1], Friedland [7], and Voorhoeve [15]. Bang and Friedland in fact showed that the permanent of any n × n doubly stochastic matrix is at least e−n. Note that limn→∞(n!/nn)1/n = e−1, so this may be seen as an asymptotic proof of the van der Waerden conjecture. It also implies that the number αk of Erdős and Rényi is at least k/e; in particular, α3 ≥ 3/e > 1. The proof of Voorhoeve gives a better bound: α3 ≥ 4/3. In fact, this bound is best possible. Indeed, it follows from a theorem of Wilf [18] that α3 ≤ 4/3, and more generally αk ≤ (k − 1) k−1 kk−2 , (3) and Schrijver and Valiant [14] conjectured that equality holds for each k. For k = 1, 2, this is trivial, and for k = 3 this follows from Voorhoeve’s theorem. The proof of Voorhoeve that α3 ≥ 4/3 is very short and elegant, and it seduces one to search for similar arguments for general k. However, it was only at the cost of frightening technicalities that Schrijver [13] found a proof that equality indeed holds in (3) for each k. This amounts to a lower bound for permanents of doubly stochastic matrices in which all entries are integer multiples of 1/k. Under this restriction, this bound is larger than the van der Waerden bound. In fact, both the bound of Falikman-Egorychev and that of Schrijver are best possible, in different asymptotic directions. Let μ(k, n) denotes the minimum permanent of n × n doubly stochastic matrices with all entries being integer multiples of 1/k. Then the two bounds state μ(k, n) ≥ n! nn and μ(k, n) ≥ ( k − 1 k )(k−1)n . (4) They are best possible in the following sense: inf k μ(k, n) = n! 1/n n and inf n μ(k, n) = ( k − 1 k )k−1 . (5) The proof of Falikman and Egorychev requires some nontrivial theorems, and the proof of Schrijver is combinatorially complex. It was a big surprise when Leonid Gurvits [8] gave an amazingly short proof of the two bounds. En route, he extended Schrijver’s theorem to: each doubly stochastic n × n matrix with at most k nonzeros in each column has permanent at least ((k − 1)/k)(k−1)n. In fact, Gurvits proved that each doubly stochastic n × n matrix A satisfies perA ≥ n ∏ i=1 g(min{i, λA(i)}) (Gurvits’s inequality), (6) where λA(i) is the number of nonzeros in the i th column of A, and where g(0) := 1 and g(k) := ( k − 1 k )k−1 for k = 1, 2, . . . (7) 904 c © THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 117 Integre Technical Publishing Co., Inc. American Mathematical Monthly 117:10 July 23, 2010 10:00 a.m. schrijver.tex page 905 (setting 00 = 1). Gurvits’s bound implies both the bound of Falikman and Egorychev and the bound of Schrijver—see Section 4. We give here a proof based on Gurvits’s proof. The building blocks of the proof are from Gurvits [8], but we take a few shortcuts. 2. DESCRIPTION OF GURVITS’S APPROACH. As usual, let R+ := {x ∈ R | x ≥ 0}. Recall the geometric-arithmetic mean inequality, saying that if λ1, . . . , λn, x1, . . . , xn ∈ R+ with ∑n i=1 λi = 1, then n ∑ i=1 λi xi ≥ n ∏ i=1 xi i . (8) It amounts to the concavity of the log function. For any n × n matrix A, define the following multivariate polynomial pA in the variables x1, . . . , xn: pA(x1, . . . , xn) := n ∏ i=1 ai x = n ∏ i=1 n ∑ j=1 ai, j x j , (9) where ai denotes the i th row of A (in ai x we take ai as a row vector and x = (x1, . . . , xn) as a column vector). So pA is homogeneous of degree n. Then the coefficient of the monomial x1 · · · xn in pA is equal to perA. This can also be stated in terms of partial derivatives as perA = ∂ n pA ∂x1 · · · ∂xn . (10) Note that the latter expression is a constant function. The crux of the method is to consider more generally the following derivatives of pA, for any i = 0, . . . , n: qi (x1, . . . , xi ) := ∂ n−i pA ∂xi+1 · · · ∂xn ∣∣∣ xi+1=···=xn=0 . (11) So qi ∈ R[x1, . . . , xi ]. Then qn = pA and q0 = perA. The polynomials qi will be related through the following concept of “capacity” of a polynomial. The capacity cap(p) of a polynomial p ∈ R[x1, . . . , xn] is defined as cap(p) := inf p(x), (12) where the infimum ranges over all x ∈ R+ with ∏n j=1 x j = 1. So cap(q0) = perA. Moreover, we have: Proposition 1. If A is doubly stochastic, then cap(pA) = 1. Proof. For any x ∈ R+ with ∏n j=1 x j = 1 we have, using the geometric-arithmetic mean inequality (8): pA(x) = ∏