Gorav Jindal

Given polynomials $f,g,h ∈𝔽[x_1,\ldots,x_n]$such that $f = g/h$, where both $g$and $h$are computable by arithmetic circuits of size $s$, we show that $f$can be computed by a circuit of size $poly(s,\deg(h))$. This solves a special case of division elimination for high-degree circuits (Kaltofen’87 & WACT’16). The result is an exponential improvement over Strassen’s classic result (Strassen’73) when $\deg(h)$is $poly(s)$and $\deg(f)$is $\exp(s)$, since the latter gives an upper bound of $poly(s, \deg(f))$. Further, we show that any univariate polynomial family $(f_d)_d$, defined by the initial segment of the power series expansion of rational function $g_d(x)/h_d(x)$up to degree $d$ (i.e. $f_d = g_d/h_d \bmod (x^d+1)$), where circuit size of $g$is $s_d$and degree of $g_d$is at most $d$, can be computed by a circuit of size $poly(s_d,\deg(h_d),\log(d))$. We also show a hardness result when the degrees of the rational functions are high (i.e. $Ω(d)$), assuming hardness of the integer factorization problem. Finally, we extend this conditional hardness to simple algebraic functions as well, and show that for every prime $p$, there is an integral algebraic power series with its minimal polynomial satisfying a degree p polynomial equation, such that its initial segment is hard to compute unless integer factoring is easy, or a multiple of $n!$is easy to compute. Both, integer factoring and computation of multiple of $n!$, are believed to be notoriously hard. In contrast, we show examples of transcendental power series whose initial segments are easy to compute.

We investigate the number of real zeros of a univariate k-sparse polynomial f over the reals, when the coefficients of f come from independent standard normal distributions. Recently Bürgisser, Ergür and Tonelli-Cueto showed that the expected number of real zeros of f in such cases is bounded by [EQUATION]. In this work, we improve the bound to [EQUATION] and also show that this bound is tight by constructing a family of sparse support whose expected number of real zeros is lower bounded by [EQUATION]. Our main technique is an alternative formulation of the Kac integral by Edelman-Kostlan which allows us to bound the expected number of zeros of f in terms of the expected number of zeros of polynomials of lower sparsity. Using our technique, we also recover the O (log n) bound on the expected number of real zeros of a dense polynomial of degree n with coefficients coming from independent standard normal distributions.

We consider the problem of commutative rank computation of a given matrix space. A matrix space is a (linear) subspace of the (linear) space of n x n matrices over a given field. The problem is fundamental, as it generalizes several computational problems from algebra and combinatorics. For instance, checking if the commutative rank of the space is n, subsumes problems such as testing perfect matching in graphs and identity testing of algebraic branching programs. An efficient deterministic computation of the commutative rank is a major open problem, although there is a simple and efficient randomized algorithm for it. Recently, there has been a series of results on computing the non-commutative rank of matrix spaces in deterministic polynomial time. Since the non-commutative rank of any matrix space is at most twice the commutative rank, one immediately gets a deterministic 1/2-approximation algorithm for the computation of the commutative rank. This leads to a natural question of whether this approximation ratio can be improved. In this paper, we answer this question affirmatively. We present a deterministic Polynomial-time approximation scheme (PTAS) for computing the commutative rank of a given matrix space B. More specifically, given a matrix space and a rational number e > 0, we give an algorithm, that runs in time O(n^(4 + 3/e)) and computes a matrix A in the given matrix space B such that the rank of A is at least (1-e) times the commutative rank of B. The algorithm is the natural greedy algorithm. It always takes the first set of k matrices that will increase the rank of the matrix constructed so far until it does not find any improvement, where the size of the set k depends on e.

Algebraic natural proofs were recently introduced by Forbes, Shpilka and Volk (Proc. of the 49th Annual ACM SIGACT Symposium on Theory of Computing (STOC), pages 653–664, 2017) and independently by Grochow, Kumar, Saks and Saraf (CoRR, abs/1701.01717, 2017) as an attempt to transfer Razborov and Rudich’s famous barrier result (J. Comput. Syst. Sci., 55(1): 24–35, 1997) for Boolean circuit complexity to algebraic complexity theory. Razborov and Rudich’s barrier result relies on a widely believed assumption, namely, the existence of pseudo-random generators. Unfortunately, there is no known analogous theory of pseudo-randomness in the algebraic setting. Therefore, Forbes et al. use a concept called succinct hitting sets instead. This assumption is related to polynomial identity testing, but it is currently not clear how plausible this assumption is. Forbes et al. are only able to construct succinct hitting sets against rather weak models of arithmetic circuits. Generalized matrix completion is the following problem: Given a matrix with affine linear forms as entries, find an assignment to the variables in the linear forms such that the rank of the resulting matrix is minimal. We call this rank the completion rank. Computing the completion rank is an NP-hard problem. As our first main result, we prove that it is also NP-hard to determine whether a given matrix can be approximated by matrices of completion rank ≤ b. The minimum quantity b for which this is possible is called border completion rank (similar to the border rank of tensors). Naturally, algebraic natural proofs can only prove lower bounds for such border complexity measures. Furthermore, these border complexity measures play an important role in the geometric complexity program. Using our hardness result above, we can prove the following barrier: We construct a small family of matrices with affine linear forms as entries and a bound b, such that at least one of these matrices does not have an algebraic natural proof of polynomial size against all matrices of border completion rank b, unless coNP ⊆ ∃ BPP. This is an algebraic barrier result that is based on a well-established and widely believed conjecture. The complexity class ∃ BPP is known to be a subset of the more well known complexity class in the literature. Thus ∃ BPP can be replaced by MA in the statements of all our results. With similar techniques, we can also prove that tensor rank is hard to approximate. Furthermore, we prove a similar result for the variety of matrices with permanent zero. There are no algebraic polynomial size natural proofs for the variety of matrices with permanent zero, unless P#P ⊆ ∃ BPP. On the other hand, we are able to prove that the geometric complexity theory approach initiated by Mulmuley and Sohoni (SIAM J. Comput. 31(2): 496–526, 2001) yields proofs of polynomial size for this variety, therefore overcoming the natural proofs barrier in this case.