A Deterministic Linear Program Solver in Current Matrix Multiplication Time

Abstract

Interior point algorithms for solving linear programs have been studied extensively for a long time [e.g. Karmarkar 1984; Lee, Sidford FOCS'14; Cohen, Lee, Song STOC'19]. For linear programs of the form min Ax=b,x≥0 c ⊤ x with n variables and d constraints, the generic case d = Ω(n) has recently been settled by Cohen, Lee and Song [STOC'19]. Their algorithm can solve linear programs in O(n ω log(n/δ)) expected time 1 , where δ is the relative accuracy. This is essentially optimal as all known linear system solvers require up to O(n ω ) time for solving Ax = b. However, for the case of deterministic solvers, the best upper bound is Vaidya's 30 years old O(n 2.5 log(n/δ)) bound [FOCS'89]. In this paper we show that one can also settle the deterministic setting by derandomizing Cohen et al.'s O(n ω log(n/δ)) time algorithm. This allows for a strict O(n ω log(n/δ)) time bound, instead of an expected one, and a simplified analysis, reducing the length of their proof of their central path method by roughly half. Derandomizing this algorithm was also an open question asked in Song's PhD Thesis.The main tool to achieve our result is a new data-structure that can maintain the solution to a linear system in subquadratic time. More accurately we are able to maintain √ U A ⊤ (AU A ⊤ ) −1 A √ U v in subquadratic time under ℓ 2 multiplicative changes to the diagonal matrix U and the vector v. This type of change is common for interior point algorithms. Previous algorithms [e.g. Vaidya STOC'89; Lee, Sidford FOCS'15; Cohen, Lee, Song STOC'19] required Ω(n 2 ) time for this task. In [Cohen, Lee, Song STOC'19] they managed to maintain the matrix √ U A ⊤ (AU A ⊤ ) −1 A √ U in subquadratic time, but multiplying it with a dense vector to solve the linear system still required Ω(n 2 ) time. To improve the complexity of their linear program solver, they restricted the solver to only multiply sparse vectors via a random sampling argument. In comparison, our data-structure maintains the entire productadditionally to just the matrix. Interestingly, this can be viewed as a simple modification of Cohen et al.'s data-structure, but it significantly simplifies their analysis of their central path method and makes their whole algorithm deterministic.1 Here O hides polylog(n) factors and O(n ω ) is the time required to multiply two n × n matrices. The stated O(n ω log(n/δ)) bound holds for the current bound on ω with ω ≈ 2.38 [V. Williams, STOC'12; Le Gall, ISSAC'14]. The upper bound for the solver will become larger than O(n ω log(n/δ)), if ω < 2 + 1/6.

A scaling-invariant algorithm for linear programming whose running time depends only on the constraint matrix

“…For this last line, there has been substantial progress in improving IPM by amortizing the cost of the iterative updates, and working with approximate computations, see e.g. [Ren88,Vai89,CLS19,vdB20]. Very recently, Cohen, Lee and Song [CLS19] developed a new inverse maintenance scheme to get a randomizedÕ(n 2.37 log(1/ε))-time algorithm for ε-approximate LP, which was derandomized by van den Brand [vdB20].…”

Revisiting Tardos's Framework for Linear Programming: Faster Exact Solutions using Approximate Solvers

“…Harnessing the progress in approximate solvers The complexity of fast approximate LP algorithms has seen substantial improvements in recent years [LS19,CLS19,vdB20,vdBLSS20,LSZ19,JSWZ20]. Taking the recent algorithm [vdB20], given a feasible LP min c, x , Ax = b, x ≥ 0, having an optimal solution of 2 norm at most R, for ε > 0 it computes a pointx ≥ 0 satisfying c,x ≤ min in deterministic time O(n ω+o(1) log(n/ε)), where ω < 2.38 is the matrix multiplication exponent.…”

“…In Theorem 5.2 we use it for the special case c = 0; we now explain the more general case. We note that, instead of using the final output of Theorem 5.1, one could obtain more direct algorithms by using the interior-point methods of [LS19,vdB20,vdBLSS20] directly on our extended system. We now give a black-box argument using Theorem 5.1, to demonstrate the compatibility of our results with any approximate LP solver, not just interior-point methods.…”

Multiplexing Services in 5G and Beyond: Optimal Resource Allocation Based on Mixed Numerology and Mini-Slots

“…In Fig. 14 we compare the commercial LP 5 optimizer embedded in the Matlab optimization toolbox and the Dantzing-Wolfe decomposition followed by the column generation algorithm. The computational complexity of both solvers is the same and it is O(m 2,38 ), where m is the input size of the optimization problem [49]. The gain visible in Fig.…”