On Convexity of Polynomials over a Box Georgina Hall Decision Sciences, INSEAD Joint work with Amir Ali Ahmadi ORFE, Princeton University 1
Convexity over a box โข A box ๐ช is a set of the form: ๐ถ = ๐ฆ โ โ ๐ ๐ ๐ โค ๐ฆ ๐ โค ๐ฃ ๐ , ๐ = 1, โฆ , ๐} where ๐ 1 , โฆ , ๐ ๐ , ๐ฃ 1 , โฆ , ๐ฃ ๐ โ โ with ๐ ๐ โค ๐ฃ ๐ . โข A function ๐ is convex over ๐ช if ๐ ๐๐ฆ + 1 โ ๐ ๐ง โค ๐๐ ๐ฆ + 1 โ ๐ ๐(๐ง) for any ๐ฆ, ๐ง โ ๐ถ and ๐ โ [0,1]. โข If ๐ช is full dimensional (i.e., ๐ ๐ < ๐ฃ ๐ , ๐ = 1, โฆ , ๐ ), this is equivalent to ๐ผ 2 ๐ ๐ฆ โฝ 0, โ๐ฆ โ ๐ถ. 2
Complexity questions Goal: study the complexity of testing convexity of a function over a box โข Restrict ourselves to polynomial functions. โข Related work: Theorem [Ahmadi, Olshevsky, Parrilo, Tsitsiklis] It is strongly NP-hard to test (global) convexity of polynomials of degree 4. โข One may hope that adding the restriction to a box could make things easier. 3
Our theorem Theorem [Ahmadi, H.] It is strongly NP-hard to test convexity of polynomials of degree 3 over a box. Why are we interested in convexity over a box? Detecting Imposing โข Nonconvex optimization: branch-and-bound โข Control theory: convex Lyapunov functions [Ahmadi and Jungers] [Chesi and Hung] โข Prior work: โข Sufficient conditions for convexity [Orban et โข Statistics: convex regression al.], [Grant et al.] โข In practice, BARON, CVX, Gurobi check convexity of quadratics and computationally tractable sufficient conditions for convexity 4
Proof of the theorem Theorem [Ahmadi, H.] It is strongly NP-hard to test convexity of polynomials of degree 3 over a box. How to prove this? Construct Generic instance I Instance J of J from I In general: of a known problem we are Reduction NP-hard problem interested in ๐ผ 2 ๐(๐ฆ) is a matrix Question: What to Idea: A cubic polynomial ๐ is convex do a reduction over a (full-dimensional) box ๐ถ if and with entries affine only if ๐ผ 2 ๐ ๐ฆ โฝ 0 , โ๐ฆ โ ๐ถ from? in ๐ Theorem [Nemirovski]: Let ๐(๐ฆ) be a matrix with entries affine in ๐ฆ . It is NP-hard to test whether ๐ ๐ฆ โฝ 0 for all ๐ฆ in a full-dimensional box ๐ถ. 5
Are we done? No! Issue 1: We want to show strong NP-hardness. Nemirovskiโs result shows weak NP- hardness. Issue 2: Not every affine polynomial matrix is a valid Hessian! 10 2๐ฆ 1 + 1 ๐๐ 11 (๐ฆ) ๐๐ 12 (๐ฆ) Example: ๐ ๐ฆ 1 , ๐ฆ 2 = . We have โ ๐๐ฆ 1 . 2๐ฆ 1 + 1 10 ๐๐ฆ 2 6
Dealing with Issue 1 (1/5) Reminder: weak vs strong NP-hardness โข Distinction only concerns problems where input is numerical โข Max(I) : largest number in magnitude that appears in the input of instance I (numerator or denominator) โข Length(I) : number of bits it takes to write down input of instance I Strong Weak โข There are instances ๐ฝ that are hard with โข The instances that are hard may contain Max( ๐ฝ ) โค ๐ (Length( ๐ฝ )) ( ๐ is a polynomial) numbers of large magnitude (e.g., 2 ๐ ) . โข No pseudo-polynomial algorithm possible โข Pseudo-polynomial algorithms possible โข Examples: โข Examples: Knapsack Sat Partition Max-Cut 7
Dealing with Issue 1 (2/5) Theorem [Nemirovski]: INTERVAL-PSDNESS Let ๐(๐ฆ) be a matrix with entries affine in ๐ฆ . It is (weakly) NP-hard to test whether ๐ ๐ฆ โฝ 0 for all ๐ฆ in a full-dimensional box ๐ถ. Why weakly NP-hard? INTERVAL PSDNESS PARTITION: Construct : ๐ท = ๐ฝ ๐ โ ๐๐ ๐ โ1 , REDUCTION Input : ๐ โ โ ๐ such that ๐ 2 โค 0.1 ๐ = ๐ โ ๐ โ2 ๐ , where ๐ ๐ = smallest cd of ๐. ๐ท ๐ฆ Take: ๐ถ = โ1,1 ๐ and ๐ ๐ฆ = Test: does there exist ๐ข โ โ1,1 ๐ ๐ . ๐ฆ ๐ such that ฯ ๐ ๐ ๐ ๐ข ๐ = 0? Test: Is ๐ ๐ฆ โฝ 0 โ๐ฆ โ ๐ถ? Show: No to PARTITION โ Yes to INTERVAL PSDNESS Weakly NP-hard Operation that can make the numbers in the instance blow up 1 0 0 0 โ1 1 0 0 ๐ 1 ๐ 2 ๐ 4 but one of the entries of ๐ต โ1 is 2 ๐โ2 ! ๐ 10 = Example: ๐ต = ๐ 8 โฑ โฑ โฎ 0 ๐ 3 8 โ1 โ1 โ1 1
Dealing with Issue 1 (3/5) Theorem [Ahmadi, H.]: INTERVAL-PSDNESS Let ๐(๐ฆ) be a matrix with entries affine in ๐ฆ . It is strongly NP-hard to test whether ๐ ๐ฆ โฝ 0 for all ๐ฆ in a full-dimensional box ๐ถ. INTERVAL PSDNESS MAX-CUT: 1 Construct : ๐ฝ = ๐+1 3 , ๐ท = 4๐ฝ(๐ฝ ๐ + ๐ฝ๐ต) ๐ = ๐ 4๐ฝ + ๐ โ 1 โ 1 Input : simple graph G=(V,E) with 4 ๐ ๐ ๐ต๐ REDUCTION ๐ = ๐ and adj. matrix A, and a ๐ท ๐ฆ positive integer ๐ โค ๐ 2 Take: ๐ถ = โ1,1 ๐ and ๐ ๐ฆ = Preserves strong ๐ . ๐ฆ ๐ NP-hardness Test: does there exist a cut in the Test: Is ๐ ๐ฆ โฝ 0 โ๐ฆ โ ๐ถ? graph of size greater or equal to ๐? Show: No to MAX-CUT โ Yes to INTERVAL PSDNESS Taylor series of 4๐ฝ ๐ฝ โ ๐ฝ๐ต โ1 truncated at the first term Strongly NP-hard Scaling needed so that ๐ฝ ๐ โ ๐ฝ๐ต โ1 โ ๐ฝ ๐ + ๐ฝ๐ต 9
Dealing with Issue 1 (4/5) In more detail: No to MAX-CUT โ Yes to INTERVAL PSDNESS 1 4 ฯ ๐,๐ ๐ต ๐๐ (1 โ ๐ฆ ๐ ๐ฆ ๐ )] โค ๐ โ 1 [max โ No cut in ๐ป of size โฅ ๐ ๐ฆโ โ1,1 ๐ โ Convex Size of largest cut in ๐ป ๐ ๐+1 3 1 1 1 4 ๐ฆ ๐ ๐ต๐ฆ ] โค โ 4 ๐ ๐ ๐ต๐ + ๐ โ 1 4 ๐ฆ ๐ ๐ + 1 3 ๐ฝ ๐ โ ๐ต ๐ฆ ] โค โ ๐ฆโ โ1,1 ๐ โ max max โ [ [ ๐ฆโ โ1,1 ๐ 4 1 4 ๐ ๐ ๐ต๐ + ๐ โ 1 โ ๐ โ ๐ฝ = ๐ + 1 3 1 4 ๐ฆ ๐ ๐ฝ๐ฝ ๐ โ ๐ต ๐ฆ ] โค ๐ 1 4 ๐ฆ ๐ ๐ฝ๐ฝ ๐ โ ๐ต ๐ฆ โค ๐ , โ๐ฆ โ โ1,1 ๐ โ [ max ๐ฆโ[โ1,1] ๐ Approximation ๐ท โ1 โ 1 4 (๐ฝ๐ฝ โ ๐ต) โ ๐ท ๐ฆ 1 โฝ 0, โ๐ฆ โ โ1,1 ๐ ๐ฆ ๐ ๐ท โ1 ๐ฆ โค ๐ + 4 , โ๐ฆ โ โ1,1 ๐ โ ๐ ๐ฆ = 1 ๐ฆ ๐ ๐ + Schur 4 Approximation error complement 10
Dealing with Issue 1 (5/5) For converse: Yes to MAX-CUT โ No to INTERVAL PSDNESS There is a cut of size โฅ ๐: ๐ฆ โฅ ๐ + 3 4 > ๐ + 1 Similar steps ๐ฆ ๐ ๐ท โ1 เท โ โ เท ๐ฆ ๐ = แ 1 if node ๐ on one side of cut Let เท to previously 4 โ1 if node ๐ on other side of cut โ ๐ฆ โ โ1,1 ๐ s.t. ๐ เท โ เท ๐ฆ 0 Corollary [Ahmadi, H.]: Let ๐ be an integer and let เท ๐ ๐๐ , เดค ๐ ๐๐ be rational numbers with เท ๐ ๐๐ โค เดค ๐ ๐๐ and เท ๐ ๐๐ = เท ๐ ๐๐ and เดค ๐ ๐๐ = เดค ๐ ๐๐ for all ๐ = 1, โฆ , ๐ and ๐ = 1, โฆ , ๐. It is strongly NP-hard to test whether all symmetric matrices with entries in [เท ๐ ๐๐ ; เดค ๐ ๐๐ ] are positive semidefinite. โข Initial problem studied by Nemirovski โข Of independent interest in robust control 11
Dealing with Issue 2 (1/3) Theorem [Ahmadi, H.] CONV3BOX It is strongly NP-hard to test convexity of polynomials of degree 3 over a box. Proof: Reduction from INTERVAL PSDNESS INTERVAL PSDNESS Problem: How to construct a cubic polynomial ๐ from ๐(๐ฆ) ? Input : ๐ ๐ฆ , เท Idea: Want ๐ผ 2 ๐ ๐ฆ = ๐ ๐ฆ . ๐ถ Test: Is ๐ ๐ฆ โฝ 0, โ๐ฆ โ เท ๐ถ? Issue: Not all ๐(๐ฆ) are valid Hessians! Key ideas for the construction of ๐: ๐ ๐ ๐ ๐ผ ๐ด ๐ ๐ โข Start with ๐ ๐, ๐ = 1 1 1 โข For ๐ผ 2 ๐ ๐ฆ, ๐ง to be able to be psd when ๐ ๐ฆ โฝ 0 , we need to have ๐ท๐ฑ ๐ ๐ท๐ฑ ๐ 0 2 ๐ผ(๐ง) 2 ๐ผ(๐ง) 2 ๐ผ(๐ง) ๐ผ 2 ๐ ๐ฆ, ๐ง = ๐ผ 2 ๐ ๐ฆ, ๐ง = ๐ผ 2 ๐ ๐ฆ, ๐ง = ๐ท ๐ ๐ ๐ผ ๐ to ๐ ๐ฆ, ๐ง . 1 1 1 a nonzero diagonal: add 2 ๐ผ ๐ง ๐ 2 ๐ผ ๐ง ๐ 2 ๐ผ ๐ง ๐ ๐ ๐ฆ + ๐๐ฝ ๐+1 ๐ ๐ฆ ๐ ๐ฆ โข ๐ ๐ฆ and ๐ผ(๐ง) do not depend on the same variable: what if โ(๐ฆ, ๐ง) s.t. ๐ ๐ฆ = 0 but ๐ผ ๐ง is not? The matrix cannot be psd: add ๐ 2 ๐ง ๐ ๐ง to ๐ ๐ฆ, ๐ง . โ ๐ ๐ฆ = 1 2 ๐ง ๐ ๐ ๐ฆ ๐ง + ๐ฝ 2 ๐ฆ ๐ ๐ฆ + ๐ 2 ๐ง ๐ ๐ง, ๐ถ = โ1,1 2๐+1 12
าง Dealing with Issue 2 (2/3) Show NO to INTERVAL PSDNESS โ NO to CONV3BOX. This is equivalent to: ๐ง โ โ1,1 2๐+1 , ๐จ s.t. ๐จ ๐ ๐ผ 2 ๐ เท ๐ฆ โ โ1,1 ๐ s.t. ๐ โ าง ๐ฆ โฝ 0 โ โ เท ๐ฆ, เท ๐ฆ, เท ๐ง ๐จ < 0 ๐ท ๐ฆ Need to leverage extra structure of ๐ ๐ฆ : ๐ ๐ฆ = 1 ๐ฆ ๐ ๐ + 4 0 ๐ 0 โ๐ท โ1 าง โ๐ท โ1 าง โ๐ซ โ๐ เดฅ Candidates: เท Candidates : เท Candidates: เท ๐ = เดฅ ๐ฆ = าง ๐ฆ = าง ๐ฆ, ๐ฆ, ๐, ๐ง = 0, ๐ง = 0, ๐ = ๐, เท เท เท ๐ = ๐จ = ๐จ = ๐ฆ ๐ ๐ฆ 1 ๐ 1 ๐ฝ๐ฝ ๐ ๐ ๐ ๐ ๐ฝ๐ฝ ๐ ๐ฝ๐ฝ ๐ ๐ ๐ผ(๐ง) ๐ 0 0 = ๐ + 1 ๐ ๐ซ + ๐๐ฝ ๐ ๐ เดฅ 2 ) ๐ฆ ๐ ๐ท โ1 าง ๐ฆ + ๐(1 + ๐ท โ1 าง โ๐ท โ1 าง โ๐ท โ1 าง ๐จ ๐ ๐ผ 2 ๐ เท ๐ฆ, เท ๐ง ๐จ = ๐ ๐ซ + ๐๐ฝ ๐ ๐ท + ๐๐ฝ ๐ เดฅ ๐ ๐ฆ 4 โ าง ๐ฆ 2 ๐ฆ ๐ฆ ๐ผ 2 ๐ ๐ฆ, ๐ง = ๐ผ 2 ๐ เท ๐ + ๐ ๐ฆ, เท ๐ง = ๐ผ ๐ง ๐ ๐ ๐ผ ๐ + ๐ ๐ + 1 1 1 ๐ เดฅ ๐ + ๐ ๐ ๐ผ ๐ฆ ๐ ๐ เดฅ ๐ + ๐ 4 + ๐ Appropriately scaled so that < ๐ as ๐ด เดฅ ๐ โฝ ๐ ๐จ ๐ ๐ผ 2 ๐ เท ๐ฆ, เท ๐ง ๐จ remains <0. 13
Recommend
More recommend