of polynomials over a box
play

of Polynomials over a Box Georgina Hall Decision Sciences, INSEAD - PowerPoint PPT Presentation

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 A box is a set of the form: =


  1. On Convexity of Polynomials over a Box Georgina Hall Decision Sciences, INSEAD Joint work with Amir Ali Ahmadi ORFE, Princeton University 1

  2. Convexity over a box โ€ข A box ๐‘ช is a set of the form: ๐ถ = ๐‘ฆ โˆˆ โ„ ๐‘œ ๐‘š ๐‘— โ‰ค ๐‘ฆ ๐‘— โ‰ค ๐‘ฃ ๐‘— , ๐‘— = 1, โ€ฆ , ๐‘œ} where ๐‘š 1 , โ€ฆ , ๐‘š ๐‘œ , ๐‘ฃ 1 , โ€ฆ , ๐‘ฃ ๐‘œ โˆˆ โ„ with ๐‘š ๐‘— โ‰ค ๐‘ฃ ๐‘— . โ€ข A function ๐’ˆ is convex over ๐‘ช if ๐‘” ๐œ‡๐‘ฆ + 1 โˆ’ ๐œ‡ ๐‘ง โ‰ค ๐œ‡๐‘” ๐‘ฆ + 1 โˆ’ ๐œ‡ ๐‘”(๐‘ง) for any ๐‘ฆ, ๐‘ง โˆˆ ๐ถ and ๐œ‡ โˆˆ [0,1]. โ€ข If ๐‘ช is full dimensional (i.e., ๐‘š ๐‘— < ๐‘ฃ ๐‘— , ๐‘— = 1, โ€ฆ , ๐‘œ ), this is equivalent to ๐›ผ 2 ๐‘” ๐‘ฆ โ‰ฝ 0, โˆ€๐‘ฆ โˆˆ ๐ถ. 2

  3. Complexity questions Goal: study the complexity of testing convexity of a function over a box โ€ข Restrict ourselves to polynomial functions. โ€ข Related work: Theorem [Ahmadi, Olshevsky, Parrilo, Tsitsiklis] It is strongly NP-hard to test (global) convexity of polynomials of degree 4. โ€ข One may hope that adding the restriction to a box could make things easier. 3

  4. 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 โ€ข Nonconvex optimization: branch-and-bound โ€ข Control theory: convex Lyapunov functions [Ahmadi and Jungers] [Chesi and Hung] โ€ข Prior work: โ€ข Sufficient conditions for convexity [Orban et โ€ข Statistics: convex regression al.], [Grant et al.] โ€ข In practice, BARON, CVX, Gurobi check convexity of quadratics and computationally tractable sufficient conditions for convexity 4

  5. 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 ๐›ผ 2 ๐‘”(๐‘ฆ) is a matrix Question: What to Idea: A cubic polynomial ๐‘” is convex do a reduction over a (full-dimensional) box ๐ถ if and with entries affine only if ๐›ผ 2 ๐‘” ๐‘ฆ โ‰ฝ 0 , โˆ€๐‘ฆ โˆˆ ๐ถ from? in ๐’š Theorem [Nemirovski]: Let ๐‘€(๐‘ฆ) be a matrix with entries affine in ๐‘ฆ . It is NP-hard to test whether ๐‘€ ๐‘ฆ โ‰ฝ 0 for all ๐‘ฆ in a full-dimensional box ๐ถ. 5

  6. Are we done? No! Issue 1: We want to show strong NP-hardness. Nemirovskiโ€™s result shows weak NP- hardness. Issue 2: Not every affine polynomial matrix is a valid Hessian! 10 2๐‘ฆ 1 + 1 ๐œ–๐‘€ 11 (๐‘ฆ) ๐œ–๐‘€ 12 (๐‘ฆ) Example: ๐‘€ ๐‘ฆ 1 , ๐‘ฆ 2 = . We have โ‰  ๐œ–๐‘ฆ 1 . 2๐‘ฆ 1 + 1 10 ๐œ–๐‘ฆ 2 6

  7. Dealing with Issue 1 (1/5) Reminder: weak vs strong NP-hardness โ€ข Distinction only concerns problems where input is numerical โ€ข Max(I) : largest number in magnitude that appears in the input of instance I (numerator or denominator) โ€ข Length(I) : number of bits it takes to write down input of instance I Strong Weak โ€ข There are instances ๐ฝ that are hard with โ€ข The instances that are hard may contain Max( ๐ฝ ) โ‰ค ๐‘ž (Length( ๐ฝ )) ( ๐‘ž is a polynomial) numbers of large magnitude (e.g., 2 ๐‘œ ) . โ€ข No pseudo-polynomial algorithm possible โ€ข Pseudo-polynomial algorithms possible โ€ข Examples: โ€ข Examples: Knapsack Sat Partition Max-Cut 7

  8. Dealing with Issue 1 (2/5) Theorem [Nemirovski]: INTERVAL-PSDNESS Let ๐‘€(๐‘ฆ) be a matrix with entries affine in ๐‘ฆ . It is (weakly) NP-hard to test whether ๐‘€ ๐‘ฆ โ‰ฝ 0 for all ๐‘ฆ in a full-dimensional box ๐ถ. Why weakly NP-hard? INTERVAL PSDNESS PARTITION: Construct : ๐ท = ๐ฝ ๐‘œ โˆ’ ๐‘๐‘ ๐‘ˆ โˆ’1 , REDUCTION Input : ๐‘ โˆˆ โ„ ๐‘œ such that ๐‘ 2 โ‰ค 0.1 ๐œˆ = ๐‘œ โˆ’ ๐‘’ โˆ’2 ๐‘ , where ๐‘’ ๐‘ = smallest cd of ๐‘. ๐ท ๐‘ฆ Take: ๐ถ = โˆ’1,1 ๐‘œ and ๐‘€ ๐‘ฆ = Test: does there exist ๐‘ข โˆˆ โˆ’1,1 ๐‘œ ๐œˆ . ๐‘ฆ ๐‘ˆ such that ฯƒ ๐‘— ๐‘ ๐‘— ๐‘ข ๐‘— = 0? Test: Is ๐‘€ ๐‘ฆ โ‰ฝ 0 โˆ€๐‘ฆ โˆˆ ๐ถ? 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 โˆ’1 1 0 0 ๐‘ 1 ๐‘ 2 ๐‘ 4 but one of the entries of ๐ต โˆ’1 is 2 ๐‘œโˆ’2 ! ๐‘ 10 = Example: ๐ต = ๐‘ 8 โ‹ฑ โ‹ฑ โ‹ฎ 0 ๐‘ 3 8 โˆ’1 โˆ’1 โˆ’1 1

  9. Dealing with Issue 1 (3/5) Theorem [Ahmadi, H.]: INTERVAL-PSDNESS Let ๐‘€(๐‘ฆ) be a matrix with entries affine in ๐‘ฆ . It is strongly NP-hard to test whether ๐‘€ ๐‘ฆ โ‰ฝ 0 for all ๐‘ฆ in a full-dimensional box ๐ถ. INTERVAL PSDNESS MAX-CUT: 1 Construct : ๐›ฝ = ๐‘œ+1 3 , ๐ท = 4๐›ฝ(๐ฝ ๐‘œ + ๐›ฝ๐ต) ๐œˆ = ๐‘œ 4๐›ฝ + ๐‘™ โˆ’ 1 โˆ’ 1 Input : simple graph G=(V,E) with 4 ๐‘“ ๐‘ˆ ๐ต๐‘“ REDUCTION ๐‘Š = ๐‘œ and adj. matrix A, and a ๐ท ๐‘ฆ positive integer ๐‘™ โ‰ค ๐‘œ 2 Take: ๐ถ = โˆ’1,1 ๐‘œ and ๐‘€ ๐‘ฆ = Preserves strong ๐œˆ . ๐‘ฆ ๐‘ˆ NP-hardness Test: does there exist a cut in the Test: Is ๐‘€ ๐‘ฆ โ‰ฝ 0 โˆ€๐‘ฆ โˆˆ ๐ถ? graph of size greater or equal to ๐‘™? Show: No to MAX-CUT โ‡” Yes to INTERVAL PSDNESS Taylor series of 4๐›ฝ ๐ฝ โˆ’ ๐›ฝ๐ต โˆ’1 truncated at the first term Strongly NP-hard Scaling needed so that ๐ฝ ๐‘œ โˆ’ ๐›ฝ๐ต โˆ’1 โ‰ˆ ๐ฝ ๐‘œ + ๐›ฝ๐ต 9

  10. Dealing with Issue 1 (4/5) In more detail: No to MAX-CUT โ‡’ Yes to INTERVAL PSDNESS 1 4 ฯƒ ๐‘—,๐‘˜ ๐ต ๐‘—๐‘˜ (1 โˆ’ ๐‘ฆ ๐‘— ๐‘ฆ ๐‘˜ )] โ‰ค ๐‘™ โˆ’ 1 [max โ‡” No cut in ๐ป of size โ‰ฅ ๐‘™ ๐‘ฆโˆˆ โˆ’1,1 ๐‘œ โ‡” Convex Size of largest cut in ๐ป ๐‘œ ๐‘œ+1 3 1 1 1 4 ๐‘ฆ ๐‘ˆ ๐ต๐‘ฆ ] โ‰ค โˆ’ 4 ๐‘“ ๐‘ˆ ๐ต๐‘“ + ๐‘™ โˆ’ 1 4 ๐‘ฆ ๐‘ˆ ๐‘œ + 1 3 ๐ฝ ๐‘œ โˆ’ ๐ต ๐‘ฆ ] โ‰ค โ‡” ๐‘ฆโˆˆ โˆ’1,1 ๐‘œ โˆ’ max max โˆ’ [ [ ๐‘ฆโˆˆ โˆ’1,1 ๐‘œ 4 1 4 ๐‘“ ๐‘ˆ ๐ต๐‘“ + ๐‘™ โˆ’ 1 โ‰” ๐œˆ โ‡” ๐›ฝ = ๐‘œ + 1 3 1 4 ๐‘ฆ ๐‘ˆ ๐›ฝ๐ฝ ๐‘œ โˆ’ ๐ต ๐‘ฆ ] โ‰ค ๐œˆ 1 4 ๐‘ฆ ๐‘ˆ ๐›ฝ๐ฝ ๐‘œ โˆ’ ๐ต ๐‘ฆ โ‰ค ๐œˆ , โˆ€๐‘ฆ โˆˆ โˆ’1,1 ๐‘œ โ‡” [ max ๐‘ฆโˆˆ[โˆ’1,1] ๐‘œ Approximation ๐ท โˆ’1 โ‰ˆ 1 4 (๐›ฝ๐ฝ โˆ’ ๐ต) โ‡’ ๐ท ๐‘ฆ 1 โ‰ฝ 0, โˆ€๐‘ฆ โˆˆ โˆ’1,1 ๐‘œ ๐‘ฆ ๐‘ˆ ๐ท โˆ’1 ๐‘ฆ โ‰ค ๐œˆ + 4 , โˆ€๐‘ฆ โˆˆ โˆ’1,1 ๐‘œ โ‡’ ๐‘€ ๐‘ฆ = 1 ๐‘ฆ ๐‘ˆ ๐œˆ + Schur 4 Approximation error complement 10

  11. Dealing with Issue 1 (5/5) For converse: Yes to MAX-CUT โ‡’ No to INTERVAL PSDNESS There is a cut of size โ‰ฅ ๐‘™: ๐‘ฆ โ‰ฅ ๐œˆ + 3 4 > ๐œˆ + 1 Similar steps ๐‘ฆ ๐‘ˆ ๐ท โˆ’1 เทœ โ‡’ โ‡’ เทœ ๐‘ฆ ๐‘— = แ‰Š 1 if node ๐‘— on one side of cut Let เทœ to previously 4 โˆ’1 if node ๐‘— on other side of cut โ‡’ ๐‘ฆ โˆˆ โˆ’1,1 ๐‘œ s.t. ๐‘€ เทœ โˆƒ เทœ ๐‘ฆ 0 Corollary [Ahmadi, H.]: Let ๐‘œ be an integer and let เทœ ๐‘Ÿ ๐‘—๐‘˜ , เดค ๐‘Ÿ ๐‘—๐‘˜ be rational numbers with เทœ ๐‘Ÿ ๐‘—๐‘˜ โ‰ค เดค ๐‘Ÿ ๐‘—๐‘˜ and เทœ ๐‘Ÿ ๐‘—๐‘˜ = เทœ ๐‘Ÿ ๐‘˜๐‘— and เดค ๐‘Ÿ ๐‘—๐‘˜ = เดค ๐‘Ÿ ๐‘˜๐‘— for all ๐‘— = 1, โ€ฆ , ๐‘œ and ๐‘˜ = 1, โ€ฆ , ๐‘œ. It is strongly NP-hard to test whether all symmetric matrices with entries in [เทœ ๐‘Ÿ ๐‘—๐‘˜ ; เดค ๐‘Ÿ ๐‘—๐‘˜ ] are positive semidefinite. โ€ข Initial problem studied by Nemirovski โ€ข Of independent interest in robust control 11

  12. 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‘€(๐‘ฆ) ? Input : ๐‘€ ๐‘ฆ , เท  Idea: Want ๐›ผ 2 ๐‘” ๐‘ฆ = ๐‘€ ๐‘ฆ . ๐ถ Test: Is ๐‘€ ๐‘ฆ โ‰ฝ 0, โˆ€๐‘ฆ โˆˆ เท  ๐ถ? Issue: Not all ๐‘€(๐‘ฆ) are valid Hessians! Key ideas for the construction of ๐’ˆ: ๐Ÿ ๐Ÿ‘ ๐’› ๐‘ผ ๐‘ด ๐’š ๐’› โ€ข Start with ๐’ˆ ๐’š, ๐’› = 1 1 1 โ€ข For ๐›ผ 2 ๐‘” ๐‘ฆ, ๐‘ง to be able to be psd when ๐‘€ ๐‘ฆ โ‰ฝ 0 , we need to have ๐œท๐‘ฑ ๐’ ๐œท๐‘ฑ ๐’ 0 2 ๐ผ(๐‘ง) 2 ๐ผ(๐‘ง) 2 ๐ผ(๐‘ง) ๐›ผ 2 ๐‘” ๐‘ฆ, ๐‘ง = ๐›ผ 2 ๐‘” ๐‘ฆ, ๐‘ง = ๐›ผ 2 ๐‘” ๐‘ฆ, ๐‘ง = ๐œท ๐Ÿ‘ ๐’š ๐‘ผ ๐’š to ๐‘” ๐‘ฆ, ๐‘ง . 1 1 1 a nonzero diagonal: add 2 ๐ผ ๐‘ง ๐‘ˆ 2 ๐ผ ๐‘ง ๐‘ˆ 2 ๐ผ ๐‘ง ๐‘ˆ ๐‘€ ๐‘ฆ + ๐œƒ๐ฝ ๐‘œ+1 ๐‘€ ๐‘ฆ ๐‘€ ๐‘ฆ โ€ข ๐‘€ ๐‘ฆ and ๐ผ(๐‘ง) do not depend on the same variable: what if โˆƒ(๐‘ฆ, ๐‘ง) s.t. ๐‘€ ๐‘ฆ = 0 but ๐ผ ๐‘ง is not? The matrix cannot be psd: add ๐œƒ 2 ๐‘ง ๐‘ˆ ๐‘ง to ๐‘” ๐‘ฆ, ๐‘ง . โ‡’ ๐‘” ๐‘ฆ = 1 2 ๐‘ง ๐‘ˆ ๐‘€ ๐‘ฆ ๐‘ง + ๐›ฝ 2 ๐‘ฆ ๐‘ˆ ๐‘ฆ + ๐œƒ 2 ๐‘ง ๐‘ˆ ๐‘ง, ๐ถ = โˆ’1,1 2๐‘œ+1 12

  13. าง Dealing with Issue 2 (2/3) Show NO to INTERVAL PSDNESS โ‡’ NO to CONV3BOX. This is equivalent to: ๐‘ง โˆˆ โˆ’1,1 2๐‘œ+1 , ๐‘จ s.t. ๐‘จ ๐‘ˆ ๐›ผ 2 ๐‘” เทœ ๐‘ฆ โˆˆ โˆ’1,1 ๐‘œ s.t. ๐‘€ โˆƒ าง ๐‘ฆ โ‰ฝ 0 โ‡’ โˆƒ เทœ ๐‘ฆ, เทœ ๐‘ฆ, เทœ ๐‘ง ๐‘จ < 0 ๐ท ๐‘ฆ Need to leverage extra structure of ๐‘€ ๐‘ฆ : ๐‘€ ๐‘ฆ = 1 ๐‘ฆ ๐‘ˆ ๐œˆ + 4 0 ๐Ÿ 0 โˆ’๐ท โˆ’1 าง โˆ’๐ท โˆ’1 าง โˆ’๐‘ซ โˆ’๐Ÿ เดฅ Candidates: เท Candidates : เทœ Candidates: เทœ ๐’š = เดฅ ๐‘ฆ = าง ๐‘ฆ = าง ๐‘ฆ, ๐‘ฆ, ๐’š, ๐‘ง = 0, ๐‘ง = 0, ๐’› = ๐Ÿ, เท เทœ เทœ ๐’œ = ๐‘จ = ๐‘จ = ๐‘ฆ ๐’š ๐‘ฆ 1 ๐Ÿ 1 ๐›ฝ๐ฝ ๐‘œ ๐Ÿ ๐Ÿ ๐‘ˆ ๐›ฝ๐ฝ ๐‘œ ๐›ฝ๐ฝ ๐‘œ ๐Ÿ ๐ผ(๐‘ง) ๐Ÿ 0 0 = ๐œˆ + 1 ๐Ÿ ๐‘ซ + ๐œƒ๐ฝ ๐‘œ ๐’š เดฅ 2 ) ๐‘ฆ ๐‘ˆ ๐ท โˆ’1 าง ๐‘ฆ + ๐œƒ(1 + ๐ท โˆ’1 าง โˆ’๐ท โˆ’1 าง โˆ’๐ท โˆ’1 าง ๐‘จ ๐‘ˆ ๐›ผ 2 ๐‘” เทœ ๐‘ฆ, เทœ ๐‘ง ๐‘จ = ๐Ÿ ๐‘ซ + ๐œƒ๐ฝ ๐‘œ ๐ท + ๐œƒ๐ฝ ๐‘œ เดฅ ๐’š ๐‘ฆ 4 โˆ’ าง ๐‘ฆ 2 ๐‘ฆ ๐‘ฆ ๐›ผ 2 ๐‘” ๐‘ฆ, ๐‘ง = ๐›ผ 2 ๐‘” เทœ ๐‚ + ๐Ÿ ๐‘ฆ, เทœ ๐‘ง = ๐ผ ๐‘ง ๐‘ˆ ๐’š ๐‘ผ ๐‚ + ๐Ÿ ๐œˆ + 1 1 1 ๐Ÿ เดฅ ๐Ÿ“ + ๐œƒ ๐’š ๐‘ผ ๐‘ฆ ๐‘ˆ ๐Ÿ เดฅ ๐Ÿ“ + ๐œƒ 4 + ๐œƒ Appropriately scaled so that < ๐Ÿ as ๐‘ด เดฅ ๐’š โ‰ฝ ๐Ÿ ๐‘จ ๐‘ˆ ๐›ผ 2 ๐‘” เทœ ๐‘ฆ, เทœ ๐‘ง ๐‘จ remains <0. 13

Recommend


More recommend