Facial Reduction and Geometry on Conic Programming

Transcription

1 Facial Reduction and Geometry on Conic Programming Masakazu Muramatsu UEC Bruno F. Lourenço, and Takashi Tsuchiya Seikei University GRIPS Based on the papers of LMT: 1. A structural geometrical analysis of weakly infeasible SDPs (Journal of Operations Research Society of Japan 2016) 2. Weak infeasibility in second order cone programming (Optimization Letters 2015) 3. Facial Reductions and Partial Polyhedrality (Under Review) 4. (Under preparation) NOTE: This talk is a `Random Walk within these works. 2016/8/13 WAO@Shinagawa

2 Contents 1. Conic Programming (CP) and Duality 2. Feasibility Statuses of CP 3. Facial Reduction Algorithm 4. Distance to Polyhedrality and FRA-Poly 5. Cone Expansion and Feasibility Transition Theorems 6. Nasty Problems and FRA dual D =sup{by : c A T y 2 K } $ P =inf{cx : Ax = b, x 2 K} primal/dual dual/primal

3 1. CP and Duality dual D =sup{by : c A T y 2 K } $ P =inf{cx : Ax = b, x 2 K} A = {c A T y : y 2 R m } A = {x : Ax = b} A \ K : Feasible Region

4 Conic Programming K K A A K : Closed Convex Cone A : Affine Subspace CP: Minimizing Linear Fn. over A \ K Example LP, SOCP, SDP,

5 Conic Programming K K A A def x 2 A \ relk, x is an interior feasible point. relative interior

6 Duality Theorem and Nasty Cases Duality Theorem in CP If an interior feasible point exists for Primal 1. Zero Duality Gap 2. Dual has an optimal solution. No interior feasible point 1. Positive Duality Gap 2. Optimal value may not be attained Hard to compute optimal value/solution Both Primal and Dual need interior feasible solutions to ensure existence of optimal solutions in both sides

7 Corruption of Computation r r Waki, Nakata, and M (2012) : Computed optimal values by SeDuMi of SDP relaxations for Polynomial Optimization indexed by relaxation order r =2, 3, 4, r Fact: The Optimal Value is zero for all r. One of the primal or dual does not have interior feasible solutions. Significant Difference

8 2. Feasibility Statuses of CP D =sup{by : c A T y 2 K } $ P =inf{cx : Ax = b, x 2 K} A = {c A T y : y 2 R m } A = {x : Ax = b} A \ K : Feasible Region

9 Four Feasibility Statuses of Conic LP I. K K A A Kmin : minimal cone A \ relk 6= ; A \ K 6= ;, buta \ relk = ; Strongly Feasible Weakly Feasible

10 Four Feasibility Statuses of Conic LP II. "!? See the next slide dist(a, K) > 0 dist(a, K) =0butA \ K = ; Strongly Infeasible Weakly Infeasible Impossible in LP

11 Weakly Infeasible CP K A K! A dist(a, K) =0butA \ K = ; : Weakly Infeasible

12 3.Facial Reduction Algorithm How to obtain a well-behaved problem

13 Facial Reduction Algorithm(FRA) 1%2#*+#/34%&45&6&)%7&'(%#)*&+,-+.)/#! K! A! K! A! Reducing Direction!! Weakly feasible instance "#$%#&'(%#)*&+,-+.)/#&-0&! Find w and take intersection of K and A. Repeat this until the `minimal cone is found

14 FRA Details D =sup{by : c A T y 2 K } $ P =inf{cx : Ax = b, x 2 K} FRA applied to D Iterate the following steps 1. Find a reducing direction w 2. Replace K in by K \ {w}? Properties D 1. The iteration number is bounded by the length of the longest chain of faces of K 2. When it stops, then we find either: - strongly feasible instance whose objective value is - strongly infeasible instance, showing is infeasible D D

15 Example: SOCP FRA K n = {(x 0, x) 2 R n : x 0 k xk} w : reducing direction If w 2 relk n, then K n \ {w}? = {0}.

16 Example: SOCP FRA K n = {(x 0, x) 2 R n : x 0 k xk} K n \ {w}? 1-dim cone (half line) w : reducing direction

17 Example: SOCP FRA K n 1 K n 2 K n m 1. One FRA iteration makes at least one SOC to polyhedral 2. At most m iteration is needed to obtain polyhedral cone - Enough to have a good property of duality FRA-Poly

18 4. Distance to Polyhedrality and FRA-Poly K n 1 K n 2 K n m

19 Distance to Polyhedrality Observation: No Nasty LPs even if not strongly feasible If we reach a polyhedral cone, we are happy. ( if needed, just one more FRA is enough to obtain a strongly feasible LP) 1. F 1 F l = K 2. F 1 is polyhedral 3. Others are non-polyhedral 4. Suppose that this chain is the longest one l 1 is called Distance to Polyhedrality

20 Partial Polyhedral Slater s (PPS) Condition K = K 1 K 2 where K 2 is polyhedral. CP satisfies PPS Condition, 9(x 1,x 2 ) 2 A \ K s.t. x 1 2 relk 1 Theorem. If CP satisfies PPS Condition, 1. No duality gap 2. Dual is attained.

21 FRA-Poly SOC PSD We can construct FRA in such a way to reduce non-polyhedral cone (FRA-Poly). Distance-to-polyhedrality is an upper bound of the number of iterations of FRA-Poly. Upper bounds of FRA predicted by the longest chain of Faces Distance to Polyhedrality 2 1 n+1 n DNN n(n+1)/2+1 n

22 5. Cone Expansion and Feasibility Transition Theorems

23 Cone Expansion K cl(k + l(w)) K l(w) Cone Expansion K 7! cl(k + l(w)) w: reducing direction of FRA applied to the dual program l(w) ={ w : 2 R} linear subspace spanned by w

24 Cone Expansion (CE) D =sup{by : c A T y 2 K } $ P =inf{cx : Ax = b, x 2 K} FRA Project the primal cone dual CE Expand the dual cone 0 D =sup{by : c A T y 2 K \ w? } $ 0 P =inf{cx : Ax = b, x 2 cl(k + l(w))} (Specially chosen w corresponds to Luo, Sturm, Zhang; Waki, M)

25 Example: SOCP CE 1. Dual of SOC is SOC (Self Dual) SOC is full-dimensional w : reducing direction If w 2 relk n, then K n + l(w) =R n

26 Example: SOCP CE 2. K n \ {w}? cl(k n + l(w)) is a half space above this hyperplane l(w) w: reducing direction The relation between SOC and w is close to weak infeasibility

27 FRA, CE and Feasibility Primal Problem Dual Problem D P FRA CE 0 D 0 P 1. FRA does not change the feasible region 2. Feasible region of P 0 could be larger than that of D 0 = D, P 0 apple P P FRA D = D 0 = D 1 =...= p D P = P 0 apple P 1 apple...apple p P CE

28 Feasibility Transition by FRA D Strongly Feasible Weakly Feasible Weakly Infeasible Strongly Infeasible FRA 0 D Strongly Feasible Weakly Feasible Weakly Infeasible Strongly Infeasible As long as the problem is in weak status, we can apply FRA. Final status: strongly feasible or infeasible instance

29 Feasibility Transition by CE LTM P Strongly Feasible Weakly Feasible Weakly Infeasible Strongly Infeasible CE 0 P Strongly Feasible Weakly Feasible Weakly Infeasible Strongly Infeasible As long as the problem weakly infeasible, we can apply CE. Final status: Feasible, or strongly infeasible instance.

30 Strongly Feasible but Non-attained problem D =sup{by : c A T y 2 K } $ P =inf{cx : Ax = b, x 2 K} Suppose that D is strongly feasible but not attained. Apply FRA to P to obtain the final problem p P (Equivalently, Apply CE to D to obtain p D) Strongly feasible P = 0 P D = 0 D FRA FRA FRA! P 1!! p P CE! 1 CE CE D!! p D Strongly feasible but not attained Strongly feasible and attained

31 6. Nasty Problems and FRA

32 Computing an approximate optimal solution Aim. Given > 0 find an feasible solution of whose obj. value > D D p D Since is strongly feasible by FTT for CE, 9ŷ, c A T ŷ 2 rel ˆK Let y be an optimal solution of p D the cone of p D It is easy to compute a feasible solution of p D whose obj. value > D using the above. Let ŷ be such a solution.

33 How to compute an approximate optimal solution Let w 1,...,w p be the reducing directions. There exists positive numbers 1,..., p such that px c A T ŷ + i w i 2 K. i=1 Cone Expansion w K 7! cl(k + l(w)) c A T ŷ

34 Properties of Reducing Direction Let w 1,...,w p 2 K be reducing directions of FRA applied to D. (p apple n) If P is weakly infeasible, then (c + span(w 1,...,w p )) \ K is also weakly infeasible. A directions approaching the cone In case of SOCP or SDP, given a positive number, we can explicitly compute a point on A whose distance from the cone is less than

35 Misleading Picture of Weak Infeasibility K A K! A We need p>0 directions to approach K in general. These directions are `reducing directions.

36 Thank you and Happy Birthday, Mizuno Sensei The papers by Lourenço, M. and Tsuchiya: 1. A structural geometrical analysis of weakly infeasible SDPs (Journal of Operations Research Society of Japan 2016) 2. Weak infeasibility in second order cone programming (Optimization Letters 2015) 3. Facial Reductions and Partial Polyhedrality (Under Review) 4. (Under preparation)

37 Example: FRA and CE on SDP w = O O O : reducing direction FRA CE S n + \ {w}? = O O O cl(s n + + l(w)) = NOTE: The resulting problems are again SDP