Primal-Dual Geometry of Level Sets and their Explanatory Value of the Practical Performance of Interior-Point Methods for Conic Optimization

Transcription

1 Primal-Dual Geometry of Level Sets and their Explanatory Value of the Practical Performance of Interior-Point Methods for Conic Optimization Robert M. Freund M.I.T. June, 2010 from papers in SIOPT, Mathematics of Operations Research, and Mathematical Programming 1

2 Outline Primal-Dual Geometry of Level Sets in Conic Optimization A Geometric Measure of Feasible Regions and Interior-Point Method (IPM) Complexity Theory Geometric Measures and their Explanatory Value of the Practical Performance of IPMs 2

3 Primal and Dual Linear Optimization Problems P : VAL := min x c T x D : VAL := max y,z b T y s.t. Ax = b s.t. A T y + z = c x 0 z 0 A IR m n x 0 is x R n + 3

4 Primal and Dual Conic Problem P : VAL := min x c T x D : VAL := max y,z b T y s.t. Ax = b s.t. A T y + z = c x C z C C X is a regular cone: closed, convex, pointed, with nonempty interior C := {z : z T x 0 x C} 4

5 The Semidefinite Cone S k denotes the set of symmetric k k matrices S+ k denotes the set of positive semi-definite k k symmetric matrices S k ++ denotes the set of positive definite k k symmetric matrices X 0 denotes that X is symmetric positive semi-definite X Y denotes that X Y 0 X 0 to denote that X is symmetric positive definite, etc. Remark: S+ k = {X Sk X 0} is a regular convex cone. Furthermore, (S+ k ) = S+ k, i.e., Sk + is self-dual. 5

6 Primal and Dual Semidefinite Optimization P : VAL := min x c T x D : VAL := max y,z b T y s.t. Ax = b s.t. A T y + z = c x S+ k z S+ k This is slightly awkward as we are not used to conceptualizing x as a matrix. 6

7 Primal and Dual Semidefinite Optimization, again min X C X max y,z mi=1 y i b i s.t. A i X = b i, i = 1,..., m s.t. mi=1 y i A i + Z = C X 0 Z 0 Here C X := k k i=1 j=1 C ij X ij is the trace inner product, since C X = trace(c T X) 7

8 Back to Linear Optimization P : VAL := min x c T x D : VAL := max y,z b T y s.t. Ax = b s.t. A T y + z = c x 0 z 0 8

9 Five Meta-Lessons from Interior-Point Theory/Methods 1. Linear optimization is not much more special than conic convex optimization 2. A problem is ill-conditioned if VAL is finite but the primal or dual objective function level sets are unbounded 3. ε-optimal solutions are important objects in their own right 4. Choice of norm is important; some norms are more natural for certain settings 9

10 Meta-Lessons from Interior-Point Theory/Methods, continued 5. All the important activity is in the (regular) cones Indeed, we could eliminate the y-variable and re-write P and D as: P : min x c T x D : VAL := min z (x 0 ) T z s.t. x x 0 L s.t. z c L x 0 z 0 where x 0 satisfies Ax 0 = b and L = null(a). But we won t. 10

11 Primal and Dual Near-Optimal Level Sets P : VAL := min x c T x D : VAL := max y,z b T y s.t. Ax = b s.t. A T y + z = c x 0 z 0 P ε := {x : Ax = b, x 0, c T x VAL + ε} D δ := {z : y satisfying A T y + z = c, z 0, b T y VAL δ} 11

12 Level Set Geometry Measures Let e := (1, 1,..., 1) T. Define for ε, δ > 0: R P ε := max x x 1 r D δ := max y,z,r r s.t. Ax = b s.t. A T y + z = c x 0 z 0 c T x VAL + ε b T y VAL δ z r e R P ε is the norm of the largest primal ε-optimal solution rδ D measures the largest distance to the boundary of R n + among all dual δ-optimal solutions z 12

13 Level Set Geometry Measures, continued R P ε := max x x 1 r D δ := max z,r r s.t. x P ε s.t. z D δ z r e P ε := {x : Ax = b, x 0, c T x VAL + ε} D δ := {z : y satisfying A T y + z = c, z 0, b T y VAL δ} 13

14 R P ε Measures Large Near-Optimal Solutions 14

15 R P ε Measures Large Near-Optimal Solutions 15

16 r D δ Measures Nicely Interior Near-Optimal Solutions 16

17 r D δ Measures Nicely Interior Near-Optimal Solutions 17

18 r D δ Measures Nicely Interior Near-Optimal Solutions 18

19 Main Result: R P ε and r D δ are Reciprocally Related R P ε := max x x 1 r D δ := max z,r r s.t. x P ε s.t. z D δ z r e Main Theorem: Suppose VAL is finite. If R P ε is positive and finite, then min{ε, δ} R P ε r D δ ε + δ. Otherwise {R P ε, r D δ } = {, 0}. 19

20 Comments min{ε, δ} R P ε r D δ ε + δ R P ε, r D δ each involves primal and dual information each inequality can be tight (and cannot be improved) setting δ = ε, we obtain ε Rε P rε D 2ε, showing these two measures are inversely proportional (to within a factor of 2) 20

21 Comments, continued min{ε, δ} R P ε r D δ ε + δ exchanging the roles of P and D how to prove 21

22 Comments, continued R P ε := max x x 1 r D δ := max z,r r s.t. x P ε s.t. z D δ z r e ε R P ε r D ε 2ε The maximum norms of the primal objective level sets are almost exactly inversely proportional to the maximum distances to the boundary of the dual objective level sets 22

23 Relation to LP Non-Regularity Property Standard LP Non-Regularity Property: If VAL is finite, the set of primal optimal solutions is unbounded iff every dual feasible z lies in the boundary of R n +. R P ε := max x x 1 r D δ := max z,r r s.t. x P ε s.t. z D δ z r e In our notation, this is Rε P = iff rd δ part of the Main Theorem = 0, which is the second 23

24 Relation to LP Non-Regularity, continued R P ε := max x x 1 r D δ := max z,r r s.t. x P ε s.t. z D δ z r e The first part of the main theorem is: if Rε P then is finite and positive, min{ε, δ} R P ε r D δ ε + δ This then is a generalization to nearly-non-regular problems, where R P ε is finite and r D δ is non-zero 24

25 Question about Main Result R P ε := max x x 1 r D δ := max z,r r s.t. x P ε s.t. z D δ z r e Q: Why the 1 norm? A: Because f(x) := x 1 is a linear function on the cone R n +. The linearity gives R P ε nice properties. If is not linear on R n + then we have to slightly weaken the main theorem as we will see

26 Primal and Dual Conic Problem P : VAL := min x c T x D : VAL := max y,z b T y s.t. Ax = b s.t. A T y + z = c x C z C C X is a regular cone: closed, convex, pointed, with nonempty interior C := {z : z T x 0 x C} 26

27 Primal and Dual Level Sets P : VAL := min x c T x D : VAL := max y,z b T y s.t. Ax = b s.t. A T y + z = c x C z C P ε := {x : Ax = b, x C, c T x VAL + ε} D δ := {z : y satisfying A T y + z = c, z C, b T y VAL δ} 27

28 Level Set Geometry Measures Fix a norm x for the space of the x variables. The dual norm is z := max{z T x : x 1} for the z variables. Define for ε, δ > 0: R P ε := max x x r D δ := max y,z dist (z, C ) s.t. Ax = b s.t. A T y + z = c x C z C c T x VAL + ε b T y VAL δ dist (z, C ) denotes the distance from z to C in the dual norm z 28

29 Level Set Geometry Measures, continued Rε P := max x x rδ D := max y,z dist (z, C ) s.t. Ax = b s.t. A T y + z = c x C z C c T x VAL + ε b T y VAL δ R P ε is the norm of the largest primal ε-optimal solution rδ D measures the largest distance to the boundary of C among all dual δ-optimal solutions z 29

30 Level Set Geometry Measures, continued R P ε := max x x r D δ := max z dist (z, C ) s.t. x P ε s.t. z D δ P ε := {x : Ax = b, x C, c T x VAL + ε} D δ := {z : y satisfying A T y + z = c, z C, b T y VAL δ} 30

31 R P ε Measures Large Near-Optimal Solutions 31

32 R P ε Measures Large Near-Optimal Solutions 32

33 r D δ Measures Nicely Interior Near-Optimal Solutions 33

34 r D δ Measures Nicely Interior Near-Optimal Solutions 34

35 r D δ Measures Nicely Interior Near-Optimal Solutions 35

36 Main Result, Again: R P ε and r D δ are Reciprocally Related R P ε := max x x r D δ := max z dist (z, C ) s.t. x P ε s.t. z D δ Main Theorem: Suppose VAL is finite. If Rε P finite, then τ C min{ε, δ} R P ε r D δ ε + δ. is positive and If R P ε = 0, then r D δ =. If R P ε = and VAL is finite, then r D δ = 0. Here τ C denotes the width of the cone C

37 On the Width of a Cone Let K be a convex cone with nonempty interior τ K := max{dist(x, K) : x K, x 1} x If K is a regular cone, then τ K (0, 1] τ K generalizes Goffin s inner measure 37

38 A Cone with small Width τ K τ K << 1 38

39 Equivalence of Norm Linearity and Width of Polar Cone Proposition: Let K be a regular cone. The following statements are equivalent: τ K α, and there exists w for which α w T x x w T x for all x K Corollary: τ K = 1 implies f(x) := x is linear on K 39

40 Main Result, Again: R P ε and r D δ are Reciprocally Related R P ε := max x x r D δ := max z dist (z, C ) s.t. x P ε s.t. z D δ Main Theorem: Suppose VAL is finite. If Rε P finite, then is positive and τ C min{ε, δ} R P ε r D δ ε + δ. If R P ε = 0, then r D δ =. If R P ε = and VAL is finite, then r D δ = 0. 40

41 Comments τ C min{ε, δ} R P ε r D δ ε + δ R P ε, r D δ each involves primal and dual information each inequality can be tight (and cannot be improved) many naturally arising norms have τ C = 1 setting δ = ε, we obtain ε Rε P rε D 2ε, showing these two measures are inversely proportional (to within a factor of 2) 41

42 Application: Robust Optimization [J.Vera] Amended format: P : z (b) := max x c T x D : min y b T y s.t. b Ax K s.t. A T y = c y K For a given tolerance ε > 0, what is the limit on the size of a perturbation b so that z (b + b) z (b) ε? 42

43 Application: Robust Optimization, continued P : z (b) := max x c T x D : min y b T y s.t. b Ax K s.t. A T y = c y K Theorem [Vera]: Let ε > 0 and b satisfy: ( ε b τ K Rε D Then z (b + b) z (b) ε. ). The result says that τ K ε/rε D is the required bound on the perturbation of the RHS needed to guarantee a change of no more than ε in the value of the problem. 43

44 τ K for Self-Scaled Cones Nonnegative Orthant: K = K = IR n +, define x p := p nj=1 x j p Then τ K = n (1/p 1), whereby τ K = 1 for p = 1 Semidefinite Cone: K = K = S k +, Define X p := λ(x) p := p kj=1 λ j (X) p Then τ K = k (1/p 1), whereby τ K = 1 for p = 1 Second-Order Cone: K = K = {x R n : (x 1,..., x n 1 ) 2 x n } Define x := max{ (x 1,..., x n 1 ) 2, x n }, then τ K = 1 44

45 Some Relations with Renegar s Condition Number For ε c it holds that: Rε P C 2 (d) + C(d) ε c r P ε ετ C 3 c (C 2 (d) + C(d)) 45

46 Outline, again Primal-Dual Geometry of Level Sets in Conic Optimization A Geometric Measure of Feasible Regions and Interior-Point Method (IPM) Complexity Theory Geometric Measures and their Explanatory Value of the Practical Performance of IPMs 46

47 Primal and Dual Conic Problem P : VAL := min x c T x D : VAL := max y,z b T y s.t. Ax = b s.t. A T y + z = c x C z C C X is a regular cone: closed, convex, pointed, with nonempty interior C := {z : z T x 0 x C} 47

48 Geometric Measure of Primal Feasible Region G P := minimum x s.t. max { Ax = b x C x dist(x, C), x, 1 dist(x, C) } G P is smaller to the extent that there is a feasible solution that is not too large and that is not too close to C G P is smaller if the primal has a well-conditioned feasible solution 48

49 Geometric Complexity Theory of Conic Optimization G P := minimum x s.t. max { Ax = b x C x dist(x, C), x, 1 dist(x, C) } [F 04] Using a (theoretical) interior-point method that solves a primal-phase-i followed by a primal-phase-ii, one can bound the IPM iterations to compute an ε-optimal solution by ( ( ) ( O ( ϑ C ln R P ε + ln G P ) + ln(1/ε) )) 49

50 Geometric Complexity Theory of Conic Optimization Computational complexity of solving primal problem P is: ( ( ) ( O ( ϑ C ln R P ε + ln G P ) + ln(1/ε) )) Is this just a pretty theory? Are R P ε and G P correlated with the performance of IPMs on conic problems in practice, say from the SDPLIB suite of SDP problems? IPMs in practice are interchangeable insofar as role of primal versus dual. Therefore let us replicate the above theory for the dual problem. 50

51 Geometric Measure of Dual Feasible Region G D := minimum y,z s.t. max { z dist (z, C ), z, A T y + z = c z C 1 dist (z, C ) } G D is smaller to the extent that there is a feasible dual solution that is not too large and that is not too close to C G D is smaller if the dual has a well-conditioned feasible solution 51

52 Geometric Complexity Theory of Conic Optimization G D := minimum z s.t. max { z dist (z, C ), z, A T y + z = c z C 1 dist (z, C ) } [F 04] Using a (theoretical) interior-point method that solves a dual-phase-i followed by a dual-phase-ii, one can bound the IPM iterations to compute an ε-optimal solution by ( ( ) ( O ( ϑ C ln R D ε + ln G D ) + ln(1/ε) )) 52

53 Aggregate Geometry Measure for Primal and Dual Define the aggregate geometry measure: G A := ( R P ε R D ε G P G D) 1/4 G A aggregates the primal and dual level set measures and the primal and dual feasible region geometry measures Let us compute G A for the SDPLIB suite and see if G A is correlated with IPM iterations. 53

54 Outline, again Primal-Dual Geometry of Level Sets in Conic Optimization A Geometric Measure of Feasible Regions and Interior-Point Method (IPM) Complexity Theory Geometric Measures and their Explanatory Value of the Practical Performance of IPMs 54

55 Semi-Definite Programming (SDP) broad generalization of LP emerged in early 1990 s as the most significant computationally tractable generalization of LP independently discovered by Alizadeh and Nesterov-Nemirovskii applications of SDP are vast, encompassing such diverse areas as integer programming and control theory 55

56 Partial List of Applications of SDP LP, Convex QP, Convex QCQP tighter relaxations of IP ( 12% of optimality for MAXCUT) static structural (truss) design, dymamic truss design, antenna array filter design, other engineered systems problems control theory shape optimization, geometric design, volume optimization problems D-optimal experimental design, outlier identification, data mining, robust regression eigenvalue problems, matrix scaling/design sensor network localization optimization or near-optimization with large classes of non-convex polynomial constraints and objectives (Parrilo, Lasserre, SOS methods) robust optimization methods for standard LP, QCQP 56

57 IPM Set-up for SDP min X C X max mi=1 y,z y i b i s.t. A i X = b i, i = 1,..., m s.t. mi=1 y i A i + Z = C X 0 Z 0 57

58 IPM for SDP, Central Path Central path: X(µ) := argmin X C X µ n j=1 ln(λ j (X)) s.t. A i X = b i, i = 1,..., m X 0 Central path: X(µ) := argmin X C X µ ln(det(x)) s.t. A i X = b i, i = 1,..., m X 0 58

59 IPM for SDP, Central Path, continued Central path: X(µ) := argmin X C X + µ ln(det(x)) s.t. A i X = b i, i = 1,..., m X 0 59

60 IPM for SDP, Central Path, continued Central path: X(µ) := argmin X C X + µ ln(det(x)) s.t. A i X = b i, i = 1,..., m X 0 Optimality gap property of the central path: C X(µ) VAL n µ Algorithm strategy: trace the central path X(µ) for a decreasing sequence of values of µ 0 60

61 IPM Strategy for SDP 61

62 IPM for SDP: Computational Reality Alizadeh, Nesterov and Nemirovski - IPM theory for SDP software for SOCP, SDP iterations on SDPLIB suite, typically 30 iterations Each IPM iteration is expensive to solve: ( H(x k ) A T ) ( x A 0 y ) = ( r1 r 2 ) O(n 6 ) work per iteration, managing sparsity and numerical stability are tougher bottlenecks most IPM computational research since 1996 has focused on work per iteration, sparsity, numerical stability, lower-order methods, etc. 62

63 92 problems SDPLIB Suite standard equality block form, SDP variables and LP variables no linear dependent equations Work with 85 problems: removed 4 infeasible problems: infd1, infd2, infp1, infp2 removed 3 very large problems: maxg55 ( ), maxg60 ( ), thetag51 ( ) m : , n :

64 Histogram of IPM Iterations for SDPLIB problems Frequency IPM Iterations SDPT3-3.1 default settings used throughout SDPT3-3.1 solves the 85 SDPLIB problem instances in iterations 64

65 IPM Iterations versus n IPM Iterations n Scatter Plot of IPM iterations and n := n s + n l 65

66 IPM Iterations versus m IPM Iterations m Scatter Plot of IPM iterations and m 66

67 Computing the Aggregate Geometry Measure G A G A := ( R P ε R D ε G P G D) 1/4 are maximum norm problems, so are generally non- Rε P, RD ε convex. Computing G P, G D involves working with dist(x, C), dist (z, C ) which is not efficiently computable in general A judicious choice of norms allows us to compute all four quantities efficiently via one associated SDP for each quantity. 67

68 Geometry Measure Results G A was computed for 85 SDPLIB problems: G A Rε P Rε D G P G D Finite Infinite Total % of problems have finite G A The pattern in the table is no coincidence... G P = R D ε = and G D = R P ε = 68

69 60 50 IPM Iterations log(g m ) IPM iterations versus log(g A ) CORR ( log(g A ), IPM Iterations ) = (53 problems) 69

70 What About Other Behavioral Measures? How well does Renegar s condition measure C(d) explain the practical performance of IPMs on the SDPLIB suite? 70

71 C(d): Renegar s Condition Measure (P d ) : min c T x (D d ) : max b T y s.t. Ax = b s.t. A T y + z = c x C z C d = (A, b, c) is the data for the instance P d and D d d = max { A, b, c } 71

72 Distance to Primal and Dual Infeasibility (P d ) : min c T x (D d ) : max b T y s.t. Ax = b s.t. A T y + z = c x C z C Distance to primal infeasibility: ρ P (d) = min { d : P d+ d is infeasible } Distance to dual infeasibility: ρ D (d) = min { d : D d+ d is infeasible } The condition measure is: d C(d) = min{ρ P (d), ρ D (d)} 72

73 C(d): Renegar s Condition Measure C(d) = d min{ρ P (d), ρ D (d)} In theory, C(d) has been shown to be connected to: bounds on sizes of feasible solutions and aspect ratios of inscribed balls in feasible regions bounds on sizes of optimal solutions and objective values bounds on rates of deformation of feasible regions as data is modified bounds on deformation of optimal solutions as data is modified bounds on the complexity of a variety of algorithms 73

74 [Renegar 95] Using a (theoretical) IPM that solves a primalphase-i followed by a primal-phase-ii, one can bound IPM iterations needed to compute an ε-optimal solution by ( ) O ϑ C (ln (C(d)) + ln(1/ε)) ϑ C = n s + n l for SDP Is C(d) just really nice theory? Is C(d) correlated with IPM iterations among SDPLIB problems? 74

75 Condition Measure Results Computed C(d) for 80 (out of 85) problems: Unable to compute ρ p (d) for 5 problems: control11, equalg51, maxg32, theta6, thetag11 (m = 1596, 1001, 2000, 4375, 2401, respectively) 60% are well-posed 40% are almost primal infeasible ρ D (d) 0 > 0 Total ρ P (d) > Total

76 60 50 IPM Iterations log C(d) IPM iterations versus log (C(d)). CORR(log (C(d)), IPM Iterations) = (48 problems) 76

77 Some Conclusions 62% of 85 SDPLIB problems have finite aggregate geometry measure G A CORR(log ( G A), IPM Iterations) = among the SDPLIB problems with finite geometry measure G A 32 of 80 SDPLIB problems are almost primal infeasible, i.e. C(d) = + CORR(log (C(d)), IPM Iterations) = among the 42 problems with finite C(d) 77

78 60 50 IPM Iterations log(g m ) IPM iterations versus log(g A ) CORR ( log(g A ), IPM Iterations ) = (53 problems) 78