Approximate Optimal Designs for Multivariate Polynomial Regression

Transcription

1 Approximate Optimal Designs for Multivariate Polynomial Regression Fabrice Gamboa Collaboration with: Yohan de Castro, Didier Henrion, Roxana Hess, Jean-Bernard Lasserre Universität Potsdam 16th of February 2018

2 Short bibliography Convex design Polynomial regression Moment spaces Moment formulation SDP relaxation Examples Agenda 1 Short bibliography 2 Convex design 3 Polynomial regression 4 Moment spaces 5 Moment formulation 6 SDP relaxation 7 Examples F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

3 Overview 1 Short bibliography 2 Convex design 3 Polynomial regression 4 Moment spaces 5 Moment formulation 6 SDP relaxation 7 Examples F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

4 Short bibliography : Optimal design H. Dette and W. J. Studden. The theory of canonical moments with applications in statistics, probability, and analysis, volume 338. John Wiley & Sons, J. Kiefer. General equivalence theory for optimum designs (approximate theory). The annals of Statistics, pages , I. Molchanov and S. Zuyev. Optimisation in space of measures and optimal design. ESAIM : Probability and Statistics, 8 :12 24, F. Pukelsheim. Optimal design of experiments. SIAM, G. Sagnol and R. Harman. Computing exact D-optimal designs by mixed integer second-order cone programming. The Annals of Statistics, 43(5) : , F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

5 Short bibliography : Optimization J. B. Lasserre. Moments, positive polynomials and their applications, volume 1 of Imperial College Press Optimization Series. Imperial College Press, London, J. B. Lasserre and T. Netzer. SOS approximations of nonnegative polynomials via simple high degree perturbations. Mathematische Zeitschrift, 256(1) :99 112, A. S. Lewis. Convex analysis on the Hermitian matrices. SIAM Journal on Optimization, 6(1) : , L. Vandenberghe, S. Boyd, and S.-P. Wu. Determinant maximization with linear matrix inequality constraints. SIAM journal on matrix analysis and applications, 19(2) : , F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

6 Overview 1 Short bibliography 2 Convex design 3 Polynomial regression 4 Moment spaces 5 Moment formulation 6 SDP relaxation 7 Examples F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

7 Classical linear model What are you dealing with : Regression model F : X R n R p, continuous. Noisy linear model observed at t i X, i = 1,..., N z i = θ, F(t i ) + ε i. θ R p unknown, ε second order homoscedastic centred white noise Best choice for t i X, i = 1,..., N? F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

8 Information matrix Normalized inverse covariance matrix of the optimal linear unbiased estimate of θ (Gauss Markov) M(ξ) = 1 N N F(t i )F T (t i ) = i=1 l w i F(x i )F T (x i ). t i, i = 1,, N are picked Nw i times within x i, i = 1,, l, (l < N) ( ) x1 x ξ = 2 x l w 1 w 2 w l i=1 w j = n j N, Simplification of the frame (if not discrete optimisation) 0 < w j < 1, w j = 1 Best choice for the design ξ? F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

9 Concave matricial criteria Wish to maximize the information matrix with respect to ξ M(ξ) = l w i F(x i )F T (x i ) = i=1 ( ) x1 x ξ = 2 x l w 1 w 2 w l X σ ξ (dx) := F(x)F T (x)dσ ξ. l w j δ xj (dx). Concave criteria for symmetric p p non negative matrix : φ q (q [, 1]) ( 1 p trace(mq )) 1/q if q, 0 M > 0, φ q (M) := det(m) 1/p if q = 0 λ min (M) if q = { ( 1 det M = 0, φ q (M) := p trace(mq )) 1/q if q (0, 1] 0 if q [, 0]. j=1 F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

10 Optimal design Wish to maximize φ q (M(ξ)) ( with respect to ξ) ( ) x1 x ξ = 2 x l w 1 w 2 w l l σ ξ (dx) := w j δ xj (dx). j=1 φ q (M) concave with respect to M and positively homogenous, φ q (M) is isotonic with respect to Loewner ordering Main idea : extend the optimization problem to all probability measures M(P) = F(x)F T (x)dp(x), P P(X). Within the solutions build one with finite support X F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

11 Overview 1 Short bibliography 2 Convex design 3 Polynomial regression 4 Moment spaces 5 Moment formulation 6 SDP relaxation 7 Examples F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

12 Our frame : polynomial regression R[x] real polynomials in the variables x = (x 1,..., x n ) d N R[x] d := {p R[x] : deg p d} deg p := total degree of p Assumption F = (f 1,..., f p ) (R[x] d ) p. X R n is a given closed basic semi-algebraic set X := {x R m : g j (x) 0, j = 1,..., m} (1) g j R[x], deg g j = d j, j = 1,..., m, and X compact e.g : g 1 (x) := R 2 x 2 F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

13 Overview 1 Short bibliography 2 Convex design 3 Polynomial regression 4 Moment spaces 5 Moment formulation 6 SDP relaxation 7 Examples F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

14 Some facts and notations x α 1 1 xα n n, α = (α 1,..., α n ) N n basis of R[x] (x α := x α 1 1 xα n R[x] d has dimension s(d) := ( ) n+d n. Basis (x α ) α d, α := α α n. n ) v d (x) := ( 1, x }{{} 1,..., x n, x 2 1 }{{}, x 1x 2,..., x 1 x n, x 2 2,..., x2 n,...,..., x d 1 }{{},..., xd n ) }{{} degree 0 degree 1 degree 2 degree d There exists a unique matrix A of size p ( ) n+d n such that x X, F(x) = A v d (x). (2) M + (X) is the cone of nonnegative Borel measures supported on X (the dual of cone of nonnegative elements of C(X)) F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

15 Moments, the moment cone and the moment matrix µ M + (X)α N n, y α = y α (µ) = X x α dµ y = y α (µ) = (y α ) α N n moment sequence of µ M d (X) moment cone=convex cone of truncated sequences M d (X) := {y R (n+d n ) : µ M+ (X) s.t. (3) } y α = x α dµ, α N n, α d. P d (X) convex cone of nonnegative polynomials of degree d (may be also viewed as a vector of coefficients) M d (X) = P d (X) and P d (X) = M d (X) If X = [a, b], M d (X) is representable using positive semidefinitness of Hankel matrices. No more true in general F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31 X

16 Generalized Hankel matrix I Sequence y = (y α ) α N n is given, build linear form L y : R[x] R mapping a polynomials f = α N n f αx α to L y (f) = α N n f αy α. Theorem A sequence y = (y α ) α N n has a representing measure µ supported on X if and only if L y (f) 0 for all polynomials f R[x] nonnegative on X. The generalized Hankel matrix associated to a truncated sequence y = (y α ) α 2d M d (y)(α, β) = L y (x α x β ) = y α+β, ( α, β d). (M d (y)(α, β)) is symmetric (M d (y)(α, β)) is linear in y If y has a representing measure, then M d (y) 0 F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

17 Generalized Hankel matrix II Generalized Hankel matrix associated to polynomial f = α r f αx α R[x] r and a sequence y = (y α ) α 2d+r M d (fy)(α, β) = L y (f(x) x α x β ) = γ r f γ y γ+α+β, ( α, β d) If y has a representing measure µ, and Suppµ {x R n : f(x) 0} then M d (fy) 0 The converse statement holds in the infinite case y = (y α ) α N n has a representing measure µ M + (X) d N, M d (y) 0 and M d (g j y) 0, j = 1,..., m F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

18 Approximations of the moment cone Set v j := d j /2, j = 1,..., m, (half the degree of the g j ). For δ N, M 2d (X) can be approximate by { M SDP 2(d+δ) (X) := y d,δ R (n+2d n ) : yδ R (n+2(d+δ) n ) such that (4) Good approximation y d,δ = (y δ,α ) α 2d and } M d+δ (y δ ) 0, M d+δ vj (g j y δ ) 0, j = 1,..., m. M 2d (X) M SDP 2(d+2) (X) MSDP 2(d+1) (X) MSDP 2d (X). This hierarchy converges M 2d (X) = δ=0 MSDP 2(d+δ) (X) F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

19 Overview 1 Short bibliography 2 Convex design 3 Polynomial regression 4 Moment spaces 5 Moment formulation 6 SDP relaxation 7 Examples F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

20 Optimal design moment formulation For l ( ) n+2d n and W the simplex Carathéodory s theorem gives { } y M 2d (X) : y 0 = 1 = {y R (n+2d n ) : yα = x α dµ α 2d, X l } µ = w i δ xi, x i X, w W, i=1 Our design problem becomes : Step 1 find y solution ρ = max φ q (M) (5) s.t. M = A γ y γ 0, γ 2d y M 2d (X), y 0 = 1, Step 2 find an atomic measure µ representing y F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

21 Equivalence Theorem Theorem (Equivalence theorem) Let q (, 1), Step 1 is is a convex optimization problem with a unique optimal solution y M 2d (X). Denote by p d the polynomial x p d (x) := v d(x) M d (y ) q 1 v d (x) = M d (y ) q 1 2 v d (x) 2 2. (6) y is the vector of moments of a discrete measure µ supported on at least ( ) ( n+d n and at most n+2d ) n points in the set (Step 2) } Ω := {x X : trace(m d (y ) q ) p d (x) = 0. y { M 2d (X) is the unique solution } to Step 1 iff y y M 2d (X) : y 0 = 1 and p :=trace(m d (y ) q ) p d 0 on X. F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

22 Overview 1 Short bibliography 2 Convex design 3 Polynomial regression 4 Moment spaces 5 Moment formulation 6 SDP relaxation 7 Examples F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

23 The SDP relaxation scheme Ideal moment Problem on M 2d (X) is not SDP representable Use the outer approximations of Step 1 ρ δ = max y s.t. φ q (M d (y)) y M SDP 2(d+δ) (X), y 0 = 1. ( δ > 0, ρ δ ρ.) Theorem (Equivalence theorem for SDP relaxation) Let q (, 1), 1 SDP relaxed Problem has a unique optimal solution y R (n+2d n ). 2 M d (y ) is positive definite. p := trace(m d (y ) q ) p d is non-negative on X and L y (p ) = 0. If y is coming from a measure µ then ρ δ = ρ and yd,δ is the solution of unrelaxed Step 1 F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

24 Asymptotics on δ Theorem Let q (, 1), y d,δ optimal solution of relaxed, p d,δ R[x] 2d dual polynomial associated. Then, 1 ρ δ ρ as δ, 2 For every α, α 2d, lim δ y d,δ,α = y α, 3 p d,δ p d as δ, 4 If the dual polynomial p := trace(m d (y ) q ) p d to the unrelaxed Step 1 belongs to P SOS 2(d+δ) (X) for some δ, then finite convergence takes place. (X) is the set of sum of squares of polynomial of degrees less than d + δ P SOS 2(d+δ) F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

25 Overview 1 Short bibliography 2 Convex design 3 Polynomial regression 4 Moment spaces 5 Moment formulation 6 SDP relaxation 7 Examples F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

26 Example I : interval X = [ 1, 1], polynomial regression model d j=0 θ jx j D-optimal design : for d = 5 and δ = 0 we obtain the sequence y (1, 0, 0.56, 0, 0.45, 0, 0.40, 0, 0.37, 0, 0.36). Recover the corresponding atomic measure from the sequence y : supported by -1, , , 0.285, and 1 (for d = 5, δ=0). The points match with the known analytic solution to the problem (critical points of the Legendre polynomial) y x FIGURE: Polynomial p D-optimality, n = 1 F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

27 Example II : Wynn s polygon Polygon given by the vertices ( 1, 1), ( 1, 1), (1, 1) and (2, 2), scaled to fit the unit circle, i.e., we consider the design space { X = x R 2 : x 1, x , x1 1 3 (x 2 + 2), x (x 1 + } 2), x x D-optimal design x x 1 x x 1 x x 1 FIGURE: The polynomial p d ( ) 2+d 2 where for d = 1, d = 2, d = 3. The red points correspond to the good level set F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

28 Example III : Ring of ellipses An ellipse with a hole in the form of a smaller ellipse X = {x R 2 : 9x x , 5x x 2 2 2}. D-optimal design x 2 0 x 2 0 x x 1 x 1 x 1 FIGURE: The boundary in bold black. The support of the optimal design measure (red points). Size of the points corresponds to the respective weights for d = 1 (left), d = 2 (middle), d = 3 (right) and δ = 3. F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

29 Example IV : Folium Zero set of f(x) = x 1 (x 2 1 2x2 2 )(x2 1 + x2 2 )2 is a curve with a triple singular point at the origin called a folium, X = {x R 2 : f(x) 0, x x 2 2 1}. D-optimal design x 2 0 x 2 0 x x 1 x 1 x 1 FIGURE: Boundary in bold black The support of the optimal design measure (red points). The good level set in thin blue d = 1 (left), d = 2 (middle), d = 3 (right), δ = 3 F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

30 The 3-dimensional unit sphere Polynomial regression α d θ αx α on the unit sphere X = {x R 3 : x x x 2 3 = 1}. D-optimal design FIGURE: Optimal design in red d = 1 (left), d = 2 (middle), d = 3 (right) and δ = 0 F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31

31 The 3-dimensional unit sphere+constraint on moments Fix y 020 := 2ω, y 002 := ω, y 110 := 0.01ω and y 101 := 0.95ω. ω chosen such that the problem is feasible D-optimal design x x 1 FIGURE: Support points d = 1 without constraint in red and constrained in blue F. Gamboa (IMT Toulouse) Polynomial design 16th of February / 31