Robust estimation of scale and covariance with P n and its application to precision matrix estimation

Transcription

1 Robust estimation of scale and covariance with P n and its application to precision matrix estimation Garth Tarr, Samuel Müller and Neville Weber USYD 2013 School of Mathematics and Statistics THE UNIVERSITY OF SYDNEY

2 Outline Robust Scale Estimator, P n Covariance Covariance Matrix Autocovariance Precision Matrix Long Range Dependence Short Range Dependence

3 Outline Robust scale estimation Robust covariance estimation Robust covariance matrices Summary and key references

4 History of relative efficiencies at the Gaussian (n = 20) 1.00 SD 0.85 P n Relative Efficiency IQR MAD Q n S n Year

5 U-quantile statistics Given data X = (X 1,..., X n ) and a symmetric kernel h : R 2 R a U-statistic of order 2 is defined as: ( ) n 1 U n (X) := h(x i, X j ). (1) 2 i<j Let H(t) = P (h(x i, X j ) t) be the cdf of the kernels with corresponding empirical distribution function, H n (t) := ( ) n 1 I{h(X i, X j ) t}, for t R. (2) 2 i<j For 0 < p < 1, the corresponding sample U-quantile is: H 1 n (p) := inf{t : H n (t) p}. (3)

6 Generalised L-statistics A generalised linear (GL) statistic can be defined as where T n (H n ) = I J(p)H 1 n (p)dp + d j=1 J is function for smooth weighting of Hn 1 (p) I [0, 1] is some interval a j are discrete coefficients for H 1 n (p j ) a j H 1 n (p j ). (Serfling, 1984)

7 Examples of GL-statistics Interquartile range: h(x) = x, IQR = Hn 1 (0.75) Hn 1 (0.25) Variance: h(x, y) = 1 2 (x y)2, 1 0 Hn 1 (p)dp Winsorized variance: h(x, y) = 1 2 (x y)2, Hn 1 (p)dp Hn 1 (0.75) Rousseeuw and Croux s Q n : h(x, y) = x y, H 1 n (0.25)

8 Pairwise mean scale estimator: P n Consider the set of ( n 2) pairwise means: {h(x i, X j ), 1 i < j n} where h(x 1, X 2 ) = (X 1 + X 2 )/2. Let H n be the corresponding empirical distribution function: H n (t) := ( ) n 1 I{h(X i, X j ) t}, for t R. 2 i<j Definition P n = c [ Hn 1 (0.75) Hn 1 (0.25) ], where c is a correction factor to make P n consistent for the standard deviation when the underlying observations are Gaussian.

9 Why pairwise means? Consider 10 observations from N (0, 1). Density Original Data Pairwise Means Observations

10 Influence curve The influence curve for a functional T at distribution F is T ((1 ɛ)f + ɛδ x ) T (F ) IC(x; T, F ) = lim ɛ 0 ɛ where δ x has all its mass at x. Serfling (1984) outlines the IC for GL-statistics. Influence curve for P n Assuming that F has derivative f > 0 on [F 1 (ɛ), F 1 (1 ɛ)] for all ɛ > 0, [ 0.75 F (2H 1 F (0.75) x) IC(x; P n, F ) = c f(2h 1 F (0.75) x)f(x)dx ] 0.25 F (2H 1 F (0.25) x) f(2h 1. F (0.25) x)f(x)dx

11 Influence curves when F = Φ SD P n IC(x; T, F ) Q n MAD x

12 Asymptotic normality The empirical U-process is ( n (Hn (t) H(t)) ) t R. Silverman (1976) proved that in this context, n(hn ( ) H( )) converges weakly to an almost sure continuous zero-mean Gaussian process W with covariance function: EW (s)w (t) = 4 P(h(X 1, X 2 ) s, h(x 1, X 3 ) t) 4H(s)H(t). For 0 < p < q < 1, if H, the derivative of H, is strictly positive on the interval [H 1 (p) ε, H 1 (q) + ε] for some ε > 0, then we can use the inverse map to show n ( H 1 n ( ) H 1 ( ) ) D W (H 1 ( )) H (H 1 ( )).

13 Asymptotic normality Recall, Hence, P n = c [ Hn 1 (3/4) Hn 1 (1/4) ]. n(pn θ) D N (0, V ) where θ = c ( H 1 (3/4) H 1 (1/4) ) and V both depend on the underlying distribution. When the underlying data are Gaussian, V = IC(x, P n, Φ) 2 d Φ(x) = This equates to an asymptotic efficiency of 0.86 as compared with 0.82 for Q n and 0.37 for the MAD at the normal.

14 Breakdown value Definition The breakdown value is the smallest fraction of contamination that can cause arbitrarily corruption to an estimator. P n will breakdown if 25% of the pairwise means are contaminated. Arbitrarily changing m of the original observations leaves n m uncontaminated and ( ) n m 2 pairwise means remain uncontaminated. P n will remain bounded so long as more than 75% of the pairwise means are uncontaminated, i.e. ( ) ( ) n m n > The asymptotic breakdown value of P n is 13.4%.

15 Why the interquartile range? Asymptotic breakdown point Consider, P n (τ) = H 1 n ( 1 + τ 2 ) H 1 n ( 1 τ The asymptotic breakdown point is: ε τ Asymptotic efficiency ). 2 1+τ τ

16 Asymptotic relative efficiency when f = t ν for ν [1, 10] Asymptotic relative efficiency P n Q n MAD Degrees of freedom

17 Adaptive trimming: Pn For preliminary high breakdown location and scale estimates, m(x) and s(x) respectively, an observation, X i, is trimmed if X i m(x) s(x) > d, (4) where d is the tuning parameter. Achieves the best possible breakdown value for a sensible choice of tuning parameter. Definition Denote P n as the adaptively trimmed P n with d = 5.

18 Efficiency of P n in finite samples Following Randal (2008) efficiencies are estimated over m independent samples as êff(t ) = For each i = 1, 2,..., m, Var (ln ˆσ 1,..., ln ˆσ m ) Var (ln T (X 1 ),..., ln T (X m )). (5) X i are independent samples of size n, ˆσ i is the ML scale estimate, and T (X i ) is the proposed scale estimate. Distributions considered t distributions with degrees of freedom between 1 and 10. Configural polysampling using Tukey s 3 corners: Gaussian, One-wild and Slash.

19 Aside: the slash distribution If Z N (0, 1) and U U(0, 1) then X = µ + σz/u follows the slash distribution. When µ = 0 and σ = 1 we have the standard slash distribution. N (0, 1) Cauchy Slash

20 Relative efficiencies: f = t ν for ν [1, 10] and n = 20 Estimated relative efficiency Degrees of freedom P n P n Q n

21 Relative efficiencies at the Slash corner (n = 20) P n 48% P n 75% 10% trim P n Q n 95% S n MAD 87% IQR Estimated relative efficiency

22 Relative efficiencies at the One-wild corner (n = 20) P n 84% P n 72% 10% trim P n Q n 68% S n MAD 40% IQR Estimated relative efficiency

23 Relative efficiencies at the Gaussian corner (n = 20) P n P n 81% 85% 10% trim P n Q n 67% S n MAD 38% IQR Estimated relative efficiency

24 Outline Robust scale estimation Robust covariance estimation Robust covariance matrices Summary and key references

25 From scale to covariance: the GK device Gnanadesikan and Kettenring (1972) relate scale and covariance using the following identity, cov(x, Y ) = 1 [var(αx + βy ) var(αx βy )], 4αβ where X and Y are random variables. In general, X and Y can have different units, so we set α = 1/ var(x) and β = 1/ var(y ). Replacing variance with P 2 n we can similarly construct, γ P (X, Y ) = 1 [ P 2 4αβ n (αx + βy ) Pn(αX 2 βy ) ], where α = 1/P n (X) and β = 1/P n (Y ).

26 Robustness properties IC((x,y),P n,φ) γ P inherits the 13.4% breakdown value from P n. Following Genton and Ma (1999), we have shown that influence curve and therefore the gross error sensitivity of γ P can be derived from the IC of P n. IC(x,y) x y 2 4! IC is bounded and hence the gross error sensitivity is finite.

27 Asymptotic variance Genton and Ma (1999) show that the asymptotic variance is directly proportional to that of the underlying scale estimator: V (γ S, Φ ρ ) = 2(1 + ρ 2 )V (S, Φ) V (γ S, Φ ρ ) 5 γ MAD 1 1 γ P γ SD! Asymptotic efficiency of γ P relative to the standard deviation at the Gaussian remains 86%. ρ

28 Bivariate t ν for ν [3, 10], n = 20 and ρ = 0.5. Estimated efficiency relative to MLE ρ P ρ Q ρ MAD Degrees of freedom

29 Bivariate t ν for ν [3, 10], n = 20 and ρ = 0.9. Estimated relative efficiency ρ P ρ Q ρ MAD Degrees of freedom

30 Outline Robust scale estimation Robust covariance estimation Robust covariance matrices Summary and key references

31 Existing robust covariance matrix estimation methods M-estimates have BDP 1/(1 + p) Minimum volume ellipsoid (MVE) converges at n 1/3 S-estimates behave like MLE for large p Minimum covariance determinant (MCD) requires subsampling, so slow for large p! Try constructing a covariance matrix using pairwise robust covariance estimates.

32 Covariance matrices Good covariance matrices should be: 1. Positive-semidefinite. 2. Affine equivariant. That is, if Ĉ is a covariance matrix estimator, A and a are constants and X is random then, Ĉ(AX + a) = AĈ(X)A.! How can we ensure that a matrix full of pairwise covariances is a good covariance matrix? Maronna and Zamar (2002) outline the OGK routine which ensures that the resulting covariance matrix is positive-definite and approximately affine-equivariant matrix.

33 OGK procedure (Maronna and Zamar, 2002) Let s( ) be a scale function and X = [x 1,..., x n ] R n p be a data matrix. 1. Let D = diag(s(x 1 ),..., s(x p )) and define Y = XD Compute the correlation matrix applying s( ) to the columns of Y: { 1 U = [u jk ] = 4 (s(y j + y k ) 2 s(y j y k ) 2 ) j k 1 j = k 3. Compute the eigendecomposition U = EΛE. 4. Project the data onto the basis eigenvectors: Z = YE. 5. Estimate the variances in the coordinate directions: Γ = diag(s(z 1 ) 2,..., s(z p ) 2 ). 6. The estimated covariance matrix is then, ˆΣ = D 2 EΓE.

34 Precision matrices In many practical applications, the covariance matrix is not what is really required. PCA, Mahalanobis distance, LDA, etc. use the inverse covariance matrix: the precision matrix, Θ = Σ 1. Precision matrices are also of interest in modelling Gaussian Markov random fields, where zeros in the correspond to conditional independence between variables. Sparsity! In many applications, it is often useful to impose a level of sparsity on the estimated precision matrix.

35 Regularisation techniques Graphical lasso (glasso) (Friedman, Hastie, Tibshirani, 2007) minimises the penalised negative Gaussian log-likelihood: f(θ) = tr( ˆΣΘ) log Θ + λ Θ 1, where Θ 1 is the L 1 norm and λ is a tuning parameter for the amount of shrinkage. Constrained l 1 -minimisation for inverse matrix estimation (CLIME) (Cai, Liu and Luo, 2011) solves the following objective function: min Θ 1 subject to: ˆΣΘ I λ. Quadratic Inverse Covariance (QUIC) (Hsieh, et. al. 2011) solves the same minimisation problem as the glasso but uses a second order approach.

36 QUadratic Inverse Covariance (QUIC) Given a regularisation parameter λ 0, the l 1 regularised Gaussian MLE for Θ = Σ 1 can be estimated by solving a regularised log determinant program: argmin Θ 0 f(θ) = argmin Θ 0 {tr( ˆΣΘ) log Θ + λ Θ }{{} 1 } }{{} g(θ) h(θ) where ˆΣ is the sample covariance matrix and Θ 1 = i,j θ ij. Key features The l 1 regularisation promotes sparsity in Θ g(θ) is twice differentiable and strictly convex quadratic approximation h(θ) is convex but not differentiable

37 Algorithm 1 QUadratic Inverse Covariance Input: symmetric matrix ˆΣ, scalar λ, stopping tolerance ɛ 1: for i = 0, 1,... do 2: Compute Σ t = Θ 1 t. 3: Find second order approximation f( ; Θ t ) to f(θ t + ). 4: Partition variables into free and fixed sets. 5: Find Newton direction D t using coordinate descent over the free variable set (Lasso problem). 6: Determine step-size α such that Θ t+1 = Θ t +αd t is positive definite and the objective function sufficiently decreases. 7: end for Output: sequence of Θ t R package QUIC

38 Randomly scattered contamination % Observation vectors 50 Rows contaminated % 50% 5 25% p=30 1 0% Variables 0% 5% 10% Cells contaminated

39 Simulation design Two scenarios: sparse and dense precision matrices, Θ. sample size n = 120 variables p = 30 data N (0, Σ) where Σ = Θ 1 sparsity parameter λ = 0.1 contamination 0% to 3% randomly scattered replications N = 1000 Performance metric The entropy loss L(Θ, ˆΘ) = tr(θ 1 ˆΘ) log Θ 1 ˆΘ p.

40 Simulation design Sparse Θ Dense Θ

41 Sparse Θ, outliers at 5 6 Entropy loss 4 2 Classic P n OGK P n Q n OGK Q n τ scale OGK τ OGKw τ MCD 0% 1% 2% 3% Random contamination

42 Dense Θ, outliers at 5 6 Entropy loss 4 2 Classic P n OGK P n Q n OGK Q n τ scale OGK τ OGKw τ MCD 0% 1% 2% 3% Random contamination

43 Outline Robust scale estimation Robust covariance estimation Robust covariance matrices Summary and key references

44 Summary 1. Aim Efficient, robust and widely applicable scale and covariance estimators. 2. Method P n scale estimator with the GK identity and an orthogonalisation procedure for pairwise covariance matrices or QUIC procedure for (sparse) precision matrix estimation. 3. Results 86% asymptotic efficiency at the Gaussian distribution. Robustness properties flow through to covariance estimates and beyond.

45 References Hampel, F. (1974). The influence curve and its role in robust estimation. Journal of the American Statistical Association, 69(346): Randal, J. (2008). A reinvestigation of robust scale estimation in finite samples. Computational Statistics & Data Analysis, 52(11): Rousseeuw, P. and Croux, C. (1993). Alternatives to the median absolute deviation. Journal of the American Statistical Association, 88(424): Serfling, R. J. (1984). Generalized L-, M-, and R-statistics. The Annals of Statistics, 12(1): Genton, M. and Ma, Y. (1999). Robustness properties of dispersion estimators. Statistics & Probability Letters, 44(4):

46 References Hsieh, C-J., Sustik, M.A., Dhillon I.S. and Ravikumar, P.K. (2011). Sparse inverse covariance matrix estimation using quadratic approximation. Advances in Neural Information Processing Systems 24, Gnanadesikan, R. and Kettenring J. R. (1972). Robust estimates, residuals and outlier detection with multiresponse data. Biometrics, 28(1): Maronna, R. and Zamar, R., (2002). Robust estimates of location and dispersion for high-dimensional datasets. Technometrics, 44(4): Tarr, G., Müller, S. and Weber, N.C., (2012). A robust scale estimator based on pairwise means. Journal of Nonparametric Statistics, 24(1):