Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Adaptive estimation of the copula correlation matrix for semiparametric elliptical copulas Marten Wegkamp Department of Mathematics Department of Statistical Science Cornell University London, January 7, 2016 Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Joint work with Yue Zhao Marten Wegkamp Supported by NSF Grant DMS 1310119 Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Outline 1 Introduction Copula, elliptical copula, copula correlation matrix Proposed research 2 b −Σ Bounding the operator norm of Σ b Bounding kT − T k2 b − Σk2 Bounding kΣ 3 Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Copula, elliptical copula, copula correlation matrix Proposed research Outline 1 Introduction Copula, elliptical copula, copula correlation matrix Proposed research 2 b −Σ Bounding the operator norm of Σ b Bounding kT − T k2 b − Σk2 Bounding kΣ 3 Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Copula, elliptical copula, copula correlation matrix Proposed research Copula Let the continuous random vector X = (X1 , . . ., Xd )0 ∈Rd have marginal distribution functions Fj (x) = P Xj ≤ x . The copula function C(u1 , . . . , ud ) is the joint cumulative distribution function of the transformation U = (F1 (X1 ), . . . , Fd (Xd ) )0 , i.e., for u = (u1 , . . . , ud )0 ∈ [0, 1]d , C(u1 , . . . , ud ) = P {F1 (X1 ) ≤ u1 , . . . , Fd (Xd ) ≤ ud } . Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Copula, elliptical copula, copula correlation matrix Proposed research Copula The copula C always describes a distribution on the unit square. Copula is invariant under strictly increasing transformations Tj of the marginals Xj — allows separate specifications of the marginals and the dependence structure. In this talk we mostly study semiparametric copula models. The dependence structure is given by a parametric copula, while the marginals are left unspecified (nonparametric). Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Copula, elliptical copula, copula correlation matrix Proposed research Elliptical distribution A random vector X ∈ Rd has an elliptical distribution if its characteristic function E(exp(it 0 X )) can be written as 0 eit µ Ψ(t 0 Σt) with parameters µ ∈ Rd , covariance matrix Σ ∈ Rd×d , and characteristic generator Ψ : [0, ∞) → R. For instance, when Ψ(t) = e−t/2 , X ∼ N(µ, Σ). Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Copula, elliptical copula, copula correlation matrix Proposed research Semiparametric elliptical distributions and the copula correlation matrix Elliptical copulas are the copulas of elliptical distributions. Elliptical copula is characterized by copula parameters: Ψ and the copula correlation matrix Σ, with Σij Σij = q . Σii Σjj Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Copula, elliptical copula, copula correlation matrix Proposed research Semiparametric Gaussian graphical copula model Let X ∼ N(0, Σ), and Ω = Σ−1 be the precision matrix. If [Ω]k ` = 0, then Xk ⊥ X` |X\{k ,`} . Accurate estimator of Ω can be obtained by feeding the b into the graphical lasso, sample covariance matrix Σ CLIME, graphical Dantzig selector, etc. Assume instead that we merely have f (X ) ∼ N(0, Σ), for some strictly increasing, unknown function f (i.e., X has a Gaussian copula with copula correlation matrix Σ), and again we let Ω = Σ−1 . Now, if [Ω]k ` = 0, then f (Xk ) ⊥ f (X` )|f (X )\{k ,`} =⇒ Xk ⊥ X` |X\{k ,`} . Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Copula, elliptical copula, copula correlation matrix Proposed research Semiparametric Gaussian graphical copula model b rank-based (using Kendall’s Using a rank-based estimator Σ tau or Spearman’s rho) of Σ, Liu et al 2012, Xue & Zou 2012 recovers the same optimal parametric rate (obtained under the Gaussian model) under the semiparametric model. b rank-based − Σkmax at This crucially p depends on estimating kΣ the rate of log(d)/n (w.h.p.), which is in turn based on Hoeffding’s classical bound for scalar U-statistic. Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Copula, elliptical copula, copula correlation matrix Proposed research Kendall’s tau e ∈ Rd be an independent copy of X . Kendall’s tau Let X between the k th and `th coordinates is h i ek )(X` − X e` ) τk ` = E sgn (Xk − X Let T be the matrix of Kendall’s tau: [T ]k ` = τk ` . Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Copula, elliptical copula, copula correlation matrix Proposed research Kendall’s tau Kendall’s tau statistic is, for samples X 1 , . . . , X n that are independent copies of X , τbk ` = XX 2 sgn (Xki − Xkj )(X`i − X`j ) . n(n − 1) 1≤i<j≤d b be the matrix of Kendall’s tau statistics: Let T b ]k ` = τbk ` . [T Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Copula, elliptical copula, copula correlation matrix Proposed research Kendall’s tau matrix as plug-in estimator For elliptical copulas, independent of the characteristic generator, π Σ = sin T 2 with the sine function acting component-wise (Kruskal 1958, Kendall & Gibbons 1990, Fang, Fang & Kotz 2002). Hence, the natural plug-in estimator for Σ is b . b = sin π T Σ 2 Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Copula, elliptical copula, copula correlation matrix Proposed research Outline 1 Introduction Copula, elliptical copula, copula correlation matrix Proposed research 2 b −Σ Bounding the operator norm of Σ b Bounding kT − T k2 b − Σk2 Bounding kΣ 3 Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Copula, elliptical copula, copula correlation matrix Proposed research For the rest of the talk, we assume X ∈ Rd follows a semiparametric elliptical distribution, The copula correlation matrix of X is Σ ∈ Rd×d , b =Σ b n of Σ is constructed from n The plug-in estimator Σ independent copies of X , (e.g., Demarta & McNeil 2004, Klüppelberg & Kuhn 2009) Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Copula, elliptical copula, copula correlation matrix Proposed research Some recent advanced in elliptical/Gaussian copulas b in the Klüppelberg & Kuhn 2009 study the property of Σ asymptotic regime, i.e., d fixed, n → ∞. Liu et al 2012, Xue & Zou 2012 study estimation of the precision matrix Σ−1 of Gaussian copulas in high dimensions. b − Σkmax is a key step of the study. Sharp bound on kΣ Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Copula, elliptical copula, copula correlation matrix Proposed research Proposed Research: Part I First, we establish a sharp bound of the operator norm b − Σk2 . (See also Han & Liu 2013.) kΣ It has often been observed (e.g., Demarta & McNeil 2004, b is Klüppelberg & Kuhn 2009) that the plug-in estimator Σ not always positive semidefinite. b − Σk2 quantifies the extent to which the A bound on kΣ non-positive semidefinite problem may happen. b is not positive semidefinite, we can project it onto the If Σ set of correlation matrices in some specific way to obtain b + . Furthermore, Σ b + − Σk2 ≤ 2kΣ b − Σk2 . kΣ Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix Copula, elliptical copula, copula correlation matrix Proposed research Proposed Research: Part II Later, we study a factor model for Σ. The factor model assumes that Σ = Θ + V for a low-rank (or nearly low-rank) matrix Θ and a diagonal matrix V . X and [L, V 1/2 ]Y have the same copula for some elliptical distribution with Σ = Ir +d . Here Θ = LL0 for some L ∈ Rd×r . Within this context, the effective number of parameters is O(r · d), for r = rank(Θ), instead of O(d 2 ) under the generic model. e for Σ using a penalized We propose a refined estimator Σ least square procedure. b − Σk2 serves to scale the penalty term. The bound on kΣ Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix b − Tk Bounding kT 2 b − Σk Bounding kΣ 2 b − Σk2 Outline for bounding kΣ b − T k2 (matrix counterpart of Hoeffding 1963’ Bounding kT classical bound for scalar U-statistic). b − T k2 (matrix b − Σk2 in terms of kT Bounding kΣ counterpart of the Lipschitz property). Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix b − Tk Bounding kT 2 b − Σk Bounding kΣ 2 Outline 1 Introduction Copula, elliptical copula, copula correlation matrix Proposed research 2 b −Σ Bounding the operator norm of Σ b Bounding kT − T k2 b − Σk2 Bounding kΣ 3 Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix b − Tk Bounding kT 2 b − Σk Bounding kΣ 2 Matrix U-statistic Theorem With probability at least 1 − α, we have np o 2 b − T k2 < max kT k2 fn,d,α , fn,d,α kT q 2 2 b k2 f 2 ≤ kT n,d,α + fn,d,α /4 + fn,d,α /2 o np 2 2 + fn,d,α kT k2 fn,d,α , fn,d,α < max with s fn,d,α = 16 d · log(2α−1 d) · . 3 n The above bounds hold for all n, d, α. Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix b − Tk Bounding kT 2 b − Σk Bounding kΣ 2 Matrix U-statistic: proof We need a Bernstein-type inequality, e.g.,oTropp (2012), to n b bound the tail probability P kT − T k2 ≥ t . That theorem, as many other results on bounding the tail probability of the operator norm of a sum of random matrices, requires that the summands be independent. b does not satisfy this condition. The U-statistics T Many results on tail bound rely on the Laplace transform technique to convert the tail probability into an expectation b − T. of some convex function of T A technique pointed out by Hoeffding (1963) allows us to b − T k2 into a problem convert the problem of bounding kT involving a sum of independent random matrices. Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix b − Tk Bounding kT 2 b − Σk Bounding kΣ 2 Outline 1 Introduction Copula, elliptical copula, copula correlation matrix Proposed research 2 b −Σ Bounding the operator norm of Σ b Bounding kT − T k2 b − Σk2 Bounding kΣ 3 Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix b − Tk Bounding kT 2 b − Σk Bounding kΣ 2 Matrix Lipschitz property Theorem We have, with probability at least 1 − α, that b − T k2 + b − Σk2 ≤ πkT kΣ 3π 2 2 ·f . 16 n,d,α b − T k2 , we have, with Combining earlier bound for kT probability at least 1 − 2α, that b − Σk2 < π · max kΣ o 3π 2 np 2 kT k2 fn,d,α , fn,d,α + · f2 . 16 n,d,α In addition, 2 kΣk2 ≤ kT k2 ≤ kΣk2 . π Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix b − Tk Bounding kT 2 b − Σk Bounding kΣ 2 Potential improvement We can obtain from the corollary that, for n & d, r b − Σk2 . kΣk2 · d · log(d) . EkΣ n b its sample covariance In contrast, for Gaussian X and Σ matrix, Koltchinskii & Lounici 2015 shows that, for n & d, r b − Σk2 kΣk2 · d . EkΣ n Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Outline 1 Introduction Copula, elliptical copula, copula correlation matrix Proposed research 2 b −Σ Bounding the operator norm of Σ b Bounding kT − T k2 b − Σk2 Bounding kΣ 3 Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations A factor model for Σ From now on, we assume that the copula correlation matrix Σ satisfies a factor model: Σ=Θ+V for some low-rank (or nearly low-rank), positive semidefinite, symmetric matrix Θ and diagonal matrix V . Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Outline 1 Introduction Copula, elliptical copula, copula correlation matrix Proposed research 2 b −Σ Bounding the operator norm of Σ b Bounding kT − T k2 b − Σk2 Bounding kΣ 3 Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Prelude: an elementary factor model for Σ In an elementary factor model for Σ = Θ + V , we assume that rank(Θ) = r < d, and V = σ 2 Id . b have the eigen-decomposition Let the plug-in estimator Σ d X bk u bk u bk0 , λ b1 ≥ · · · ≥ λ bd . λ k =1 Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Prelude: an elementary factor model for Σ Construct the following closed-form estimators for r , σ 2 , Θ and Σ, based on some threshold µ: b r = d n o X 1 λbk ≥ λbd + µ , k =1 σ b2 = 1 Xb λj , d −b r j>b r b = Θ b r X bk − σ bk u bk0 (λ b2 )u k =1 e = Θ b − diag(Θ) b + I. Σ Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Prelude: an elementary factor model for Σ, continued Theorem n o b − Σk2 < µ , we have Assume λr (Θ) ≥ 2µ. On the event 2kΣ b r = r, e − Σk2 ≤ kΘ b − Θk2 ≤ 8r kΣ b − Σk2 , kΣ F F 2 2 2 b |b σ − σ | ≤ kΣ − Σk2 . Here, k · kF denotes the Frobenius norm. Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Prelude: an elementary factor model for Σ, continued Corollary Set q 2 4 b k2 f 2 µ = 8 kT n,d,α + fn,d,α /4 + 8fn,d,α p 2 . µ̄ = 8 kT k2 fn,d,α + 12fn,d,α 2 Assume λr (Θ) ≥ 2µ̄, and kT k2 ≥ fn,d,α . Then, e − Σk2 ≤ 2r µ̄2 , kΣ F hold with probability exceeding 1 − 2α. Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Outline 1 Introduction Copula, elliptical copula, copula correlation matrix Proposed research 2 b −Σ Bounding the operator norm of Σ b Bounding kT − T k2 b − Σk2 Bounding kΣ 3 Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations e of Σ = Θ + V The refined estimator Σ In the factor model we can write Σ = Θ − diag(Θ) + Id := Θo + Id . e as the This motivates us to estimate the low-rank matrix Θ by Θ solution to the convex program 1 0 0 2 e = argmin b o k + µkΘ k∗ . Θ kΘ − Σ F 2 o Θ0 ∈Rd×d Here, k · k∗ denotes the nuclear norm. e of the copula correlation Finally, we let the refined estimator Σ matrix Σ be e =Θ e o + Id . Σ Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations e with existing estimators Comparison of Σ Saunderson et al 2012 study a factor model for Σ in the noiseless setting using minimum trace factor analysis: 0 Σ = Θ + V (Θ, V ) = argmin tr(Θ0 ) subject to Θ0 % 0 V diagonal If Θ has column space U such that its coherence √ max kPU ei k2 < 1/ 2, 1≤i≤d then the above algorithm recovers Θ, V exactly. Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations e with existing estimators, continued Comparison of Σ For Θ with effective low-rank and without sparse corruption, Lounici 2013 proposes 1 0 b 2 0 e Θ = argmin kΘ − ΘkF + µkΘ k∗ 2 Θ0 ∈Rd×d b of Θ (which we lack). for an unbiased initial estimator Θ For Σ = Θ + S with low-rank matrix Θ and generic sparse matrix S, Chandrasekaran et al 2012, Hsu et al 2011 propose e = argmin 1 kΘ0 + S − Σk e S) b 2 + µkΘ0 k∗ + λkSk` . (Θ, F 1 2 Θ0 ,S∈Rd×d This formulation requires λ > 0. Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations e with existing estimators, continued Comparison of Σ We are in between the above scenarios. We will extend the result of the former and relax the condition of the latter with more precise analysis. Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations e Oracle inequality for the refined estimator Σ Theorem Let Θr be best rank-r approximation to Θ, with reduced SVD Θr = Ur Dr Ur0 . Set q 2 4 b k2 f 2 µ = 10 kT n,d,α + fn,d,α /4 + 10fn,d,α , i hp 2 2 kT k2 fn,d,α , fn,d,α + 15fn,d,α . µ̄ = 10 max Then, with probability at least 1 − 2α, X e − Σk2 ≤ kΣ λ2j (Θ) + 18r µ̄2 F j>r for all r with γr = kUr Ur0 kmax ≤ 1/9. Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Outline 1 Introduction Copula, elliptical copula, copula correlation matrix Proposed research 2 b −Σ Bounding the operator norm of Σ b Bounding kT − T k2 b − Σk2 Bounding kΣ 3 Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Simulation result b − Σk2 kΣ F , which our e − Σk2 kΣ F theory predicts to be roughly proportional to d/r as n → ∞. The object of interest is the ratio We (somewhat arbitrarily) set C = 0.1, α = 0.1, and forgo the log(d) factor when choosing the regularization parameter. We estimate this ratio by simulation. Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Simulation setup kΣk2 = 2. 10 3 (Larger is better) e ! 'k2 b ! 'k2 =j' k' F F 10 2 d=128, r=1 10 1 d=32, r=1 10 0 d=128, r=4 d=8, r=1 d=32, r=4 10 -1 d=8, r=4 10 -2 10 2 10 3 Sample size Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Summary We have obtained a sharp bound on the operator norm of b − Σ, for the plug-in estimator Σ b of the copula correlation Σ matrix Σ. We have applied the above bound to the setting of a factor e a sharp oracle model of Σ to obtain a refined estimator Σ; e inequality for Σ is established. Marten Wegkamp Copula correlation matrix for elliptical copulas Introduction b−Σ Bounding the operator norm of Σ Analyzing a factor model for the copula correlation matrix The factor model for Σ Study of an elementary factor model Analysis of the refined estimator Simulations Thanks! Marten Wegkamp Copula correlation matrix for elliptical copulas