A Matrix-Variate Log-Normal Model for Covariance Matrices

We propose a modeling framework for time-varying covariance matrices based on the assumption that the logarithm of a realized covariance matrix follows a matrix-variate oNrmal distribution. By operating in the space of symmetric matrices, the approach guarantees positive definiteness without imposing parameter constraints beyond stationarity. The conditional mean of the logarithmic covariance matrix is specified through a BEKK-type structure that can be rewritten as a diagonal vector representation, yielding a parsimonious specification that mitigates the curse of dimensionality. Estimation is performed by maximum likelihood exploiting properties of matrix-variate Normal distributions and expressing the scale parameter matrix as a function of the location matrix. The covariance matrix is recovered via the matrix exponential. Since this transformation induces an upward bias, an approximate, time-specific bias correction based on a second-order Taylor expansion is proposed. The framework is flexible and applicable to a wide class of problems involving symmetric positive definite matrices.

💡 Research Summary

The paper introduces a novel framework for modeling time‑varying covariance matrices by assuming that the matrix logarithm of a realized covariance matrix follows a matrix‑variate Normal distribution. By operating in the space of symmetric real matrices, the model automatically guarantees positive definiteness without imposing any constraints on the parameters beyond the usual stationarity conditions. The conditional mean of the log‑covariance matrix is specified through a BEKK‑type recursion, but unlike the classic BEKK model it does not require any positivity constraints on the coefficient matrices because the exponential back‑transformation will always produce a positive‑definite matrix.

Formally, for each time t the realized covariance Cₜ is decomposed as Cₜ = VₜΛₜVₜ′, and the log‑covariance is defined as Γₜ = Vₜ log(Λₜ) Vₜ′. The assumption Γₜ ∼ MN(Mₜ, Uₜ, Uₜ) leads to the vectorized representation vec(Γₜ) ∼ N(vec(Mₜ), Uₜ⊗Uₜ). By extracting only the lower‑triangular elements (vech), the dimensionality is reduced from n² to n* = n(n + 1)/2. The BEKK recursion is rewritten in this reduced space as a diagonal VEC model:
μₜ = (I* − A* − B*)γ* + Aγₜ₋₁ + Bμₜ₋₁,
where μₜ and γ are the vech of Mₜ and Γₜ respectively, and A*, B* are diagonal matrices containing the original BEKK coefficients. Stationarity simply requires |aᵢ + bᵢ| < 1 for each diagonal element.

Maximum‑likelihood estimation proceeds by first estimating the covariance of the log‑matrices, Σ̂ₜ = (1/T)∑(Γₜ − M̂ₜ)(Γₜ − M̂ₜ)′, and then scaling it to obtain the row/column scale matrix Ûₜ = Σ̂ₜ · tr(Σ̂ₜ)⁻¹. This functional relationship eliminates the need to estimate a full Uₜ, thereby avoiding the explosion of parameters that would otherwise accompany a matrix‑variate Normal likelihood. An iterative algorithm alternates between (i) fixing Ûₜ and maximizing the likelihood with respect to the diagonal VEC parameters, and (ii) updating Ûₜ using the residuals from the current M̂ₜ. Convergence yields estimates of the log‑means M̂ₜ.

The estimated log‑means are exponentiated to recover the covariance matrices: Ĉₜ = exp(M̂ₜ) = Wₜ exp(Lₜ) Wₜ′, where Lₜ contains the eigenvalues of M̂ₜ and Wₜ the corresponding eigenvectors. Because the exponential map is convex, Ĉₜ is upward biased (Jensen’s inequality). The authors propose a time‑specific bias correction based on a second‑order Taylor expansion (the delta method). For each variance element i, they approximate
E