Efficient estimation of high-dimensional matrices--including covariance and precision matrices--is a cornerstone of modern multivariate statistics. Most existing studies have focused primarily on the theoretical properties of the estimators (e.g., consistency and sparsity), while largely overlooking the computational challenges inherent in high-dimensional settings. Theoretically, we first prove the convergence of LADMM, and then establish the convergence, convergence rate, and monotonicity of its reparameterized counterpart; importantly, we show that the reparameterized LADMM enjoys a faster convergence rate. Notably, the proposed reparameterization theory and methodology are applicable to the estimation of both high-dimensional covariance and precision matrices. Keywords: ADMM; High-dimensional; Learning-based optimization; Matrix estimation. 1. Introduction High-dimensional matrix estimation--covering both covariance and precision matrix estimation--constitutes a cornerstone of modern statistics and data science [1, 2, 3]. Accurate covariance estimation enables the characterization of dependence structures among a large number of variables [4, 5, 6], which is indispensable in diverse domains such as genomics [7, 8], neuroscience [9], finance [10, 11, 12], and climate science [13, 14]. Over the past two decades, substantial progress has been made in the statistical theory of high-dimensional matrix estimation, particularly with respect to the accuracy of estimators, including properties such as sparsistency and consistency [5, 15, 16]. However, in empirical studies, the dimensionality is often only on the order of tens to hundreds, and in many cases is comparable to the sample size [21, 22, 23, 24]. This observation highlights a notable gap between the statistical theory of estimators and the practical challenges of their computational implementation.

SVM for "after kernels" which are the CK and NTK matrices of trained NNs.

We observe that for the self-attention layers, the correlation of weights for the same head is stronger. Additionally, the best grouping might depend on the type of the layer (e.g., key, query, value, or To simplify the implementation, we treat all the different kernels in the self-attention as a type of fully-connected layer. We down-sample along each dimension to make the computation feasible. To relate with the Frobenius norm, we compute the square of each element and normalize the value. In Figure 5, we show the approximation error comparison for different approximation methods.