Matrix Norm

The Frobenius norm of an $m\times n$ matrix $\mathbf{A}$ is defined as $\Vert \mathbf{A}{\Vert }_F=\sqrt{{\sum }_{i=1}^m{\sum }_{j=1}^n{a}_{ij}^2}$, the square root of the sum of every entry squared, equivalently $\Vert \mathbf{A}{\Vert }_F=\sqrt{\text{trace}({\mathbf{A}}^{\top }\mathbf{A})}$. Other matrix norms exist -- the spectral norm (the largest singular value, measuring the maximum stretching factor the matrix applies to any unit vector) and various induced operator norms -- but the Frobenius norm's simple entrywise definition makes it the most commonly reported measure of a matrix's overall magnitude or the discrepancy between two matrices, such as an estimated covariance matrix and its target.

• The Frobenius norm is simply the Euclidean vector norm applied to the matrix's entries treated as one long vector -- this makes it the most computationally direct matrix norm, requiring no eigenvalues or singular values, just summing every entry's square and taking the square root.
• The spectral norm (largest singular value) and Frobenius norm agree for a rank-1 matrix but diverge for higher-rank matrices, the Frobenius norm accounts for the total energy across all singular values, while the spectral norm reports only the single largest one -- a distinction that matters when the two norms are used to bound different kinds of numerical error.
• Matrix norms are used to quantify estimation error in matrix-valued statistical quantities, e.g. reporting $\Vert \hat{\mathbf{\Sigma }}-\mathbf{\Sigma }{\Vert }_F$ as a single summary of how far an estimated covariance matrix $\hat{\mathbf{\Sigma }}$ is from its true target $\mathbf{\Sigma }$, playing the same summarizing role for matrices that mean squared error plays for scalar estimators.