View on GitHub

memo

Computing a nearest symmetric positive semidefinite matrix

Computing a nearest symmetric positive semidefinite matrix

spectra radius $\rho: \mathbb{R}^{n\times n} \rightarrow \mathbb{R}$

\[\rho(A) := \max \{ |\lambda| \mid \mathrm{det}(A - \lambda I) = 0 \}\]

The 2-norm

\[\|A\|_{2} := \rho(A^{\mathrm{T}}A)^{1/2}\]

The Frobenius norm

\[\|A\|_{F} := \left( \sum_{i=1}^{n} \sum_{j=1}^{n} (a_{i}^{j})^{2} \right)^{1/2}\]

We denote positive semidefinite matrix $A$ by \(A \geqslant 0\).

The distance in the norm \(\| \cdot \|_{2}\) is denoted by

\[\delta_{2}(A) := \min_{X \in \mathcal{S}} \| A - X\|_{2}\]

Positive semidefinite matrix $X$ satisfying \(\|A - X\| = \delta_{2}(A)\) is termed a positve approximant of $A$. For a positive semidefinite matrix $A$, \(A^{1/2}\) denotes $X \in \mathcal{S}$ which satisfies

\[X^{2} = A .\]

2. The Frobenius norm Positive Approximant

Theorem 2.1.

Then \(X_{F} := (B + H)/2\) is unique positive approximant of $A$ in the frobenius norm.

\[\min_{X \in \mathcal{S}}\|A - X\|_{F}^{2} = \sum_{\lambda_{i}(B) < 0} \lambda_{i}(B)^{2} + \|C\|_{F}^{2}\]

where \(\lambda_{i}(B)\) is $i$-th largest eigenvalue of $B$. Moreover,

\[\begin{eqnarray} X^{*} & := & Z \mathrm{diag}(d_{1}, \ldots, d_{n}) Z^{\mathrm{T}} \nonumber \\ & = & \arg \min_{X \in \mathcal{S}} \|A - X\|_{F}^{2} \end{eqnarray}\]

where $Z \in \mathbb{R}^{n \times n }$ is unitary matrix satisfying \(B = Z \mathrm{diag}(\lambda_{1}(B), \ldots, \lambda_{n}(B))Z^{\mathrm{T}}\) and

\[d_{i} := \begin{cases} \lambda_{i} & \lambda_{i} \ge 0 \\ 0 & \lambda_{i} < 0 \end{cases} .\]

proof.

Let $X$ be positive semidefnite. $B$ is symmetric and $C$ is skew-symmetric so that by proposition we have

\[\begin{eqnarray} \|A - X \|_{F}^{2} & = & \|B + C - X \|_{F}^{2} \nonumber \\ & = & \|B - X\|_{F} + \|C\|_{F}^{2} \nonumber . \end{eqnarray}\]

The problem reduces to that of approximating $B$. Since $B$ is symmetric, by eigenvalue decomposition there exists unitary marix $Z$ and diagonal matrix \(\Lambda := \mathrm{diag}(\lambda_{1}, \ldots, \lambda_{n})\) such that \(B = Z \Lambda Z^{\mathrm{T}}\). Let \(Y := (y_{j}^{i})_{i,j = 1, \ldots, n} := Z^{\mathrm{T}}XZ\). By theorem, $Y$ is positive semidefinite. Then

\[\begin{eqnarray} \|B - X\|_{F}^{2} & = & \|Z \Lambda Z^{\mathrm{T}} - X\|_{F}^{2} \nonumber \\ & = & \|Z (\Lambda - Z^{\mathrm{T}}XZ )Z^{\mathrm{T}}\|_{F}^{2} \nonumber \\ & = & \|\Lambda - Y\|_{F}^{2} \quad (\because Z \text{ is unitary}) \nonumber \\ & = & \sum_{i \neq j} (y_{j}^{i})^{2} + \sum_{i = 1}^{n} (\lambda_{i} - y_{i}^{i})^{2} \nonumber \\ & \ge & \sum_{i = 1}^{n} (\lambda_{i} - y_{i}^{i})^{2} \nonumber \\ & \ge & \sum_{i: \lambda_{i} < 0} (\lambda_{i} - y_{i}^{i})^{2} \quad (\because Y \text{ is positive semi-definite so that } y_{i}^{i} \ge 0) \nonumber \\ & \ge & \sum_{i: \lambda_{i} < 0} (\lambda_{i})^{2} \nonumber \end{eqnarray}\]

The lower bound is attaind uniquly by \(Y := \mathrm{diag}(d_{1}, \ldots, d_{n})\), where

\[d_{i} := \begin{cases} \lambda_{i} & \lambda_{i} \ge 0 \\ 0 & \lambda_{i} < 0 \end{cases} .\]

Hence

\[X^{*} := Z \mathrm{diag}(d_{1}, \ldots, d_{n}) Z^{\mathrm{T}}\]

is the minimizer.

$\Box$

Reference