对矩阵det和logdet求导

吴炳洋

已于 2022-04-15 13:13:27 修改

阅读量6.3k

点赞数 4

分类专栏：数学基础文章标签：线性代数数字通信 mimo

于 2022-02-12 13:19:41 首次发布

本文链接：https://blog.csdn.net/wubyatseu/article/details/122894982

版权

数学基础专栏收录该内容

20 篇文章

订阅专栏

这篇博客探讨了矩阵导数的概念，特别是对于行列式的导数。内容包括矩阵的转置和伴随矩阵，以及当矩阵可逆时，利用矩阵的导数计算对数行列式的导数。文章通过一个例子说明了如何在信息科学和技术领域，如通信和信号处理中，计算矩阵B+rC的对数行列式的导数，这对于理解MIMO信道容量的变化率至关重要。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

Derivative of det

Assume matrix $A=[a_{ij}], i=1,2,...,n, j=1,2,...,n$

$\frac {d\det(A)}{d\alpha }=\sum_{i,j=1}^{n}\frac{d \det(A)}{a_{ij}}\frac{d a_{ij}}{d\alpha}$

$\det(A)=\sum_{j=1}^{n}a_{ij}\textrm {adj}(A_{ij})$

for any i,where $\textrm {adj}(A_{ij})$ denotes the adjugate of A for $a_{ij}$ . So

$\begin {align*} \label{ref1}\frac{d \det(A)}{a_{ik}}& = \textrm {adj}(A_{ik}) + \sum_{j\neq k}\frac {\operatorname {d} \textrm {adj}(A_{ij})}{d a_{ik}} \\ &= \textrm {adj}(A_{ik}) \end{align*}$

the second part of rhs of the first equation is 0, for $\textrm {adj}(A_{ij})$ do not include $a_{ik}$ .

$\begin {align} \displaystyle {\frac {d\det(A)}{d\alpha }}&=\sum_{i,j=1}^{n}\textrm {adj}(A_{ij})\frac{d a_{ij}}{d\alpha} \nonumber\\ &=\operatorname {tr} \left(\operatorname {adj} (A){\frac {dA^T}{d\alpha }}\right) \nonumber\end{align}$

where $adj(A)$ denotes the adjugate of A,and $()^T$ denotes transpose. If A is invertible, considering $\operatorname {adj} (A)= \det(A)\operatorname A^{-1}$ , so

${\frac {d\det(A)}{d\alpha }}=\det(A)\operatorname {tr} \left(A^{-1}{\frac {dA^T}{d\alpha }}\right).$

Derivative of logdet

Using

$\operatorname {d}\left ( \det(\Sigma) \right )=\mathrm {tr}\left ( \Sigma^{-1}\operatorname {d}\Sigma \right )\det(\Sigma)$

It is easy to obtain

$\begin {align} \label{ref2} \operatorname {d}\left (\log \det(\Sigma) \right )&=\frac{\operatorname {d}\left (\log \det(\Sigma) \right )} {d\det(\Sigma)}\cdot {d\det(\Sigma)} \nonumber\\ &=\frac{1} {\det(\Sigma)}\cdot {d\det(\Sigma)}\nonumber \\ &=\mathrm {tr}\left ( \Sigma^{-1}d\Sigma\right ) \nonumber\end{align}$

Example

We often meet the problem to derivate a logdet of a matrix like B+rC in information science and technology, such as communications and signal processing, because the capacity of a MIMO channel is logdet(X), if the channel is the form of B+rC, the capacity is logdet(B+rC). To investigat the change rate of capacity corresponding to r, it necessary to calculate the derivation.

Using the above result, we obtain

$\begin {align} \frac{d}{d \lambda} \left (\log \det (B+\lambda C) \right ) =\mathrm {tr}\left ( \left ( B+\lambda C \right )^{-1}C \right ) \nonumber\end{align}$