Show $U = V$ in SVD of Hermitian, positive definite matrice $A_{n \times n}$


Since eigenvalues and singular values coincide in Hermitian matrices,

$$ AV = U \Sigma = U \Lambda $$

where $\Sigma$ is the singular value diagonal matrix with decreasing diagonal, and $\Lambda$ is the eigenvalue diagonal matrix, i.e. $\Sigma = \Lambda$.

This implies that

$$ Av_i = \lambda_i u_i \ \ (\ast) $$

where $i=1,\cdots, n$, which does not necessarily imply that $v_i$ are eigenvectors, and $v_i = u_i$, $\forall i$.


$$ v_i = c_{i1}\xi_1 + \cdots + c_{in}\xi_n \\ u_i = d_{i1}\xi_1 + \cdots + d_{in}\xi_n \\ $$

where $\xi_i$ is the eigenvectors corresponding to $\lambda_i$.

The $(\ast)$ implies that, $\forall i$,

$$ c_{i1}\lambda_1\xi_1 + \cdots + c_{in}\lambda_n\xi_n = d_{i1}\lambda_i\xi_1 + \cdots + d_{in}\lambda_i\xi_n $$

thus, since $[\xi_1 | \cdots | \xi_n]$ : invertible by spectral theorem,

$$ \begin{bmatrix} c_{11} & c_{12} & \cdots & c_{1n} \\ c_{21} & c_{22} & \cdots & c_{2n} \\ \vdots & \vdots & \ddots & \vdots \\ c_{n1} & c_{n2} & \cdots & c_{nn} \\ \end{bmatrix} \begin{bmatrix} \lambda_1 \\ \lambda_2 \\ \vdots \\ \lambda_n \end{bmatrix} = \begin{bmatrix} d_{11} & d_{12} & \cdots & d_{1n} \\ d_{21} & d_{22} & \cdots & d_{2n} \\ \vdots & \vdots & \ddots & \vdots \\ d_{n1} & d_{n2} & \cdots & d_{nn} \\ \end{bmatrix}^T \begin{bmatrix} \lambda_1 \\ \lambda_2 \\ \vdots \\ \lambda_n \end{bmatrix} $$

I notice that the matrices of $c_{ij}$ and $d_{ij}$ are invertible by the invertibility of $U$ and $V$, but I'm stuck at showing the identity.

Any help will be appreciated.

  • 2
    $\begingroup$ $U$ contains the eigenvectors of $AA^H$ and $V$ has those of $A^HA$. Since $A$ is hermitian, $A^HA = AA^H$ implying that the matrices $U$ and $V$ are identical. $\endgroup$ – sudeep5221 May 29 at 1:29
  • $\begingroup$ @sudeep5221 But when there exists $\sigma_i = \sigma_j$, $i \neq j$? $\endgroup$ – Moreblue May 29 at 3:57
  • $\begingroup$ @sudeep5221 I don't see how your argument works. The equality $A^HA=AA^H$ is also satisfied by every normal matrix $A$, but unless $A$ is positive semidefinite, we cannot possibly have $U=V$ in the SVD of a normal matrix $A$. $\endgroup$ – user1551 May 29 at 5:02
  • $\begingroup$ @user1551 Oh I see what you are trying to say. I had in mind what you wrote in your answer and took it for granted forgetting to point that out clearly. Sorry about that and thanks for pointing that out. :) $\endgroup$ – sudeep5221 May 29 at 16:56

Here is one way to prove it. Note that every positive semidefinite matrix $P$ has a unique positive semidefinite square root $P^{1/2}$. In fact, $P^{1/2}=f(P)$ where $f$ is the Lagrange interpolation polynomial that maps each eigenvalue of $P$ to its square root.

Now, in your case, since $A^2=AA^\ast=US^2U^\ast$, by the uniqueness of positive definite square root, we have $A=USU^\ast$. Hence $U=V$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.