If $T$ is a self-adjoint linear map on a $n$-dimensional inner product space $X$ (either real or complex) then we know by the spectral theorem that there is an orthonormal basis of $X$, call it $v = \{v_1, v_2, \dots, v_n\}$, under which the matrix of $T$ is diagonal.

Even though this is not true for arbitrary linear maps, we can still do something. In fact, by the singular value decomposition, $T$ is still guaranteed to have a diagonal matrix if we are willing to use two different orthonormal bases for $X$, call them $e = \{e_1, e_2, \dots, e_n\}$ and $f = \{f_1, f_2, \dots, f_n\}$.

At first, I thought that if we performed a SVD to a self-adjoint operator, then we would get, up to a reordering, $v = e = f$ and the same diagonal matrices.

Actually this is not true, because the SVD is even stronger that that: the elements of the diagonal matrix are positive!

So I am left wondering if there exists a less strong SVD where the elements of the diagonal matrix are not constrained to be positive, so that if we performed this decomposition to a self-adjoint operator we would get the same of an eigenvalue-eigenvector decomposition..


2 Answers 2


Let $T = U \Sigma V^*$ be an SVD of $T$. Then

$$U \Sigma V^* = T = T^* = V \Sigma U^*,$$


$$\Sigma = (U^*V) \Sigma (U^*V)^*.$$

Obviously, $U^*V$ can be a diagonal, in which case it is trivial to change signs in $\Sigma$ in a way that $U$ and $V$ become equal.

But if, for example, $T$ has two same singular values, i.e., $\Sigma$ has two same diagonal elements, then $U^*V$ may be a permutation (maybe also multiplied by a unitary diagonal), so the fix is harder.

Even worse, if $T$ is singular or rank $r$, then

$$T = \begin{bmatrix} U_1 & U_2 \end{bmatrix} \begin{bmatrix} \Sigma_1 \\ & 0 \end{bmatrix} = \begin{bmatrix} V_1^* \\ V_2^* \end{bmatrix} = \begin{bmatrix} U_1 & 0 \end{bmatrix} \begin{bmatrix} \Sigma_1 \\ & 0 \end{bmatrix} = \begin{bmatrix} V_1^* \\ 0 \end{bmatrix} = U_1 \Sigma V_1^*.$$

Here, $U_2$ and $V_2$ are pretty arbitrary (their columns still have to be normal, and orthogonal to each other and those of $U_1$ and $V_1$, respectively), which makes them quite unrelated. However, this can be fixed by computing some form of reduced SVD (i.e., computing only rectangular orthonormal $U_1$ and $V_1$).

So, changing the SVD to behave the way you want is not all that trivial. I think it is much easier to check if the matrix is Hermitian/symmetric and, if it is, compute Schur decomposition instead of the SVD. The check may not be as simple if you want more generality ($\Sigma$ not necessarily real, i.e., $T$ any normal matrix).


I think the question turns on how much information the singular values $\sigma_i$ carry for an operator which may not be normal.

Suppose we want to factor $T = U \Gamma V^*$, and we're allowed to choose $\Gamma$ diagonal with non-positive (or even non-real) entries. The entries would have to have moduli $|\gamma_i | = \sigma _i$ in some ordering, by an argument I'll give below. And beyond that, I'm not sure what kind of canonical choice we could make for the arguments of the $\gamma_i$.

We know what $U$ and $V$ have to be: if $T: W \to W$ is a complex linear operator and $T = U \Gamma V^*$ for $\Gamma$ diagonal, then $T^*T = V \Gamma^* \Gamma V^*$. This has to be the orthogonal diagonalization of the positive operator $T^* T$, so $V$ must be the eigenvectors for $T^* T$. By a similar argument $U$ must be the eigenvectors for $TT^*$. And we can see that $\overline {\gamma_i}{\gamma_i} = |\gamma_i|^2 $ have to be the eigenvalues of $T^* T$, so we know the modulus of $\gamma_i$. But which argument for $\gamma_i$ do we choose, if not zero? In order to get what you're asking for, we'd like to make some choice that reflects something about the original operator $T$ (like, say, something to do with its eigenvalues). But I'm not sure what that would be.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .