MathOverflow is a question and answer site for professional mathematicians. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

I want to find the unitary $N \times N$ matrix U from the following data. Let $M$ be an integer $(1< M<N-1)$ and let $\mathcal S$ be the space of all the possible subsets of $\{1,2,\dots, N\}$ with exactly $M$ elements.

Let $U(s,s')$ be the $M\times M$ submatrix of U whose rows are specified by the elements of $s\in\mathcal S$ and whose columns are specified by the elements of $s'\in\mathcal S$.

My data is a set of $I(s,s')\in\mathbb Q$ such that $$ \det U(s,s') = e^{i \pi I(s,s')}.$$

Can I find the full matrix $U$? I would be interested also in understand why this problem might not have a solution. However, if the solution exists (for certain values of $M,N$), than I would like to find the algorithm to obtain $U$.

share|cite|improve this question
Homework assignment...? – Wlodek Kuperberg Feb 3 '14 at 17:10
No, I can solve a research problem if there is a solution to this "homework" :-/ – benkj Feb 3 '14 at 17:21
Do you know that the adjoint is the case when M=N-1? I suspect Cayley or Sylvester handled cases for other M. Gerhard "Does More Solitary Than Unitary" Paseman, 2014.02.03 – Gerhard Paseman Feb 3 '14 at 17:31
You are right, thanks... but unfortunately when M=1 or M=N-1 the original problem is quite trivial – benkj Feb 3 '14 at 17:38
Evidently, I misinterpreted the question. Sorry. – Wlodek Kuperberg Feb 3 '14 at 20:22

$\newcommand{\bC}{\mathbb{C}}$ View the unitary matrix as a unitary operator $U:\bC^N\to\bC^N$. For any $1\leq M\leq N$ it canonically induces a unitary map

$$\Lambda^MU: \Lambda^M\bC^N\to\Lambda^M\bC^N, $$

where $\Lambda^M$ denotes the $M$-th exterior power of a vector space. Your problem can be rephrased as follows: how much of $U$ can one recover given that we have complete knowledge of $\Lambda^M U$.

Immediately one ses that we can recover quite a bit. If $\lambda_1,\dotsc,\lambda_N$ are the eigenvalues of $U$ then the eigenvalues of $\Lambda^M U$ are

$$ \lambda(S)=\prod_{k\in S}\lambda_k, $$

where $S$ runs through the subsets of $\{1,\dotsc, N\}$ of cardinality $M$. Observe that if $M< N$, then the collection of eigenvalues

$$\{ \lambda(S);\;\;|S|=M\} $$

determines the collection of ratios

$$\frac{\lambda_i}{\lambda_j},\;\; 1\leq i<j\leq N. $$

With this extra knowledge we can compute the powers $\lambda_1^M,\dotsc, \lambda_N^M$, and thus the eigenvalues of $U$ up to an overall multiplicative ambiguity $\rho$ which is an $M$-th root of $1$. You cannot do better than this because

$$\Lambda^M U=\Lambda^M(\rho U). $$

share|cite|improve this answer
Thanks for your answer. Do you think that one can reconstruct also the entire matrix from the conditions? For example, let us consider the first non-trivial case: M=2 (or M=N-2). The constraints that I have are N(N-1)/2, a number which is smaller than the number of independent coefficients of U. So I hope that, at least in this case, there is a way for determining U (not in a unique way) – benkj Feb 4 '14 at 8:23
The real dimension of $U$ is $N^2$ and if you are given $N(N-1)/2$ complex constraints, this indicates that the space of solutions of $\Lambda^2 U=T$ is a real variety of dimension $N$. – Liviu Nicolaescu Feb 4 '14 at 13:29
Cannot edit my previous comment. You actually have $\binom{N}{2}^2$ constraints so my above comment is misleading. One case that seems hard to decide is when $U$ is an involution and $\dim \ker(1\pm U)>1$. – Liviu Nicolaescu Feb 4 '14 at 13:41
Here is an even more challenging case. Assume that the spectrum of $U$ as a set consists of all the $2k$-th roots of $1$ for some $k\geq 1$ and each eigenvalue has multiplicity $>1$. It seems difficult to recover the the eigenspaces of $U$ from those of $\Lambda^2 U$. – Liviu Nicolaescu Feb 4 '14 at 13:50

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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