Every locally compact group has left-invariants haar measures. In particular, the compact groups O(n) and U(n) have them.

I was wondering if there is a realization of such a measure on these groups, or its integral operator. Of course, right invariant ones are as good.

  • $\begingroup$ Compact groups are unimodular (ie the left invariant measure is also right invariant) $\endgroup$ Sep 1 '13 at 2:49
  • $\begingroup$ Ok, but I still would like a concrete realization. $\endgroup$ Sep 1 '13 at 2:53
  • 2
    $\begingroup$ This thread may be interesting: mathoverflow.net/questions/36025/… $\endgroup$
    – John M
    Sep 1 '13 at 3:23
  • $\begingroup$ Wow! This thread is great John M $\endgroup$ Sep 1 '13 at 3:45

I recently gave an explicit parameterization of $SO(n)$ and corresponding explicit formula for the Haar measure as an answer to this question. Here I will do the same thing for $U(n)$. I highly recommend reading my other answer on $SO(n)$ before reading this one.

Let $S^{2n-1}$ denote the unit sphere in $\mathbb{C}^n$, and note that $U(n)$ acts transitively on $S^{2n-1}$. I will use this action give an explicit inductive description of a parameterization of $U(n)$. The idea will be to first perform an arbitrary unitary transformation of the first $n-1$ complex coordinates, and then perform a unitary transformation that maps $\textbf{e}_n$ to any possible location on $S^{2n-1}$.

For convenience, we will use the following notation. If $\textbf{v}\in\mathbb{C}^n$ is a vector, let $\textbf{v}^a\in\mathbb{C}^{n+1}$ be the vector obtained by augmenting $\textbf{v}$ with a zero, i.e. $$ (v_1,\ldots,v_n)^a \;=\; (v_1,\ldots,v_n,0). $$ Similarly, if $M$ is an $n\times n$ complex matrix, let $M^a$ be the $(n+1)\times(n+1)$ complex matrix with the following block diagonal form: $$ M^a \;=\; \begin{bmatrix}M & \textbf{0} \\ \textbf{0}^T & 1\end{bmatrix}. $$ We will also use the notation $\textbf{e}_1,\ldots,\textbf{e}_n$ for the standard basis vectors in $\mathbb{C}^n$.

Complex Hyperspherical Coordinates

These are a coordinate system for specifying a point $\boldsymbol{\Sigma}_n(\phi_1,\ldots,\phi_n,\theta_1,\ldots,\theta_{n-1})$ on $S^{2n-1}$ given $2n+1$ angles $\phi_1,\ldots,\phi_n,\theta_1,\ldots,\theta_{n-1}$. The first few complex hyperspherical coordinate systems are given by \begin{align*} \boldsymbol{\Sigma}_1(\phi_1) &\;=\; e^{i\phi_1}, \\[3pt] \boldsymbol{\Sigma}_2(\phi_1,\phi_2,\theta_1) &\;=\; (e^{i\phi_1}\sin\theta_1,\,e^{i\phi_2}\cos\theta_1), \\[3pt] \text{and }\boldsymbol{\Sigma}_3(\phi_1,\phi_2,\phi_3,\theta_1,\theta_2) &\;=\; (e^{i\phi_1}\sin\theta_1\sin\theta_2,\,e^{i\phi_2}\cos\theta_1\sin\theta_2,\,e^{i\phi_3}\cos\theta_2), \end{align*} and in general the $k$th complex Cartesian coordinate $\Sigma_{n,k}$ of $\boldsymbol{\Sigma}_n$ is given by the formula $$ \Sigma_{n,k}(\phi_1,\ldots,\phi_n,\theta_1,\ldots,\theta_{n-1}) \;=\; \begin{cases}e^{i\phi_1}\sin \theta_1 \cdots \sin \theta_{n-1} & \text{if } k=1, \\[3pt] e^{i\phi_k}\cos \theta_{k-1} \sin \theta_k \cdots \sin \theta_{n-1} & \text{if }2\leq k \leq n.\end{cases} $$ The function $\boldsymbol{\Sigma}_n$ can also be defined inductively by the formula $$ \boldsymbol{\Sigma}_n(\phi_1,\ldots,\theta_{n-1}) \;=\; (\sin \theta_{n-1})\,\bigl(\boldsymbol{\Sigma}_{n-1}(\phi_1,\ldots,\phi_{n-1},\theta_1,\ldots,\theta_{n-2})\bigr)^a \,+\, (e^{i\phi_n}\cos \theta_{n-1})\,\textbf{e}_n. $$ with base case $\boldsymbol{\Sigma}_1$.

Domain and Volume Form

If we let $D_n$ be the subset of the parameter space defined by $$ 0\leq\phi_k\leq 2\pi, \qquad\text{and}\qquad 0\leq \theta_k\leq \pi/2 $$ for all $k$, then $\boldsymbol{\Sigma}_n$ maps $D_n$ onto $S^{2n-1}$, and is one-to-one on the interior of $D_n$. The $(2n-1)$-dimensional volume form with respect to $\boldsymbol{\Sigma}_n$ is $$ dV \;=\; \left(\prod_{k=1}^{n-1} \cos \theta_k \sin^{2k-1} \theta_k\right)\,d\phi_1\cdots d\phi_n\,d\theta_1 \cdots d\theta_{n-1}. $$

An Orthonormal Basis

Before writing down the parameterization of $U(n)$, we need to extend $\{\boldsymbol{\Sigma}_n(\phi_1,\ldots,\phi_n,\theta_1,\ldots,\theta_{n-1})\}$ to an orthonormal basis of $\mathbb{C}^n$. The basis is $$ \bigl\{\textbf{U}_{n,1},\ldots,\textbf{U}_{n,n-1},\boldsymbol{\Sigma}_n\bigr\} $$ where $$ \textbf{U}_{n,k}(\phi_1,\ldots,\phi_n,\theta_1,\ldots,\theta_{n-1}) \;=\; \frac{1}{(\sin \theta_{k+1}) \cdots (\sin \theta_{n-1})}\frac{\partial \boldsymbol{\Sigma}(\phi_1,\ldots,\phi_n,\theta_1,\ldots,\theta_{n-1})}{\partial \theta_k}. $$ That is, $\textbf{U}_{n,k}$ is the unit vector tangent to $S^{2n-1}$ in the direction of increasing $\theta_k$.

The vectors $\textbf{U}_{n,k}(\phi_1,\ldots,\theta_{n-1})$ can also be defined inductively by the formula $$ \textbf{U}_{n,n-1}(\phi_1,\ldots,\theta_{n-1}) \;=\; (\cos \theta_{n-1})\,\bigl(\boldsymbol{\Sigma}_{n-1}(\phi_1,\ldots,\phi_{n-1},\theta_1,\ldots\theta_{n-2})\bigr)^a \,-\, (e^{i\phi_n}\sin \theta_{n-1})\,\textbf{e}_{n} $$ and $\textbf{U}_{n,k}(\phi_1,\ldots,\phi_n,\theta_1,\ldots,\theta_{n-1}) = \bigl(\textbf{U}_{n-1,k}(\phi_1,\ldots,\phi_{n-1},\theta_1,\ldots,\theta_{n-2})\bigr)^a$ for $k<n-1$.

Let $M_n(\phi_1,\ldots,\phi_n,\theta_1,\ldots,\theta_{n-1})$ denote the $n\times n$ unitary matrix whose columns are the vectors of this orthonormal basis: $$ M_n \;=\; \begin{bmatrix}\textbf{U}_{n,1} & \cdots & \textbf{U}_{n,n-1} & \boldsymbol{\Sigma}_n\end{bmatrix}. $$ So $M_n(\phi_1,\ldots,\phi_n,\theta_1,\ldots,\theta_{n-1})$ maps $\textbf{e}_n$ to an arbitrary point $\boldsymbol{\Sigma}_{n}(\phi_1,\ldots,\phi_n,\theta_1,\ldots,\theta_{n-1})$ on the unit $(2n-1)$-sphere.

Parameterization of $U(n)$

Our parameterization for $U(n)$ will be an inductively defined function $\Phi_n$, which will take the $n^2$ angles $\{\phi_{jk}\}_{1\leq j\leq k\leq n}\cup\{\theta_{jk}\}_{1\leq j < k\leq n}$ as input, and output an $n\times n$ matrix in $U(n)$. It is defined inductively by the rule $$ \Phi_1(\phi_{11}) \;=\; e^{i\phi_{11}}. $$ and $$ \Phi_n\bigl(\phi_{jk}\text{'s},\theta_{jk}\text{'s}\bigr) \;=\; M_n(\phi_{1,n},\ldots,\phi_{n,n},\theta_{1,n},\ldots,\theta_{n-1,n})\, \bigl(\Phi_{n-1}(\phi_{jk}\text{'s},\theta_{jk}\text{'s})\bigr)^a $$ where the product is a matrix product, and the inputs to $\Phi_{n-1}$ include all variables for which $k\leq n-1$. Conceptually, the second factor performs an arbitrary unitary transformation on the first $n-1$ coordinates, and then the first factor performs a specific unitary transformation that maps $\textbf{e}_n$ to an arbitrary point on $S^{2n-1}$.

Again, if we let $E_n$ be the subset of parameter space defined by $0\leq \phi_{jk}\leq 2\pi$ and $0\leq \theta_{jk} \leq \pi/2$, then $\Phi_n$ maps $E_n$ onto $U(n)$ and $\Phi_n$ is one-to-one on the interior of $E_n$.

The volume form on $U(n)$ corresponding to Haar measure is $$ dV \;=\; \left(\prod_{1\leq j < k \leq n} \cos \theta_{jk} \sin^{2j-1} \theta_{jk} \right) d\phi_{11} \cdots d\theta_{n-1,n}. $$ Note that this measure isn't normalized. Instead, the total volume of $U(n)$ is the product $$ \prod_{k=1}^{n} \mathrm{Vol}(S^{2k-1}), $$ where $\mathrm{Vol}(S^{2k-1})$ denotes the $(2k-1)$-dimensional volume (i.e. surface area) of the unit sphere in $\mathbb{C}^k$.

Some Examples

For $n=2$, we are parameterizing $U(2)$ using $4$ variables $\phi_{11},\phi_{12},\phi_{22},\theta_{12}$, where $0\leq\phi_{jk}\leq 2\pi$ and $0\leq \theta_{12} \leq \pi/2$. The parameteriztion $\Phi_2(\phi_{11},\phi_{12},\phi_{22},\theta_{12})$ is the following matrix product $$ \begin{bmatrix} e^{i\phi_{12}} \cos\theta_{12} & e^{i\phi_{12}}\sin \theta_{12} \\ -e^{i\phi_{22}}\sin\theta_{12} & e^{i\phi_{22}}\cos \theta_{12} \end{bmatrix} \begin{bmatrix}e^{i\phi_{11}} & 0 \\ 0 & 1\end{bmatrix}. $$ The volume form is $$ dV \;=\; \cos \theta_{12} \sin \theta_{12} \,d\phi_{11}\,d\phi_{12}\,d\phi_{22}\,d\theta_{12}, $$ and the total volume of $U(2)$ is $(2\pi)(2\pi^2) = 4\pi^3$.

For $n=3$, we are parameterizing $U(3)$ with nine parameters $\phi_{11},\phi_{12},\phi_{22},\phi_{13},\phi_{23},\phi_{33},\theta_{12},\theta_{13},\theta_{23}$, where $0\leq\phi_{jk}\leq 2\pi$ and $0\leq \theta_{jk} \leq \pi/2$. The parameterization $\Phi_3(\phi_{11},\phi_{12},\phi_{22},\phi_{13},\phi_{23},\phi_{33},\theta_{12},\theta_{13},\theta_{23})$ is the product of the matrix $$ \begin{bmatrix} e^{i\phi_{13}} \cos \theta_{13} & e^{i\phi_{13}} \sin \theta_{13} \cos \theta_{23} & e^{i\phi_{13}} \sin \theta_{13} \sin \theta_{23} \\ -e^{i\phi_{23}} \sin \theta_{13} & e^{i\phi_{23}} \cos \theta_{13} \cos \theta_{23} & e^{i\phi_{23}} \cos \theta_{13} \sin \theta_{23} \\ 0 & -e^{i\phi_{33}} \sin \theta_{23} & e^{i\phi_{33}} \cos \theta_{23} \end{bmatrix} $$ with $\begin{bmatrix}\Phi_2(\phi_{11},\phi_{12},\phi_{22},\theta_{12}) & \textbf{0} \\ \textbf{0}^T & 1\end{bmatrix}$. The volume form is $$ dV \;=\; \bigl(\cos\theta_{12}\sin \theta_{12}\bigr) \bigl(\cos\theta_{13}\sin \theta_{13}\bigr) \bigl(\cos \theta_{23}\sin^3 \theta_{23}\bigr)\,d\phi_{11}\cdots d\theta_{23}, $$ and the total volume of $U(3)$ is $(2\pi)(2\pi^2)(\pi^3) = 4\pi^6$.

  • $\begingroup$ can one divide this parametrization of a random matrix from U(N) by its determinant to get a random matrix from SU(N)? $\endgroup$ Mar 5 '20 at 0:58

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.