A matrix model of a non-Hermitian \beta-ensemble

Let us take T\in \mathcal{T}'(n) and let

be the principal submatrix of T obtained by removing the last n-k rows and columns. Note that T_n=T. Let \lambda_j^{(k)}, j=1,\dotsc,k, denote the eigenvalues of T_k and write

The characteristic polynomials \Phi_k obey the recurrence relation

with initial conditions

and b_n undetermined. Equation (Proof) implies that \lambda_j =\lambda_j^{(n)} is an eigenvalue of T and that

is the right-eigenvector corresponding to \lambda_j. Furthermore, by definition, the matrix

belongs to R'(n). If there is no degeneracy in the spectrum, the eigenvectors are unique up to a multiplication by a constant. Since \Phi_0\bigl(\lambda_j\bigr)=1, it follows that r_j\neq 0 for j=1,\dotsc,n.

In order to show that v_j\neq 0, suppose that

Since b_{j}\neq 0, j=1,\dotsc,n-1, it follows by induction that all the entries of (Proof) are zero and that \det R = 0, which contradicts the assumption that the spectrum is simple.

We now want to show that

is a set of representatives of R'(n). Suppose that the elements of the first column of R^{-1} are all different from zero. Then, Eq. (2.1) follows from the identity RR^{-1}=I and from the fact that matrices in the same equivalent class of \mathcal{R}'(n) differ by the right multiplication R \mapsto RD, where D\in D(n). The proof that the elements of the first column of R^{-1} are all different from zero follows similar lines to those that we have used to prove Eq. (2.1).

Take the principal submatrix of T obtained by removing the first n-k rows and columns,

Let \mu_j^{(k)}, j=1,\dotsc,k, denote the eigenvalues of U_k and write

We have that \lambda_{j}= \mu^{(n)}_{j}. The polynomials \chi_k obey the recurrence relation

with initial conditions

and b_0 undetermined. Equation (?) and

imply that

is the left eigenvector corresponding to \lambda_j and is proportional to the j-th row of R^{-1}. The same arguments that led to the conclusion that \Phi_{n-1}\bigl(\lambda_j\bigr)\neq 0 give \chi_{n-1}\bigl(\lambda_j\bigr)\neq 0.

Take R from the set of representative of \mathcal{R}'(n) defined in Lemma 2.1. Let us write the matrix equations

as a set of vector equations:

for j=1,\dotsc,n, where we have adopted the notation (?). Let \mathbf{r}_1=\mathbf{r} and

We want to show that given \Lambda and \mathbf{r}_1^t with \mathbf{r}^t_1\mathbf{l}_1=1, where

we can reconstruct T, R and R^{-1} uniquely from the recurrence relations (Proof) with boundary conditions b_n=0 and \mathbf{l}_0=\mathbf{0}, \mathbf{r}_{n+1}^t=\mathbf{0}. The vectors \mathbf{r}_0^t and \mathbf{l}_{n+1} are undetermined.

Write

For any fixed \Lambda, the set exceptional points \mathbf{r}_1 such that b_{n-1}=0 has measure zero in \mathcal{H}_n and can be neglected. The quantities a_n, b_{n-1}, \mathbf{r}_1^t and \mathbf{l}_1 solve (Proof) for j=1. We need to check that they are consistent with the condition RR^{-1} = I. Eqs.(?) and (?) give \mathbf{r}_2^t \mathbf{l}_1=0. Similarly, by Eqs. (?) and (?) we have \mathbf{r}_1^t \mathbf{l}_2 =0. Finally, we have

Note that combining the orthogonality relations \mathbf{r}_1^t \mathbf{l}_2 = 0 and \mathbf{r}_2^t \mathbf{l}_2=1 with Eq. (?) gives

as it should be.

When 1<j<n take

subject to the conditions

Define

For any fixed \Lambda, we have b_{n-j}=0 for \mathbf{r}_{j+1} and \mathbf{l}_j belonging to a set of measure zero in {\mathbb C}^n. In turn, \mathbf{r}_{j+1} and \mathbf{l}_j are rational functions of the elements of \mathbf{r}_1; therefore, b_{n-j}=0 only for \mathbf{r}_1 belonging to a set of measure zero in \mathcal{H}_n. Combining (?) with (?) and proceeding in a similar way as when j=1, we can show that

Furthermore, we have

The orthogonality relations

follow by induction and from the fact that by construction \mathbf{r}^t_j \Lambda \mathbf{l}_k = 0 if \left| j-k \right|>1.

The boundary condition \mathbf{r}^t_{n+1}=\mathbf{0} terminates the recurrence relations.

One can prove by induction that

It follows that for any \psi \in \mathcal{P}_j[\mathbb{C}] we have

This lower bound is sharp if there exist a \psi\in \mathcal{P}_j[\mathbb{C}] such that

Note that the first j components of \widetilde{T}^j\mathbf{e}_1 contain only elements from \widetilde{T}_j, as do the vectors \widetilde{T}^k\mathbf{e}_1 for 1\le k<j. Let

Then, we have

where, with an abuse of notation, \mathbf{e}_1 on the left-hand side belongs to {\mathbb C}^n, while on the right-hand side \mathbf{e}_1\in {\mathbb C}^j. The Cayley-Hamilton theorem states that \chi_j(\widetilde{T}_j)=0. It follows that

This proves statements 1. and 2. The proof of 3. and 4. is analogous.

The definition of T and Proposition 3.2 give

for j=1,2,\dotsc,n-1. Using the spectral decomposition T = R\Lambda R^{-1} and the definition of R given in (?), Eqs. (Proof) become

Multiplying these equations side by side gives

which can be rewritten as

for j,k=0,\dotsc, n-1, where we have used \mathbf{l}_1^t = (1,\dotsc,1) and the fact that by definition \chi_0(z)=1.

Write

and

Eq. (Proof) is the matrix product

written in terms of matrix elements. Since \det(W) = \Delta(\boldsymbol{\lambda}), taking the determinant of both sides of this equation gives

Write

and take the Taylor expansion of both sides of this equation. By equating the coefficients of the powers of z up to z^{2n-1} we arrive at

The symbol * in Eqs. (Proof) indicates polynomials in which only the a_js and b_ks contained in lower order terms appear. Therefore, if we choose the ordering

the matrix of the coefficients of the differentials of the left-hand sides of Eqs. (Proof) is triangular. Taking the wedge product line by line gives

Differentiating the right-hand sides of Eqs. (Proof) leads to

where we have used the relation

which follows from the first equation in (Proof). Taking the wedge product of the right-hand side of Eq. (?) for j=1,\dotsc, 2n-1 we arrive at

The determinant in this equation can be calculated explicitly and is found as Eq. (1.175) in the book by Forrester [For10]. We have

Eqs. (?) and (?) give

which leads to (3.4).

The canonical vector

is cyclic for T; therefore,

are linearly independent. As a consequence, T^n\mathbf{e}_1 has a unique representation as a linear combination of the vectors (Proof). Equivalently, there exists a unique monic polynomial \chi_n of degree n such that \chi_n(T)\mathbf{e}_1=0. Furthermore, \chi_n(T) is the polynomial in T of least degree that annihilates \mathbf{e}_1. Next, let \phi be the minimal polynomial of T. It satisfies the following properties:

1. \phi(T)=0 \implies \phi(T)\mathbf{e}_1=0;

2. by the Calyley-Hamilton theorem, the degree of \phi cannot exceed n;

3. \phi is a divisor of the characteristic polynomial of T.

Thus, it follows that \chi_n = \phi is both the minimal and the characteristic polynomial of T. Thus, the spectrum of T is either simple, or, if \lambda is a multiple root of \chi_n, then it has only one block in the Jordan normal form of T.

Let \mu\neq\lambda_j, for all j=1,\dotsc,n, so that the resolvent (\mu I-\widetilde{T})^{-1} is well defined. We have

where \Delta(\mu)=\operatorname{diag}(\delta_1(\mu),\dotsc,\delta_n(\mu)) and

Taking the limit as \mu\rightarrow\lambda_j of both sides of \eqref{eq:appendix:limit} yields

where

The right hand side of Eq. (C.2) is obtained by computing the matrix elements of \operatorname{adj}\left(\lambda_jI-\widetilde{T}\right) using the Laplace expansion. The left hand side is given by Lemma C.1.