Introduction and summary
The one-channel white-noise Schrödinger operator is the textbook example of a disordered spectral problem allowing a very detailed statistical analysis for arbitrary disorder beyond perturbation theory. In particular, Frisch and Lloyd and then Halperin obtained its eigenvalue density in the infinite-length limit through Airy functions [35 , 52 ] , and Thouless [85 ] related that density to the Lyapunov exponent characterizing the localization length in the system. These two quantities are the real and imaginary parts of a single “complex characteristic exponent”, see e.g. [15 ] . Many early results have been reviewed in the book [65 ] , with subsequent work providing a detailed characterisation of eigenfunction statistics [61 , 56 , 57 ] . Complementary mathematical work has revealed a rich rigorous probabilistic picture of localized eigenfunctions and spectral fluctuations [22 , 23 , 24 , 26 , 25 ] .
Physically, strictly one-dimensional systems are highly idealised models, raising the natural question: how much of the solvability survives for an N N N -channel, matrix-valued microscopic Schrödinger operator which describes with growing N N N the gradual transition to quasi-one-dimensional rather than truly one-dimensional localization? Investigating this problem as fully as possible remains not only a challenging theoretical undertaking, but is also motivated by continuing experimental relevance of systems displaying multichannel localization, see e.g. [81 , 73 , 17 , 18 , 54 ]
In the recent paper Ref. [46 ] it was shown that in the unitary class β = 2 \beta= 2 β = 2 the N N N -channel problem in many aspects remains exactly integrable for any length L L L of the sample. The method of solution exploited the matrix Riccati variable and discovered that the associated Dyson gas of Riccati eigenvalues can be exactly fermionized. In the limit L → ∞ L \to\infty L → ∞ the fermionization produced the analogue of complex characteristic exponent exactly for every N N N , through an Airy Hankel τ \tau τ -function, which has been demonstrated to be equivalent to an outgoing-contour version of the Kontsevich matrix Airy integral [62 ] (the reason for calling the contour “outgoing” will be justified below). Such a representation turned out to be very powerful for analysing the associated Lifshitz tails and further helped to identify the large-N N N eigenvalue density edge profile conjectured to be universal across a class of β = 2 \beta= 2 β = 2 quasi-one-dimensional systems including random band matrices.
The mapping to Kontsevich integral through fermionization relied on cancellations specific to β = 2 \beta= 2 β = 2 and not available for the other two standard symmetry classes β = 1 , 4 \beta= 1, 4 β = 1 , 4 . The aim of the present paper is to suggest an alternative method which does not depend on fermionization and hence is equally applicable to all three classes, with the price of being implementable only in the limit of infinite sample length. The organizing object is the full stationary distribution of the Riccati matrix at complex energy . The Kontsevich integral then appears as the holomorphic factor of its normalization, most useful in the limit of real energy.
Complex energy is not just a mathematical device: E + i η E+i\eta E + i η describes a quasi-one-dimensional wire with uniform absorption. Physically, the Riccati matrix is the admittance matrix of a semi-infinite absorbing wire, whose probability distribution is given by our stationary law. The underlying geometry is a matrix analogue of the hyperbolic half-plane, expressed in terms of a Siegel domain. The connections with waveguide scattering, hyperbolic diffusion and transport theory are discussed in Appendix § 23 .
In the present paper, the first in a planned series, we concentrate mainly on the spectral characteristics of N N N -channel system of infinite length, namely its normalization factor and holomorphic sector, which yields the density of states and the sum of Lyapunov exponents for all classical values β = 1 , 2 , 4 \beta=1,2,4 β = 1 , 2 , 4 . Scattering applications of the full distribution are deferred to Paper II [44 ] . Appendix § 23 collects the relevant physical interpretation and boundary identities, and outlines that programme.
Outline of the main results
Let E = E + i η \mathcal{E}=E+i\eta E = E + i η with η > 0 \eta>0 η > 0 and F = R , C , H \mathbb{F}=\mathbb{R},\mathbb{C},\mathbb{H} F = R , C , H for β = 1 , 2 , 4 \beta=1,2,4 β = 1 , 2 , 4 . We consider ψ : [ 0 , L ] → F N \psi:[0,L]\to\mathbb{F}^{N} ψ : [ 0 , L ] → F N and the (quasi-)one-dimensional multichannel Schrödinger operator with Dirichlet boundary conditions:
H L = − ∂ x 2 I N + W ( x ) , H L ψ = E ψ , ψ ( 0 ) = ψ ( L ) = 0 , (1) \mathcal{H}_{L}=-\partial_{x}^{2}I_{N}+W(x),\qquad\mathcal{H}_{L}\psi=E\psi,\qquad\psi(0)=\psi(L)=0, \tag*{(1)} H L = − ∂ x 2 I N + W ( x ) , H L ψ = E ψ , ψ ( 0 ) = ψ ( L ) = 0 , (1)
where I N I_{N} I N is the N × N N\times N N × N identity matrix and W ( x ) = W ( x ) † W(x)=W(x)^{\dagger} W ( x ) = W ( x ) † is real symmetric, complex Hermitian or quaternion self-dual Gaussian white noise with the weight defined in (10) . Its variance is controlled by the parameter T = J 2 / ( 2 N ) > 0 T=J^{2}/(2N)>0 T = J 2 / ( 2 N ) > 0 .
Let Ψ ( x ) \Psi(x) Ψ ( x ) be the N × N N\times N N × N (F \mathbb{F} F -valued) fundamental solution of the N N N -channel Schrödinger equation: − Ψ ′ ′ + W Ψ = E Ψ -\Psi''+W\Psi=E\Psi − Ψ ′′ + W Ψ = E Ψ with Ψ ( 0 ) = 0 \Psi(0)=0 Ψ ( 0 ) = 0 , Ψ ′ ( 0 ) = I N \Psi'(0)=I_{N} Ψ ′ ( 0 ) = I N . The Riccati matrix Z = Ψ ′ Ψ − 1 Z=\Psi'\Psi^{-1} Z = Ψ ′ Ψ − 1 obeys
Z ′ = − E − Z 2 + W ( x ) . (2) Z'=-E-Z^{2}+W(x). \tag*{(2)} Z ′ = − E − Z 2 + W ( x ) . (2)
Write the Riccati matrix as Z = A − i P Z=A-iP Z = A − i P with A , P A,P A , P Hermitian of class β \beta β , and set Y = η P − 1 > 0 Y=\eta P^{-1}>0 Y = η P − 1 > 0 .
The stationary joint probability density (§ 14 ). In the limit of a semi-infinite sample, L → ∞ L\to\infty L → ∞ , the pair of matrices ( A , Y ) (A,Y) ( A , Y ) acquires a stationary joint probability density, independent of the initial conditions. It is given explicitly by
P E , η ( A , Y ) = 1 Z N ( E , η ) exp { 1 T [ Tr U E ( A − Y ) − Tr U E ( A ) − η 2 Tr Y − 1 ] } , U E ( λ ) = E λ + λ 3 3 . (3) \mathcal{P}_{E,\eta}(A,Y)=\frac{1}{Z_{N}(E,\eta)}\exp\left\{\frac{1}{T}\left[\operatorname{Tr}U_{E}(A-Y)-\operatorname{Tr}U_{E}(A)-\eta^{2}\operatorname{Tr}Y^{-1}\right]\right\},
\qquad
U_{E}(\lambda)=E\lambda+\frac{\lambda^{3}}{3}. \tag*{(3)} P E , η ( A , Y ) = Z N ( E , η ) 1 exp { T 1 [ Tr U E ( A − Y ) − Tr U E ( A ) − η 2 Tr Y − 1 ] } , U E ( λ ) = E λ + 3 λ 3 . (3)
defined for A ∈ Herm N ( F ) A\in\operatorname{Herm}_{N}(\mathbb{F}) A ∈ Herm N ( F ) and Y > 0 Y>0 Y > 0 . The normalization is fixed by ∫ P E , η ( A , Y ) d A d Y = 1 \int\mathcal{P}_{E,\eta}(A,Y)\,\mathrm{d}A\,\mathrm{d}Y=1 ∫ P E , η ( A , Y ) d A d Y = 1 , where d A \mathrm{d}A d A and d Y \mathrm{d}Y d Y are the flat measures, i.e. products of the differentials of all independent real components of A A A and Y Y Y . For N = 1 N=1 N = 1 it reduces to the explicit scalar density of Sec. 3.
Factorization (§ 15 ). Let s = − E / T 2 / 3 s=-E/T^{2/3} s = − E / T 2/3 and s = − E / T 2 / 3 \mathfrak{s}=-\mathcal{E}/T^{2/3} s = − E / T 2/3 , and let C − \mathcal{C}_{-} C − be the outgoing (see the explanation after (8) ) Airy contour, running from e i π / 3 ∞ e^{i\pi/3}\infty e iπ /3 ∞ to e i π ∞ e^{i\pi}\infty e iπ ∞ , which produces A − = Ai − i Bi \mathcal{A}_{-}=\operatorname{Ai}-i\operatorname{Bi} A − = Ai − i Bi . Define
τ N ( β ) ( s ) = ∫ w 1 ≺ w 2 ≺ ⋯ ≺ w N ∈ C − ∏ i < j ( w j − w i ) β ∏ j = 1 N e w j 3 / 3 − s w j d w j , (4) \tau_{N}^{(\beta)}(s)=\int_{w_{1}\prec w_{2}\prec\cdots\prec w_{N}\in\mathcal{C}_{-}}\prod_{i<j}(w_{j}-w_{i})^{\beta}\prod_{j=1}^{N}e^{w_{j}^{3}/3-sw_{j}}\,\mathrm{d}w_{j}, \tag*{(4)} τ N ( β ) ( s ) = ∫ w 1 ≺ w 2 ≺ ⋯ ≺ w N ∈ C − i < j ∏ ( w j − w i ) β j = 1 ∏ N e w j 3 /3 − s w j d w j , (4)
where ≺ \prec ≺ denotes the order along the contour.
A contour deformation, justified by the physical argument in § 15 , gives
Z N ( E , η ) = C N , β T N β , β ∣ τ N ( β ) ( s ) ∣ 2 . (5) Z_{N}(E,\eta)=C_{N,\beta}T^{N_{\beta,\beta}}\left|\tau_{N}^{(\beta)}(\mathfrak{s})\right|^{2}. \tag*{(5)} Z N ( E , η ) = C N , β T N β , β τ N ( β ) ( s ) 2 . (5)
Here C N , β C_{N,\beta} C N , β and κ N , β \kappa_{N,\beta} κ N , β are constants. In words, the normalization of the stationary distribution (3) is proportional to the squared modulus of the outgoing Kontsevich-type integral with Dyson index β \beta β , defined in (4) . Note that although we use it only for β = 1 , 2 , 4 \beta=1,2,4 β = 1 , 2 , 4 , the definition makes sense for every β > 0 \beta>0 β > 0 : for ordered points w j − w i ≠ 0 w_j-w_i\ne0 w j − w i = 0 , and we take the continuous branch of ( w j − w i ) β (w_j-w_i)^\beta ( w j − w i ) β on the ordered configuration space (on the deformed contour e i π / 3 R → 0 → − R e^{i\pi/3}R\to0\to-R e iπ /3 R → 0 → − R one has arg ( w j − w i ) ∈ [ π , 4 π / 3 ] \arg(w_j-w_i)\in[\pi,4\pi/3] arg ( w j − w i ) ∈ [ π , 4 π /3 ] ). The remaining global phase drops out of ∣ τ N ( β ) ∣ 2 \lvert\tau_N^{(\beta)}\rvert^2 ∣ τ N ( β ) ∣ 2 and of ∂ s log τ N ( β ) \partial_s\log\tau_N^{(\beta)} ∂ s log τ N ( β ) .
Holomorphic statistics
At fixed Y Y Y , the cubic terms in A A A cancel in the exponent of (3) , leaving a Gaussian density in A A A . Averaging e − i k Tr Z e^{-ik\operatorname{Tr}Z} e − ik Tr Z , with Z = A − i η Y − 1 Z=A-i\eta Y^{-1} Z = A − i η Y − 1 and k = q / T 1 / 3 ≥ 0 k=q/T^{1/3}\ge0 k = q / T 1/3 ≥ 0 , over this Gaussian distribution at fixed Y Y Y leaves an integral over Y Y Y of exactly the same form as the normalization integral Z N ( E , η ) \mathcal{Z}_N(E,\eta) Z N ( E , η ) as in (5) , but with shifted parameters
E ′ = E + i k T 2 , η ′ = η + k T 2 . E'=E+\frac{ikT}{2},\qquad\eta'=\eta+\frac{kT}{2}. E ′ = E + 2 ik T , η ′ = η + 2 k T .
so that ⟨ e − i k Tr Z ⟩ E , η = Z N ( E ′ , η ′ ) / Z N ( E , η ) \left\langle e^{-ik\operatorname{Tr}Z}\right\rangle_{E,\eta}=\mathcal{Z}_N(E',\eta')/\mathcal{Z}_N(E,\eta) ⟨ e − ik Tr Z ⟩ E , η = Z N ( E ′ , η ′ ) / Z N ( E , η ) , where ⟨ ⋅ ⟩ E , η \langle\cdot\rangle_{E,\eta} ⟨ ⋅ ⟩ E , η denotes averaging with respect to the stationary density (3) . Here the complex value of E ′ E' E ′ denotes analytic continuation of the normalization integral, not a change of the physical real energy. To apply the factorization (5) , write it as
Z N ( E , η ) ∝ τ N ( β ) ( − E + i η T 2 / 3 ) τ N ( β ) # ( − E − i η T 2 / 3 ) , τ N ( β ) # ( z ) : = τ N ( β ) ( z ˉ ) ‾ . \mathcal{Z}_N(E,\eta)\propto\tau_N^{(\beta)}\left(-\frac{E+i\eta}{T^{2/3}}\right)\tau_N^{(\beta)\#}\left(-\frac{E-i\eta}{T^{2/3}}\right),\qquad\tau_N^{(\beta)\#}(z):=\overline{\tau_N^{(\beta)}(\bar z)}. Z N ( E , η ) ∝ τ N ( β ) ( − T 2/3 E + i η ) τ N ( β ) # ( − T 2/3 E − i η ) , τ N ( β ) # ( z ) := τ N ( β ) ( z ˉ ) .
For real E E E and η \eta η this product is ∣ τ N ( β ) ( s ) ∣ 2 \lvert\tau_N^{(\beta)}(s)\rvert^2 ∣ τ N ( β ) ( s ) ∣ 2 ; for complex E ′ E' E ′ it is understood as the product of the two analytically continued factors. Since
E ′ + i η ′ = E + i η + i k T , E ′ − i η ′ = E − i η , E'+i\eta'=E+i\eta+ikT,\qquad E'-i\eta'=E-i\eta, E ′ + i η ′ = E + i η + ik T , E ′ − i η ′ = E − i η ,
the argument of the first factor shifts from s s s to s − i q s-iq s − i q , while the second factor remains unchanged and cancels in the ratio. Hence (§ 15 )
⟨ e − i q Tr Z / T 1 / 3 ⟩ E , η = τ N ( β ) ( s − i q ) τ N ( β ) ( s ) , q ≥ 0. (6) \left\langle e^{-iq\operatorname{Tr}Z/T^{1/3}}\right\rangle_{E,\eta}=\frac{\tau_N^{(\beta)}(s-iq)}{\tau_N^{(\beta)}(s)},\qquad q\ge0. \tag*{(6)} ⟨ e − i q Tr Z / T 1/3 ⟩ E , η = τ N ( β ) ( s ) τ N ( β ) ( s − i q ) , q ≥ 0. (6)
Real energy
As η → 0 + \eta\to0^+ η → 0 + , for every N ≥ 1 N\ge1 N ≥ 1 and β ∈ { 1 , 2 , 4 } \beta\in\{1,2,4\} β ∈ { 1 , 2 , 4 } ,
Ξ N ( β ) ( E ) : = Γ N ( E ) − i π ν N ( E ) = T 1 / 3 ∂ s log τ N ( β ) ( s ) ∣ s = − E / T 2 / 3 , ρ N ( E ) = 1 π T 1 / 3 Im ∂ s 2 log τ N ( β ) ( s ) . (7) \Xi_N^{(\beta)}(E):=\Gamma_N(E)-i\pi\nu_N(E)=T^{1/3}\left.\partial_s\log\tau_N^{(\beta)}(s)\right|_{s=-E/T^{2/3}},\qquad\rho_N(E)=\frac{1}{\pi T^{1/3}}\operatorname{Im}\partial_s^2\log\tau_N^{(\beta)}(s). \tag*{(7)} Ξ N ( β ) ( E ) := Γ N ( E ) − iπ ν N ( E ) = T 1/3 ∂ s log τ N ( β ) ( s ) s = − E / T 2/3 , ρ N ( E ) = π T 1/3 1 Im ∂ s 2 log τ N ( β ) ( s ) . (7)
Here ν N \nu_N ν N is the integrated density of states (IDOS) per unit length of the multichannel Schrödinger operator (1) , ρ N = ν N ′ \rho_N=\nu_N' ρ N = ν N ′ is the density of states, and Γ N = ∑ j = 1 N γ j \Gamma_N=\sum_{j=1}^{N}\gamma_j Γ N = ∑ j = 1 N γ j is the sum of the positive Lyapunov exponents.
In the zero-absorption limit, the stationary joint probability density takes the form of a constrained two-matrix model with weight e − Tr U E ( Λ ) / T e + Tr U E ( M ) / T e^{-\operatorname{Tr}U_E(\Lambda)/T}e^{+\operatorname{Tr}U_E(M)/T} e − Tr U E ( Λ ) / T e + Tr U E ( M ) / T and Λ − M > 0 \Lambda-M>0 Λ − M > 0 . Further integrating out M M M yields a matrix generalization of Halperin’s stationary Riccati density; the IDOS ν N \nu_N ν N and the sum of Lyapunov exponents Γ N \Gamma_N Γ N admit then explicit representations as ratios of positive integrals, Eqs. (68) and (65) .
Rescaling w j = − z j / T 1 / 3 w_j=-z_j/T^{1/3} w j = − z j / T 1/3 turns Eq. (4) into the partition function of a β \beta β -ensemble in the cubic potential U E U_E U E , with all integration variables on the outgoing contour Γ − : e − 2 π i / 3 ∞ → + ∞ \Gamma_-:e^{-2\pi i/3}\infty\to+\infty Γ − : e − 2 π i /3 ∞ → + ∞ :
Z N ( β ) ( E ) = ∫ Γ − , o r d e r e d ∏ i < j ( z j − z i ) β ∏ j = 1 N e − U E ( z j ) / T d z j ∝ τ N ( β ) ( − E / T 2 / 3 ) , (8) \mathcal{Z}_N^{(\beta)}(E)=\int_{\Gamma_-,\ \mathrm{ordered}}\prod_{i<j}(z_j-z_i)^\beta\prod_{j=1}^{N}e^{-U_E(z_j)/T}\,dz_j\ \propto\ \tau_N^{(\beta)}\left(-E/T^{2/3}\right), \tag*{(8)} Z N ( β ) ( E ) = ∫ Γ − , ordered i < j ∏ ( z j − z i ) β j = 1 ∏ N e − U E ( z j ) / T d z j ∝ τ N ( β ) ( − E / T 2/3 ) , (8)
Ξ N ( β ) ( E ) = − T ∂ E log Z N ( β ) ( E ) . \Xi_N^{(\beta)}(E)=-T\,\partial_E\log\mathcal{Z}_N^{(\beta)}(E). Ξ N ( β ) ( E ) = − T ∂ E log Z N ( β ) ( E ) .
The Riccati observables are thus obtained from the energy derivative of the free energy of a cubic β \beta β -ensemble: its formal Gibbs weight is not normalizable on the real axis, but its partition function becomes convergent on the outgoing contour. The positive physical normalization is proportional to ∣ τ N ( β ) ∣ 2 |\tau_N^{(\beta)}|^2 ∣ τ N ( β ) ∣ 2 , with τ N ( β ) \tau_N^{(\beta)} τ N ( β ) as its holomorphic factor.
Why “outgoing”. The name can be understood in several equivalent ways. In terms of waves, the cubic-exponential integral along Γ − \Gamma_- Γ − , after the rescaling w = − z / T 1 / 3 w=-z/T^{1/3} w = − z / T 1/3 , is proportional to the Airy combination A − = Ai − i Bi A_-=\operatorname{Ai}-i\operatorname{Bi} A − = Ai − i Bi . Whereas Ai \operatorname{Ai} Ai and Bi \operatorname{Bi} Bi are standing waves on the oscillatory side, Ai ∓ i Bi \operatorname{Ai}\mp i\operatorname{Bi} Ai ∓ i Bi are the two travelling waves, the Airy analogues of the Hankel functions H ( 1 , 2 ) = J ∓ i Y H^{(1,2)}=J\mp iY H ( 1 , 2 ) = J ∓ iY ; for x → + ∞ x\to+\infty x → + ∞ , A − ( − x ) ≃ − i π − 1 / 2 x − 1 / 4 e i ( 2 3 x 3 / 2 + π 4 ) A_-(-x)\simeq-i\pi^{-1/2}x^{-1/4}e^{i(\frac{2}{3}x^{3/2}+\frac{\pi}{4})} A − ( − x ) ≃ − i π − 1/2 x − 1/4 e i ( 3 2 x 3/2 + 4 π ) . Physically, the choice is made by absorption: at E + i η E+i\eta E + i η with η > 0 \eta>0 η > 0 , the wave entering the semi-infinite wire from its end is absorbed inside and never comes back. This is the radiation condition of scattering theory, i.e. the retarded prescription E + i 0 E+i0 E + i 0 . The conjugate contour Γ + \Gamma_+ Γ + gives the incoming combination A + = Ai + i Bi A_+=\operatorname{Ai}+i\operatorname{Bi} A + = Ai + i Bi and the advanced prescription E − i 0 E-i0 E − i 0 . In terms of observables, the two choices differ in the sign of the imaginary part of Ξ N = Γ N − i π ν N \Xi_N=\Gamma_N-i\pi\nu_N Ξ N = Γ N − iπ ν N . Only the outgoing one gives a positive integrated density of states, i.e. a winding current of Riccati eigenvalues running in the physical direction, from + ∞ +\infty + ∞ down to − ∞ -\infty − ∞ . Finally, from the point of view of the matrix integral, the cubic Gibbs weight e − Tr U ( E ) / T e^{-\operatorname{Tr}U(E)/T} e − Tr U ( E ) / T cannot be normalized on the real axis, and among the contours on which it can be, Γ − \Gamma_- Γ − is the one that runs out to infinity along the positive real axis from below, in the lower half-plane where Z Z Z lives; see Figure 1 . Its complex conjugate supplies the second factor of the positive normalization ∣ τ N ∣ 2 |\tau_N|^2 ∣ τ N ∣ 2 .
Figure 1. Decay sectors of e − λ 3 / ( 3 T ) e^{-\lambda^3/(3T)} e − λ 3 / ( 3 T ) (shaded) and a schematic outgoing contour Γ − \Gamma_- Γ − . The contour runs from infinity along arg λ = − 2 π / 3 \arg\lambda=-2\pi/3 arg λ = − 2 π /3 to infinity just below the positive real axis.
The tables below provides a compact summary of content of the most essential formulas of the paper.
Table 1. Summary. Here S = 4 ( − E ) 3 / 2 3 T S = \frac{4(-E)^{3/2}}{3T} S = 3 T 4 ( − E ) 3/2 is the Halperin action, K β ( N ) = 3 N 4 [ 4 β 2 N 2 − 9 β 2 N + 5 β 2 + 18 β N − 18 β + 20 ] K_\beta(N) = \frac{3N}{4}[4\beta^2N^2 - 9\beta^2N + 5\beta^2 + 18\beta N - 18\beta+ 20] K β ( N ) = 4 3 N [ 4 β 2 N 2 − 9 β 2 N + 5 β 2 + 18 β N − 18 β + 20 ] and c N ( β ) = − 21 β 2 N 2 − 60 β 2 N + 54 β N + 39 β 2 − 54 β + 20 96 c_N^{(\beta)} = -\frac{21\beta^2N^2-60\beta^2N+54\beta N+39\beta^2-54\beta+20}{96} c N ( β ) = − 96 21 β 2 N 2 − 60 β 2 N + 54 β N + 39 β 2 − 54 β + 20 .
Table 2. Lifshitz tails, large-N N N spectral edges, and the conjectured correspondence with random band matrices (Appendix § 22 and [46 ] ). Here ξ = ( 2 − λ ) W 4 / 5 \xi=(2-\lambda)W^{4/5} ξ = ( 2 − λ ) W 4/5 , σ β ( ξ ) \sigma_\beta(\xi) σ β ( ξ ) is the cumulative edge intensity of band matrices of bandwidth W W W , and δ σ β \delta\sigma_\beta δ σ β is σ β \sigma_\beta σ β with its allowed-side terms of the first band row removed. The β = 1 \beta=1 β = 1 entries agree with the band-matrix diagram expansion, the β = 4 \beta=4 β = 4 entries are predictions.
In the deep tail,
ν N ( β ) ( E ) ≃ N Γ ( 1 + β / 2 ) π Γ ( 1 + N β / 2 ) ( − E ) 1 / 2 ( 8 ( − E ) 3 / 2 T ) β ( N − 1 ) 2 e − 4 ( − E ) 3 / 2 3 T [ 1 + c N ( β ) T ( − E ) 3 / 2 + ⋯ ] . (9) \nu_N^{(\beta)}(E)\simeq\frac{N\Gamma(1+\beta/2)}{\pi\Gamma(1+N\beta/2)}(-E)^{1/2}\left(\frac{8(-E)^{3/2}}{T}\right)^{\frac{\beta(N-1)}{2}}e^{-\frac{4(-E)^{3/2}}{3T}}\left[1+c_N^{(\beta)}\frac{T}{(-E)^{3/2}}+\cdots\right]. \tag*{(9)} ν N ( β ) ( E ) ≃ π Γ ( 1 + N β /2 ) N Γ ( 1 + β /2 ) ( − E ) 1/2 ( T 8 ( − E ) 3/2 ) 2 β ( N − 1 ) e − 3 T 4 ( − E ) 3/2 [ 1 + c N ( β ) ( − E ) 3/2 T + ⋯ ] . (9)
The leading exponential factor is independent of β \beta β and of N N N at fixed T T T . The reason is that the optimal rare fluctuation (instanton) producing a deeply bound state has the rank-one form W ∗ ( x ) = w ∗ ( x ) u u † W_*(x)=w_*(x)uu^\dagger W ∗ ( x ) = w ∗ ( x ) u u † , where u u u is a unit vector in channel space. The bound-state wavefunction is proportional to u u u , so the binding problem reduces to the scalar one and has the same optimal-fluctuation action.
The dependence on N N N and β \beta β enters the leading prefactor through the freedom to orient u u u in channel space and through fluctuations about the optimal potential. Since u u † uu^\dagger u u † is unchanged by multiplying u u u by a sign, a complex phase, or a unit quaternion, the distinct orientations form R P N − 1 \mathbb{R}\mathbb{P}^{N-1} R P N − 1 , C P N − 1 \mathbb{C}\mathbb{P}^{N-1} C P N − 1 , or H P N − 1 \mathbb{H}\mathbb{P}^{N-1} H P N − 1 for β = 1 , 2 , 4 \beta= 1, 2, 4 β = 1 , 2 , 4 , respectively. Integration over this orientation space, together with the collective-coordinate Jacobians and transverse fluctuation determinants, reproduces the N N N - and β \beta β -dependent leading prefactor in Eq. (9) (§ 9 .
The mechanisms behind the approach
We now summarize the physical mechanisms behind these results.
At real energy, the eigenvalues λ 1 , … , λ N \lambda_1,\ldots,\lambda_N λ 1 , … , λ N of the Riccati matrix Z Z Z perform Dyson Brownian motion in the non-confining cubic potential U E ( λ ) = E λ + λ 3 / 3 U_E(\lambda)=E\lambda+\lambda^3/3 U E ( λ ) = E λ + λ 3 /3 . They run to − ∞ -\infty − ∞ and are reinjected at + ∞ +\infty + ∞ , with each such winding counting one level. The formal Gibbs weight ∣ Δ ( λ ) ∣ β e − ∑ i = 1 N U E ( λ i ) / T |\Delta(\lambda)|^\beta e^{-\sum_{i=1}^N U_E(\lambda_i)/T} ∣Δ ( λ ) ∣ β e − ∑ i = 1 N U E ( λ i ) / T , where Δ ( λ ) = ∏ i < j ( λ j − λ i ) \Delta(\lambda)=\prod_{i<j}(\lambda_j-\lambda_i) Δ ( λ ) = ∏ i < j ( λ j − λ i ) , is not normalizable.
In contrast, at complex energy, i.e. for η > 0 \eta>0 η > 0 , the Riccati matrix Z = A − i P Z=A-iP Z = A − i P remains in the Siegel half-space P > 0 P>0 P > 0 , so the real-energy passages through infinity are regularized. This confinement does not imply detailed balance: the stationary state supports circulating probability currents. Absorption also makes solutions driven by the same realization of the random potential contract towards one another, providing the mechanism for loss of memory of initial conditions and uniqueness of the stationary law.
An important role in finding the exact stationary solution is played by a hidden gradient structure. Namely, the variables ( A , Y = η P − 1 ) (A,Y=\eta P^{-1}) ( A , Y = η P − 1 ) , the noise acts only on A A A , while the equation for Y Y Y contains no noise term. The drift has the form ( D + A ) ∇ Φ (D+A)\nabla\Phi ( D + A ) ∇Φ , where D D D is the diffusion matrix, A A A is a constant antisymmetric matrix, and Φ \Phi Φ is the exponent in (3) . This structure gives a stationary density proportional to e Φ e^\Phi e Φ , with divergence-free probability current proportional to A ∇ e Φ A\nabla e^\Phi A ∇ e Φ and no probability flux through the boundary. The two identities underlying this representation of the drift are trace identities, valid for non-commuting matrices and for all three values of β \beta β .
The central role in the large-N N N analysis is played by revealing the role of Kontsevich integrals as normalizing factors of the stationary density. The product τ ( s ˉ ) τ ( s ) \tau(\bar{s})\tau(s) τ ( s ˉ ) τ ( s ) turn out to be a double matrix-contour integral of a Gibbs weight for one matrix and an anti-Gibbs weight for the other. The anti-Gibbs weight appears because U E U_E U E is odd. Changing to centre-of-mass and relative coordinates and completing the square in the centre-of-mass coordinate reproduces the stationary joint probability density (3) after a contour deformation. On the deformed contour, the Gibbs weight is e − Tr U E ( Z ) / T e^{-\operatorname{Tr}U_E(Z)/T} e − Tr U E ( Z ) / T , so holomorphic observables of the Riccati matrix depend only on the factor τ \tau τ . The outgoing contour is selected by the requirement that the contour integral and its complex conjugate reproduce the positive stationary probability density after deformation of the double integration contour.
After the revealing detour via complex energies one may come back to real energy, setting η → 0 + \eta\to0^+ η → 0 + . Then the stationary probability density approaches the real-energy stationary density with nonzero probability current. Taking this limit also yields the characteristic function of Tr Z \operatorname{Tr}Z Tr Z , whose one-sided expansion as q → 0 + q\to0^+ q → 0 + gives Γ N − i π ν N \Gamma_N-i\pi\nu_N Γ N − iπ ν N .
Finally, it is worth commenting on the mathematical status of the results. Most of what follows consists of exact calculations, done by hand or in exact symbolic algebra. Three steps go beyond that.
(i) Uniqueness of the stationary probability density. At finite absorption it follows from the exact contraction identity (43) , valid for each realization of the random potential, combined with an ergodic argument based on the loss of memory of initial conditions. At real energy the same loss of memory follows, at the physics level, from the positivity of all N N N Lyapunov exponents (§ 6 C). We do not give a complete mathematical argument for the present matrix dynamics in either case.
(ii) The factorization (5) rests on a statement about integration surfaces. We give a physics-level argument for it (§ 15 ): the deformed integration surface is an integer combination of products of single-matrix contours, and matching at high energy selects the outgoing one. (iii) The identification of the zero-absorption distribution (63) with the current-carrying stationary state at real energy is exact for β = 2 \beta= 2 β = 2 , see Appendix § 20 where it is shown algebraically that (63) coincides with the fermionic stationary state of Ref. [46 ] , so only the uniqueness question of (i) remains. For β = 1 , 4 \beta= 1,4 β = 1 , 4 it is argued at the physics level (§ 6 C): the limit η → 0 \eta\to0 η → 0 is followed through the Riccati poles, and the flux leaving at − ∞ -\infty − ∞ is checked to equal the flux reinjected at + ∞ +\infty + ∞ . The real-energy formulas (7), (65) and (68) do not depend on this identification.
Main reasoning line and layout of the paper
The derivation follows through the following steps:
explicit JPD P E , η ( A , Y ) ⟶ Z N ∝ ∣ τ N ( β ) ∣ 2 ⟶ Ξ N = T 1 / 3 ∂ s log τ N ( β ) , \text{explicit JPD }P_{E,\eta}(A,Y)\longrightarrow Z_N\propto\left|\tau_N^{(\beta)}\right|^2\longrightarrow\Xi_N=T^{1/3}\partial_s\log\tau_N^{(\beta)}, explicit JPD P E , η ( A , Y ) ⟶ Z N ∝ τ N ( β ) 2 ⟶ Ξ N = T 1/3 ∂ s log τ N ( β ) ,
and finally the evaluation of τ N ( β ) \tau_N^{(\beta)} τ N ( β ) , which completes this chain of reasoning. The outgoing Kontsevich integral is the central object of this construction: it follows from the stationary probability density of the complex Riccati matrix and provides all the explicit results for the holomorphic observables considered here. The Dyson parameter β \beta β enters only through the flat matrix integration measures.
The paper is organized as follows. Section 3 demonstrates the derivation for N = 1 N=1 N = 1 , where every formula is easy to check. § 14 –§ 6 extend the derivation to general ( N , β ) (N,\beta) ( N , β ) : the stationary probability density (§ 14 ), its normalization as an outgoing Kontsevich integral (§ 15 ), and the passage to real energy (§ 6 ). § 7 –§ 11 evaluate τ N ( β ) \tau_N^{(\beta)} τ N ( β ) : determinant and Pfaffian formulas at finite N N N , the high-energy genus expansion, the Lifshitz tail of the Schrödinger spectral density, the leading large-N N N (planar) limit, and the β = 4 \beta= 4 β = 4 and β = 1 \beta= 1 β = 1 spectral edges. § 17 summarizes the present paper and outlines Paper II. The appendices contain material that illustrates and completes the main line without being part of it: an algebraic comparison with the β = 2 \beta= 2 β = 2 fermionic stationary state (Appendix § 20 ), the complete even-N N N reduction for β = 1 \beta= 1 β = 1 (Appendix § 21 ), and the connection to intersection numbers, large-order asymptotics, and random surface gluings (Appendix § 24 ). Appendix § 23 collects the scattering interpretation and time-delay identities as preparation for Paper II.
Model, Riccati gas and observables
Three symmetry classes with a common normalization
We recall that W ( x ) = W ( x ) † W(x) = W(x)^{\dagger} W ( x ) = W ( x ) † is real symmetric, complex Hermitian or quaternion self-dual, matrix Gaussian white noise with weight explicitly defined as
P [ W ] ∝ exp [ − 1 4 T ∫ 0 L Tr W ( x ) 2 d x ] , T = J 2 2 N . (10) \mathcal{P}[W]\propto\exp\left[-\frac{1}{4T}\int_0^L\operatorname{Tr}W(x)^2\,\mathrm{d}x\right],\qquad T=\frac{J^2}{2N}. \tag*{(10)} P [ W ] ∝ exp [ − 4 T 1 ∫ 0 L Tr W ( x ) 2 d x ] , T = 2 N J 2 . (10)
For β = 4 \beta= 4 β = 4 we write quaternions as 2 × 2 2\times2 2 × 2 complex blocks, and Tr \operatorname{Tr} Tr denotes half the trace of the resulting 2 N × 2 N 2N\times2N 2 N × 2 N matrix (the “quaternion trace”), so that Tr I N = N \operatorname{Tr}I_N=N Tr I N = N in all three classes. Eq. (10) means, in all three classes each diagonal element W i i W_{ii} W ii is real, with ⟨ W i i ( x ) W i i ( x ′ ) ⟩ = 2 T δ ( x − x ′ ) \langle W_{ii}(x)W_{ii}(x')\rangle=2T\delta(x-x') ⟨ W ii ( x ) W ii ( x ′ )⟩ = 2 T δ ( x − x ′ ) . Further, each off-diagonal element W i j W_{ij} W ij (i < j i<j i < j ) has β \beta β real components, each with covariance ⟨ W i j ( x ) W i j ( x ′ ) ⟩ = β T δ ( x − x ′ ) \langle W_{ij}(x)W_{ij}(x')\rangle=\beta T\delta(x-x') ⟨ W ij ( x ) W ij ( x ′ )⟩ = β T δ ( x − x ′ ) . Hence ⟨ W i j ( x ) W i j ( x ′ ) ⟩ = β T δ ( x − x ′ ) \langle W_{ij}(x)W_{ij}(x')\rangle=\beta T\delta(x-x') ⟨ W ij ( x ) W ij ( x ′ )⟩ = β T δ ( x − x ′ ) . For β = 2 \beta= 2 β = 2 this is ⟨ W i j ( x ) W k l ( x ′ ) ⟩ = 2 T δ i l δ j k δ ( x − x ′ ) \langle W_{ij}(x)W_{kl}(x')\rangle=2T\delta_{il}\delta_{jk}\delta(x-x') ⟨ W ij ( x ) W k l ( x ′ )⟩ = 2 T δ i l δ j k δ ( x − x ′ ) , the convention of [46 ] . For β = 1 \beta= 1 β = 1 it is ⟨ W i j ( x ) W k l ( x ′ ) ⟩ = T ( δ i k δ j l + δ i l δ j k ) δ ( x − x ′ ) \langle W_{ij}(x)W_{kl}(x')\rangle=T(\delta_{ik}\delta_{jl}+\delta_{il}\delta_{jk})\delta(x-x') ⟨ W ij ( x ) W k l ( x ′ )⟩ = T ( δ ik δ j l + δ i l δ j k ) δ ( x − x ′ ) . For N = 1 N=1 N = 1 all three classes reduce to the scalar white noise ⟨ W ( x ) W ( x ′ ) ⟩ = 2 T δ ( x − x ′ ) \langle W(x)W(x')\rangle=2T\delta(x-x') ⟨ W ( x ) W ( x ′ )⟩ = 2 T δ ( x − x ′ ) of Halperin’s problem.
Matrix coordinates and the trace metric. To write the Fokker–Planck equation in components we use real coordinates on Herm N ( F ) \operatorname{Herm}_N(\mathbb{F}) Herm N ( F ) . Let e i j e_{ij} e ij denote the matrix unit with entries ( e i j ) k l = δ i k δ j l (e_{ij})_{kl}=\delta_{ik}\delta_{jl} ( e ij ) k l = δ ik δ j l . The standard real basis { e a } \{e_a\} { e a } consists of the N N N diagonal matrices e i i e_{ii} e ii together with, for each pair i < j i<j i < j , the β \beta β matrices e i j u + e j i u ˉ e_{ij}u+e_{ji}\bar{u} e ij u + e j i u ˉ , where u u u runs over the standard real basis of F \mathbb{F} F . Altogether there are d = N + β N ( N − 1 ) / 2 d = N + \beta N(N-1)/2 d = N + β N ( N − 1 ) /2 basis elements. They are orthogonal with respect to the trace metric ⟨ X , X ′ ⟩ = Tr X X ′ \langle X, X' \rangle= \operatorname{Tr}XX' ⟨ X , X ′ ⟩ = Tr X X ′ , but not all of the same length: g a : = Tr e a 2 g_a := \operatorname{Tr}e_a^2 g a := Tr e a 2 equals 1 for diagonal and 2 for off-diagonal basis elements.
These unequal lengths appear in two places, and they cancel. The first is the noise. In the real coordinates the white-noise matrix has covariance per unit length C a b = 2 T g a − 1 δ a b C_{ab} = 2Tg_a^{-1}\delta_{ab} C ab = 2 T g a − 1 δ ab . Different coordinates are therefore uncorrelated, and each diagonal entry diffuses with variance 2 T 2T 2 T per unit length, while each real component of an off-diagonal entry diffuses with variance T T T . The second is the gradient. We define the matrix gradient G = ∇ X Φ G = \nabla_X\Phi G = ∇ X Φ by δ Φ = Tr ( G δ X ) \delta\Phi= \operatorname{Tr}(G\,\delta X) δ Φ = Tr ( G δ X ) for every Hermitian variation δ X \delta X δ X . Expanding X = ∑ a X a e a X = \sum_a X_a e_a X = ∑ a X a e a and G = ∑ a G a e a G = \sum_a G_a e_a G = ∑ a G a e a and using Tr ( e a e b ) = g a δ a b \operatorname{Tr}(e_a e_b) = g_a\delta_{ab} Tr ( e a e b ) = g a δ ab , we find ∂ X a Φ = g a G a \partial_{X_a}\Phi= g_aG_a ∂ X a Φ = g a G a , so a coordinate derivative carries the same factor g a g_a g a that the diffusion coefficient divides out. Combining the two,
1 2 C a a ∂ X a Φ = T G a , with no summation over a . (11) \frac{1}{2}C_{aa}\,\partial_{X_a}\Phi= TG_a,\qquad\text{with no summation over }a. \tag*{(11)} 2 1 C aa ∂ X a Φ = T G a , with no summation over a . (11)
Thus, when the probability density is proportional to e Φ e^\Phi e Φ , the diffusive contribution to its probability current is − T e Φ ∇ A Φ -Te^\Phi\nabla_A\Phi − T e Φ ∇ A Φ in matrix notation, independently of the basis normalization. This compatibility between the noise covariance and the trace metric is the identity used repeatedly below. In eigenvalue–eigenvector coordinates, X = U diag ( λ 1 , … , λ N ) U † X = U\operatorname{diag}(\lambda_1,\ldots,\lambda_N)U^\dagger X = U diag ( λ 1 , … , λ N ) U † , the flat measure d X = ∏ a d X a dX = \prod_a dX_a d X = ∏ a d X a takes the form
d X ∝ ∣ Δ ( λ ) ∣ β ∏ j = 1 N d λ j d U H a a r . dX \propto|\Delta(\lambda)|^\beta\prod_{j=1}^{N}d\lambda_j\,dU_{\mathrm{Haar}}. d X ∝ ∣Δ ( λ ) ∣ β j = 1 ∏ N d λ j d U Haar .
Kramers degeneracy (β = 4 \beta= 4 β = 4 ). Every eigenvalue of the Schrödinger operator (1) (which we will call energy levels to distinguish from eigenvalues of the Riccati matrix) is then doubly degenerate. Throughout we count Kramers pairs : ν N \nu_N ν N is the number of distinct levels per unit length, and Γ N = ∑ j = 1 N γ j \Gamma_N = \sum_{j=1}^{N}\gamma_j Γ N = ∑ j = 1 N γ j runs over the N N N distinct positive exponents. With this convention the N = 1 N = 1 N = 1 problem is literally Halperin’s in all three classes.
For β = 4 \beta= 4 β = 4 , logarithmic determinants are normalized consistently with this counting convention: log det Ψ \log\det\Psi log det Ψ denotes 1 2 log det C Ψ \frac{1}{2}\log\det_{\mathbb{C}}\Psi 2 1 log det C Ψ in the 2 N × 2 N 2N \times2N 2 N × 2 N complex representation, with the logarithm continued along the Riccati evolution. Thus ∂ x log det Ψ = Tr Z \partial_x\log\det\Psi= \operatorname{Tr}Z ∂ x log det Ψ = Tr Z in all three classes.
Riccati matrix and its eigenvalue gas
At real energies E E E , conservation of the Wronskian Ψ † Ψ ′ − Ψ ′ † Ψ \Psi^\dagger\Psi' - \Psi'^\dagger\Psi Ψ † Ψ ′ − Ψ ′ † Ψ , see (37) , which vanishes at x = 0 x = 0 x = 0 , implies that Z = Ψ ′ Ψ − 1 Z = \Psi'\Psi^{-1} Z = Ψ ′ Ψ − 1 is Hermitian wherever Ψ \Psi Ψ is invertible. Its N N N real eigenvalues (counting each Kramers pair once for β = 4 \beta= 4 β = 4 ) are ordered as λ 1 < ⋯ < λ N \lambda_1 < \cdots< \lambda_N λ 1 < ⋯ < λ N . Their dynamics follows from second-order perturbation theory, as for Dyson’s Brownian motion [27 ] . Viewing the spatial coordinate x x x as “time”, the Riccati matrix receives the increment − ( E + Z 2 ) d x + d W -(E+Z^2)\,dx+dW − ( E + Z 2 ) d x + d W over d x dx d x . In the eigenbasis of Z Z Z , the eigenvalue increments satisfy
d λ i = − ( E + λ i 2 ) d x + d W i i + ∑ j ≠ i ∣ d W i j ∣ 2 λ i − λ j d\lambda_i = -(E+\lambda_i^2)\,dx+dW_{ii}+\sum_{j\ne i}\frac{|dW_{ij}|^2}{\lambda_i-\lambda_j} d λ i = − ( E + λ i 2 ) d x + d W ii + j = i ∑ λ i − λ j ∣ d W ij ∣ 2
and ∣ d W i j ∣ 2 = β T d x |dW_{ij}|^2 = \beta T\,dx ∣ d W ij ∣ 2 = β T d x . Hence
d λ i = [ − ( E + λ i 2 ) + β T ∑ j ≠ i 1 λ i − λ j ] d x + 2 T d B i , (12) d\lambda_i = \left[-(E+\lambda_i^2)+\beta T\sum_{j\ne i}\frac{1}{\lambda_i-\lambda_j}\right]dx+\sqrt{2T}\,dB_i, \tag*{(12)} d λ i = − ( E + λ i 2 ) + β T j = i ∑ λ i − λ j 1 d x + 2 T d B i , (12)
This is Dyson’s gas at inverse temperature β \beta β in the cubic potential U E ( λ ) = E λ + λ 3 / 3 U_E(\lambda)=E\lambda+\lambda^3/3 U E ( λ ) = E λ + λ 3 /3 , with drift − U E ′ ( λ i ) -U'_E(\lambda_i) − U E ′ ( λ i ) . It is exactly the same β \beta β -Dyson gas which appeared in Allez and Dumaz [1 ] without any reference to underlying matrix Schrödinger equation; the same model also underlies the large-deviation analysis of Ref. [41 ] . Its finite-N N N stationary density with reinjection is constructed for every β > 0 \beta> 0 β > 0 in Ref. [42 ] . Treatment of this gas displays the following subtlety: the cubic potential does not confine. The drift − λ 2 -\lambda^2 − λ 2 sends the smallest eigenvalue to − ∞ -\infty − ∞ in finite “time” x x x . This “explosion” is however benign: in the projective (Prüfer) variable λ = cot θ \lambda= \cot\theta λ = cot θ the point λ = − ∞ \lambda= -\infty λ = − ∞ is identified with λ = + ∞ \lambda= +\infty λ = + ∞ , and the eigenvalue simply reappears at the top of the spectrum. In the language of the F N \mathbb{F}^N F N -valued solutions, det Ψ \det\Psi det Ψ passes through zero. By matrix oscillation theory [3 ] each such event counts one Dirichlet level below E E E . The Riccati eigenvalue gas is thus a gas on a circle, with a nonzero stationary current around it. Such a current has profound consequences, see more on this in [46 ] .
For a confining potential the stationary probability density of Dyson’s diffusion (12) would be the Gibbs distribution ∣ Δ ( λ ) ∣ β ∏ j e − U E ( λ j ) / T |\Delta(\lambda)|^\beta\prod_j e^{-U_E(\lambda_j)/T} ∣Δ ( λ ) ∣ β ∏ j e − U E ( λ j ) / T , with Δ ( λ ) = ∏ i < j ( λ j − λ i ) \Delta(\lambda) = \prod_{i<j}(\lambda_j-\lambda_i) Δ ( λ ) = ∏ i < j ( λ j − λ i ) . Here it is not normalizable: e − λ 3 / 3 T e^{-\lambda^3/3T} e − λ 3 /3 T blows up as λ → − ∞ \lambda\to-\infty λ → − ∞ . The true stationary state P s t P_{\mathrm{st}} P st instead carries a probability current J s t w i n d J_{\mathrm{st}}^{\mathrm{wind}} J st wind of Riccati eigenvalues flowing around the projective circle. The mean integrated density of states ν N ( E ) \nu_N(E) ν N ( E ) for the random Schrödinger operator equals the stationary winding current and determines the asymptotic tails of the mean Riccati eigenvalue density, see Ref. [46 ] :
ν N ( E ) = J s t w i n d , n ( λ ) : = ⟨ ∑ j δ ( λ − λ j ) ⟩ s t ≃ ν N ( E ) λ 2 ( λ → ± ∞ ) . (13) \nu_N(E) = J_{\mathrm{st}}^{\mathrm{wind}}, \qquad n(\lambda) := \left\langle\sum_j \delta(\lambda-\lambda_j) \right\rangle_{\mathrm{st}} \simeq\frac{\nu_N(E)}{\lambda^2} \qquad(\lambda\to\pm\infty). \tag*{(13)} ν N ( E ) = J st wind , n ( λ ) := ⟨ j ∑ δ ( λ − λ j ) ⟩ st ≃ λ 2 ν N ( E ) ( λ → ± ∞ ) . (13)
Indeed, for large ∣ λ ∣ |\lambda| ∣ λ ∣ the Riccati eigenvalue motion is dominated by the drift, d λ / d x ≃ − λ 2 \mathrm{d}\lambda/\mathrm{d}x \simeq-\lambda^2 d λ / d x ≃ − λ 2 ; the interaction and diffusion terms give corrections of relative order ∣ λ ∣ − 3 |\lambda|^{-3} ∣ λ ∣ − 3 . For large ∣ λ ∣ |\lambda| ∣ λ ∣ , the mean number of Riccati eigenvalues crossing the value λ \lambda λ per unit length is therefore ∣ d λ / d x ∣ n ( λ ) = λ 2 n ( λ ) |\mathrm{d}\lambda/\mathrm{d}x|\,n(\lambda) = \lambda^2 n(\lambda) ∣ d λ / d x ∣ n ( λ ) = λ 2 n ( λ ) . In the stationary state this flux does not depend on λ \lambda λ , and every such crossing is a passage through infinity, so it equals the winding rate ν N ( E ) \nu_N(E) ν N ( E ) . Reinjection at + ∞ +\infty + ∞ after escape to − ∞ -\infty − ∞ gives the same flux, and hence the same tail coefficient, at both ends of the real axis.
For β = 2 \beta= 2 β = 2 the paper [46 ] solved the current-carrying problem exactly. The gauge transformation by the Vandermonde determinant removes the Calogero interaction and, thanks to an identity special to the quadratic drift, the system can be mapped to free fermions. For β ≠ 2 \beta\ne2 β = 2 , the same transformation leaves an inverse-square interaction with coupling proportional to β ( β − 2 ) \beta(\beta-2) β ( β − 2 ) . The transformed problem therefore remains an interacting many-particle problem in a quartic single-particle potential, rather than reducing to free fermions.
What replaces the Gibbs state is found by going off the real axis. At complex energy the Riccati matrix becomes a complex matrix with a negative imaginary part, the state space doubles in dimension, and the stationary distribution of the enlarged dynamical system is explicit for every β \beta β (Secs. 3 and § 14 ). The real-energy state is then recovered as a limit (§ 6 ), Eq. (63) .
Observables and the complex characteristic exponent
Let Ψ ( x ) \Psi(x) Ψ ( x ) be an N × N N \times N N × N matrix of independent Schrödinger solutions, with generic initial data, and let Z = Ψ ′ Ψ − 1 Z = \Psi'\Psi^{-1} Z = Ψ ′ Ψ − 1 wherever Ψ \Psi Ψ is invertible. Jacobi’s formula gives, wherever Ψ ( x ) \Psi(x) Ψ ( x ) is invertible,
∂ x log ∣ det Ψ ( x ) ∣ = Re Tr ( Ψ ′ ( x ) Ψ ( x ) − 1 ) = Tr Z ( x ) = ∑ i = 1 N λ i ( x ) , \partial_x \log|\det\Psi(x)| = \operatorname{Re}\operatorname{Tr}\left(\Psi'(x)\Psi(x)^{-1}\right) = \operatorname{Tr} Z(x) = \sum_{i=1}^{N} \lambda_i(x), ∂ x log ∣ det Ψ ( x ) ∣ = Re Tr ( Ψ ′ ( x ) Ψ ( x ) − 1 ) = Tr Z ( x ) = i = 1 ∑ N λ i ( x ) ,
where we used the Hermiticity of Z Z Z at real energy. Thus the trace of the Riccati matrix is the local logarithmic growth rate of the volume spanned by the columns of Ψ \Psi Ψ . For generic initial conditions, the long-distance growth rate of this volume is the sum Γ N ( E ) \Gamma_N(E) Γ N ( E ) of the N N N positive Lyapunov exponents.
This identity requires a prescription at the Riccati windings. At a simple zero x ∗ x_* x ∗ of det Ψ \det\Psi det Ψ , one has det Ψ ( x ) ∼ c ( x − x ∗ ) \det\Psi(x) \sim c(x-x_*) det Ψ ( x ) ∼ c ( x − x ∗ ) , and hence Tr Z ( x ) ∼ ( x − x ∗ ) − 1 \operatorname{Tr} Z(x) \sim(x-x_*)^{-1} Tr Z ( x ) ∼ ( x − x ∗ ) − 1 . The divergent contributions on the two sides of the zero cancel in the principal-value sense. Accordingly, integration of the logarithmic derivative through the windings is understood with this regularization.
In the stationary ergodic regime, the regularized spatial average can be expressed through the one-point density
n ( λ ) = ⟨ ∑ i = 1 N δ ( λ − λ i ) ⟩ , ∫ R n ( λ ) d λ = N . n(\lambda)=\left\langle\sum_{i=1}^{N}\delta(\lambda-\lambda_i)\right\rangle,\qquad\int_{\mathbb{R}} n(\lambda)\,\mathrm{d}\lambda=N. n ( λ ) = ⟨ i = 1 ∑ N δ ( λ − λ i ) ⟩ , ∫ R n ( λ ) d λ = N .
With the principal-value prescription inherited from the Riccati windings, this yields
Γ N ( E ) = lim L → ∞ 1 L log ∣ det Ψ ( L ) ∣ = PV ∫ R λ n ( λ ) d λ (14) \Gamma_N(E)=\lim_{L\to\infty}\frac{1}{L}\log\left|\det\Psi(L)\right|=\operatorname{PV}\int_{\mathbb{R}}\lambda n(\lambda)\,\mathrm{d}\lambda \tag*{(14)} Γ N ( E ) = L → ∞ lim L 1 log ∣ det Ψ ( L ) ∣ = PV ∫ R λn ( λ ) d λ (14)
where
PV ∫ R λ n ( λ ) d λ : = lim R → ∞ ∫ − R R λ n ( λ ) d λ . \operatorname{PV}\int_{\mathbb{R}}\lambda n(\lambda)\,\mathrm{d}\lambda:=\lim_{R\to\infty}\int_{-R}^{R}\lambda n(\lambda)\,\mathrm{d}\lambda. PV ∫ R λn ( λ ) d λ := R → ∞ lim ∫ − R R λn ( λ ) d λ .
The principal value is essential: the stationary Riccati density has tails associated with the flux through infinity, so its first moment need not exist as an absolutely convergent integral. Both IDOS ν N \nu_N ν N and Γ N \Gamma_N Γ N are contained in the one-sided characteristic function of the Riccati trace,
F N ( q ; E ) = ⟨ e − i q Tr Z / T 1 / 3 ⟩ s t , q ≥ 0. (15) \mathcal{F}_N(q;E)=\left\langle e^{-\mathrm{i}q\operatorname{Tr}Z/T^{1/3}}\right\rangle_{\mathrm{st}},\qquad q\geq0. \tag*{(15)} F N ( q ; E ) = ⟨ e − i q Tr Z / T 1/3 ⟩ st , q ≥ 0. (15)
At this point it is worth discussing the subtlety of this characteristic function. A large value of Tr Z \operatorname{Tr}Z Tr Z comes from a single large eigenvalue, so by (13) the distribution of Tr Z \operatorname{Tr}Z Tr Z inherits the tails ν N / λ 2 \nu_N/\lambda^2 ν N / λ 2 of n ( λ ) n(\lambda) n ( λ ) . Hence Tr Z \operatorname{Tr}Z Tr Z has no absolutely convergent first moment, and F N \mathcal{F}_N F N is not differentiable at q = 0 q=0 q = 0 . Its one-sided expansion is nevertheless meaningful and well-defined. At first order in q q q only configurations with one large eigenvalue matter, so 1 − F N ≃ ∫ n ( λ ) ( 1 − e − i q λ / T 1 / 3 ) d λ 1-\mathcal{F}_N\simeq\int n(\lambda)(1-e^{-\mathrm{i}q\lambda/T^{1/3}})\,\mathrm{d}\lambda 1 − F N ≃ ∫ n ( λ ) ( 1 − e − i q λ / T 1/3 ) d λ . The even part of the tail gives ν N ∫ ( 1 − cos u ) u − 2 d u = π ν N \nu_N\int(1-\cos u)u^{-2}\,\mathrm{d}u=\pi\nu_N ν N ∫ ( 1 − cos u ) u − 2 d u = π ν N per unit q / T 1 / 3 q/T^{1/3} q / T 1/3 . The odd part, thanks to the equal tails at ± ∞ \pm\infty ± ∞ in (13) , reduces to the principal-value mean. Thus, at the physics level,
F N ( q ; E ) = 1 − q T 1 / 3 [ π ν N + i Γ N ] + o ( q ) , q → 0 + . (16) \mathcal{F}_N(q;E)=1-\frac{q}{T^{1/3}}\left[\pi\nu_N+\mathrm{i}\Gamma_N\right]+o(q),\qquad q\to0^+. \tag*{(16)} F N ( q ; E ) = 1 − T 1/3 q [ π ν N + i Γ N ] + o ( q ) , q → 0 + . (16)
Any closed formula for F N \mathcal{F}_N F N therefore gives a possibility to evaluate the complex characteristic exponent Ξ N : = Γ N − i π ν N \Xi_N:=\Gamma_N-\mathrm{i}\pi\nu_N Ξ N := Γ N − i π ν N . This step is identical for all β \beta β .
The complex characteristic exponent Ξ N \Xi_N Ξ N has a direct spectral interpretation. For Im E > 0 \operatorname{Im}\mathcal{E}>0 Im E > 0 ,
Ξ N ( E ) = lim L → ∞ 1 L log det Ψ ( L ; E ) , \Xi_N(\mathcal{E})=\lim_{L\to\infty}\frac{1}{L}\log\det\Psi(L;\mathcal{E}), Ξ N ( E ) = L → ∞ lim L 1 log det Ψ ( L ; E ) ,
with the branch of the logarithm fixed by continuation along x x x . Up to the normalization associated with the chosen determinant regularization, det Ψ ( L ; E ) \det\Psi(L;\mathcal{E}) det Ψ ( L ; E ) is the Dirichlet spectral determinant of H L − E \mathcal{H}_L-\mathcal{E} H L − E , as follows from the matrix Gelfand–Yaglom construction [32 , 76 ] . Thus Ξ N \Xi_N Ξ N is the self-averaging logarithm of the spectral determinant per unit length, or equivalently a spectral free energy. At the real-energy boundary, − π Im Ξ N ( E + i 0 ) = ν N ( E ) -\pi\operatorname{Im}\Xi_N(E+\mathrm{i}0)=\nu_N(E) − π Im Ξ N ( E + i 0 ) = ν N ( E ) counts the levels below E E E per unit length, while Re Ξ N ( E + i 0 ) = Γ N ( E ) \operatorname{Re}\Xi_N(E+\mathrm{i}0)=\Gamma_N(E) Re Ξ N ( E + i 0 ) = Γ N ( E ) . The two are not independent. Since det Ψ ( L ; E ) \det\Psi(L;\mathcal{E}) det Ψ ( L ; E ) is the spectral determinant, ∂ E Ξ N \partial_{\mathcal{E}}\Xi_N ∂ E Ξ N is the mean trace of the resolvent per unit length,
∂ E Ξ N ( E ) = ∫ ρ N ( E ′ ) d E ′ E − E ′ , Im E > 0. (17) \partial_{\mathcal{E}}\Xi_N(\mathcal{E})=\int\frac{\rho_N(E')\,\mathrm{d}E'}{\mathcal{E}-E'},\qquad\operatorname{Im}\mathcal{E}>0. \tag*{(17)} ∂ E Ξ N ( E ) = ∫ E − E ′ ρ N ( E ′ ) d E ′ , Im E > 0. (17)
Integrating once gives the multichannel Thouless formula [85 , 63 ] ,
Γ N ( E ) − Γ N ( E 0 ) = ∫ log ∣ E − E ′ E 0 − E ′ ∣ ρ N ( E ′ ) d E ′ . \Gamma_N(E)-\Gamma_N(E_0)=\int\log\left|\frac{E-E'}{E_0-E'}\right|\rho_N(E')\,\mathrm{d}E'. Γ N ( E ) − Γ N ( E 0 ) = ∫ log E 0 − E ′ E − E ′ ρ N ( E ′ ) d E ′ .
i.e. the sum of the positive Lyapunov exponents is the logarithmic potential of the density of states. The subtraction at a reference energy E 0 E_0 E 0 is needed because ν N ( E ′ ) ≃ N E ′ / π \nu_N(E') \simeq N\sqrt{E'}/\pi ν N ( E ′ ) ≃ N E ′ / π grows at high energy. Equivalently,
Γ N ′ ( E ) = PV ∫ R ρ N ( E ′ ) E − E ′ d E ′ \Gamma'_N(E)=\operatorname{PV}\int_{\mathbb{R}}\frac{\rho_N(E')}{E-E'}\,\mathrm{d}E' Γ N ′ ( E ) = PV ∫ R E − E ′ ρ N ( E ′ ) d E ′
so Γ N \Gamma_N Γ N and − π ν N -\pi\nu_N − π ν N form a Kramers–Kronig pair: they are the real and imaginary parts of the boundary value of a single function holomorphic in Im E > 0 \operatorname{Im}E>0 Im E > 0 , with the dispersion relation understood with the appropriate high-energy subtraction.
Warm-up: the scalar case N = 1 N=1 N = 1 in detail
For one channel all three β = 1 , 2 , 4 \beta=1,2,4 β = 1 , 2 , 4 classes coincide, and the final real-energy answer is one of Halperin [52 ] . It is however instructive to go through the complete construction for N = 1 N=1 N = 1 for two reasons. First, every formula is elementary and can be checked by hand, including the full two-dimensional stationary probability density at complex energy, which seems to be new as a closed formula. Second, the general case (§ 14 –§ 6 ) repeats the same six steps in the same order. Only two things change: the variables become matrices, and β \beta β enters through flat matrix measures. Table 3 in Section 3G gives the dictionary.
Table 3. The six steps for N = 1 N=1 N = 1 and their counterparts for general ( N , β ) (N,\beta) ( N , β ) .
For N = 1 N=1 N = 1 we write the Riccati variable at complex energy E + i η E+\mathrm{i}\eta E + i η , η > 0 \eta>0 η > 0 , in terms of the real variables u u u and v v v as
Z = T 1 / 3 ( u − i v ) , ϵ = E T 2 / 3 , a = η T 2 / 3 , t = T 1 / 3 x . (18) Z=T^{1/3}(u-\mathrm{i}v),\qquad\epsilon=\frac{E}{T^{2/3}},\qquad a=\frac{\eta}{T^{2/3}},\qquad t=T^{1/3}x. \tag*{(18)} Z = T 1/3 ( u − i v ) , ϵ = T 2/3 E , a = T 2/3 η , t = T 1/3 x . (18)
Then d Z = − ( E + i η + Z 2 ) d x + 2 T d B x \mathrm{d}Z=-(E+\mathrm{i}\eta+Z^2)\,\mathrm{d}x+\sqrt{2T}\,\mathrm{d}B_x d Z = − ( E + i η + Z 2 ) d x + 2 T d B x becomes
d u = ( − ϵ − u 2 + v 2 ) d t + 2 d B t , d v = ( a − 2 u v ) d t . (19) \mathrm{d}u=(-\epsilon-u^2+v^2)\,\mathrm{d}t+\sqrt{2}\,\mathrm{d}B_t,\qquad\mathrm{d}v=(a-2uv)\,\mathrm{d}t. \tag*{(19)} d u = ( − ϵ − u 2 + v 2 ) d t + 2 d B t , d v = ( a − 2 uv ) d t . (19)
All densities in this section are dimensionless. We also use s = − ϵ − i a = − ( E + i η ) / T 2 / 3 \mathfrak{s}=-\epsilon-\mathrm{i}a=-(E+\mathrm{i}\eta)/T^{2/3} s = − ϵ − i a = − ( E + i η ) / T 2/3 , U ϵ ( u ) = ϵ u + u 3 / 3 U_\epsilon(u)=\epsilon u+u^3/3 U ϵ ( u ) = ϵ u + u 3 /3 and A ± = Ai ± i Bi \mathcal{A}_{\pm}=\operatorname{Ai}\pm\mathrm{i}\operatorname{Bi} A ± = Ai ± i Bi .
Step 1: complex energy, absorption and contraction
Confinement and absence of poles. At v = 0 v=0 v = 0 , Eq. (19) gives d v = a d t > 0 \mathrm{d}v=a\,\mathrm{d}t>0 d v = a d t > 0 , so the dynamics preserves v > 0 v>0 v > 0 and Z Z Z remains in the lower half-plane, provided Z Z Z stays finite. At real energy Z Z Z can blow up at finite x x x : these Riccati poles, where Z Z Z escapes to − ∞ -\infty − ∞ and re-enters from + ∞ +\infty + ∞ , are the windings of § 14 . To show that at η > 0 \eta>0 η > 0 no such pole occurs at any finite position x > 0 x>0 x > 0 along the wire, consider the Dirichlet solution ψ ( x ) \psi(x) ψ ( x ) at energy E + i η E+\mathrm{i}\eta E + i η , with ψ ( 0 ) = 0 \psi(0)=0 ψ ( 0 ) = 0 and ψ ′ ( 0 ) = 1 \psi'(0)=1 ψ ′ ( 0 ) = 1 . Since the random potential is real, its Wronskian satisfies
d d x ( ψ ‾ ψ ′ − ψ ′ ‾ ψ ) = − 2 i η ∣ ψ ∣ 2 . \frac{\mathrm{d}}{\mathrm{d}x}\left(\overline{\psi}\,\psi'-\overline{\psi'}\,\psi\right)=-2\mathrm{i}\eta\lvert\psi\rvert^2. d x d ( ψ ψ ′ − ψ ′ ψ ) = − 2 i η ∣ ψ ∣ 2 .
Integrating from 0 0 0 to x x x gives
ψ ‾ ( x ) ψ ′ ( x ) − ψ ′ ‾ ( x ) ψ ( x ) = − 2 i η ∫ 0 x ∣ ψ ( s ) ∣ 2 d s . \overline{\psi}(x)\psi'(x)-\overline{\psi'}(x)\psi(x)=-2\mathrm{i}\eta\int_0^x\lvert\psi(s)\rvert^2\,\mathrm{d}s. ψ ( x ) ψ ′ ( x ) − ψ ′ ( x ) ψ ( x ) = − 2 i η ∫ 0 x ∣ ψ ( s ) ∣ 2 d s .
If ψ ( x 0 ) = 0 \psi(x_0)=0 ψ ( x 0 ) = 0 at some x 0 > 0 x_0>0 x 0 > 0 , the left-hand side would vanish, whereas the right-hand side is nonzero: η > 0 \eta>0 η > 0 and the solution is nontrivial. Thus, for every finite x > 0 x>0 x > 0 , ψ ( x ) ≠ 0 \psi(x)\ne0 ψ ( x ) = 0 and Z ( x ) = ψ ′ ( x ) / ψ ( x ) Z(x)=\psi'(x)/\psi(x) Z ( x ) = ψ ′ ( x ) / ψ ( x ) is finite. Absorption therefore removes the Riccati poles at finite positive positions: there is no escape to infinity and reinjection of the kind occurring at real energy. This does not preclude circulating probability currents in the stationary distribution of Z Z Z .
More explicitly, at a fixed complex energy E + i η E + i\eta E + i η , consider the solution satisfying ψ ( 0 ) = 0 \psi(0) = 0 ψ ( 0 ) = 0 and ψ ′ ( L ) = b ≠ 0 \psi^{\prime}(L) = b \ne0 ψ ′ ( L ) = b = 0 . This solution is a rescaling of the fundamental solution used above, leaving Z Z Z unchanged. These conditions define an inhomogeneous boundary-value problem, not an eigenvalue problem. The prescribed boundary amplitude fixes the normalization, so the integrated wavefunction intensity is not constrained to unity. The imaginary part ℑ Z ( x ) \Im Z(x) ℑ Z ( x ) determines the probability current. Indeed, using ψ ′ = Z ψ \psi^{\prime} = Z\psi ψ ′ = Z ψ ,
j ( x ) = ℏ m ℑ [ ψ ∗ ( x ) ψ ′ ( x ) ] = ℏ m ∣ ψ ( x ) ∣ 2 ℑ Z ( x ) = − ℏ m Γ 1 / 3 ∣ ψ ( x ) ∣ 2 v ( x ) . j(x) = \frac{\hbar}{m}\Im\left[\psi^{*}(x)\psi^{\prime}(x)\right] = \frac{\hbar}{m}|\psi(x)|^{2}\Im Z(x) = -\frac{\hbar}{m}\Gamma^{1/3}|\psi(x)|^{2}v(x). j ( x ) = m ℏ ℑ [ ψ ∗ ( x ) ψ ′ ( x ) ] = m ℏ ∣ ψ ( x ) ∣ 2 ℑ Z ( x ) = − m ℏ Γ 1/3 ∣ ψ ( x ) ∣ 2 v ( x ) .
This local expression is consistent with the integrated flux balance: the Wronskian identity above gives
∣ ψ ( x ) ∣ 2 ℑ Z ( x ) = − η ∫ 0 x ∣ ψ ( s ) ∣ 2 d s . |\psi(x)|^{2}\Im Z(x) = -\eta\int_{0}^{x}|\psi(s)|^{2}\,\mathrm{d}s. ∣ ψ ( x ) ∣ 2 ℑ Z ( x ) = − η ∫ 0 x ∣ ψ ( s ) ∣ 2 d s .
Thus v v v measures the flux per unit boundary intensity, with the sign fixed by the orientation of the boundary. The conversion from boundary admittance to scattering observables is discussed in Appendix § 23 .
d z j = − ( ϵ + i a + z j 2 ) d t + 2 d B t , j = 1 , 2. \mathrm{d}z_j = -(\epsilon+ \mathrm{i}a + z_j^2)\,\mathrm{d}t + \sqrt{2}\,\mathrm{d}B_t,\qquad j = 1,2. d z j = − ( ϵ + i a + z j 2 ) d t + 2 d B t , j = 1 , 2.
The common noise cancels in their difference δ = z 1 − z 2 \delta= z_1-z_2 δ = z 1 − z 2 , giving δ ′ = − ( z 1 + z 2 ) δ \delta^{\prime} = -(z_1+z_2)\delta δ ′ = − ( z 1 + z 2 ) δ . For an infinitesimal separation about a trajectory z = u − i v z = u-\mathrm{i}v z = u − i v , this becomes δ ′ = − 2 z δ \delta^{\prime} = -2z\delta δ ′ = − 2 z δ , and the real part of δ ′ / δ \delta^{\prime}/\delta δ ′ / δ gives d d t log ∣ δ ∣ = − 2 u \frac{\mathrm{d}}{\mathrm{d}t}\log|\delta| = -2u d t d log ∣ δ ∣ = − 2 u . The v v v -equation in (19) contains no noise and gives directly d d t log v = a v − 2 u \frac{\mathrm{d}}{\mathrm{d}t}\log v = \frac{a}{v}-2u d t d log v = v a − 2 u . Subtracting the two gives
d d t log ∣ δ ∣ v = − a v < 0 ( a > 0 ) , hence ∣ δ ( t ) ∣ v ( t ) = ∣ δ ( 0 ) ∣ v ( 0 ) exp [ − a ∫ 0 t d s v ( s ) ] , (20) \frac{\mathrm{d}}{\mathrm{d}t}\log\frac{|\delta|}{v}=-\frac{a}{v}<0\quad(a>0),
\qquad\text{hence}\qquad
\frac{|\delta(t)|}{v(t)}=\frac{|\delta(0)|}{v(0)}\exp\left[-a\int_{0}^{t}\frac{\mathrm{d}s}{v(s)}\right], \tag*{(20)} d t d log v ∣ δ ∣ = − v a < 0 ( a > 0 ) , hence v ( t ) ∣ δ ( t ) ∣ = v ( 0 ) ∣ δ ( 0 ) ∣ exp [ − a ∫ 0 t v ( s ) d s ] , (20)
with v ( s ) v(s) v ( s ) taken along the reference trajectory. Here ∣ δ ∣ / v |\delta|/v ∣ δ ∣/ v is the infinitesimal length in the Poincaré metric of the lower half-plane,
d ℓ 2 = d u 2 + d v 2 v 2 , \mathrm{d}\ell^{2}=\frac{\mathrm{d}u^{2}+\mathrm{d}v^{2}}{v^{2}}, d ℓ 2 = v 2 d u 2 + d v 2 ,
which measures separations relative to the distance from the real axis. In this metric the real-energy part of the dynamics drops out (the terms − 2 u -2u − 2 u cancel), and only absorption shrinks separations, at the instantaneous rate a / v a/v a / v . Thus, infinitesimal separations contract exponentially whenever 1 / v 1/v 1/ v has a strictly positive long-time average. This loss of sensitivity to initial conditions provides the physical mechanism underlying uniqueness of the stationary probability density. A rigorous proof also requires control of the long-time dynamics, including existence of a stationary state; we do not pursue these steps here.
Step 2: hidden gradient structure
Lemma 1. Given the Itô diffusion
d X t = b ( X t ) d t + σ d B t (21) \mathrm{d}X_t=b(X_t)\,\mathrm{d}t+\sigma\,\mathrm{d}B_t \tag*{(21)} d X t = b ( X t ) d t + σ d B t (21)
on an open domain Ω ⊂ R n \Omega\subset\mathbb{R}^{n} Ω ⊂ R n , where B t B_t B t is a standard Brownian motion and σ \sigma σ is a constant matrix, consider the corresponding Fokker–Planck equation
∂ t ρ = − ∇ ⋅ J , J = b ρ − D ∇ ρ , D = 1 2 σ σ T . (22) \partial_t \rho= -\nabla\cdot J,\qquad J = b\rho- D\nabla\rho,\qquad D = \frac{1}{2}\sigma\sigma^{\mathsf{T}}. \tag*{(22)} ∂ t ρ = − ∇ ⋅ J , J = b ρ − D ∇ ρ , D = 2 1 σ σ T . (22)
Suppose that the drift has the form
b = ( D + A ) ∇ Φ , (23) b = (D + A)\nabla\Phi, \tag*{(23)} b = ( D + A ) ∇Φ , (23)
where A A A is a constant antisymmetric matrix. Then ρ s t ∝ e Φ \rho_{\mathrm{st}} \propto e^{\Phi} ρ st ∝ e Φ solves the stationary Fokker–Planck equation in the interior of Ω \Omega Ω , with probability current
J s t = A ∇ ρ s t . (24) J_{\mathrm{st}} = A\nabla\rho_{\mathrm{st}}. \tag*{(24)} J st = A ∇ ρ st . (24)
It defines a stationary probability density provided that it is normalizable and satisfies the boundary conditions of the dynamics.
Derivation. Since ∇ ρ s t = ρ s t ∇ Φ \nabla\rho_{\mathrm{st}} = \rho_{\mathrm{st}}\nabla\Phi ∇ ρ st = ρ st ∇Φ , the diffusive contribution cancels the symmetric part of the drift:
J s t = ( D + A ) ρ s t ∇ Φ − D ∇ ρ s t = A ∇ ρ s t . J_{\mathrm{st}} = (D + A)\rho_{\mathrm{st}}\nabla\Phi- D\nabla\rho_{\mathrm{st}} = A\nabla\rho_{\mathrm{st}}. J st = ( D + A ) ρ st ∇Φ − D ∇ ρ st = A ∇ ρ st .
The remaining current is divergence-free, ∇ ⋅ J = ∑ i j A i j ∂ i ∂ j ρ = 0 \nabla\cdot J = \sum_{ij} A_{ij}\partial_i\partial_j\rho= 0 ∇ ⋅ J = ∑ ij A ij ∂ i ∂ j ρ = 0 , because A A A is antisymmetric.
□ \square □
A few remarks are due.
(i.) For A = 0 A = 0 A = 0 this is the familiar gradient (detailed-balance) case. For A ≠ 0 A \ne0 A = 0 the stationary state carries divergence-free circulating currents. So the absence of detailed balance does not prevent an explicit solution.
(ii.) The lemma establishes stationarity in the interior of the domain. To define a stationary probability density, ρ s t ∝ e Φ \rho_{\mathrm{st}} \propto e^{\Phi} ρ st ∝ e Φ must also be normalizable and satisfy the boundary conditions of the dynamics. For a boundary through which no probability can pass, the required condition is
n ⋅ J s t = n ⋅ A ∇ ρ s t = 0 on ∂ Ω , n \cdot J_{\mathrm{st}} = n \cdot A\nabla\rho_{\mathrm{st}} = 0 \qquad\text{on } \partial\Omega, n ⋅ J st = n ⋅ A ∇ ρ st = 0 on ∂ Ω ,
where n n n is the outward unit normal. Other boundary prescriptions, such as periodic identification or reinjection, require the corresponding matching conditions rather than zero normal current. For an unbounded domain, probability leakage at infinity must also be excluded. These conditions must be checked in each application.
(iii.) Condition (23) is a property of the drift alone: ( D + A ) − 1 b (D + A)^{-1}b ( D + A ) − 1 b must be a gradient. It is a strong condition, and it is exactly what the Riccati flow satisfies in the right variables.
Step 3: the explicit stationary density
Set y = a / v > 0 y = a/v > 0 y = a / v > 0 . Since d v = ( a − 2 u v ) d t dv = (a - 2uv)\,dt d v = ( a − 2 uv ) d t contains no Brownian increment, v v v is differentiable along each realization of the noise. The ordinary chain rule therefore applies to y = a / v y = a/v y = a / v , with no Itô correction. Then
d u = ( − ϵ − u 2 + a 2 y 2 ) d t + 2 d B t , d y = ( 2 u y − y 2 ) d t . (25) du = \left(-\epsilon- u^2 + \frac{a^2}{y^2}\right)dt + \sqrt{2}\,dB_t,\qquad dy = (2uy - y^2)\,dt. \tag*{(25)} d u = ( − ϵ − u 2 + y 2 a 2 ) d t + 2 d B t , d y = ( 2 u y − y 2 ) d t . (25)
Recalling U ϵ ( u ) = ϵ u + u 3 / 3 U_{\epsilon}(u) = \epsilon u + u^3/3 U ϵ ( u ) = ϵ u + u 3 /3 define
Φ a ( u , y ) = U ϵ ( u − y ) − U ϵ ( u ) − a 2 y = − y ( u − y 2 ) 2 − ϵ y − y 3 12 − a 2 y . (26) \Phi_a(u,y) = U_{\epsilon}(u-y) - U_{\epsilon}(u) - \frac{a^2}{y} = -y\left(u-\frac{y}{2}\right)^2 - \epsilon y - \frac{y^3}{12} - \frac{a^2}{y}. \tag*{(26)} Φ a ( u , y ) = U ϵ ( u − y ) − U ϵ ( u ) − y a 2 = − y ( u − 2 y ) 2 − ϵy − 12 y 3 − y a 2 . (26)
Its derivatives are ∂ u Φ a = − 2 u y + y 2 \partial_u\Phi_a=-2uy+y^2 ∂ u Φ a = − 2 u y + y 2 and ∂ y Φ a = − ϵ − ( u − y ) 2 + a 2 / y 2 \partial_y\Phi_a=-\epsilon-(u-y)^2+a^2/y^2 ∂ y Φ a = − ϵ − ( u − y ) 2 + a 2 / y 2 . Hence, exactly,
b u = − ϵ − u 2 + a 2 y 2 = ∂ u Φ a + ∂ y Φ a , b y = 2 u y − y 2 = − ∂ u Φ a . (27) b_u=-\epsilon-u^2+\frac{a^2}{y^2}=\partial_u\Phi_a+\partial_y\Phi_a,\qquad b_y=2uy-y^2=-\partial_u\Phi_a. \tag*{(27)} b u = − ϵ − u 2 + y 2 a 2 = ∂ u Φ a + ∂ y Φ a , b y = 2 u y − y 2 = − ∂ u Φ a . (27)
This is (23) with D = diag ( 1 , 0 ) D=\operatorname{diag}(1,0) D = diag ( 1 , 0 ) and A = ( 0 1 − 1 0 ) \mathcal{A}=\begin{pmatrix}0&1\\-1&0\end{pmatrix} A = ( 0 − 1 1 0 ) . By Lemma 1, the stationary joint probability density is given by
ρ a ( u , y ) = 1 Z ( ϵ , a ) e Φ a ( u , y ) , u ∈ R , y > 0 , (28) \rho_a(u,y)=\frac{1}{\mathcal{Z}(\epsilon,a)}e^{\Phi_a(u,y)},\qquad u\in\mathbb{R},\quad y>0, \tag*{(28)} ρ a ( u , y ) = Z ( ϵ , a ) 1 e Φ a ( u , y ) , u ∈ R , y > 0 , (28)
with probability current
J u = ∂ y ρ a , J y = − ∂ u ρ a . J_u=\partial_y\rho_a,\qquad J_y=-\partial_u\rho_a. J u = ∂ y ρ a , J y = − ∂ u ρ a .
It remains to check that there is no probability flux through the boundary. Indeed, as a > 0 a>0 a > 0 the factor e − a 2 / y e^{-a^2/y} e − a 2 / y suppresses the density as y → 0 y\to0 y → 0 , the factor e − y 3 / 12 e^{-y^3/12} e − y 3 /12 suppresses it as y → ∞ y\to\infty y → ∞ , and the Gaussian factor e − y ( u − y / 2 ) 2 e^{-y(u-y/2)^2} e − y ( u − y /2 ) 2 suppresses it as ∣ u ∣ → ∞ \lvert u\rvert\to\infty ∣ u ∣ → ∞ at fixed y > 0 y>0 y > 0 . The decay after Gaussian integration over u u u also excludes probability flux in joint limits where ∣ u ∣ → ∞ \lvert u\rvert\to\infty ∣ u ∣ → ∞ while y → 0 y\to0 y → 0 or y → ∞ y\to\infty y → ∞ .
In the original variables ( u , v ) (u,v) ( u , v ) the same density takes the form
P a ( u , v ) = a Z ( ϵ , a ) v 2 exp [ − a ( u 2 + ϵ ) v + a 2 u v 2 − a 3 3 v 3 − a v ] , v > 0. (29) P_a(u,v)=\frac{a}{\mathcal{Z}(\epsilon,a)v^2}\exp\left[-\frac{a(u^2+\epsilon)}{v}+\frac{a^2u}{v^2}-\frac{a^3}{3v^3}-av\right],\qquad v>0. \tag*{(29)} P a ( u , v ) = Z ( ϵ , a ) v 2 a exp [ − v a ( u 2 + ϵ ) + v 2 a 2 u − 3 v 3 a 3 − a v ] , v > 0. (29)
The factor v − 2 v^{-2} v − 2 is the hyperbolic area density. Thus, relative to the invariant volume element d u d v / v 2 du\,dv/v^2 d u d v / v 2 on the lower half-plane, the stationary probability density is proportional to e Φ a ( u , a / v ) e^{\Phi_a(u,a/v)} e Φ a ( u , a / v ) . The loss-of-memory argument of Step 1 identifies this stationary distribution with that of the absorbing-wire admittance, subject to the long-distance relaxation assumptions stated there. The relation of this geometry to waveguide scattering and hyperbolic diffusion is discussed in Appendix § 23 .
Step 4: Gaussian structure and the Airy normalization
The exponent (26) is quadratic in u u u . Consequently, at fixed y y y , the Gaussian density of u u u is
p ( u ∣ y ) = y π exp [ − y ( u − y 2 ) 2 ] , p(u\mid y)=\sqrt{\frac{y}{\pi}}\exp\left[-y\left(u-\frac{y}{2}\right)^2\right], p ( u ∣ y ) = π y exp [ − y ( u − 2 y ) 2 ] ,
with mean y / 2 y/2 y /2 and variance 1 / ( 2 y ) 1/(2y) 1/ ( 2 y ) . Integrating out the u u u -variable gives the marginal density of y y y :
f a ( y ) = π Z ( ϵ , a ) y − 1 / 2 e − ϵ y − y 3 / 12 − a 2 / y , Z ( ϵ , a ) = π ∫ 0 ∞ y − 1 / 2 e − ϵ y − y 3 / 12 − a 2 / y d y , (30) f_a(y)=\frac{\sqrt{\pi}}{\mathcal{Z}(\epsilon,a)}y^{-1/2}e^{-\epsilon y-y^3/12-a^2/y},\qquad\mathcal{Z}(\epsilon,a)=\sqrt{\pi}\int_0^\infty y^{-1/2}e^{-\epsilon y-y^3/12-a^2/y}\,dy, \tag*{(30)} f a ( y ) = Z ( ϵ , a ) π y − 1/2 e − ϵy − y 3 /12 − a 2 / y , Z ( ϵ , a ) = π ∫ 0 ∞ y − 1/2 e − ϵy − y 3 /12 − a 2 / y d y , (30)
and the integral for Z \mathcal{Z} Z turns out to be an Airy modulus squared:
Z ( ϵ , a ) = π 2 ∣ A − ( s ) ∣ 2 , s = − ϵ − i a . (31) \mathcal{Z}(\epsilon,a)=\pi^2\lvert A_-(s)\rvert^2,\qquad s=-\epsilon-ia. \tag*{(31)} Z ( ϵ , a ) = π 2 ∣ A − ( s ) ∣ 2 , s = − ϵ − ia . (31)
We derive this by writing Z \mathcal{Z} Z as a double contour integral. The same steps give the general-N N N factorization of § 15 .
Derivation. Let Γ − \Gamma_- Γ − run from e − 2 π i / 3 ∞ e^{-2\pi i/3}\infty e − 2 π i /3 ∞ to + ∞ +\infty + ∞ , below the real axis, and put
U E ( λ ) = E λ + λ 3 / 3 , E = ϵ + i a , s = − E . U_E(\lambda)=E\lambda+\lambda^3/3,\qquad E=\epsilon+ia,\qquad s=-E. U E ( λ ) = E λ + λ 3 /3 , E = ϵ + ia , s = − E .
Under λ = − w \lambda=-w λ = − w , the induced contour C − \mathcal{C}_- C − runs from e π i / 3 ∞ e^{\pi i/3}\infty e π i /3 ∞ to − ∞ -\infty − ∞ , giving
J ( E ) : = ∫ Γ − e − U E ( λ ) d λ = − ∫ C − e w 3 / 3 − s w d w = i π A − ( s ) , (32) \mathcal{J}(E):=\int_{\Gamma_-}e^{-U_E(\lambda)}\,d\lambda=-\int_{\mathcal{C}_-}e^{w^3/3-sw}\,dw=i\pi A_-(s), \tag*{(32)} J ( E ) := ∫ Γ − e − U E ( λ ) d λ = − ∫ C − e w 3 /3 − s w d w = iπ A − ( s ) , (32)
where A − ( s ) = Ai ( s ) − i Bi ( s ) \mathcal{A}_{-}(s)=\operatorname{Ai}(s)-i\operatorname{Bi}(s) A − ( s ) = Ai ( s ) − i Bi ( s ) .
In the conjugate integral, substitute λ ′ = − μ \lambda'=-\mu λ ′ = − μ and use the oddness of U U U . Define Γ + \Gamma_{+} Γ + to run from − ∞ -\infty − ∞ to e − π i / 3 ∞ e^{-\pi i/3}\infty e − π i /3 ∞ in the lower half-plane: its orientation is opposite to that induced on − Γ ‾ − -\overline{\Gamma}_{-} − Γ − , absorbing the minus sign from d λ ′ = − d μ d\lambda'=-d\mu d λ ′ = − d μ . Thus
∣ J ( E ) ∣ 2 = ∫ Γ − d λ ∫ Γ + d μ e − U E ( λ ) + U E ‾ ( μ ) . (33) |\mathcal{J}(\mathcal{E})|^{2}=\int_{\Gamma_{-}}d\lambda\int_{\Gamma_{+}}d\mu\ e^{-U_{\mathcal{E}}(\lambda)+U_{\overline{\mathcal{E}}}(\mu)}. \tag*{(33)} ∣ J ( E ) ∣ 2 = ∫ Γ − d λ ∫ Γ + d μ e − U E ( λ ) + U E ( μ ) . (33)
We deform the product contour onto
Σ a = { ( λ , μ ) = ( z , z − y ) : z = u − i a y , u ∈ R , y > 0 } . \Sigma_{a}=\left\{(\lambda,\mu)=(z,z-y):z=u-\frac{ia}{y},\ u\in\mathbb{R},\ y>0\right\}. Σ a = { ( λ , μ ) = ( z , z − y ) : z = u − y ia , u ∈ R , y > 0 } .
The justification for this deformation is given for general N N N in § 15 (step 4 of Result 2). The Jacobian satisfies d λ ∧ d μ = d y ∧ d u d\lambda\wedge d\mu=dy\wedge du d λ ∧ d μ = d y ∧ d u . Writing c = u − y / 2 c=u-y/2 c = u − y /2 , the exponent on Σ a \Sigma_{a} Σ a becomes
− U E ( λ ) + U E ‾ ( μ ) = − ϵ y − 2 i a ( c − i a y ) − y ( c − i a y ) 2 − y 3 12 = − y ( u − y 2 ) 2 − ϵ y − y 3 12 − a 2 y = Φ a ( u , y ) . -U_{\mathcal{E}}(\lambda)+U_{\overline{\mathcal{E}}}(\mu)=-\epsilon y-2ia\left(c-\frac{ia}{y}\right)-y\left(c-\frac{ia}{y}\right)^{2}-\frac{y^{3}}{12}=-y\left(u-\frac{y}{2}\right)^{2}-\epsilon y-\frac{y^{3}}{12}-\frac{a^{2}}{y}=\Phi_{a}(u,y). − U E ( λ ) + U E ( μ ) = − ϵy − 2 ia ( c − y ia ) − y ( c − y ia ) 2 − 12 y 3 = − y ( u − 2 y ) 2 − ϵy − 12 y 3 − y a 2 = Φ a ( u , y ) .
With the inherited orientation, we therefore obtain
∣ J ( E ) ∣ 2 = ∫ 0 ∞ d y ∫ R d u e Φ a ( u , y ) . |\mathcal{J}(\mathcal{E})|^{2}=\int_{0}^{\infty}dy\int_{\mathbb{R}}du\ e^{\Phi_{a}(u,y)}. ∣ J ( E ) ∣ 2 = ∫ 0 ∞ d y ∫ R d u e Φ a ( u , y ) .
Performing the Gaussian integral
∫ R e − y ( u − y / 2 ) 2 d u = π y , \int_{\mathbb{R}}e^{-y(u-y/2)^{2}}\,du=\sqrt{\frac{\pi}{y}}, ∫ R e − y ( u − y /2 ) 2 d u = y π ,
we find
∣ J ( E ) ∣ 2 = π ∫ 0 ∞ d y y exp ( − ϵ y − y 3 12 − a 2 y ) = Z ( ϵ , a ) , |\mathcal{J}(\mathcal{E})|^{2}=\sqrt{\pi}\int_{0}^{\infty}\frac{dy}{\sqrt{y}}\exp\left(-\epsilon y-\frac{y^{3}}{12}-\frac{a^{2}}{y}\right)=\mathcal{Z}(\epsilon,a), ∣ J ( E ) ∣ 2 = π ∫ 0 ∞ y d y exp ( − ϵy − 12 y 3 − y a 2 ) = Z ( ϵ , a ) ,
which proves Eq. (31) .
At a = 0 a=0 a = 0 , Gaussian integration over u u u recovers the classical representation
Ai 2 ( x ) + Bi 2 ( x ) = π − 3 / 2 ∫ 0 ∞ t − 1 / 2 e x t − t 3 / 12 d t \operatorname{Ai}^{2}(x)+\operatorname{Bi}^{2}(x)=\pi^{-3/2}\int_{0}^{\infty}t^{-1/2}e^{xt-t^{3}/12}\,dt Ai 2 ( x ) + Bi 2 ( x ) = π − 3/2 ∫ 0 ∞ t − 1/2 e x t − t 3 /12 d t
[74 Eq. 9.11.4] , with x = − ϵ x=-\epsilon x = − ϵ , providing an independent check of the factor π 2 \pi^{2} π 2 .
The derivation also identifies the variables. On Σ a \Sigma_{a} Σ a the first coordinate z = u − i a / y z=u-ia/y z = u − ia / y is the Riccati variable itself, and the density (28) is precisely the normalized integrand of ∣ J ∣ 2 |\mathcal{J}|^{2} ∣ J ∣ 2 : a Gibbs weight in z z z times an anti-Gibbs weight in z − y z-y z − y . At a = 0 a=0 a = 0 , Σ a \Sigma_{a} Σ a becomes the real region { ( λ , μ ) ∈ R 2 : λ > μ } \{(\lambda,\mu)\in\mathbb{R}^{2}:\lambda>\mu\} {( λ , μ ) ∈ R 2 : λ > μ } appearing in Step 6.
Step 5: holomorphic statistics from the stationary distribution
Take z = u − i v = u − i a / y z=u-iv=u-ia/y z = u − i v = u − ia / y and q ≥ 0 q\geq0 q ≥ 0 . Throughout this subsection, ⟨ ⋅ ⟩ \langle\cdot\rangle ⟨ ⋅ ⟩ denotes averaging with respect to the stationary density (28) :
⟨ F ( z ) ⟩ = ∫ 0 ∞ d y ∫ R d u ρ a ( u , y ) F ( u − i a / y ) . \langle F(z)\rangle=\int_{0}^{\infty}dy\int_{\mathbb{R}}du\ \rho_{a}(u,y)F(u-ia/y). ⟨ F ( z )⟩ = ∫ 0 ∞ d y ∫ R d u ρ a ( u , y ) F ( u − ia / y ) .
Then the conditional Gaussian average over u u u at fixed y y y gives
⟨ e − i q u ∣ y ⟩ = e − i q y / 2 − q 2 / ( 4 y ) . \langle e^{-iqu}\mid y\rangle=e^{-iqy/2-q^{2}/(4y)}. ⟨ e − i q u ∣ y ⟩ = e − i q y /2 − q 2 / ( 4 y ) .
Multiplying by e − q a / y e^{-qa/y} e − q a / y and by the weight e − ϵ y − y 3 / 12 − a 2 / y e^{-\epsilon y-y^3/12-a^2/y} e − ϵy − y 3 /12 − a 2 / y yields
e − ( ϵ + i q / 2 ) y − y 3 / 12 − ( a + q / 2 ) 2 / y . e^{-(\epsilon+iq/2)y-y^3/12-(a+q/2)^2/y}. e − ( ϵ + i q /2 ) y − y 3 /12 − ( a + q /2 ) 2 / y .
This is the same weight at the shifted parameters ϵ ′ = ϵ + i q / 2 \epsilon'=\epsilon+iq/2 ϵ ′ = ϵ + i q /2 and a ′ = a + q / 2 a'=a+q/2 a ′ = a + q /2 . Now, (31) gives Z ( ϵ , a ) = π 2 A − ( − ϵ − i a ) A + ( − ϵ + i a ) \mathcal{Z}(\epsilon,a)=\pi^2\mathcal{A}_{-}(-\epsilon-ia)\mathcal{A}_{+}(-\epsilon+ia) Z ( ϵ , a ) = π 2 A − ( − ϵ − ia ) A + ( − ϵ + ia ) which continues analytically in ( ϵ , a ) (\epsilon,a) ( ϵ , a ) . Under the shift, − ϵ ′ − i a ′ = s − i q -\epsilon'-ia'=s-iq − ϵ ′ − i a ′ = s − i q , while − ϵ ′ + i a ′ = − ϵ + i a -\epsilon'+ia'=-\epsilon+ia − ϵ ′ + i a ′ = − ϵ + ia is unchanged. Therefore
⟨ e − i q z ⟩ = A − ( s − i q ) A − ( s ) , ⟨ z ⟩ = A − ′ ( s ) A − ( s ) . (34) \left\langle e^{-iqz}\right\rangle=\frac{\mathcal{A}_{-}(s-iq)}{\mathcal{A}_{-}(s)},\qquad\langle z\rangle=\frac{\mathcal{A}_{-}'(s)}{\mathcal{A}_{-}(s)}. \tag*{(34)} ⟨ e − i q z ⟩ = A − ( s ) A − ( s − i q ) , ⟨ z ⟩ = A − ( s ) A − ′ ( s ) . (34)
In physical units Ξ 1 ( E + i η ) = ⟨ Z ⟩ = T 1 / 3 A − ′ ( s ) / A − ( s ) \Xi_1(E+i\eta)=\langle Z\rangle=T^{1/3}\mathcal{A}_{-}'(s)/\mathcal{A}_{-}(s) Ξ 1 ( E + i η ) = ⟨ Z ⟩ = T 1/3 A − ′ ( s ) / A − ( s ) . The outgoing Airy function is the holomorphic factor of the positive normalization (31) .
Put w = v / a = 1 / y w=v/a=1/y w = v / a = 1/ y . From d v = ( a − 2 u v ) d t dv=(a-2uv)dt d v = ( a − 2 uv ) d t , the variable w w w obeys d w = ( 1 − 2 u w ) d t dw=(1-2uw)\,dt d w = ( 1 − 2 u w ) d t , which is also the equation of − ∂ ϵ u -\partial_\epsilon u − ∂ ϵ u at real energy. So as a → 0 a\to0 a → 0 , w w w becomes the energy derivative of the Riccati variable, w = − ∂ ϵ z w=-\partial_\epsilon z w = − ∂ ϵ z , a positive energy-derivative variable. To see this directly, let ψ \psi ψ solve the real-energy Schrödinger equation
− ψ ′ ′ + V ψ = ϵ ψ , ψ ( 0 ) = 0 , ψ ′ ( 0 ) = 1. -\psi''+V\psi=\epsilon\psi,\qquad\psi(0)=0,\qquad\psi'(0)=1. − ψ ′′ + V ψ = ϵ ψ , ψ ( 0 ) = 0 , ψ ′ ( 0 ) = 1.
where primes denote derivatives with respect to t t t . Differentiating with respect to ϵ \epsilon ϵ gives
d d t [ ψ ∂ ϵ ψ ′ − ψ ′ ∂ ϵ ψ ] = − ψ 2 . \frac{d}{dt}\left[\psi\,\partial_\epsilon\psi'-\psi'\,\partial_\epsilon\psi\right]=-\psi^2. d t d [ ψ ∂ ϵ ψ ′ − ψ ′ ∂ ϵ ψ ] = − ψ 2 .
The expression in brackets vanishes at t = 0 t=0 t = 0 , since the initial data are independent of energy. Hence, wherever ψ ( t ) ≠ 0 \psi(t)\ne0 ψ ( t ) = 0 ,
w ( t ) = − ∂ ϵ ( ψ ′ ( t ) ψ ( t ) ) = ∫ 0 t ψ 2 ( s ) d s ψ 2 ( t ) > 0. w(t)=-\partial_\epsilon\left(\frac{\psi'(t)}{\psi(t)}\right)=\frac{\int_0^t\psi^2(s)\,ds}{\psi^2(t)}>0. w ( t ) = − ∂ ϵ ( ψ ( t ) ψ ′ ( t ) ) = ψ 2 ( t ) ∫ 0 t ψ 2 ( s ) d s > 0.
The corresponding matrix identity is derived below in Eqs. (37) –(38) .
Further, Eq. (28) for the joint probability density at a = 0 a=0 a = 0 in the variables ( λ , μ ) = ( u , u − 1 / w ) (\lambda,\mu)=(u,u-1/w) ( λ , μ ) = ( u , u − 1/ w ) takes the form of Gibbs weight for λ \lambda λ times an anti-Gibbs weight for μ \mu μ , with an order constraint:
ρ 0 = 1 Z ( ϵ , 0 ) e − U ϵ ( λ ) + U ϵ ( μ ) , μ < λ . (35) \rho_0=\frac{1}{\mathcal{Z}(\epsilon,0)}e^{-U_\epsilon(\lambda)+U_\epsilon(\mu)},\qquad\mu<\lambda. \tag*{(35)} ρ 0 = Z ( ϵ , 0 ) 1 e − U ϵ ( λ ) + U ϵ ( μ ) , μ < λ . (35)
Integrating over μ \mu μ yields Halperin’s stationary density [52 ] :
p H ( u ) = j e − U ϵ ( u ) ∫ − ∞ u e U ϵ ( v ′ ) d v ′ , j = 1 Z ( ϵ , 0 ) = 1 π 2 [ Ai ( − ϵ ) 2 + Bi ( − ϵ ) 2 ] . (36) p_{\mathrm{H}}(u)=j\,e^{-U_\epsilon(u)}\int_{-\infty}^{u}e^{U_\epsilon(v')} \,dv',\qquad j=\frac{1}{\mathcal{Z}(\epsilon,0)}=\frac{1}{\pi^2[\operatorname{Ai}(-\epsilon)^2+\operatorname{Bi}(-\epsilon)^2]}. \tag*{(36)} p H ( u ) = j e − U ϵ ( u ) ∫ − ∞ u e U ϵ ( v ′ ) d v ′ , j = Z ( ϵ , 0 ) 1 = π 2 [ Ai ( − ϵ ) 2 + Bi ( − ϵ ) 2 ] 1 . (36)
Here j > 0 j>0 j > 0 is the magnitude of the stationary probability current flowing towards u = − ∞ u=-\infty u = − ∞ . Thus integrating out μ \mu μ in the two-point Gibbs/anti-Gibbs distribution gives Halperin’s formula, with the auxiliary variable μ = u − 1 / w \mu=u-1/w μ = u − 1/ w expressed in terms of the energy derivative.
We now extract the complex characteristic exponent, whose real and imaginary parts determine the Lyapunov exponent and the integrated density of states. In the dimensionless variables used here, it is
Ω 1 ( ϵ ) = lim a → 0 + ⟨ z ⟩ = Γ 1 ( ϵ ) − i π ν 1 ( ϵ ) . \Omega_1(\epsilon)=\lim_{a\to0+}\langle z\rangle=\Gamma_1(\epsilon)-i\pi\nu_1(\epsilon). Ω 1 ( ϵ ) = a → 0 + lim ⟨ z ⟩ = Γ 1 ( ϵ ) − iπ ν 1 ( ϵ ) .
Recalling that v = a / y v = a/y v = a / y and that the conditional Gaussian mean is ⟨ u ∣ y ⟩ = y / 2 \langle u \mid y\rangle= y/2 ⟨ u ∣ y ⟩ = y /2 , we obtain from (30)
⟨ v ⟩ = a ⟨ y − 1 ⟩ = − 1 2 ∂ a log Z , ⟨ u ⟩ = 1 2 ⟨ y ⟩ = − 1 2 ∂ ϵ log Z . \langle v\rangle= a\langle y^{-1}\rangle= -\frac{1}{2}\partial_a \log Z,\qquad\langle u\rangle= \frac{1}{2}\langle y\rangle= -\frac{1}{2}\partial_\epsilon\log Z. ⟨ v ⟩ = a ⟨ y − 1 ⟩ = − 2 1 ∂ a log Z , ⟨ u ⟩ = 2 1 ⟨ y ⟩ = − 2 1 ∂ ϵ log Z .
The term linear in a a a arises from the region y → 0 y \to0 y → 0 :
∫ 0 ∞ y − 1 / 2 ( e − a 2 / y − 1 ) d y = − 2 π a , a > 0. \int_0^\infty y^{-1/2}\left(e^{-a^2/y}-1\right)\,dy = -2\sqrt{\pi}\,a,\qquad a>0. ∫ 0 ∞ y − 1/2 ( e − a 2 / y − 1 ) d y = − 2 π a , a > 0.
Consequently,
Z ( ϵ , a ) = Z ( ϵ , 0 ) − 2 π a + O ( a 2 ) , Z(\epsilon,a) = Z(\epsilon,0) - 2\pi a + O(a^2), Z ( ϵ , a ) = Z ( ϵ , 0 ) − 2 π a + O ( a 2 ) ,
and hence
ν 1 ( ϵ ) = j = 1 Z ( ϵ , 0 ) , Γ 1 ( ϵ ) = − 1 2 ∂ ϵ log Z ( ϵ , 0 ) , \nu_1(\epsilon) = j = \frac{1}{Z(\epsilon,0)},\qquad\Gamma_1(\epsilon) = -\frac{1}{2}\partial_\epsilon\log Z(\epsilon,0), ν 1 ( ϵ ) = j = Z ( ϵ , 0 ) 1 , Γ 1 ( ϵ ) = − 2 1 ∂ ϵ log Z ( ϵ , 0 ) ,
where Γ 1 \Gamma_1 Γ 1 is the principal-value mean of the real-energy Riccati variable. Together these expressions reproduce the classical Airy formula for the complex characteristic exponent [52 , 15 ] :
Ω 1 ( ϵ ) = A − ′ ( s ) A − ( s ) , s = − ϵ . \Omega_1(\epsilon) = \frac{A_{-}'(s)}{A_{-}(s)},\qquad s=-\epsilon. Ω 1 ( ϵ ) = A − ( s ) A − ′ ( s ) , s = − ϵ .
Dictionary to the general case
Below we summarize in the table the modifications required to extend the N = 1 N=1 N = 1 case to general number of channels N ≥ 1 N\geq1 N ≥ 1 .
The explicit stationary distribution for general ( N , β ) (N,\beta) ( N , β )
This section repeats Steps 1–3 of Sec. 3 for matrices. Throughout, E = E + i η \mathcal{E}=E+i\eta E = E + i η with η > 0 \eta>0 η > 0 .
Complex energy confines Z Z Z to a Siegel half-space
Write Z = A − i P Z=A-iP Z = A − i P , where A = ( Z + Z † ) / 2 A=(Z+Z^\dagger)/2 A = ( Z + Z † ) /2 and P = i ( Z − Z † ) / 2 P=i(Z-Z^\dagger)/2 P = i ( Z − Z † ) /2 are F-Hermitian. Because W W W is Hermitian the Wronskian-type identity
d d x ( Ψ † Ψ ′ − Ψ ′ † Ψ ) = Ψ † ( E ‾ − E ) Ψ = − 2 i η Ψ † Ψ (37) \frac{d}{dx}\left(\Psi^\dagger\Psi' - \Psi'^\dagger\Psi\right)=\Psi^\dagger(\overline{\mathcal{E}}-\mathcal{E})\Psi=-2i\eta\,\Psi^\dagger\Psi \tag*{(37)} d x d ( Ψ † Ψ ′ − Ψ ′ † Ψ ) = Ψ † ( E − E ) Ψ = − 2 i η Ψ † Ψ (37)
holds for each noise realization. Starting from Ψ ( 0 ) = 0 \Psi(0)=0 Ψ ( 0 ) = 0 , it gives
P = η S , S : = Ψ † − 1 [ ∫ 0 x Ψ † Ψ ] Ψ − 1 > 0. (38) P=\eta S,\qquad S:=\Psi^{\dagger-1}\left[\int_0^x\Psi^\dagger\Psi\right]\Psi^{-1}>0. \tag*{(38)} P = η S , S := Ψ †− 1 [ ∫ 0 x Ψ † Ψ ] Ψ − 1 > 0. (38)
Taking the Hermitian and anti-Hermitian parts of (2) at energy E \mathcal{E} E yields the system of two matrix Itô equations
d A = [ − E − A 2 + P 2 ] d x + d W , P ′ = η I − ( A P + P A ) . (39) dA = \left[-\mathcal{E}-A^2+P^2\right]dx+dW,\qquad P'=\eta I-(AP+PA). \tag*{(39)} d A = [ − E − A 2 + P 2 ] d x + d W , P ′ = η I − ( A P + P A ) . (39)
The noise does not appear in the equation for P P P because W W W is Hermitian and only shifts A A A . Thus Z Z Z lives in the Siegel half-space S β = { Z = A − i P : P > 0 } \mathcal{S}_\beta=\{Z=A-iP:P>0\} S β = { Z = A − i P : P > 0 } , the matrix analogue of the lower half-plane. In particular, det Ψ ( x ) ≠ 0 \det\Psi(x)\ne0 det Ψ ( x ) = 0 for every finite x > 0 x>0 x > 0 , so Z ( x ) = Ψ ′ ( x ) Ψ ( x ) − 1 Z(x)=\Psi'(x)\Psi(x)^{-1} Z ( x ) = Ψ ′ ( x ) Ψ ( x ) − 1 remains finite along the wire, and a continuous branch of log det Ψ ( x ) \log\det\Psi(x) log det Ψ ( x ) can be chosen for x > 0 x>0 x > 0 . Hence
Ξ N ( E ) = lim L → ∞ 1 L log det Ψ ( L ; E ) = ⟨ Tr Z ⟩ μ E , η , η > 0. (40) \Xi_N(\mathcal{E})=\lim_{L\to\infty}\frac{1}{L}\log\det\Psi(L;\mathcal{E})=\langle\operatorname{Tr}Z\rangle_{\mu_{\mathcal{E},\eta}},\qquad\eta>0. \tag*{(40)} Ξ N ( E ) = L → ∞ lim L 1 log det Ψ ( L ; E ) = ⟨ Tr Z ⟩ μ E , η , η > 0. (40)
Physical meaning: boundary admittance and internal norm. The Riccati matrix Z ( L ; E ) Z(L;\mathcal{E}) Z ( L ; E ) maps a prescribed boundary amplitude b = ψ ( L ) b=\psi(L) b = ψ ( L ) to the boundary derivative: ψ ′ ( L ) = Z ( L ; E ) b \psi'(L)=Z(L;\mathcal{E})b ψ ′ ( L ) = Z ( L ; E ) b , with ψ ( 0 ) = 0 \psi(0)=0 ψ ( 0 ) = 0 . It is therefore the Dirichlet-to-Neumann map, or boundary admittance, of the sample. For E = E + i η \mathcal{E}=E+i\eta E = E + i η with η > 0 \eta>0 η > 0 , the matrix Ψ ( L ; E ) \Psi(L;\mathcal{E}) Ψ ( L ; E ) is invertible. Defining
F ( x ; E ) = Ψ ( x ; E ) Ψ ( L ; E ) − 1 , F(x;\mathcal{E})=\Psi(x;\mathcal{E})\Psi(L;\mathcal{E})^{-1}, F ( x ; E ) = Ψ ( x ; E ) Ψ ( L ; E ) − 1 ,
the solution with prescribed boundary values ψ ( 0 ) = 0 \psi(0)=0 ψ ( 0 ) = 0 and ψ ( L ) = b \psi(L)=b ψ ( L ) = b is ψ ( x ) = F ( x ; E ) b \psi(x)=F(x;\mathcal{E})b ψ ( x ) = F ( x ; E ) b . The integrated Wronskian identity defines the internal-norm matrix
N ( E , η ) : = P ( E , η ) η = ∫ 0 L F ( x ; E ) † F ( x ; E ) d x . (41) \mathcal{N}(E,\eta):=\frac{P(E,\eta)}{\eta}=\int_0^L F(x;\mathcal{E})^\dagger F(x;\mathcal{E})\,dx. \tag*{(41)} N ( E , η ) := η P ( E , η ) = ∫ 0 L F ( x ; E ) † F ( x ; E ) d x . (41)
Thus b † N ( E , η ) b b^\dagger\mathcal{N}(E,\eta)b b † N ( E , η ) b is the integrated wavefunction intensity inside the sample for the imposed boundary amplitude b b b . This identity holds for every real E E E when η > 0 \eta>0 η > 0 . If E E E is not a Dirichlet eigenvalue, the zero-absorption limit is finite and satisfies
lim η → 0 + P ( E , η ) η = T ( E ) : = − ∂ E Z ( L ; E ) . \lim_{\eta\to0^+}\frac{P(E,\eta)}{\eta}=\mathcal{T}(E):=-\partial_E Z(L;E). η → 0 + lim η P ( E , η ) = T ( E ) := − ∂ E Z ( L ; E ) .
The relation of these quantities to reflection, dwell times and the Wigner–Smith matrix is explained in Appendix § 23 . Their statistics will be developed in Paper II [44 ] .
Loss of memory: an exact contraction identity
Consider two solutions of (2) driven by the same noise , and let δ \delta δ be their (infinitesimal) difference. The noise cancels, yielding
δ ′ = − ( Z δ + δ Z ) . (42) \delta'=-(Z\delta+\delta Z). \tag*{(42)} δ ′ = − ( Z δ + δ Z ) . (42)
Measure the difference δ \delta δ with the natural distance of the Siegel half-space, ℓ 2 = Tr ( P − 1 δ P − 2 δ † ) \ell^2=\operatorname{Tr}(P^{-1}\delta P^{-2}\delta^\dagger) ℓ 2 = Tr ( P − 1 δ P − 2 δ † ) ; for N = 1 N=1 N = 1 this is the Poincaré length of Eq. (20) . Using (39) and (42) we get the exact identity valid for each realization of the random potential:
d ℓ 2 d x = − η Tr ( P − 2 δ P − 1 δ † + P − 1 δ P − 2 δ † ) ≤ − 2 η ∥ P ∥ 2 ℓ 2 . (43) \frac{d\ell^2}{dx}=-\eta\operatorname{Tr}\left(P^{-2}\delta P^{-1}\delta^\dagger+P^{-1}\delta P^{-2}\delta^\dagger\right)\le-\frac{2\eta}{\lVert P\rVert^2}\ell^2. \tag*{(43)} d x d ℓ 2 = − η Tr ( P − 2 δ P − 1 δ † + P − 1 δ P − 2 δ † ) ≤ − ∥ P ∥ 2 2 η ℓ 2 . (43)
The inequality follows from P − 1 ≥ ∥ P ∥ − 1 I P^{-1}\ge\lVert P\rVert^{-1}I P − 1 ≥ ∥ P ∥ − 1 I , where ∥ P ∥ \lVert P\rVert ∥ P ∥ denotes the operator norm, and positivity of the trace pairing.
Equation (43) expresses the contraction produced by absorption. At η = 0 \eta=0 η = 0 , Z = Ψ ′ Ψ − 1 Z=\Psi'\Psi^{-1} Z = Ψ ′ Ψ − 1 is built from solutions of the Schrödinger equation with Hermitian potential and real energy. The pair ( Ψ , Ψ ′ ) (\Psi,\Psi') ( Ψ , Ψ ′ ) evolves along the wire through a transfer matrix. More explicitly, propagation from a reference position x 0 x_0 x 0 to x x x is described by
( Ψ ( x ) Ψ ′ ( x ) ) = ( A B C D ) ( Ψ ( x 0 ) Ψ ′ ( x 0 ) ) , \begin{pmatrix}\Psi(x)\\ \Psi'(x)\end{pmatrix}
=
\begin{pmatrix}A&B\\ C&D\end{pmatrix}
\begin{pmatrix}\Psi(x_0)\\ \Psi'(x_0)\end{pmatrix}, ( Ψ ( x ) Ψ ′ ( x ) ) = ( A C B D ) ( Ψ ( x 0 ) Ψ ′ ( x 0 ) ) ,
where the blocks depend on x x x , x 0 x_0 x 0 and the realization of the random potential. This transfer matrix preserves the conserved Wronskian form Ψ † Ψ ′ − Ψ ′ † Ψ \Psi^\dagger\Psi' - \Psi'^\dagger\Psi Ψ † Ψ ′ − Ψ ′ † Ψ ; for β = 1 \beta= 1 β = 1 it is a real symplectic matrix. Consequently the Riccati matrix itself evolves by matrix Möbius (linear-fractional) transformations,
Z ⟼ ( C + D Z ) ( A + B Z ) − 1 Z \longmapsto(C + DZ)(A + BZ)^{-1} Z ⟼ ( C + D Z ) ( A + B Z ) − 1
which map the Siegel half-space onto itself and preserve the metric ℓ 2 \ell^2 ℓ 2 . This holds for the deterministic flow and for each noise increment separately. For η > 0 \eta> 0 η > 0 the Wronskian form is no longer conserved, and the additional term − i η -i\eta − i η makes the evolution contractive: the metric length of every nonzero infinitesimal variation decreases. In particular,
ℓ 2 ( x ) ≤ ℓ 2 ( x 0 ) exp [ − 2 η ∫ x 0 x d s ∥ P ( s ) ∥ ] . \ell^2(x) \leq\ell^2(x_0) \exp\left[-2\eta\int_{x_0}^{x}\frac{ds}{\lVert P(s)\rVert}\right]. ℓ 2 ( x ) ≤ ℓ 2 ( x 0 ) exp [ − 2 η ∫ x 0 x ∥ P ( s )∥ d s ] .
Exponential contraction follows provided the long-time average of 1 / ∥ P ∥ 1/\lVert P\rVert 1/ ∥ P ∥ is bounded away from zero. Together with suitable control of long-distance fluctuations, this contraction provides the basis for establishing uniqueness of the stationary distribution and loss of dependence on the initial condition. We do not supply the full ergodic argument here.
The variables ( A , Y ) (A,Y) ( A , Y )
Set Y = η P − 1 > 0 Y = \eta P^{-1} > 0 Y = η P − 1 > 0 . Set Y = η P − 1 > 0 Y = \eta P^{-1} > 0 Y = η P − 1 > 0 . As in the scalar case, this change of variables requires no Itô correction, since P P P has finite variation. From (39) , Y ′ = − η P − 1 P ′ P − 1 Y' = -\eta P^{-1}P'P^{-1} Y ′ = − η P − 1 P ′ P − 1 and P 2 = η 2 Y − 2 P^2 = \eta^2Y^{-2} P 2 = η 2 Y − 2 . Hence
d A = [ − E − A 2 + η 2 Y − 2 ] d x + d W , d Y = [ A Y + Y A − Y 2 ] d x . (44) dA = \left[-E - A^2 + \eta^2Y^{-2}\right]dx + dW,\qquad dY = \left[AY + YA - Y^2\right]dx. \tag*{(44)} d A = [ − E − A 2 + η 2 Y − 2 ] d x + d W , d Y = [ A Y + Y A − Y 2 ] d x . (44)
These are the scalar equations (25) with u → A u \to A u → A and y → Y y \to Y y → Y , in the only order that keeps them Hermitian. In the limit η → 0 \eta\to0 η → 0 we have Y → S − 1 Y \to S^{-1} Y → S − 1 , where S = P / η S = P/\eta S = P / η is the positive matrix of (38) . At real energy S S S coincides with the energy derivative T : = − ∂ E Z T := -\partial_E Z T := − ∂ E Z , because both obey X ′ = I − ( Z X + X Z ) X' = I - (ZX + XZ) X ′ = I − ( Z X + X Z ) with X ( x ) ≃ ( x / 3 ) I N X(x) \simeq(x/3)I_N X ( x ) ≃ ( x /3 ) I N as x → 0 + x \to0^+ x → 0 + .
The potential, two trace identities, and stationarity
Define
Φ ( A , Y ) = 1 T [ Tr U E ( A − Y ) − Tr U E ( A ) − η 2 Tr Y − 1 ] = − 1 T [ Tr ( B Y B ) + E Tr Y + 1 12 Tr Y 3 + η 2 Tr Y − 1 ] , (45) \begin{aligned}
\Phi(A,Y) = \frac{1}{T}\left[\operatorname{Tr}U_E(A-Y) - \operatorname{Tr}U_E(A) - \eta^2\operatorname{Tr}Y^{-1}\right] \\
&= -\frac{1}{T}\left[\operatorname{Tr}(BYB) + E\operatorname{Tr}Y + \frac{1}{12}\operatorname{Tr}Y^3 + \eta^2\operatorname{Tr}Y^{-1}\right], \tag*{(45)}
\end{aligned} Φ ( A , Y ) = T 1 [ Tr U E ( A − Y ) − Tr U E ( A ) − η 2 Tr Y − 1 ] = − T 1 [ Tr ( B Y B ) + E Tr Y + 12 1 Tr Y 3 + η 2 Tr Y − 1 ] , (45)
with B = A − 1 2 Y B = A - \frac{1}{2}Y B = A − 2 1 Y . The second form uses only cyclicity of the trace: Tr ( A − Y ) 3 − Tr A 3 = − 3 Tr A 2 Y + 3 Tr A Y 2 − Tr Y 3 \operatorname{Tr}(A-Y)^3 - \operatorname{Tr}A^3 = -3\operatorname{Tr}A^2Y + 3\operatorname{Tr}AY^2 - \operatorname{Tr}Y^3 Tr ( A − Y ) 3 − Tr A 3 = − 3 Tr A 2 Y + 3 Tr A Y 2 − Tr Y 3 and Tr ( B Y B ) = Tr A 2 Y − Tr A Y 2 + 1 4 Tr Y 3 \operatorname{Tr}(BYB) = \operatorname{Tr}A^2Y - \operatorname{Tr}AY^2 + \frac{1}{4}\operatorname{Tr}Y^3 Tr ( B Y B ) = Tr A 2 Y − Tr A Y 2 + 4 1 Tr Y 3 .
Trace gradients. For any polynomial f f f , we have δ Tr f ( X ) = Tr ( f ′ ( X ) δ X ) \delta\operatorname{Tr}f(X) = \operatorname{Tr}(f'(X)\,\delta X) δ Tr f ( X ) = Tr ( f ′ ( X ) δ X ) . With U E ′ ( X ) = E + X 2 U_E'(X) = E + X^2 U E ′ ( X ) = E + X 2 , we then have
T ∇ A Φ = ( A − Y ) 2 − A 2 , T ∇ Y Φ = − E − ( A − Y ) 2 + η 2 Y − 2 . (46) T\nabla_A\Phi= (A-Y)^2 - A^2,\qquad T\nabla_Y\Phi= -E - (A-Y)^2 + \eta^2Y^{-2}. \tag*{(46)} T ∇ A Φ = ( A − Y ) 2 − A 2 , T ∇ Y Φ = − E − ( A − Y ) 2 + η 2 Y − 2 . (46)
Comparing with the drifts in (44) , ( A − Y ) 2 − A 2 = − A Y − Y A + Y 2 (A-Y)^2 - A^2 = -AY - YA + Y^2 ( A − Y ) 2 − A 2 = − A Y − Y A + Y 2 , so
b Y = A Y + Y A − Y 2 = − T ∇ A Φ , b A = − E − A 2 + η 2 Y − 2 = T ( ∇ A Φ + ∇ Y Φ ) . (47) b_Y = AY + YA - Y^2 = -T\nabla_A\Phi,\qquad b_A = -E - A^2 + \eta^2Y^{-2} = T(\nabla_A\Phi+ \nabla_Y\Phi). \tag*{(47)} b Y = A Y + Y A − Y 2 = − T ∇ A Φ , b A = − E − A 2 + η 2 Y − 2 = T ( ∇ A Φ + ∇ Y Φ ) . (47)
These are the matrix versions of the two scalar identities of Sec. 3C. They hold for non-commuting A A A and Y Y Y , because Φ \Phi Φ is a trace.
Proposition 1 (Explicit stationary distribution). For every N ≥ 1 N \ge1 N ≥ 1 , β ∈ { 1 , 2 , 4 } \beta\in\{1,2,4\} β ∈ { 1 , 2 , 4 } , E ∈ R E \in\mathbb{R} E ∈ R and η > 0 \eta> 0 η > 0 , the density e Φ ( A , Y ) e^{\Phi(A,Y)} e Φ ( A , Y ) with respect to the flat measure d A d Y \mathrm{d}A\,\mathrm{d}Y d A d Y on Herm N ( F ) × Herm N + ( F ) \operatorname{Herm}_{N}(\mathbb{F}) \times\operatorname{Herm}_{N}^{+}(\mathbb{F}) Herm N ( F ) × Herm N + ( F ) is stationary for (44) . It is integrable. The contraction and relaxation argument of § 14 B identifies its normalized form, Eq. (3) , as the stationary distribution reached at long distances; this identification uses the relaxation assumptions stated there.
Derivation. Use the real coordinates A a A_a A a , Y a Y_a Y a of Sec. 2. The noise acts only on A A A , with D A A = 1 2 C = T g − 1 D_{AA} = \frac{1}{2}C = T g^{-1} D AA = 2 1 C = T g − 1 , and D = 0 D = 0 D = 0 in all other blocks. By (11) , T ( ∇ A Φ ) a = T g a − 1 ∂ A a Φ T(\nabla_A\Phi)_a = T g_a^{-1}\partial_{A_a}\Phi T ( ∇ A Φ ) a = T g a − 1 ∂ A a Φ , and likewise for Y Y Y . So (47) reads
b = ( T g − 1 T g − 1 − T g − 1 0 ) ∇ Φ = ( D + A ) ∇ Φ , A = T ( 0 g − 1 − g − 1 0 ) . (48) b = \begin{pmatrix} T g^{-1} & T g^{-1} \\ -T g^{-1} & 0 \end{pmatrix}\nabla\Phi= (D+\mathcal{A})\nabla\Phi,\qquad\mathcal{A}=T\begin{pmatrix} 0 & g^{-1} \\ -g^{-1} & 0 \end{pmatrix}. \tag*{(48)} b = ( T g − 1 − T g − 1 T g − 1 0 ) ∇Φ = ( D + A ) ∇Φ , A = T ( 0 − g − 1 g − 1 0 ) . (48)
g − 1 g^{-1} g − 1 is diagonal and constant, so A \mathcal{A} A is constant and antisymmetric. The noise is additive, so there is no Itô drift. Lemma 1 then gives stationarity of e Φ e^{\Phi} e Φ , with circulating current J = A ∇ e Φ J=\mathcal{A}\nabla e^{\Phi} J = A ∇ e Φ , i.e. J A = T ∇ Y e Φ J_A=T\nabla_Y e^{\Phi} J A = T ∇ Y e Φ and J Y = − T ∇ A e Φ J_Y=-T\nabla_A e^{\Phi} J Y = − T ∇ A e Φ in trace-gradient form. Now we also need to check that there is no flux through the boundary of the domain. An eigenvalue of Y Y Y tending to 0 is suppressed by e − η 2 Tr Y − 1 / T e^{-\eta^{2}\operatorname{Tr}Y^{-1}/T} e − η 2 Tr Y − 1 / T , faster than any power. Large Y Y Y is suppressed by e − Tr Y 3 / 12 T e^{-\operatorname{Tr}Y^{3}/12T} e − Tr Y 3 /12 T , since Tr ( B Y B ) ≥ 0 \operatorname{Tr}(BYB)\ge0 Tr ( B Y B ) ≥ 0 for Y > 0 Y>0 Y > 0 . Large A A A at fixed Y > 0 Y>0 Y > 0 is suppressed by the Gaussian factor. For integrability see Eq. (50) below. The integral converges even at η = 0 \eta=0 η = 0 . Uniqueness follows by § 14 B, at the level of the long-distance relaxation argument given there □.
The Dyson parameter β \beta β enters the construction only through the flat measures and through g g g , which drops out of the final stationary distribution. For N = 1 N=1 N = 1 , Result 1 is Eq. (28) .
Structure of the stationary probability distribution
Inversion on the cone of positive F \mathbb{F} F -Hermitian matrices has Jacobian d ( X − 1 ) = det X − ( β ( N − 1 ) + 2 ) d X \mathrm{d}(X^{-1})=\det X^{-(\beta(N-1)+2)}\mathrm{d}X d ( X − 1 ) = det X − ( β ( N − 1 ) + 2 ) d X . In terms of Z = A − i P Z=A-\mathrm{i}P Z = A − i P , the stationary distribution can be written in Siegel-invariant form as
μ E , η ( d Z ) = η d Z N ( E , η ) exp { 1 T [ Tr U E ( A − η P − 1 ) − Tr U E ( A ) − η Tr P ] } d A d P det P β ( N − 1 ) + 2 , (49) \mu_{E,\eta}(\mathrm{d}Z)=\frac{\eta^{d}}{\mathcal{Z}_{N}(E,\eta)}\exp\left\{\frac{1}{T}\left[\operatorname{Tr}U_E(A-\eta P^{-1})-\operatorname{Tr}U_E(A)-\eta\operatorname{Tr}P\right]\right\}\frac{\mathrm{d}A\,\mathrm{d}P}{\det P^{\beta(N-1)+2}}, \tag*{(49)} μ E , η ( d Z ) = Z N ( E , η ) η d exp { T 1 [ Tr U E ( A − η P − 1 ) − Tr U E ( A ) − η Tr P ] } det P β ( N − 1 ) + 2 d A d P , (49)
with d = dim Herm N ( F ) = N + β N ( N − 1 ) / 2 d=\dim\operatorname{Herm}_{N}(\mathbb{F})=N+\beta N(N-1)/2 d = dim Herm N ( F ) = N + β N ( N − 1 ) /2 . The last factor is the invariant volume of the Siegel half-space. Relative to this volume element, the probability density is proportional to e Φ e^{\Phi} e Φ . For N = 1 N=1 N = 1 this is Eq. (29) .
We now determine the statistics of the internal-norm matrix N = P / η \mathcal{N}=P/\eta N = P / η defined in (41) . Since N \mathcal{N} N is positive definite for η > 0 \eta>0 η > 0 , its inverse is well defined. It is convenient to work with
Y = N − 1 = η P − 1 , Y=\mathcal{N}^{-1}=\eta P^{-1}, Y = N − 1 = η P − 1 ,
because the joint density is Gaussian in B = A − Y / 2 B=A-Y/2 B = A − Y /2 . Integrating out B B B therefore gives an explicit eigenvalue distribution for Y Y Y . The eigenvalues τ i \tau_i τ i of the internal-norm matrix are recovered by τ i = 1 / y i \tau_i=1/y_i τ i = 1/ y i ; in the zero-absorption limit they become the eigenvalues of T = − O E Z T=-\mathcal{O}_E Z T = − O E Z .
Next, given Y Y Y , the matrix B = A − 1 2 Y B=A-\frac{1}{2}Y B = A − 2 1 Y is Gaussian with the weight e − Tr ( B Y B ) / T e^{-\operatorname{Tr}(BYB)/T} e − Tr ( B Y B ) / T . In the eigenbasis of Y Y Y we have Tr ( B Y B ) = ∑ i y i B i i 2 + ∑ i < j ( y i + y j ) ∣ B i j ∣ 2 \operatorname{Tr}(BYB)=\sum_i y_i B_{ii}^{2}+\sum_{i<j}(y_i+y_j)|B_{ij}|^{2} Tr ( B Y B ) = ∑ i y i B ii 2 + ∑ i < j ( y i + y j ) ∣ B ij ∣ 2 . Integrating out B B B and the eigenvectors of Y Y Y gives the joint probability density of the eigenvalues y i > 0 y_i>0 y i > 0 of Y Y Y :
p ( y ) ∝ ∣ Δ ( y ) ∣ β ∏ i y i − 1 / 2 ∏ i < j ( y i + y j ) − β / 2 exp [ − 1 T ∑ i ( E y i + y i 3 12 + η 2 y i ) ] . (50) p(y)\propto|\Delta(y)|^{\beta}\prod_i y_i^{-1/2}\prod_{i<j}(y_i+y_j)^{-\beta/2}\exp\left[-\frac{1}{T}\sum_i\left(Ey_i+\frac{y_i^{3}}{12}+\frac{\eta^{2}}{y_i}\right)\right]. \tag*{(50)} p ( y ) ∝ ∣Δ ( y ) ∣ β i ∏ y i − 1/2 i < j ∏ ( y i + y j ) − β /2 exp [ − T 1 i ∑ ( E y i + 12 y i 3 + y i η 2 ) ] . (50)
Thus Eq. (50) describes the inverse internal-norm eigenvalues. Their reciprocals τ i = 1 / y i \tau_i=1/y_i τ i = 1/ y i give the integrated norm inside the sample when the imposed boundary amplitude b = ψ ( L ) b=\psi(L) b = ψ ( L ) is a normalized eigenvector of P / η P/\eta P / η .
Let Y N ( E , η ) \mathcal{Y}_N(E,\eta) Y N ( E , η ) denote the integral over R + N \mathbb{R}_{+}^{N} R + N of the unnormalized eigenvalue weight displayed in Eq. (50) . Then Z N = c N , β T d / 2 Y N \mathcal{Z}_N=c_{N,\beta}T^{d/2}\mathcal{Y}_N Z N = c N , β T d /2 Y N . We now discuss its properties in parallel with the N = 1 N=1 N = 1 case.
Conditional Gaussian structure. The joint probability density of the Hermitian matrices A A A and P P P in Z = A − i P Z=A-iP Z = A − i P is equivalently specified by the eigenvalue density (50) of Y = η P − 1 Y=\eta P^{-1} Y = η P − 1 , the invariant angular measure for its eigenvectors, and the conditional Gaussian density of B = A − 1 2 Y B=A-\frac{1}{2}Y B = A − 2 1 Y at fixed Y Y Y . In the eigenbasis of Y Y Y , the independent real components of B B B have zero mean and variances T / ( 2 y i ) T/(2y_i) T / ( 2 y i ) for diagonal entries and T / [ 2 ( y i + y j ) ] T/[2(y_i+y_j)] T / [ 2 ( y i + y j )] for off-diagonal entries. The admittance matrix is then
Z = B + 1 2 Y − i η Y − 1 . Z=B+\frac{1}{2}Y-i\eta Y^{-1}. Z = B + 2 1 Y − i η Y − 1 .
For N = 1 N=1 N = 1 , in the scaled variables of Sec. 3C (where T = 1 T=1 T = 1 , η → a \eta\to a η → a and y = a / v y=a/v y = a / v ), this is z = u − i a / y z=u-ia/y z = u − ia / y with u = b + y 2 u=b+\frac{y}{2} u = b + 2 y and, at fixed y y y , b b b has zero mean and variance 1 / ( 2 y ) 1/(2y) 1/ ( 2 y ) , reproducing the Gaussian density of Sec. 3D.
An alternative route to large-N N N asymptotics. Equation (50) represents the normalization as an integral with a positive weight over R + N \mathbb{R}_+^N R + N . It therefore offers an alternative to analysing the complex-contour Kontsevich representation: one may instead study a real interacting eigenvalue ensemble using Coulomb-gas, saddle-point, or loop-equation methods.
Besides the usual logarithmic repulsion, the ensemble contains a pair interaction through y i + y j y_i+y_j y i + y j . For β = 2 \beta=2 β = 2 , this interaction has the Bures, or Schur–Pfaffian, structure: for even N N N ,
∏ i < j y i − y j y i + y j = Pf [ y i − y j y i + y j ] . \prod_{i<j}\frac{y_i-y_j}{y_i+y_j}
=\operatorname{Pf}\left[\frac{y_i-y_j}{y_i+y_j}\right]. i < j ∏ y i + y j y i − y j = Pf [ y i + y j y i − y j ] .
This relates the ensemble to those studied in [75 ] .
To recover the physical density of states, however, the large-N N N analysis must also resolve the small-absorption behaviour, including contributions from configurations in which an eigenvalue approaches y = 0 y=0 y = 0 . Whether this positive representation simplifies the analysis of the density-of-states edge remains to be investigated. We do not pursue this route here.
The normalization is ∣ τ ∣ 2 \lvert\tau\rvert^2 ∣ τ ∣ 2 : three Kontsevich integrals
The outgoing integration surface
For T > 0 T>0 T > 0 , the factor e − λ 3 / ( 3 T ) e^{-\lambda^3/(3T)} e − λ 3 / ( 3 T ) decays as ∣ λ ∣ → ∞ \lvert\lambda\rvert\to\infty ∣ λ ∣ → ∞ along rays in three sectors of the complex λ \lambda λ -plane, each of opening π / 3 \pi/3 π /3 :
− π 6 < arg λ < π 6 , π 2 < arg λ < 5 π 6 , − 5 π 6 < arg λ < − π 2 . -\frac{\pi}{6}<\arg\lambda<\frac{\pi}{6},\qquad
\frac{\pi}{2}<\arg\lambda<\frac{5\pi}{6},\qquad
-\frac{5\pi}{6}<\arg\lambda<-\frac{\pi}{2}. − 6 π < arg λ < 6 π , 2 π < arg λ < 6 5 π , − 6 5 π < arg λ < − 2 π .
The outgoing contour
Γ − : e − 2 π i / 3 ∞ ⟶ e − i ϵ ∞ ( ϵ → 0 + ) (51) \Gamma_-:e^{-2\pi i/3}\infty\longrightarrow e^{-i\epsilon}\infty\quad(\epsilon\to0^+) \tag*{(51)} Γ − : e − 2 π i /3 ∞ ⟶ e − i ϵ ∞ ( ϵ → 0 + ) (51)
joins two of them, see Figure 1 , and lies in the lower half-plane, where Z Z Z lives. Under w = − λ / T 1 / 3 w=-\lambda/T^{1/3} w = − λ / T 1/3 , it becomes C − : e i π / 3 ∞ → e i π ∞ C_-:e^{i\pi/3}\infty\to e^{i\pi}\infty C − : e iπ /3 ∞ → e iπ ∞ , which gives an integral representation of the outgoing Airy combination A − = Ai − i Bi \mathcal{A}_-=\operatorname{Ai}-i\operatorname{Bi} A − = Ai − i Bi .
Matrix integration surface. Let γ : R → Γ − \gamma:\mathbb{R}\to\Gamma_- γ : R → Γ − parametrize the contour, and apply it to Hermitian matrices by functional calculus, γ ( U D U † ) = U γ ( D ) U † \gamma(U D U^\dagger)=U\gamma(D)U^\dagger γ ( U D U † ) = U γ ( D ) U † . The image E = γ ( Herm N ( R ) ) \mathcal{E}=\gamma(\operatorname{Herm}_N(\mathbb{R})) E = γ ( Herm N ( R )) is a real d d d -dimensional surface without boundary inside S β \mathcal{S}_\beta S β . On the integration surface E \mathcal{E} E , the holomorphic volume element d Z dZ d Z takes the following form in eigenvalue coordinates:
d Z ∣ E = const × ∏ i < j ( γ ( t j ) − γ ( t i ) ) β ∏ j γ ′ ( t j ) d t j d U H a a r , t 1 < ⋯ < t N . (52) \left.dZ\right|_{\mathcal{E}}
=\operatorname{const}\times\prod_{i<j}\bigl(\gamma(t_j)-\gamma(t_i)\bigr)^\beta
\prod_j\gamma'(t_j)\,dt_j\,dU_{\mathrm{Haar}},
\qquad t_1<\cdots<t_N. \tag*{(52)} d Z ∣ E = const × i < j ∏ ( γ ( t j ) − γ ( t i ) ) β j ∏ γ ′ ( t j ) d t j d U Haar , t 1 < ⋯ < t N . (52)
A subtle point in this change of variables deserves discussion. On the real line the Jacobian is ∣ Δ ∣ β |\Delta|^{\beta} ∣Δ ∣ β . For β = 2 , 4 \beta=2,4 β = 2 , 4 the absolute value is superfluous. For β = 1 \beta=1 β = 1 it is replaced, on the contour, by the analytic function ∏ i < j ( w j − w i ) \prod_{i<j}(w_j-w_i) ∏ i < j ( w j − w i ) with the eigenvalues ordered along the contour . This is the correct analytic continuation, because on the real axis ordering removes the absolute value. It is the origin of the “ordered” prescription in Eq. (4) .
Define the matrix Airy (Kontsevich-type) integral over the outgoing integration surface:
J N ( E ) = ∫ E e − Tr U E ( Λ ) / T d Λ = c N , β ′ T κ N , β ′ τ N ( β ) ( s ) , s = − E / T 2 / 3 , (53) \mathcal{J}_N(\mathcal{E})=\int_{\mathcal{E}}e^{-\operatorname{Tr}U_{\mathcal{E}}(\Lambda)/T}\,d\Lambda=c'_{N,\beta}T^{\kappa'_{N,\beta}}\tau_N^{(\beta)}(\mathfrak{s}),\qquad\mathfrak{s}=-\mathcal{E}/T^{2/3}, \tag*{(53)} J N ( E ) = ∫ E e − Tr U E ( Λ ) / T d Λ = c N , β ′ T κ N , β ′ τ N ( β ) ( s ) , s = − E / T 2/3 , (53)
where the second equality follows from (52) and w = − λ / T 1 / 3 w=-\lambda/T^{1/3} w = − λ / T 1/3 .
Factorization
Proposition 2 (The normalization is a modulus squared). For η > 0 \eta>0 η > 0 , Z N ( E , η ) = ± J N ( E ) J N ( E ) ‾ \mathcal{Z}_N(E,\eta)=\pm\mathcal{J}_N(\mathcal{E})\overline{\mathcal{J}_N(\mathcal{E})} Z N ( E , η ) = ± J N ( E ) J N ( E ) , i.e. Eq. (5) .
Derivation. We follow essentially the same steps as in N = 1 N=1 N = 1 example.
Conjugation and reflection. J N ( E ) ‾ \overline{\mathcal{J}_N(\mathcal{E})} J N ( E ) is the integral of e − Tr U E ‾ ( Λ ′ ) / T e^{-\operatorname{Tr}U_{\overline{\mathcal{E}}}(\Lambda')/T} e − Tr U E ( Λ ′ ) / T over the conjugate integration surface E ‾ \overline{\mathcal{E}} E , in the upper half-space. Substitute Λ ′ = − M \Lambda'=-M Λ ′ = − M . Since U U U is odd, Tr U E ‾ ( − M ) = − Tr U E ‾ ( M ) \operatorname{Tr}U_{\overline{\mathcal{E}}}(-M)=-\operatorname{Tr}U_{\overline{\mathcal{E}}}(M) Tr U E ( − M ) = − Tr U E ( M ) , and M M M runs over − E ‾ -\overline{\mathcal{E}} − E , again in the lower Siegel half-space. Hence
J N ( E ) J N ( E ) ‾ = ± ∫ E × ( − E ‾ ) exp { 1 T [ − Tr U E ( Λ ) + Tr U E ‾ ( M ) ] } d Λ d M . (54) \mathcal{J}_N(\mathcal{E})\overline{\mathcal{J}_N(\mathcal{E})}=\pm\int_{\mathcal{E}\times(-\overline{\mathcal{E}})}\exp\left\{\frac{1}{T}\left[-\operatorname{Tr}U_{\mathcal{E}}(\Lambda)+\operatorname{Tr}U_{\overline{\mathcal{E}}}(M)\right]\right\}\,d\Lambda\,dM. \tag*{(54)} J N ( E ) J N ( E ) = ± ∫ E × ( − E ) exp { T 1 [ − Tr U E ( Λ ) + Tr U E ( M ) ] } d Λ d M . (54)
A Gibbs weight for Λ \Lambda Λ multiplies an anti-Gibbs weight for M M M .
Centre and relative coordinates. Put Λ = C + 1 2 Y \Lambda=C+\frac{1}{2}Y Λ = C + 2 1 Y and M = C − 1 2 Y M=C-\frac{1}{2}Y M = C − 2 1 Y ; the Jacobian is one. Using E Tr Λ − E ‾ Tr M = E Tr Y + 2 i η Tr C \mathcal{E}\operatorname{Tr}\Lambda-\overline{\mathcal{E}}\operatorname{Tr}M=\mathcal{E}\operatorname{Tr}Y+2i\eta\operatorname{Tr}C E Tr Λ − E Tr M = E Tr Y + 2 i η Tr C and 1 3 Tr ( Λ 3 − M 3 ) = Tr ( C Y C ) + 1 12 Tr Y 3 \frac{1}{3}\operatorname{Tr}(\Lambda^3-M^3)=\operatorname{Tr}(CYC)+\frac{1}{12}\operatorname{Tr}Y^3 3 1 Tr ( Λ 3 − M 3 ) = Tr ( C Y C ) + 12 1 Tr Y 3 we see that
− Tr U E ( Λ ) + Tr U E ‾ ( M ) = − Tr ( C Y C ) − 2 i η Tr C − E Tr Y − 1 12 Tr Y 3 . (55) -\operatorname{Tr}U_{\mathcal{E}}(\Lambda)+\operatorname{Tr}U_{\overline{\mathcal{E}}}(M)=-\operatorname{Tr}(CYC)-2i\eta\operatorname{Tr}C-\mathcal{E}\operatorname{Tr}Y-\frac{1}{12}\operatorname{Tr}Y^3. \tag*{(55)} − Tr U E ( Λ ) + Tr U E ( M ) = − Tr ( C Y C ) − 2 i η Tr C − E Tr Y − 12 1 Tr Y 3 . (55)
Completing the square. With C ~ = C + i η Y − 1 \widetilde{C}=C+i\eta Y^{-1} C = C + i η Y − 1 ,
Tr ( C Y C ) + 2 i η Tr C = Tr ( C ~ Y C ~ ) + η 2 Tr Y − 1 . (56) \operatorname{Tr}(CYC)+2i\eta\operatorname{Tr}C=\operatorname{Tr}(\widetilde{C}Y\widetilde{C})+\eta^2\operatorname{Tr}Y^{-1}. \tag*{(56)} Tr ( C Y C ) + 2 i η Tr C = Tr ( C Y C ) + η 2 Tr Y − 1 . (56)
The exponent of (54) is therefore exactly Φ \Phi Φ of (45) , with B → C ~ B\to\widetilde{C} B → C . The term η 2 Tr Y − 1 \eta^2\operatorname{Tr}Y^{-1} η 2 Tr Y − 1 , which suppresses the joint probability density when an eigenvalue of Y Y Y approaches zero, is produced by the Gaussian shift.
The integration surface (physics level). Steps 1–3 are identities between integrands; it remains to identify the integration surface. All steps are carried out at the matrix level, with the holomorphic volume form d Λ d M d\Lambda\,dM d Λ d M on the complexification of Herm N ( F ) × Herm N ( F ) \operatorname{Herm}_N(\mathbb{F})\times\operatorname{Herm}_N(\mathbb{F}) Herm N ( F ) × Herm N ( F ) . No absolute values or eigenvalue orderings appear. For β = 1 \beta=1 β = 1 the ordered prescription enters only when J N \mathcal{J}_N J N is reduced to eigenvalues, Eq. (52) .
The target integration surface. Take C ~ ∈ Herm N ( F ) \widetilde{C}\in\operatorname{Herm}_N(\mathbb{F}) C ∈ Herm N ( F ) and Y ∈ Herm N + ( F ) Y\in\operatorname{Herm}_N^+(\mathbb{F}) Y ∈ Herm N + ( F ) . Through Eq. (59) below, this defines a real 2 d 2d 2 d -dimensional surface Σ ′ = { ( Z , Z − Y ) } \Sigma'=\{(Z,Z-Y)\} Σ ′ = {( Z , Z − Y )} , on which the integrand is exactly e Φ e^{\Phi} e Φ . The surface Σ ′ \Sigma' Σ ′ has no boundary at finite distance. Its ends at infinity are of three types, all lying in the region where the integrand decays:
C ~ → ∞ \widetilde{C}\to\infty C → ∞ at fixed Y > 0 Y>0 Y > 0 , where the Gaussian factor e − Tr ( C ~ Y C ~ ) / T e^{-\operatorname{Tr}(\widetilde{C}Y\widetilde{C})/T} e − Tr ( C Y C ) / T decays;
Y → ∞ Y\to\infty Y → ∞ , where e − Tr Y 3 / ( 12 T ) e^{-\operatorname{Tr}Y^3/(12T)} e − Tr Y 3 / ( 12 T ) decays;
The third end requires η > 0 \eta> 0 η > 0 . At η = 0 \eta= 0 η = 0 the surface acquires a boundary Λ = M \Lambda= M Λ = M at finite distance; this is why the factorization is stated for η > 0 \eta> 0 η > 0 , and real energy is reached as a limit (Sec. 6).
Decomposition. The integrand is a product e f ( Λ ) e g ( M ) e^{f(\Lambda)}e^{g(M)} e f ( Λ ) e g ( M ) , with f = − Tr U E / T f=-\operatorname{Tr}U_E/T f = − Tr U E / T and g = + Tr U E ‾ / T g=+\operatorname{Tr}U_{\overline{E}}/T g = + Tr U E / T . For a single variable this is the familiar Airy situation: by Cauchy’s theorem, any contour along which e w 3 / 3 − s w e^{w^3/3-sw} e w 3 /3 − s w decays at both ends can be deformed, without changing the integral, into an integer combination of the two standard Airy contours [74 ] . Applied variable by variable, the same argument shows that the integral over any integration surface whose ends lie in the decay region is an integer combination of products of single-matrix integrals over such standard surfaces.
For invariant integrands the single-matrix basis can be labelled by the number k k k of eigenvalues placed on the conjugate Airy contour C + = C − ‾ C_+ = \overline{C_-} C + = C − , which produces A + = Ai + i Bi A_+ = \operatorname{Ai} + i\operatorname{Bi} A + = Ai + i Bi ; the other N − k N-k N − k eigenvalues stay on C − C_- C − . Denote the resulting holomorphic integrals by τ N [ k ] ( s ) \tau_N^{[k]}(\mathfrak{s}) τ N [ k ] ( s ) , so that τ N [ 0 ] = τ N ( β ) \tau_N^{[0]}=\tau_N^{(\beta)} τ N [ 0 ] = τ N ( β ) . Hence
Z N ( E , η ) ∝ ∑ k , l = 0 N n k l τ N [ k ] ( s ) τ N [ l ] ( s ) ‾ , n k l ∈ Z . (57) \mathcal{Z}_N(E,\eta) \propto\sum_{k,l=0}^{N} n_{kl}\,\tau_N^{[k]}(\mathfrak{s})\,\overline{\tau_N^{[l]}(\mathfrak{s})}, \qquad n_{kl}\in\mathbb{Z}. \tag*{(57)} Z N ( E , η ) ∝ k , l = 0 ∑ N n k l τ N [ k ] ( s ) τ N [ l ] ( s ) , n k l ∈ Z . (57)
The integers n k l n_{kl} n k l do not depend on E E E or on η > 0 \eta> 0 η > 0 : as the parameters vary, Σ ′ \Sigma' Σ ′ deforms continuously and its ends remain in the decay region.
Matching at high energy. The integers n k l n_{kl} n k l are fixed by letting E → + ∞ E\to+\infty E → + ∞ at fixed η > 0 \eta>0 η > 0 and comparing two independent estimates of Z N \mathcal{Z}_N Z N . In the positive integral (50) each eigenvalue is localized near y i ≃ η / E y_i\simeq\eta/\sqrt{E} y i ≃ η / E , the minimum of E y + η 2 / y Ey+\eta^2/y E y + η 2 / y , so that Y N ( E , η ) = exp [ − 2 N η E / T + O ( log E ) ] \mathcal{Y}_N(E,\eta)=\exp[-2N\eta\sqrt{E}/T+O(\log E)] Y N ( E , η ) = exp [ − 2 N η E / T + O ( log E )] : the normalization decays at the rate 2 N η E / T 2N\eta\sqrt{E}/T 2 N η E / T . On the other side of Eq. (57) , the Airy asymptotics give ∣ A ∓ ( s ) ∣ = exp [ ∓ η E / T + O ( log E ) ] |A_{\mp}(\mathfrak{s})|=\exp[\mp\eta\sqrt{E}/T+O(\log E)] ∣ A ∓ ( s ) ∣ = exp [ ∓ η E / T + O ( log E )] , so every eigenvalue moved to the conjugate contour C + C_+ C + turns a decaying factor into a growing one. Consequently ∣ τ N [ k ] τ N [ l ] ∣ = exp [ − 2 ( N − k − l ) η E / T + O ( log E ) ] |\tau_N^{[k]}\tau_N^{[l]}|=\exp[-2(N-k-l)\eta\sqrt{E}/T+O(\log E)] ∣ τ N [ k ] τ N [ l ] ∣ = exp [ − 2 ( N − k − l ) η E / T + O ( log E )] , and any term with ( k , l ) ≠ ( 0 , 0 ) (k,l)\ne(0,0) ( k , l ) = ( 0 , 0 ) would decay more slowly than the direct estimate allows and would dominate it. Such terms could only disappear by cancelling among themselves, but terms with equal decay rates carry different oscillating phases and cannot cancel. Hence all n k l n_{kl} n k l with ( k , l ) ≠ ( 0 , 0 ) (k,l)\ne(0,0) ( k , l ) = ( 0 , 0 ) vanish, only n 00 n_{00} n 00 survives, and we obtain Eq. (5) ; its sign and the constant are fixed by positivity and by the case N = 1 N=1 N = 1 .
Checks. The factorization passes several independent tests. For N = 1 N=1 N = 1 and η = 0 \eta=0 η = 0 it reduces to the classical representation Ai 2 ( x ) + Bi 2 ( x ) = π − 3 / 2 ∫ 0 ∞ t − 1 / 2 e x t − t 3 / 12 d t \operatorname{Ai}^2(x)+\operatorname{Bi}^2(x)=\pi^{-3/2}\int_0^\infty t^{-1/2}e^{xt-t^3/12}\,\mathrm{d}t Ai 2 ( x ) + Bi 2 ( x ) = π − 3/2 ∫ 0 ∞ t − 1/2 e x t − t 3 /12 d t [74 Eq. 9.11.4] . For general N N N at η = 0 \eta=0 η = 0 , both limits of the energy can be compared. As E → + ∞ E\to+\infty E → + ∞ the integral (50) scales as ( T / E ) a β (T/E)^{a_\beta} ( T / E ) a β , with a β = N 2 + β N ( N − 1 ) 4 a_\beta=\frac{N}{2}+\frac{\beta N(N-1)}{4} a β = 2 N + 4 β N ( N − 1 ) , in agreement with ∣ τ N ( β ) ∣ 2 ∝ ∣ s ∣ − a β |\tau_N^{(\beta)}|^2\propto|\mathfrak{s}|^{-a_\beta} ∣ τ N ( β ) ∣ 2 ∝ ∣ s ∣ − a β found in § 8 . As E → − ∞ E\to-\infty E → − ∞ the Laplace point of Eq. (50) moves to y i = 2 − E y_i=2\sqrt{-E} y i = 2 − E , giving Y N ∝ ( − E ) − a β e 4 3 N ( − E ) 3 / 2 / T \mathcal{Y}_N\propto(-E)^{-a_\beta}e^{\frac{4}{3}N(-E)^{3/2}/T} Y N ∝ ( − E ) − a β e 3 4 N ( − E ) 3/2 / T ; this is ∣ τ N ( β ) ∣ 2 |\tau_N^{(\beta)}|^2 ∣ τ N ( β ) ∣ 2 at the Lifshitz saddle w − w_- w − of § 9 , with both the exponent and the power reproduced. Finally, for β = 2 \beta=2 β = 2 , Appendix § 20 shows independently that the zero-absorption distribution coincides with the fermionic state of Ref. [46 ] , whose holomorphic factor is τ N ( 2 ) \tau_N^{(2)} τ N ( 2 ) .
Holomorphic statistics and a zero-free region
Holomorphic generating function. Given Proposition 2 , the holomorphic generating function of Tr Z \operatorname{Tr}Z Tr Z follows by repeating the steps of the N = 1 N=1 N = 1 example. For k ≥ 0 k\geq0 k ≥ 0 we have e − i k Tr Z = e − i k Tr A − k η Tr Y − 1 e^{-ik\operatorname{Tr}Z}=e^{-ik\operatorname{Tr}A-k\eta\operatorname{Tr}Y^{-1}} e − ik Tr Z = e − ik Tr A − k η Tr Y − 1 . At fixed Y Y Y the matrix B = A − 1 2 Y B=A-\frac{1}{2}Y B = A − 2 1 Y is Gaussian, and averaging over it produces e − i k Tr Y / 2 − k 2 T Tr Y − 1 / 4 e^{-ik\operatorname{Tr}Y/2-k^{2}T\operatorname{Tr}Y^{-1}/4} e − ik Tr Y /2 − k 2 T Tr Y − 1 /4 . Combined with e − k η Tr Y − 1 e^{-k\eta\operatorname{Tr}Y^{-1}} e − k η Tr Y − 1 , this gives the weight in Eq. (50) evaluated at
E ′ = E + i k T 2 , η ′ = η + k T 2 . E'=E+\frac{ikT}{2},\qquad\eta'=\eta+\frac{kT}{2}. E ′ = E + 2 ik T , η ′ = η + 2 k T .
Here E ′ E' E ′ is an analytically continued parameter; the shifted weight is no longer positive. Consequently,
⟨ e − i k Tr Z ⟩ μ E , η = Z N ( E ′ , η ′ ) Z N ( E , η ) . \left\langle e^{-ik\operatorname{Tr}Z}\right\rangle_{\mu_{E,\eta}}=\frac{Z_N(E',\eta')}{Z_N(E,\eta)}. ⟨ e − ik Tr Z ⟩ μ E , η = Z N ( E , η ) Z N ( E ′ , η ′ ) .
By Eq. (5) ,
Z N ( E , η ) ∝ J N ( E + i η ) J N ♯ ( E − i η ) , J N ♯ ( ζ ) : = J N ( ζ ‾ ) ‾ . Z_N(E,\eta)\propto\mathcal{J}_N(E+i\eta)\mathcal{J}^{\sharp}_N(E-i\eta),\qquad\mathcal{J}^{\sharp}_N(\zeta):=\overline{\mathcal{J}_N(\overline{\zeta})}. Z N ( E , η ) ∝ J N ( E + i η ) J N ♯ ( E − i η ) , J N ♯ ( ζ ) := J N ( ζ ) .
with the factorized expression analytically continued to the shifted parameters. Writing E = E + i η \mathcal{E}=E+i\eta E = E + i η , we have
E ′ + i η ′ = E + i k T , E ′ − i η ′ = E − i η = E ‾ . E'+i\eta'=\mathcal{E}+ikT,\qquad E'-i\eta'=E-i\eta=\overline{\mathcal{E}}. E ′ + i η ′ = E + ik T , E ′ − i η ′ = E − i η = E .
Thus the second factor cancels in the ratio. With q = k T 1 / 3 q=kT^{1/3} q = k T 1/3 and s = − E / T 2 / 3 \mathfrak{s}=-\mathcal{E}/T^{2/3} s = − E / T 2/3 , we arrive at
F N ( β ) ( q ; E ) = ⟨ e − i q Tr Z / T 1 / 3 ⟩ μ E , η = τ N ( β ) ( s − i q ) τ N ( β ) ( s ) , ⟨ Tr Z ⟩ μ E , η = T 1 / 3 ∂ s log τ N ( β ) ( s ) . (58) \mathcal{F}^{(\beta)}_N(q;\mathcal{E})=\left\langle e^{-iq\operatorname{Tr}Z/T^{1/3}}\right\rangle_{\mu_{E,\eta}}=\frac{\tau^{(\beta)}_N(\mathfrak{s}-iq)}{\tau^{(\beta)}_N(\mathfrak{s})},\qquad\left\langle\operatorname{Tr}Z\right\rangle_{\mu_{E,\eta}}=T^{1/3}\partial_{\mathfrak{s}}\log\tau^{(\beta)}_N(\mathfrak{s}). \tag*{(58)} F N ( β ) ( q ; E ) = ⟨ e − i q Tr Z / T 1/3 ⟩ μ E , η = τ N ( β ) ( s ) τ N ( β ) ( s − i q ) , ⟨ Tr Z ⟩ μ E , η = T 1/3 ∂ s log τ N ( β ) ( s ) . (58)
This is Eq. (7) at complex energy; for N = 1 N=1 N = 1 it reduces to Eq. (34) .
Why only the outgoing factor survives. The reason the answer depends on τ N ( β ) \tau^{(\beta)}_N τ N ( β ) alone becomes clear when the distribution is written in the variables of the double integral (54) . There the two matrices are
Λ = C ~ + 1 2 Y − i η Y − 1 = A − i P = Z , M = Λ − Y = Z − Y , (59) \Lambda=\widetilde{C}+\frac{1}{2}Y-i\eta Y^{-1}=A-iP=Z,\qquad M=\Lambda-Y=Z-Y, \tag*{(59)} Λ = C + 2 1 Y − i η Y − 1 = A − i P = Z , M = Λ − Y = Z − Y , (59)
and − Tr U E ( Z ) + Tr U E ‾ ( Z − Y ) = T Φ -\operatorname{Tr}U_{\mathcal{E}}(Z)+\operatorname{Tr}U_{\overline{\mathcal{E}}}(Z-Y)=T\Phi − Tr U E ( Z ) + Tr U E ( Z − Y ) = T Φ holds identically, even though A A A and Y Y Y do not commute. The stationary distribution (3) is therefore nothing but the integrand of J N J N ♯ \mathcal{J}_N\mathcal{J}^{\sharp}_N J N J N ♯ , a Gibbs weight in Z Z Z times an anti-Gibbs weight in Z − Y Z-Y Z − Y , evaluated on the deformed integration surface. Now insert into Eq. (54) an observable φ ( Z ) \varphi(Z) φ ( Z ) that is holomorphic and depends on Λ \Lambda Λ only. If φ \varphi φ is entire and grows at infinity more slowly than the cubic exponential, as e − i k Tr Z e^{-ik\operatorname{Tr}Z} e − ik Tr Z and polynomials do, it does not change the region where the integrand decays. The decomposition of step 4 then holds with the same integers n k l n_{kl} n k l , since these are a property of the integration surface Σ ′ \Sigma' Σ ′ and not of the integrand. Again only the ( 0 , 0 ) (0,0) ( 0 , 0 ) term survives, the M M M -integral still gives J N \mathcal{J}_N J N , and dividing by Eq. (5) we find
∫ φ d μ E , η = ∫ C φ e − Tr U E / T d Z ∫ C e − Tr U E / T d Z . (60) \int\varphi\,\mathrm{d}\mu_{E,\eta}=\frac{\int_{\mathcal{C}}\varphi e^{-\operatorname{Tr}U_{\mathcal{E}}/T}\,\mathrm{d}Z}{\int_{\mathcal{C}}e^{-\operatorname{Tr}U_{\mathcal{E}}/T}\,\mathrm{d}Z}. \tag*{(60)} ∫ φ d μ E , η = ∫ C e − Tr U E / T d Z ∫ C φ e − Tr U E / T d Z . (60)
For φ = e − i k Tr Z \varphi=e^{-ik\operatorname{Tr}Z} φ = e − ik Tr Z this is Eq. (58) again. Thus averages of holomorphic observables over the full joint probability density (3) can be evaluated with the normalized complex contour weight alone; for non-holomorphic observables this is not the case. This is the precise sense in which the formal, non-normalizable Gibbs weight e − Tr U E / T e^{-\operatorname{Tr}U_{\mathcal{E}}/T} e − Tr U E / T of the Riccati gas is correct: put on the outgoing integration surface, it gives every holomorphic average exactly. The full joint probability density is obtained, after normalization, from the product of the Gibbs weight of Z Z Z and the anti-Gibbs weight of Z − Y Z-Y Z − Y .
Zero-free region. Given Result 2, the left-hand side of (5) is a strictly positive integral for η > 0 \eta> 0 η > 0 . It remains finite and positive at η = 0 \eta= 0 η = 0 , and both sides are continuous as η → 0 + \eta\to0+ η → 0 + (§ 6 ). Hence
τ N ( β ) ( s ) ≠ 0 for Im s ≤ 0 , (61) \tau_N^{(\beta)}(s) \ne0 \qquad\text{for } \operatorname{Im}s \le0, \tag*{(61)} τ N ( β ) ( s ) = 0 for Im s ≤ 0 , (61)
the real axis included. Consequently the complex characteristic exponent Ξ N = T 1 / 3 ∂ s log τ N ( β ) \Xi_N = T^{1/3}\partial_s \log\tau_N^{(\beta)} Ξ N = T 1/3 ∂ s log τ N ( β ) has no poles for Im E ≥ 0 \operatorname{Im}E \ge0 Im E ≥ 0 , which is what the passage to real energy in § 6 requires.
The three Kontsevich integrals
In all three classes, τ N ( β ) \tau_N^{(\beta)} τ N ( β ) is the holomorphic factor of the positive integral Y N \mathcal{Y}_N Y N which was defined below equation (50) . The classes differ only in how τ \tau τ is evaluated (§ 7 ):
β = 2 \beta= 2 β = 2 : a Hankel determinant of Airy derivatives, det [ A − ( j + k ) ] \det[\mathcal{A}_{-}^{(j+k)}] det [ A − ( j + k ) ] which is precisely the fermionic result of [46 ] ;
β = 4 \beta= 4 β = 4 : a Pfaffian of Airy derivatives, Pf [ ( k − j ) A − ( j + k − 1 ) ] \operatorname{Pf}[(k-j)\mathcal{A}_{-}^{(j+k-1)}] Pf [( k − j ) A − ( j + k − 1 ) ] , expressible exactly in terms of the orthogonal-polynomial recurrence coefficients of the β = 2 \beta= 2 β = 2 ensemble (§ 11 );
β = 1 \beta= 1 β = 1 : a Pfaffian of ordered double integrals, expressible in terms of the orthogonal-polynomial norms and recurrence coefficients of the β = 2 \beta= 2 β = 2 ensemble at temperature T / 2 T/2 T /2 , together with one explicit Airy primitive, for even N N N (Subsec. 11D and Appendix § 21 ).
The positive integral Y N \mathcal{Y}_N Y N is the same object for all three, with β \beta β entering through ∣ Δ ∣ β \lvert\Delta\rvert^\beta ∣ Δ ∣ β and ( y i + y j ) − β / 2 (y_i+y_j)^{-\beta/2} ( y i + y j ) − β /2 .
Real energy: the non-absorptive limit
The zero-absorption distribution: a two-matrix model
In the stationary joint density of ( A , Y ) (A,Y) ( A , Y ) , the unnormalized weight depends on absorption only through the factor e − η 2 Tr Y − 1 / T e^{-\eta^2\operatorname{Tr}Y^{-1}/T} e − η 2 Tr Y − 1 / T . As η → 0 + \eta\to0+ η → 0 + , this factor increases to one, while the remaining zero-absorption weight is integrable (§ 14 E). The limit can therefore be taken directly in the joint density of ( A , Y ) (A,Y) ( A , Y ) : the normalization converges, as do the averages of all bounded observables of A A A and Y Y Y . In the limiting real-energy description, Y = T − 1 Y = T^{-1} Y = T − 1 , where T = − ∂ E Z T = -\partial_E Z T = − ∂ E Z (§ 14 ).
The linear change of variables Λ = A \Lambda= A Λ = A and M = A − Y M = A-Y M = A − Y has Jacobian of unit absolute value. In these variables the limiting distribution is
μ E , 0 ( d Λ d M ) = 1 Z N ( E , 0 ) e − Tr U E ( Λ ) / T e + Tr U E ( M ) / T d Λ d M , M < Λ . (62) \mu_{E,0}(\mathrm{d}\Lambda\,\mathrm{d}M)=\frac{1}{Z_N(E,0)}e^{-\operatorname{Tr}U_E(\Lambda)/T}e^{+\operatorname{Tr}U_E(M)/T}\,\mathrm{d}\Lambda\,\mathrm{d}M,\qquad M<\Lambda. \tag*{(62)} μ E , 0 ( d Λ d M ) = Z N ( E , 0 ) 1 e − Tr U E ( Λ ) / T e + Tr U E ( M ) / T d Λ d M , M < Λ. (62)
Here M < Λ M < \Lambda M < Λ means that Λ − M \Lambda- M Λ − M is positive definite. The joint weight consists of a Gibbs factor for Λ \Lambda Λ , an anti-Gibbs factor for M M M , and this matrix-order constraint. In the real-energy interpretation, Λ = Z \Lambda= Z Λ = Z and M = Z − T − 1 M = Z-T^{-1} M = Z − T − 1 . For N = 1 N = 1 N = 1 , the formula reduces to Eq. (35) .
Integrating out M M M gives the marginal density
P s t ( Λ ) = 1 Z N ( E , 0 ) e − Tr U E ( Λ ) / T ∫ M < Λ e Tr U E ( M ) / T d M . (63) P_{\mathrm{st}}(\Lambda)=\frac{1}{Z_N(E,0)}e^{-\operatorname{Tr}U_E(\Lambda)/T}\int_{M<\Lambda}e^{\operatorname{Tr}U_E(M)/T}\,\mathrm{d}M. \tag*{(63)} P st ( Λ ) = Z N ( E , 0 ) 1 e − Tr U E ( Λ ) / T ∫ M < Λ e Tr U E ( M ) / T d M . (63)
The integral converges. Indeed, writing M = Λ − Y M = \Lambda-Y M = Λ − Y with Y > 0 Y > 0 Y > 0 , the cubic term gives
Tr U E ( Λ − Y ) = − 1 3 Tr Y 3 + O ( Tr Y 2 + Tr Y + 1 ) \operatorname{Tr}U_E(\Lambda-Y)=-\frac{1}{3}\operatorname{Tr}Y^3+O\left(\operatorname{Tr}Y^2+\operatorname{Tr}Y+1\right) Tr U E ( Λ − Y ) = − 3 1 Tr Y 3 + O ( Tr Y 2 + Tr Y + 1 )
for fixed Λ \Lambda Λ and E E E , so the exponent decays as Y Y Y becomes large.
Here P s t P_{\mathrm{st}} P st is a density with respect to the flat matrix measure d Λ \mathrm{d}\Lambda d Λ . The density is invariant under conjugation and hence symmetric in the eigenvalues of Λ \Lambda Λ ; no restriction to an ordered eigenvalue chamber is imposed. For N = 1 N=1 N = 1 , Eq. (63) is Halperin’s density (36). We identify the density P s t ( Λ ) P_{\mathrm{st}}(\Lambda) P st ( Λ ) of Eq. (63) with the current-carrying stationary density of the real-energy Riccati matrix (§ 6 C). In this description, Riccati poles represent passages through infinity rather than termination of the evolution: a diverging eigenvalue re-enters from the opposite infinity. In the variables ( A , Y ) (A,Y) ( A , Y ) the same event is a passage of an eigenvalue of Y = T − 1 Y=\mathcal{T}^{-1} Y = T − 1 through zero. At η = 0 \eta=0 η = 0 the boundary det Y = 0 \det Y=0 det Y = 0 is therefore reached, and the appropriate boundary condition is reinjection, not zero flux. The density e Φ e^{\Phi} e Φ at η = 0 \eta=0 η = 0 does not vanish there, and the flux through this boundary is the winding current.
For β = 2 \beta=2 β = 2 , this identification implies agreement with the fermionic expression of [46 ] . A direct demonstration of the equality of the two expressions requires a considerable effort and is presented in Appendix A. For general β > 0 \beta>0 β > 0 the eigenvalue form of Eq. (63) can also be tested directly against the stationary Fokker–Planck equation of the eigenvalue gas (12), with zero flux at eigenvalue collisions and reinjection at + ∞ +\infty + ∞ of the eigenvalue escaping to − ∞ -\infty − ∞ . This is done in [42 ] , rigorously for N = 2 N=2 N = 2 and at the physics level for general N N N . That construction also yields the positive-integral representation of the normalization obtained from (50) at η = 0 \eta=0 η = 0 , together with Eqs. (65) and (68) for the principal-value mean trace and the winding current, respectively, for every β > 0 \beta>0 β > 0 .
To describe finite absorption, express both distributions in the same variables ( A , S ) (A,S) ( A , S ) , with S = Y − 1 S=Y^{-1} S = Y − 1 . At η > 0 \eta>0 η > 0 , S = P / η S=P/\eta S = P / η and Z = A − i η S Z=A-\mathrm{i}\eta S Z = A − i η S . The joint distribution then satisfies
μ E , η = Z N ( E , 0 ) Z N ( E , η ) e − η 2 Tr S / T μ E , 0 , Z = A − i η S . (64) \mu_{E,\eta}=\frac{\mathcal{Z}_N(E,0)}{\mathcal{Z}_N(E,\eta)}e^{-\eta^2\operatorname{Tr}S/T}\mu_{E,0},\qquad Z=A-\mathrm{i}\eta S. \tag*{(64)} μ E , η = Z N ( E , η ) Z N ( E , 0 ) e − η 2 Tr S / T μ E , 0 , Z = A − i η S . (64)
Thus absorption enters as an exponential reweighting of the zero-absorption distribution represented in ( A , S ) (A,S) ( A , S ) coordinates. In the real-energy limit, A = Z A=Z A = Z and S = T = − ∂ E Z S=\mathcal{T}=-\partial_E Z S = T = − ∂ E Z ; at finite absorption, S = P / η S=P/\eta S = P / η need not be identified with this energy derivative. The heavy tail of Tr S \operatorname{Tr}S Tr S prevents a term-by-term expansion of the exponential under the integral. As shown below, the normalization instead acquires a non-analytic term linear in η \eta η .
Γ N \Gamma_N Γ N and ν N \nu_N ν N as ratios of positive integrals
The results of this subsection use only the stationary joint distribution of ( A , Y ) (A,Y) ( A , Y ) in Eq. (3) , its limit as η → 0 + \eta\to0_+ η → 0 + , and the boundary values of Ξ N \Xi_N Ξ N discussed in § 6 C. They do not use the identification of Eq. (63) with the real-energy stationary state. Here unlabelled brackets denote expectation with respect to Eq. (3) at η > 0 \eta>0 η > 0 (or its corresponding marginal for observables depending only on Y Y Y ); real-energy averages are understood as the limits specified below.
Both quantities follow by differentiating the positive integral Y N ( E , η ) \mathcal{Y}_N(E,\eta) Y N ( E , η ) defined after Eq. (50) , in which E E E and η \eta η enter only through the factor e − ( E Tr Y + η 2 Tr Y − 1 ) / T e^{-(E\operatorname{Tr}Y+\eta^2\operatorname{Tr}Y^{-1})/T} e − ( E Tr Y + η 2 Tr Y − 1 ) / T .
Sum of Lyapunov exponents. Differentiation with respect to E E E brings down − Tr Y / T -\operatorname{Tr}Y/T − Tr Y / T , so − T ∂ E log Y N = ⟨ Tr Y ⟩ -T\partial_E\log\mathcal{Y}_N=\langle\operatorname{Tr}Y\rangle − T ∂ E log Y N = ⟨ Tr Y ⟩ . Since A = B + 1 2 Y A=B+\frac{1}{2}Y A = B + 2 1 Y with B B B conditionally Gaussian of zero mean, Γ N = Re ⟨ Tr Z ⟩ = ⟨ Tr A ⟩ = 1 2 ⟨ Tr Y ⟩ \Gamma_N=\operatorname{Re}\langle\operatorname{Tr}Z\rangle=\langle\operatorname{Tr}A\rangle=\frac{1}{2}\langle\operatorname{Tr}Y\rangle Γ N = Re ⟨ Tr Z ⟩ = ⟨ Tr A ⟩ = 2 1 ⟨ Tr Y ⟩ . At real energy this gives
Γ N ( E ) = − T 2 ∂ E log Y N ( E , 0 ) . (65) \Gamma_N(E)=-\frac{T}{2}\partial_E\log\mathcal{Y}_N(E,0). \tag*{(65)} Γ N ( E ) = − 2 T ∂ E log Y N ( E , 0 ) . (65)
At η = 0 \eta=0 η = 0 the mean of Tr A \operatorname{Tr}A Tr A is not absolutely convergent, because of the λ − 2 \lambda^{-2} λ − 2 tails of Eq. (13) . It is understood as the limit η → 0 + \eta\to0_+ η → 0 + of the convergent averages at η > 0 \eta>0 η > 0 , as in § 6 C. The mean ⟨ Tr Y ⟩ \langle\operatorname{Tr}Y\rangle ⟨ Tr Y ⟩ itself converges.
Integrated density of states. Differentiation with respect to η \eta η brings down − 2 η Tr Y − 1 / T -2\eta\operatorname{Tr}Y^{-1}/T − 2 η Tr Y − 1 / T . With P = η Y − 1 P=\eta Y^{-1} P = η Y − 1 ,
⟨ Tr P ⟩ = η ⟨ Tr Y − 1 ⟩ = − T 2 ∂ η log Y N ( E , η ) , η > 0. \langle\operatorname{Tr}P\rangle=\eta\langle\operatorname{Tr}Y^{-1}\rangle=-\frac{T}{2}\partial_\eta\log\mathcal{Y}_N(E,\eta),\qquad\eta>0. ⟨ Tr P ⟩ = η ⟨ Tr Y − 1 ⟩ = − 2 T ∂ η log Y N ( E , η ) , η > 0.
Since Im ⟨ Tr Z ⟩ = − ⟨ Tr P ⟩ \operatorname{Im}\langle\operatorname{Tr}Z\rangle=-\langle\operatorname{Tr}P\rangle Im ⟨ Tr Z ⟩ = − ⟨ Tr P ⟩ , Eq. (7) gives ⟨ Tr P ⟩ → π ν N ( E ) \langle\operatorname{Tr}P\rangle\to\pi\nu_N(E) ⟨ Tr P ⟩ → π ν N ( E ) as η → 0 + \eta\to0^+ η → 0 + .
The leading non-analytic dependence of Y N ( E , η ) \mathcal{Y}_N(E,\eta) Y N ( E , η ) comes from configurations in which one eigenvalue y i y_i y i of Y Y Y is of order η 2 / T \eta^2/T η 2 / T . The relevant boundary integral is
∫ 0 ∞ y − 1 / 2 ( e − η 2 / ( T y ) − 1 ) d y = − 2 π / T η , η > 0. \int_0^\infty y^{-1/2}\left(e^{-\eta^2/(Ty)}-1\right)\,dy=-2\sqrt{\pi/T}\,\eta,\qquad\eta>0. ∫ 0 ∞ y − 1/2 ( e − η 2 / ( T y ) − 1 ) d y = − 2 π / T η , η > 0.
As y 1 → 0 y_1\to0 y 1 → 0 , the pair-interaction factors in the joint eigenvalue density (50) satisfy, for each remaining eigenvalue y j y_j y j ,
∣ y j − y 1 ∣ β ( y j + y 1 ) − β / 2 ⟶ y j β / 2 . \lvert y_j-y_1\rvert^\beta(y_j+y_1)^{-\beta/2}\longrightarrow y_j^{\beta/2}. ∣ y j − y 1 ∣ β ( y j + y 1 ) − β /2 ⟶ y j β /2 .
Together with the original one-body factor y j − 1 / 2 y_j^{-1/2} y j − 1/2 , this gives the power y j ( β − 1 ) / 2 y_j^{(\beta-1)/2} y j ( β − 1 ) /2 in the boundary integral
Y N − 1 + ( E ) = ∫ R + N − 1 ∣ Δ ( y ) ∣ β ∏ i y i ( β − 1 ) / 2 ∏ i < j ( y i + y j ) − β / 2 e − 1 T ∑ i ( E y i + y i 3 / 12 ) ∏ i d y i . (66) \mathcal{Y}_{N-1}^+(E)=\int_{\mathbb{R}_+^{N-1}}\lvert\Delta(y)\rvert^\beta\prod_i y_i^{(\beta-1)/2}\prod_{i<j}(y_i+y_j)^{-\beta/2}e^{-\frac{1}{T}\sum_i(Ey_i+y_i^3/12)}\prod_i dy_i. \tag*{(66)} Y N − 1 + ( E ) = ∫ R + N − 1 ∣ Δ ( y ) ∣ β i ∏ y i ( β − 1 ) /2 i < j ∏ ( y i + y j ) − β /2 e − T 1 ∑ i ( E y i + y i 3 /12 ) i ∏ d y i . (66)
There are N N N equivalent choices of the vanishing eigenvalue. Boundary power counting shows that configurations with two or more small eigenvalues contribute only o ( η ) o(\eta) o ( η ) : when k k k eigenvalues of Y Y Y are simultaneously of order η 2 / T \eta^2/T η 2 / T , their combined boundary weight scales as η k + β k ( k − 1 ) / 2 \eta^{k+\beta k(k-1)/2} η k + β k ( k − 1 ) /2 . Consequently,
Y N ( E , η ) = Y N ( E , 0 ) − 2 π / T N η Y N − 1 + ( E ) + o ( η ) . \mathcal{Y}_N(E,\eta)=\mathcal{Y}_N(E,0)-2\sqrt{\pi/T}\,N\eta\,\mathcal{Y}_{N-1}^+(E)+o(\eta). Y N ( E , η ) = Y N ( E , 0 ) − 2 π / T N η Y N − 1 + ( E ) + o ( η ) .
Applying the same boundary rescaling directly to the differentiated integral gives
lim η → 0 + ∂ η Y N ( E , η ) = − 2 π / T N Y N − 1 + ( E ) . (67) \lim_{\eta\to0^+}\partial_\eta\mathcal{Y}_N(E,\eta)=-2\sqrt{\pi/T}\,N\mathcal{Y}_{N-1}^+(E). \tag*{(67)} η → 0 + lim ∂ η Y N ( E , η ) = − 2 π / T N Y N − 1 + ( E ) . (67)
Combining the relation
ν N ( E ) = − T 2 π lim η → 0 + ∂ η log Y N ( E , η ) \nu_N(E)=-\frac{T}{2\pi}\lim_{\eta\to0^+}\partial_\eta\log\mathcal{Y}_N(E,\eta) ν N ( E ) = − 2 π T η → 0 + lim ∂ η log Y N ( E , η )
with the limiting derivative in (67) and the convergence Y N ( E , η ) → Y N ( E , 0 ) \mathcal{Y}_N(E,\eta)\to\mathcal{Y}_N(E,0) Y N ( E , η ) → Y N ( E , 0 ) , we obtain
ν N ( E ) = T π N Y N − 1 + ( E ) Y N ( E , 0 ) . (68) \nu_N(E)=\sqrt{\frac{T}{\pi}\frac{N\mathcal{Y}_{N-1}^+(E)}{\mathcal{Y}_N(E,0)}}. \tag*{(68)} ν N ( E ) = π T Y N ( E , 0 ) N Y N − 1 + ( E ) . (68)
For N = 1 N=1 N = 1 , Y 0 + = 1 \mathcal{Y}_0^+=1 Y 0 + = 1 , and Eq. (68) reproduces Halperin’s integrated density of states,
ν 1 ( E ) = T 1 / 3 π 2 [ Ai 2 ( − E / T 2 / 3 ) + Bi 2 ( − E / T 2 / 3 ) ] . \nu_1(E)=\frac{T^{1/3}}{\pi^2\left[\operatorname{Ai}^2(-E/T^{2/3})+\operatorname{Bi}^2(-E/T^{2/3})\right]}. ν 1 ( E ) = π 2 [ Ai 2 ( − E / T 2/3 ) + Bi 2 ( − E / T 2/3 ) ] T 1/3 .
Comparison with Eq. (7) further gives
Y N − 1 + ∝ − Im ( τ N ′ τ N ‾ ) , \mathcal{Y}_{N-1}^+\propto-\operatorname{Im}\left(\tau_N'\overline{\tau_N}\right), Y N − 1 + ∝ − Im ( τ N ′ τ N ) ,
with the derivative taken in the convention of that equation. This expresses the positive boundary integral through the outgoing Kontsevich integral and its derivative.
Thus the integrated density of states is expressed as a boundary contribution from configurations with one vanishing inverse delay y i y_i y i . This is the positive-integral counterpart of the current associated with Riccati passages through infinity, generalizing the scalar pole-matching construction of Sec. 3F.
Justification of the limit
The stationary distribution (3) makes the passage to zero absorption transparent for typical configurations. As η → 0 + \eta\to0_{+} η → 0 + , the normalized density of ( A , Y ) (A,Y) ( A , Y ) approaches its integrable zero-absorption limit. Consequently, the probability of P = η Y − 1 P=\eta Y^{-1} P = η Y − 1 exceeding any fixed positive threshold tends to zero, and Z Z Z approaches the real matrix Λ \Lambda Λ . This does not imply that averages containing P P P vanish: rare configurations with small eigenvalues of Y Y Y can survive in such averages, as the finite limit ⟨ Tr P − 1 ⟩ → π ν N \langle\operatorname{Tr}P^{-1}\rangle\to\pi\nu_{N} ⟨ Tr P − 1 ⟩ → π ν N illustrates. The way this happens is fixed by the stationarity of log det P \log\det P log det P , which gives
η ⟨ Tr P − 1 ⟩ μ η = 2 Re ⟨ Tr Z ⟩ μ η → η → 0 2 Γ N ; (69) \eta\langle\operatorname{Tr}P^{-1}\rangle_{\mu_\eta}=2\operatorname{Re}\langle\operatorname{Tr}Z\rangle_{\mu_\eta}\xrightarrow{\eta\to0}2\Gamma_{N}; \tag*{(69)} η ⟨ Tr P − 1 ⟩ μ η = 2 Re ⟨ Tr Z ⟩ μ η η → 0 2 Γ N ; (69)
for the explicit distribution this is just ⟨ Tr Y ⟩ = 2 ⟨ Tr A ⟩ \langle\operatorname{Tr}Y\rangle=2\langle\operatorname{Tr}A\rangle ⟨ Tr Y ⟩ = 2 ⟨ Tr A ⟩ . Thus ⟨ Tr P − 1 ⟩ ≃ 2 Γ N / η \langle\operatorname{Tr}P^{-1}\rangle\simeq2\Gamma_{N}/\eta ⟨ Tr P − 1 ⟩ ≃ 2 Γ N / η : the imaginary part of Z Z Z vanishes on the scale set by comparing the level broadening η \eta η with the total Lyapunov growth rate Γ N \Gamma_{N} Γ N .
The characteristic function follows the same limit without difficulty. For q ≥ 0 q\ge0 q ≥ 0 the factor e − i q Tr Z / T 1 / 3 e^{-iq\operatorname{Tr}Z/T^{1/3}} e − i q Tr Z / T 1/3 is bounded and smooth up to real Z Z Z , so its average tends to the real-energy characteristic function F N ( q ; E ) \mathcal{F}_{N}(q;E) F N ( q ; E ) . On the other side of Eq. (58) , the ratio of τ \tau τ -functions is continuous up to real s s s , because τ \tau τ has no zeros there, Eq. (61) . Hence
F N ( β ) ( q ; E ) = τ N ( β ) ( s − i q ) τ N ( β ) ( s ) , E ∈ R , q > 0. (70) \mathcal{F}_{N}^{(\beta)}(q;E)=\frac{\tau_{N}^{(\beta)}(s-iq)}{\tau_{N}^{(\beta)}(s)},\qquad E\in\mathbb{R},\ q>0. \tag*{(70)} F N ( β ) ( q ; E ) = τ N ( β ) ( s ) τ N ( β ) ( s − i q ) , E ∈ R , q > 0. (70)
The right-hand side is a smooth function of q q q , and its expansion 1 − i q ∂ s log τ N ( β ) ( s ) + O ( q 2 ) 1-iq\partial_{s}\log\tau_{N}^{(\beta)}(s)+O(q^{2}) 1 − i q ∂ s log τ N ( β ) ( s ) + O ( q 2 ) for q → 0 + q\to0_{+} q → 0 + can be compared with the one-sided expansion (16) . This gives Eq. (7) . The comparison is legitimate only in this order: we first let η → 0 + \eta\to0_{+} η → 0 + at fixed q > 0 q>0 q > 0 , and only then let q → 0 + q\to0_{+} q → 0 + . The full characteristic function is not analytic at q = 0 q=0 q = 0 , since its negative-q q q branch is the complex conjugate. With this order of limits the multichannel Thouless relation Ξ N ( E + i 0 ) = Γ N − i π ν N \Xi_{N}(E+i0)=\Gamma_{N}-i\pi\nu_{N} Ξ N ( E + i 0 ) = Γ N − iπ ν N need not be assumed; it comes out as a by-product. Conversely, one may start from it: Ξ N = T 1 / 3 ∂ s log τ N ( β ) \Xi_{N}=T^{1/3}\partial_{s}\log\tau_{N}^{(\beta)} Ξ N = T 1/3 ∂ s log τ N ( β ) is holomorphic for η > 0 \eta>0 η > 0 , has no poles on the real axis, and by the Thouless formula [63 ] its boundary value is Γ N − i π ν N \Gamma_{N}-i\pi\nu_{N} Γ N − iπ ν N . The two routes thus give the same boundary values of Ξ N \Xi_{N} Ξ N .
Equations (62) and (63) give the matrix analogue of Halperin’s real-energy stationary distribution. The reinjection of Riccati eigenvalues from − ∞ -\infty − ∞ to + ∞ +\infty + ∞ requires matching tails n ( λ ) ≃ ν N / λ 2 n(\lambda)\simeq\nu_{N}/\lambda^{2} n ( λ ) ≃ ν N / λ 2 at both ends, as in Eq. (13) . These tails follow directly from the explicit distribution.
Conditioned on Y Y Y , the matrix B = A − 1 2 Y B=A-\frac{1}{2}Y B = A − 2 1 Y is Gaussian, with diagonal variances T / ( 2 y i ) T/(2y_i) T / ( 2 y i ) in the eigenbasis of Y Y Y . The leading contribution to a large eigenvalue λ \lambda λ of Λ = A \Lambda=A Λ = A comes from a single small eigenvalue y 1 y_{1} y 1 of Y Y Y , with A 11 A_{11} A 11 . Near y 1 = 0 y_{1}=0 y 1 = 0 , the density (50) behaves as y 1 − 1 / 2 y_{1}^{-1/2} y 1 − 1/2 times the integrand of Y N − 1 + \mathcal{Y}_{N-1}^{+} Y N − 1 + , as in § 6 . Hence
n ( λ ) ≃ N Y N − 1 + ( E ) Y N ( E , 0 ) ∫ 0 ∞ d y y y π T e − y λ 2 / T = T π N Y N − 1 + ( E ) Y N ( E , 0 ) 1 λ 2 , λ → ± ∞ . n(\lambda)\simeq\frac{N\mathcal{Y}_{N-1}^{+}(E)}{\mathcal{Y}_{N}(E,0)}\int_{0}^{\infty}\frac{dy}{\sqrt{y}}\sqrt{\frac{y}{\pi T}}e^{-y\lambda^{2}/T}=\sqrt{\frac{T}{\pi}}\frac{N\mathcal{Y}_{N-1}^{+}(E)}{\mathcal{Y}_{N}(E,0)}\frac{1}{\lambda^{2}},\qquad\lambda\to\pm\infty. n ( λ ) ≃ Y N ( E , 0 ) N Y N − 1 + ( E ) ∫ 0 ∞ y d y π T y e − y λ 2 / T = π T Y N ( E , 0 ) N Y N − 1 + ( E ) λ 2 1 , λ → ± ∞.
By Eq. (68) , the coefficient is ν N ( E ) \nu_{N}(E) ν N ( E ) at both ends, confirming the reinjection flux required by Eq. (13) .
Explicit finite-N N N formulas
We now evaluate the holomorphic factor τ N ( β ) \tau_{N}^{(\beta)} τ N ( β ) explicitly. All integrals run over the outgoing Airy contour C − \mathcal{C}_{-} C − , which comes in from infinity along arg w = π / 3 \arg w=\pi/3 arg w = π /3 and goes out to infinity along arg w = π \arg w=\pi arg w = π , with the weight e φ ( w ) e^{\varphi(w)} e φ ( w ) , φ ( w ) = w 3 / 3 − s w \varphi(w)=w^{3}/3-sw φ ( w ) = w 3 /3 − s w . After the angular integration, τ N ( β ) \tau_{N}^{(\beta)} τ N ( β ) is an N N N -fold integral of Δ ( w ) β ∏ i e φ ( w i ) \Delta(w)^\beta\prod_i e^{\varphi(w_i)} Δ ( w ) β ∏ i e φ ( w i ) , with the ordering and branch convention of Eq. (4) , and the three classes differ only in how the Vandermonde factor is handled. The single-variable building block is always the Airy function A − \mathcal{A}_{-} A − and its derivatives, obtained by inserting powers of w w w under the integral.
For β = 2 \beta=2 β = 2 the Vandermonde factor appears squared, and Andréief’s identity reduces the integral to a determinant of moments, τ N ( 2 ) ∝ det 0 ≤ j , k ≤ N − 1 [ A − ( j + k ) ( s ) ] \tau_N^{(2)} \propto\det_{0\le j,k\le N-1}[\mathcal{A}_{-}^{(j+k)}(s)] τ N ( 2 ) ∝ det 0 ≤ j , k ≤ N − 1 [ A − ( j + k ) ( s )] [46 ] . This is the Hankel determinant of the unitary problem.
For β = 4 \beta=4 β = 4 the same idea works once the fourth power of the Vandermonde determinant is written as a confluent Vandermonde determinant, Δ ( w ) 4 = det [ w i k , k w i k − 1 ] k = 0 , … , 2 N − 1 \Delta(w)^4=\det[w_i^k,\;kw_i^{k-1}]_{k=0,\ldots,2N-1} Δ ( w ) 4 = det [ w i k , k w i k − 1 ] k = 0 , … , 2 N − 1 . Each variable then carries a pair of columns, a power and its derivative, and de Bruijn’s integration formula [19 ] turns the integral into a Pfaffian of single-variable moments:
τ N ( 4 ) ( s ) ∝ Pf 0 ≤ j , k ≤ 2 N − 1 [ ( k − j ) A − ( j + k − 1 ) ( s ) ] . (71) \tau_N^{(4)}(s) \propto\operatorname{Pf}_{0\le j,k\le2N-1}\left[(k-j)\mathcal{A}_{-}^{(j+k-1)}(s)\right]. \tag*{(71)} τ N ( 4 ) ( s ) ∝ Pf 0 ≤ j , k ≤ 2 N − 1 [ ( k − j ) A − ( j + k − 1 ) ( s ) ] . (71)
The entry with j = k = 0 j=k=0 j = k = 0 is understood to be zero. For N = 1 N=1 N = 1 the Pfaffian is the single entry A − \mathcal{A}_{-} A − , as it must be.
For β = 1 \beta=1 β = 1 , the contour representation involves a single Vandermonde, which is antisymmetric under permutations of the integration variables. The integral must therefore retain their ordering along the oriented contour C − C_{-} C − . Applying de Bruijn’s formula to this ordered integral gives for even N N N :
τ N ( 1 ) ( s ) = Pf 0 ≤ j , k ≤ N − 1 [ a j k ( s ) ] , a j k = ∬ C − × C − sgn ( w ′ ≻ w ) w j w ′ k e φ ( w ) + φ ( w ′ ) d w d w ′ . (72) \tau_N^{(1)}(s)=\operatorname{Pf}_{0\le j,k\le N-1}[a_{jk}(s)],\qquad a_{jk}=\iint_{C_{-}\times C_{-}}\operatorname{sgn}(w'\succ w)w^j{w'}^k e^{\varphi(w)+\varphi(w')}dw\,dw'. \tag*{(72)} τ N ( 1 ) ( s ) = Pf 0 ≤ j , k ≤ N − 1 [ a j k ( s )] , a j k = ∬ C − × C − sgn ( w ′ ≻ w ) w j w ′ k e φ ( w ) + φ ( w ′ ) d w d w ′ . (72)
Here sgn ( w ′ ≻ w ) \operatorname{sgn}(w'\succ w) sgn ( w ′ ≻ w ) equals + 1 +1 + 1 if w ′ w' w ′ occurs after w w w along the oriented contour C − C_{-} C − , and − 1 -1 − 1 otherwise. It records contour ordering, not the sign of a complex number.
The simplest case is N = 2 N=2 N = 2 , where τ 2 ( 1 ) = a 01 = ∬ w ≺ w ′ ( w ′ − w ) e φ ( w ) + φ ( w ′ ) \tau_2^{(1)}=a_{01}=\iint_{w\prec w'}(w'-w)e^{\varphi(w)+\varphi(w')} τ 2 ( 1 ) = a 01 = ∬ w ≺ w ′ ( w ′ − w ) e φ ( w ) + φ ( w ′ ) . § 11 shows how for even N N N these Pfaffians can be expressed in terms of the recurrence coefficients and squared norms of the orthogonal polynomials for the unitary problem at temperature T / 2 T/2 T /2 .
For odd N N N the matrix is bordered by the single moments b j = ∫ w j e φ b_j=\int w^j e^\varphi b j = ∫ w j e φ . As usual for β = 1 \beta=1 β = 1 , these ordered double integrals do not reduce to derivatives of A − \mathcal{A}_{-} A − alone. Each of them is a single integral of an Airy-type function against its own primitive along the contour, a j k = ∫ w ′ k e φ ( w ′ ) [ 2 α j ( w ′ ) − α j ( ∞ ) ] d w ′ a_{jk}=\int w'^k e^{\varphi(w')}[2\alpha_j(w')-\alpha_j(\infty)]dw' a j k = ∫ w ′ k e φ ( w ′ ) [ 2 α j ( w ′ ) − α j ( ∞ )] d w ′ , with α j ( w ′ ) = ∫ i π / 3 ∞ w j e φ ( w ) d w \alpha_j(w')=\int_{i\pi/3}^{\infty}w^j e^{\varphi(w)}dw α j ( w ′ ) = ∫ iπ /3 ∞ w j e φ ( w ) d w . We do not treat odd-N N N case in this paper.
All three formulas reduce to the same answer for a single channel: τ 1 = A − \tau_1=\mathcal{A}_{-} τ 1 = A − and Ξ 1 = 1 3 A − ′ / A − \Xi_1=\frac{1}{3}\mathcal{A}_{-}'/\mathcal{A}_{-} Ξ 1 = 3 1 A − ′ / A − for every β \beta β . This is Halperin’s density of states [52 ] together with the associated Lyapunov exponent.
High energy: genus expansion with non-orientable surfaces
In this section, the high-energy limit is taken at fixed N N N ; the expansion is not uniform as N → ∞ N\to\infty N → ∞ .
At high energies E ≫ T 2 / 3 E\gg T^{2/3} E ≫ T 2/3 the disorder produces small corrections to the free-particle density of states, while the Lyapunov exponents describe the weak localization of the propagating channels. Both effects can be extracted from the same saddle-point expansion of the outgoing Kontsevich integral. In this regime s → − ∞ s\to-\infty s → − ∞ , and the two saddles of φ ( w ) = w 3 / 3 − s w \varphi(w)=w^3/3-sw φ ( w ) = w 3 /3 − s w are w ± = ± i ∣ s ∣ 1 / 2 w_{\pm}=\pm i|s|^{1/2} w ± = ± i ∣ s ∣ 1/2 . The outgoing contour C − C_{-} C − selects w 0 = i ∣ s ∣ 1 / 2 w_0=i|s|^{1/2} w 0 = i ∣ s ∣ 1/2 , around which all N N N integration variables fluctuate.
Writing w j = w 0 + u j w_j=w_0+u_j w j = w 0 + u j we obtain
φ ( w 0 + u ) = φ ( w 0 ) + w 0 u 2 + 1 3 u 3 . (73) \varphi(w_0+u)=\varphi(w_0)+w_0u^2+\frac{1}{3}u^3. \tag*{(73)} φ ( w 0 + u ) = φ ( w 0 ) + w 0 u 2 + 3 1 u 3 . (73)
This expansion is exact because the potential is cubic. The quadratic term defines a Gaussian β \beta β -ensemble, while the cubic term is treated perturbatively. Write
p k = ∑ j = 1 N u j k , σ = − 1 2 w 0 . p_k=\sum_{j=1}^{N}u_j^k,\qquad\sigma=-\frac{1}{2w_0}. p k = j = 1 ∑ N u j k , σ = − 2 w 0 1 .
To distinguish these averages from the stationary Riccati averages used earlier, we denote normalized Gaussian expectation by ⟨ ⋅ ⟩ G , β \langle\cdot\rangle_{\mathrm{G},\beta} ⟨ ⋅ ⟩ G , β . For σ > 0 \sigma>0 σ > 0 , its explicit definition is
⟨ F ⟩ G , β : = 1 G N , β ( σ ) ∫ R N F ( u ) ∣ Δ ( u ) ∣ β exp ( − 1 2 σ ∑ j = 1 N u j 2 ) ∏ j = 1 N d u j . (74) \langle F\rangle_{\mathrm{G},\beta}:=\frac{1}{G_{N,\beta}(\sigma)}\int_{\mathbb{R}^N}F(u)\lvert\Delta(u)\rvert^\beta\exp\left(-\frac{1}{2\sigma}\sum_{j=1}^{N}u_j^2\right)\prod_{j=1}^{N}\mathrm{d}u_j. \tag*{(74)} ⟨ F ⟩ G , β := G N , β ( σ ) 1 ∫ R N F ( u ) ∣ Δ ( u ) ∣ β exp ( − 2 σ 1 j = 1 ∑ N u j 2 ) j = 1 ∏ N d u j . (74)
G N , β ( σ ) : = ∫ R N ∣ Δ ( u ) ∣ β exp ( − 1 2 σ ∑ j = 1 N u j 2 ) ∏ j = 1 N d u j . (75) G_{N,\beta}(\sigma):=\int_{\mathbb{R}^N}\lvert\Delta(u)\rvert^\beta\exp\left(-\frac{1}{2\sigma}\sum_{j=1}^{N}u_j^2\right)\prod_{j=1}^{N}\mathrm{d}u_j. \tag*{(75)} G N , β ( σ ) := ∫ R N ∣ Δ ( u ) ∣ β exp ( − 2 σ 1 j = 1 ∑ N u j 2 ) j = 1 ∏ N d u j . (75)
Thus ⟨ 1 ⟩ G , β = 1 \langle1\rangle_{\mathrm{G},\beta}=1 ⟨ 1 ⟩ G , β = 1 ; the dependence on N N N and σ \sigma σ is suppressed in the bracket notation. For the complex value σ = − 1 / ( 2 w 0 ) \sigma=-1/(2w_0) σ = − 1/ ( 2 w 0 ) , the polynomial moments are defined by analytic continuation from σ > 0 \sigma>0 σ > 0 , with the continuation fixed by the local steepest-descent contour. The real-axis absolute Vandermonde in Eq. (75) specifies the starting measure, not an absolute-value prescription on the complex contour. The Gaussian normalization contributes a factor ( − w 0 ) − a β (-w_0)^{-a_\beta} ( − w 0 ) − a β , where
a β = N 2 + β N ( N − 1 ) 4 . a_\beta=\frac{N}{2}+\frac{\beta N(N-1)}{4}. a β = 2 N + 4 β N ( N − 1 ) .
The first-order contribution of the cubic perturbation vanishes because the cubic term is odd in the fluctuations about the stationary point: under the simultaneous sign reversal u j ↦ − u j u_j\mapsto-u_j u j ↦ − u j for all j j j , p 3 p_3 p 3 changes sign, whereas the Gaussian measure in Eq. (75) is invariant. Consequently ⟨ p 3 ⟩ G , β = 0 \langle p_3\rangle_{\mathrm{G},\beta}=0 ⟨ p 3 ⟩ G , β = 0 , and the first nonzero contribution therefore comes from two cubic vertices:
log τ N ( β ) ≃ N φ ( w 0 ) − a β log ( − w 0 ) + 1 18 ⟨ p 3 2 ⟩ G , β + ⋯ , ⟨ p 3 2 ⟩ G , β = K β ( N ) σ 3 . (76) \log\tau_N^{(\beta)}\simeq N\varphi(w_0)-a_\beta\log(-w_0)+\frac{1}{18}\langle p_3^2\rangle_{\mathrm{G},\beta}+\cdots,\qquad\langle p_3^2\rangle_{\mathrm{G},\beta}=K_\beta(N)\sigma^3. \tag*{(76)} log τ N ( β ) ≃ N φ ( w 0 ) − a β log ( − w 0 ) + 18 1 ⟨ p 3 2 ⟩ G , β + ⋯ , ⟨ p 3 2 ⟩ G , β = K β ( N ) σ 3 . (76)
An additive constant independent of s s s has been omitted since it does not affect the characteristic exponent.
The high-energy correction is thus determined by a single Gaussian moment. It can be evaluated for all three symmetry classes using integration by parts, without performing separate Wick-contraction calculations. To derive the required identity, first take σ > 0 \sigma>0 σ > 0 and write the normalized Gaussian β \beta β -ensemble density as
P β ( u ) = 1 G N , β ( σ ) ∣ Δ ( u ) ∣ β e − ∑ j u j 2 / ( 2 σ ) . P_\beta(u)=\frac{1}{G_{N,\beta}(\sigma)}\lvert\Delta(u)\rvert^\beta e^{-\sum_j u_j^2/(2\sigma)}. P β ( u ) = G N , β ( σ ) 1 ∣ Δ ( u ) ∣ β e − ∑ j u j 2 / ( 2 σ ) .
Away from coincident eigenvalues, its logarithmic derivative is
∂ i log P β ( u ) = − u i σ + β ∑ j ≠ i 1 u i − u j . \partial_i\log P_\beta(u)=-\frac{u_i}{\sigma}+\beta\sum_{j\ne i}\frac{1}{u_i-u_j}. ∂ i log P β ( u ) = − σ u i + β j = i ∑ u i − u j 1 .
For a symmetric polynomial g g g integrating the total derivative ∑ i ∂ i [ u i k g P β ] \sum_i\partial_i[u_i^k gP_\beta] ∑ i ∂ i [ u i k g P β ] gives zero. The Gaussian decay removes the boundary terms at infinity, while the Vandermonde factor removes them at coincident eigenvalues. Expanding the derivative therefore gives
0 = ⟨ ∑ i ∂ i ( u i k g ) ⟩ G , β − 1 σ ⟨ p k + 1 g ⟩ G , β + β ⟨ ∑ i ∑ j ≠ i u i k u i − u j g ⟩ G , β . 0=\left\langle\sum_i\partial_i(u_i^k g)\right\rangle_{\mathrm{G},\beta}-\frac{1}{\sigma}\langle p_{k+1}g\rangle_{\mathrm{G},\beta}+\beta\left\langle\sum_i\sum_{j\ne i}\frac{u_i^k}{u_i-u_j}g\right\rangle_{\mathrm{G},\beta}. 0 = ⟨ i ∑ ∂ i ( u i k g ) ⟩ G , β − σ 1 ⟨ p k + 1 g ⟩ G , β + β ⟨ i ∑ j = i ∑ u i − u j u i k g ⟩ G , β .
Pairing the terms with indices ( i , j ) (i,j) ( i , j ) and ( j , i ) (j,i) ( j , i ) , we obtain
1 σ ⟨ p k + 1 g ⟩ G , β = ⟨ ∑ i ∂ i ( u i k g ) ⟩ G , β + β ⟨ ∑ i < j u i k − u j k u i − u j g ⟩ G , β . (77) \frac{1}{\sigma}\langle p_{k+1}g\rangle_{\mathrm{G},\beta}=\left\langle\sum_i\partial_i(u_i^k g)\right\rangle_{\mathrm{G},\beta}+\beta\left\langle\sum_{i<j}\frac{u_i^k-u_j^k}{u_i-u_j}g\right\rangle_{\mathrm{G},\beta}. \tag*{(77)} σ 1 ⟨ p k + 1 g ⟩ G , β = ⟨ i ∑ ∂ i ( u i k g ) ⟩ G , β + β ⟨ i < j ∑ u i − u j u i k − u j k g ⟩ G , β . (77)
These integration-by-parts identities are usually called Gaussian loop equations. For the polynomial moments used below, they extend by analytic continuation to the complex value σ = − 1 / ( 2 w 0 ) \sigma=-1/(2w_0) σ = − 1/ ( 2 w 0 ) selected by the saddle. Introducing m 2 = N + β N ( N − 1 ) / 2 m_2=N+\beta N(N-1)/2 m 2 = N + β N ( N − 1 ) /2 , the moments needed here follow successively as
⟨ p 2 ⟩ G , β = σ m 2 , \langle p_2\rangle_{G,\beta}=\sigma m_2, ⟨ p 2 ⟩ G , β = σ m 2 ,
⟨ p 1 p 3 ⟩ G , β = 3 σ 2 m 2 , \langle p_1p_3\rangle_{G,\beta}=3\sigma^2m_2, ⟨ p 1 p 3 ⟩ G , β = 3 σ 2 m 2 ,
⟨ p 4 ⟩ G , β = σ 2 [ 3 m 2 + β ( ( N − 1 ) m 2 + 1 2 ( N − m 2 ) ) ] , \langle p_4\rangle_{G,\beta}=\sigma^2\left[3m_2+\beta\left((N-1)m_2+\frac{1}{2}(N-m_2)\right)\right], ⟨ p 4 ⟩ G , β = σ 2 [ 3 m 2 + β ( ( N − 1 ) m 2 + 2 1 ( N − m 2 ) ) ] ,
⟨ p 3 2 ⟩ G , β = σ [ ( 2 + β ( N − 1 ) ) ⟨ p 1 p 3 ⟩ G , β + 3 ⟨ p 4 ⟩ G , β ] . \langle p_3^2\rangle_{G,\beta}=\sigma\left[(2+\beta(N-1))\langle p_1p_3\rangle_{G,\beta}+3\langle p_4\rangle_{G,\beta}\right]. ⟨ p 3 2 ⟩ G , β = σ [ ( 2 + β ( N − 1 )) ⟨ p 1 p 3 ⟩ G , β + 3 ⟨ p 4 ⟩ G , β ] .
Combining these relations yields
K β ( N ) = 3 N 4 [ 4 β 2 N 2 − 9 β 2 N + 5 β 2 + 18 β N − 18 β + 20 ] . (78) K_\beta(N)=\frac{3N}{4}\left[4\beta^2N^2-9\beta^2N+5\beta^2+18\beta N-18\beta+20\right]. \tag*{(78)} K β ( N ) = 4 3 N [ 4 β 2 N 2 − 9 β 2 N + 5 β 2 + 18 β N − 18 β + 20 ] . (78)
In particular,
K 2 = 3 N ( 4 N 2 + 1 ) , K 1 = 3 N 4 ( 4 N 2 + 9 N + 7 ) , K 4 = 3 N ( 16 N 2 − 18 N + 7 ) . K_2=3N(4N^2+1),\qquad K_1=\frac{3N}{4}(4N^2+9N+7),\qquad K_4=3N(16N^2-18N+7). K 2 = 3 N ( 4 N 2 + 1 ) , K 1 = 4 3 N ( 4 N 2 + 9 N + 7 ) , K 4 = 3 N ( 16 N 2 − 18 N + 7 ) .
The dependence on N N N has a topological interpretation in terms of ribbon graphs generated by the two cubic vertices; see, e.g., Chapter 3 of Ref. [64 ] . In the unitary class, only orientable surfaces contribute: the N 3 N^3 N 3 and N N N terms correspond to the planar and torus contributions, respectively. For β = 1 , 4 \beta=1,4 β = 1 , 4 , an additional N 2 N^2 N 2 term appears, reflecting the contribution of non-orientable surfaces.
The real and quaternion coefficients obey the polynomial identity
K 4 ( N ) = − 2 K 1 ( − 2 N ) . K_4(N)=-2K_1(-2N). K 4 ( N ) = − 2 K 1 ( − 2 N ) .
This is the two-cubic-vertex manifestation, in our normalization, of the orthogonal–symplectic Gaussian-integral duality studied by Mulase and Waldron [72 ] . The substitution N ↦ − 2 N N\mapsto-2N N ↦ − 2 N is understood formally at the level of the diagrammatic polynomials, not as an actual negative matrix dimension. After the appropriate normalization, the two ensembles have the same surface expansion, except that non-orientable surfaces of odd Euler characteristic contribute with opposite signs. For two cubic vertices, V = 2 V=2 V = 2 and E = 3 E=3 E = 3 , so χ = V − E + F = F − 1 \chi=V-E+F=F-1 χ = V − E + F = F − 1 . The N 2 N^2 N 2 term therefore corresponds to χ = 1 \chi=1 χ = 1 , i.e. the projective plane, explaining its opposite sign in the real and quaternion classes.
We now return from the Gaussian expansion to the physical observables. Using ∂ s φ ( w 0 ) = − w 0 \partial_s\varphi(w_0)=-w_0 ∂ s φ ( w 0 ) = − w 0 and ∂ s w 0 = 1 / ( 2 w 0 ) \partial_sw_0=1/(2w_0) ∂ s w 0 = 1/ ( 2 w 0 ) , and restoring physical units with z 0 = − i E z_0=-i\sqrt{E} z 0 = − i E , we find
Ξ N ( β ) = N z 0 − a β T 2 z 0 2 − K β ( N ) T 2 96 z 0 5 + O ( T 3 ) . (79) \Xi_N^{(\beta)}=Nz_0-\frac{a_\beta T}{2z_0^2}-\frac{K_\beta(N)T^2}{96z_0^5}+O(T^3). \tag*{(79)} Ξ N ( β ) = N z 0 − 2 z 0 2 a β T − 96 z 0 5 K β ( N ) T 2 + O ( T 3 ) . (79)
The real and imaginary parts give the sum of Lyapunov exponents and the density of states:
Γ N = N ( β N + 2 − β ) T 8 E + … , ρ N N = 1 2 π E [ 1 − 5 K β ( N ) T 2 96 N E 3 + … ] . (80) \Gamma_N=\frac{N(\beta N+2-\beta)T}{8E}+\ldots,\qquad\frac{\rho_N}{N}=\frac{1}{2\pi\sqrt{E}}\left[1-\frac{5K_\beta(N)T^2}{96NE^3}+\ldots\right]. \tag*{(80)} Γ N = 8 E N ( β N + 2 − β ) T + … , N ρ N = 2 π E 1 [ 1 − 96 N E 3 5 K β ( N ) T 2 + … ] . (80)
The leading density is the free-particle result obtained by completely neglecting the random potential in the original matrix Schrödinger equation. Localization effects reflected in the positivity of Lyapunov’s appear already at order T T T , whereas the first correction to the density of states is of order T 2 T^2 T 2 .
The leading Lyapunov sum can also be written as
Γ N = T 4 E ∑ j = 1 N [ 1 + β ( j − 1 ) ] + … . \Gamma_N=\frac{T}{4E}\sum_{j=1}^{N}[1+\beta(j-1)]+\ldots. Γ N = 4 E T j = 1 ∑ N [ 1 + β ( j − 1 )] + … .
Its channel dependence is therefore consistent with the familiar isotropic, DMPK-type ladder of weak-disorder Lyapunov exponents γ j ∝ 1 + β ( j − 1 ) \gamma_j\propto1+\beta(j-1) γ j ∝ 1 + β ( j − 1 ) [70 ] . The symmetry class enters through the spacing of this ladder, while T / E T/E T / E sets its overall scale.
The Lifshitz tail: one-eigenvalue tunnelling
We now turn to the opposite regime, − E ≫ T 2 / 3 -E \gg T^{2/3} − E ≫ T 2/3 , where states arise from rare fluctuations of the random potential. Throughout this section the deep-tail limit is taken at fixed N N N . The joint large-N N N edge regime is treated separately in § 10 . In the Kontsevich representation, their exponentially small density is encoded by tunnelling between the two saddles of the cubic potential. For s → + ∞ s \to+\infty s → + ∞ these saddles lie on the real axis,
w ± = ± s , φ ( w ± ) = ± 2 3 s 3 / 2 . w_{\pm} = \pm\sqrt{s}, \qquad\varphi(w_{\pm}) = \pm\frac{2}{3}s^{3/2}. w ± = ± s , φ ( w ± ) = ± 3 2 s 3/2 .
The outgoing contour C − C_{-} C − can be deformed to come in from e i π / 3 ∞ e^{i\pi/3}\infty e iπ /3 ∞ , cross w + w_{+} w + vertically, and then run left along the real axis through the dominant saddle w − w_{-} w − before leaving to − ∞ -\infty − ∞ .
The dominant configuration has all N N N integration variables near w − w_{-} w − . Its contribution is real and gives the leading Lyapunov growth, Γ N ≃ N − E \Gamma_N \simeq N\sqrt{-E} Γ N ≃ N − E , but no density of states. The latter comes from the exponentially smaller configuration in which one integration variable moves to w + w_{+} w + while the remaining N − 1 N-1 N − 1 stay near w − w_{-} w − . We denote these two contributions to the partition function by τ N ( 0 ) \tau_N^{(0)} τ N ( 0 ) and τ N ( 1 ) \tau_N^{(1)} τ N ( 1 ) , respectively. The superscripts here count tunnelling eigenvalues, not symmetry classes.
The exponential cost of this process is
e φ ( w + ) − φ ( w − ) = e − 4 3 s 3 / 2 . e^{\varphi(w_{+})-\varphi(w_{-})} = e^{-\frac{4}{3}s^{3/2}}. e φ ( w + ) − φ ( w − ) = e − 3 4 s 3/2 .
Restoring physical units gives Halperin’s action, S = 4 3 ( − E ) 3 / 2 / T S = \frac{4}{3}(-E)^{3/2}/T S = 3 4 ( − E ) 3/2 / T . At fixed T T T , this action is independent of both N N N and β \beta β : a single tunnelling eigenvalue carries the rare state. The channel number and symmetry class enter through its prefactor.
The phase and the Stokes multiplier of the tunnelling contribution are fixed by the scalar problem. Indeed, for N = 1 N=1 N = 1 the outgoing Airy combination A − = A i − i B i A_{-} = \mathrm{Ai} - i\mathrm{Bi} A − = Ai − i Bi gives
A i ( s ) − i B i ( s ) ≃ i 2 e − 4 3 s 3 / 2 . \frac{\mathrm{Ai}(s)}{-i\mathrm{Bi}(s)} \simeq\frac{i}{2}e^{-\frac{4}{3}s^{3/2}}. − i Bi ( s ) Ai ( s ) ≃ 2 i e − 3 4 s 3/2 .
The factor 1 2 \frac{1}{2} 2 1 reflects that the positive real s s s -axis is a Stokes line. For general N N N , this scalar tunnelling amplitude must be combined with the interaction between the tunnelling eigenvalue and the remaining eigenvalues, and with the Gaussian normalizations around the two saddles.
At leading order the cross-factors in the Vandermonde weight, coupling the tunnelling eigenvalue to the remaining N − 1 N-1 N − 1 eigenvalues, contribute
( w − − w + ) β ( N − 1 ) = ( − 2 s ) β ( N − 1 ) . (w_{-}-w_{+})^{\beta(N-1)} = (-2\sqrt{s})^{\beta(N-1)}. ( w − − w + ) β ( N − 1 ) = ( − 2 s ) β ( N − 1 ) .
The ratio of Gaussian β \beta β -ensemble normalizations, with curvature 2 s 2\sqrt{s} 2 s at w − w_{-} w − , can be obtained as follows. Let G n ( a ) G_n(a) G n ( a ) denote the ordered Gaussian β \beta β -ensemble integral
G n ( a ) = ∫ u 1 < ⋯ < u n ∣ Δ ( u ) ∣ β e − a 2 ∑ j = 1 n u j 2 ∏ j = 1 n d u j , G 0 ( a ) = 1. G_n(a) = \int_{u_1<\cdots<u_n} |\Delta(u)|^\beta e^{-\frac{a}{2}\sum_{j=1}^{n}u_j^2}\prod_{j=1}^{n}du_j, \qquad G_0(a)=1. G n ( a ) = ∫ u 1 < ⋯ < u n ∣Δ ( u ) ∣ β e − 2 a ∑ j = 1 n u j 2 j = 1 ∏ n d u j , G 0 ( a ) = 1.
The Mehta formula [33 ] gives
G n ( a ) = ( 2 π ) n / 2 n ! a − n / 2 − β n ( n − 1 ) / 4 ∏ j = 1 n Γ ( 1 + j β / 2 ) Γ ( 1 + β / 2 ) . G_n(a) = \frac{(2\pi)^{n/2}}{n!}a^{-n/2-\beta n(n-1)/4}\prod_{j=1}^{n}\frac{\Gamma(1+j\beta/2)}{\Gamma(1+\beta/2)}. G n ( a ) = n ! ( 2 π ) n /2 a − n /2 − β n ( n − 1 ) /4 j = 1 ∏ n Γ ( 1 + β /2 ) Γ ( 1 + j β /2 ) .
Consequently, with a = 2 s a=2\sqrt{s} a = 2 s ,
G 1 ( a ) G N − 1 ( a ) G N ( a ) = N Γ ( 1 + β / 2 ) Γ ( 1 + N β / 2 ) ( 2 s ) β ( N − 1 ) / 2 . \frac{G_1(a)G_{N-1}(a)}{G_N(a)} = N\frac{\Gamma(1+\beta/2)}{\Gamma(1+N\beta/2)}(2\sqrt{s})^{\beta(N-1)/2}. G N ( a ) G 1 ( a ) G N − 1 ( a ) = N Γ ( 1 + N β /2 ) Γ ( 1 + β /2 ) ( 2 s ) β ( N − 1 ) /2 .
The contour orientation must also be retained. The tunnelling eigenvalue comes first along C − C_{-} C − , whereas the contour near w − w_{-} w − runs in the negative direction. In the real class, these orientation signs cancel the signs of the ordered Vandermonde. Combining the factors gives
τ N ( 1 ) τ N ( 0 ) = i 2 e − 4 3 s 3 / 2 N Γ ( 1 + β / 2 ) Γ ( 1 + N β / 2 ) ( 2 s ) 3 β ( N − 1 ) 2 [ 1 + κ N s 3 / 2 + ⋯ ] . (81) \frac{\tau_N^{(1)}}{\tau_N^{(0)}}=\frac{\mathrm{i}}{2}e^{-\frac{4}{3}s^{3/2}}N\frac{\Gamma(1+\beta/2)}{\Gamma(1+N\beta/2)}(2\sqrt{s})^{\frac{3\beta(N-1)}{2}}\left[1+\frac{\kappa_N}{s^{3/2}}+\cdots\right]. \tag*{(81)} τ N ( 0 ) τ N ( 1 ) = 2 i e − 3 4 s 3/2 N Γ ( 1 + N β /2 ) Γ ( 1 + β /2 ) ( 2 s ) 2 3 β ( N − 1 ) [ 1 + s 3/2 κ N + ⋯ ] . (81)
To obtain the first correction, we must include both the cubic fluctuations around the saddles and the variation of the cross Vandermonde. The cubic contribution follows from the Gaussian moments computed in § 8 . The fluctuation variance is + 1 / ( 2 s ) +1/(2\sqrt{s}) + 1/ ( 2 s ) at w − w_{-} w − and − 1 / ( 2 s ) -1/(2\sqrt{s}) − 1/ ( 2 s ) at w + w_{+} w + , so the two saddle sectors give the contribution
K β ( N − 1 ) − K β ( N ) − K β ( 1 ) 144 \frac{K_{\beta}(N-1)-K_{\beta}(N)-K_{\beta}(1)}{144} 144 K β ( N − 1 ) − K β ( N ) − K β ( 1 )
to κ N \kappa_N κ N .
For the Vandermonde correction, let u 1 u_1 u 1 be the fluctuation of the tunnelling eigenvalue around w + w_{+} w + and let u j u_j u j , j = 2 , … , N j=2,\ldots,N j = 2 , … , N , be the fluctuations around the dominant saddle w − w_{-} w − . Writing δ j = u j − u 1 \delta_j=u_j-u_1 δ j = u j − u 1 and L 1 = − β 2 s ∑ j = 2 N δ j L_1=-\frac{\beta}{2\sqrt{s}}\sum_{j=2}^{N}\delta_j L 1 = − 2 s β ∑ j = 2 N δ j , we expand the cross factor ∏ j = 2 N ∣ 1 − δ j 2 s ∣ β \prod_{j=2}^{N}\left|1-\frac{\delta_j}{2\sqrt{s}}\right|^{\beta} ∏ j = 2 N 1 − 2 s δ j β . Let ⟨ ⋅ ⟩ t u n \langle\cdot\rangle_{\mathrm{tun}} ⟨ ⋅ ⟩ tun denote the normalized Gaussian average around the one-eigenvalue tunnelling configuration: u 1 u_1 u 1 fluctuates around w + w_{+} w + with variance parameter ⟨ u 1 2 ⟩ t u n = − σ 2 \langle u_1^2\rangle_{\mathrm{tun}}=-\sigma^2 ⟨ u 1 2 ⟩ tun = − σ 2 and ⟨ u 1 4 ⟩ t u n = 3 σ 4 \langle u_1^4\rangle_{\mathrm{tun}}=3\sigma^4 ⟨ u 1 4 ⟩ tun = 3 σ 4 , where σ 2 = 1 / ( 2 s ) \sigma^2=1/(2\sqrt{s}) σ 2 = 1/ ( 2 s ) , while u 2 , … , u N u_2,\ldots,u_N u 2 , … , u N form an independent Gaussian β \beta β -ensemble around w − w_{-} w − with variance parameter σ 2 \sigma^2 σ 2 (with the negative parameter understood along the local steepest-descent contour). At the required order, the three contributions are
⟨ − β 8 s ∑ j = 2 N δ j 2 ⟩ t u n , 1 2 ⟨ L 1 2 ⟩ t u n , ⟨ L 1 p 3 + u 1 3 3 ⟩ t u n , \left\langle-\frac{\beta}{8s}\sum_{j=2}^{N}\delta_j^2\right\rangle_{\mathrm{tun}},\qquad\frac{1}{2}\langle L_1^2\rangle_{\mathrm{tun}},\qquad\left\langle L_1\frac{p_3+u_1^3}{3}\right\rangle_{\mathrm{tun}}, ⟨ − 8 s β j = 2 ∑ N δ j 2 ⟩ tun , 2 1 ⟨ L 1 2 ⟩ tun , ⟨ L 1 3 p 3 + u 1 3 ⟩ tun ,
where p k = ∑ j = 2 N u j k p_k=\sum_{j=2}^{N}u_j^k p k = ∑ j = 2 N u j k . The last term accounts for the interference between the linear Vandermonde correction and the cubic vertices. For the ( N − 1 ) (N-1) ( N − 1 ) -particle ensemble around w − w_{-} w − , the center of mass p 1 p_1 p 1 decouples from the translation-invariant Vandermonde factor, giving ⟨ p 1 2 ⟩ t u n = ( N − 1 ) σ 2 \langle p_1^2\rangle_{\mathrm{tun}}=(N-1)\sigma^2 ⟨ p 1 2 ⟩ tun = ( N − 1 ) σ 2 . Integrating by parts with respect to u j u_j u j in the ( N − 1 ) (N-1) ( N − 1 ) -particle Gaussian β \beta β -ensemble (via ∑ j = 2 N ∂ u j ( u j P ) = 0 \sum_{j=2}^{N}\partial_{u_j}(u_jP)=0 ∑ j = 2 N ∂ u j ( u j P ) = 0 and the translation-invariance identity ∑ j = 2 N ∂ u j ( p 3 P ) = 0 \sum_{j=2}^{N}\partial_{u_j}(p_3P)=0 ∑ j = 2 N ∂ u j ( p 3 P ) = 0 ) yields
⟨ p 2 ⟩ t u n = ( N − 1 ) σ 2 ( 1 + β 2 ( N − 2 ) ) , ⟨ p 1 p 3 ⟩ t u n = 3 σ 2 ⟨ p 2 ⟩ t u n . \langle p_2\rangle_{\mathrm{tun}}=(N-1)\sigma^2\left(1+\frac{\beta}{2}(N-2)\right),\qquad\langle p_1p_3\rangle_{\mathrm{tun}}=3\sigma^2\langle p_2\rangle_{\mathrm{tun}}. ⟨ p 2 ⟩ tun = ( N − 1 ) σ 2 ( 1 + 2 β ( N − 2 ) ) , ⟨ p 1 p 3 ⟩ tun = 3 σ 2 ⟨ p 2 ⟩ tun .
Using the independence of u 1 u_1 u 1 from { u 2 , … , u N } \{u_2,\ldots,u_N\} { u 2 , … , u N } and ⟨ u 1 ⟩ t u n = ⟨ u 1 3 ⟩ t u n = 0 \langle u_1\rangle_{\mathrm{tun}}=\langle u_1^3\rangle_{\mathrm{tun}}=0 ⟨ u 1 ⟩ tun = ⟨ u 1 3 ⟩ tun = 0 , the three contributions evaluate to
⟨ − β 8 s ∑ j = 2 N δ j 2 ⟩ t u n = − β 8 s ( ⟨ p 2 ⟩ t u n + ( N − 1 ) ⟨ u 1 2 ⟩ t u n ) = − β 2 ( N − 1 ) ( N − 2 ) 32 s 3 / 2 , \left\langle-\frac{\beta}{8s}\sum_{j=2}^{N}\delta_j^2\right\rangle_{\mathrm{tun}}=-\frac{\beta}{8s}\left(\langle p_2\rangle_{\mathrm{tun}}+(N-1)\langle u_1^2\rangle_{\mathrm{tun}}\right)=-\frac{\beta^2(N-1)(N-2)}{32s^{3/2}}, ⟨ − 8 s β j = 2 ∑ N δ j 2 ⟩ tun = − 8 s β ( ⟨ p 2 ⟩ tun + ( N − 1 ) ⟨ u 1 2 ⟩ tun ) = − 32 s 3/2 β 2 ( N − 1 ) ( N − 2 ) ,
1 2 ⟨ L 1 2 ⟩ t u n = β 2 8 s ( ⟨ p 1 2 ⟩ t u n + ( N − 1 ) 2 ⟨ u 1 2 ⟩ t u n ) = − β 2 ( N − 1 ) ( N − 2 ) 16 s 3 / 2 , \frac{1}{2}\langle L_1^2\rangle_{\mathrm{tun}}=\frac{\beta^2}{8s}\left(\langle p_1^2\rangle_{\mathrm{tun}}+(N-1)^2\langle u_1^2\rangle_{\mathrm{tun}}\right)=-\frac{\beta^2(N-1)(N-2)}{16s^{3/2}}, 2 1 ⟨ L 1 2 ⟩ tun = 8 s β 2 ( ⟨ p 1 2 ⟩ tun + ( N − 1 ) 2 ⟨ u 1 2 ⟩ tun ) = − 16 s 3/2 β 2 ( N − 1 ) ( N − 2 ) ,
⟨ L 1 p 3 + u 1 3 3 ⟩ t u n = − β 6 s ( ⟨ p 1 p 3 ⟩ t u n − ( N − 1 ) ⟨ u 1 4 ⟩ t u n ) = − β 2 ( N − 1 ) ( N − 2 ) 16 s 3 / 2 . \left\langle L_1\frac{p_3+u_1^3}{3}\right\rangle_{\mathrm{tun}}=-\frac{\beta}{6\sqrt{s}}\left(\langle p_1p_3\rangle_{\mathrm{tun}}-(N-1)\langle u_1^4\rangle_{\mathrm{tun}}\right)=-\frac{\beta^2(N-1)(N-2)}{16s^{3/2}}. ⟨ L 1 3 p 3 + u 1 3 ⟩ tun = − 6 s β ( ⟨ p 1 p 3 ⟩ tun − ( N − 1 ) ⟨ u 1 4 ⟩ tun ) = − 16 s 3/2 β 2 ( N − 1 ) ( N − 2 ) .
Their respective contributions have the relative weights 1 32 : 1 16 : 1 16 \frac{1}{32}:\frac{1}{16}:\frac{1}{16} 32 1 : 16 1 : 16 1 , and together give the coefficient − 5 32 -\frac{5}{32} − 32 5 . Thus
κ N = K β ( N − 1 ) − K β ( N ) − K β ( 1 ) 144 − 5 β 2 ( N − 1 ) ( N − 2 ) 32 . \kappa_N=\frac{K_{\beta}(N-1)-K_{\beta}(N)-K_{\beta}(1)}{144}-\frac{5\beta^2(N-1)(N-2)}{32}. κ N = 144 K β ( N − 1 ) − K β ( N ) − K β ( 1 ) − 32 5 β 2 ( N − 1 ) ( N − 2 ) .
The integrated density of states is obtained by differentiating the imaginary part of the tunnelling contribution:
ν N = − T 1 / 3 π Im ∂ s ( τ N ( 1 ) τ N ( 0 ) ) \nu_N=-\frac{T^{1/3}}{\pi}\operatorname{Im}\partial_s\left(\frac{\tau_N^{(1)}}{\tau_N^{(0)}}\right) ν N = − π T 1/3 Im ∂ s ( τ N ( 0 ) τ N ( 1 ) )
to leading exponential order. Besides differentiating the tunnelling action, we must differentiate its algebraic prefactor s 3 β ( N − 1 ) / 4 s^{3\beta(N-1)/4} s 3 β ( N − 1 ) /4 . This adds − 3 β ( N − 1 ) / 8 -3\beta(N-1)/8 − 3 β ( N − 1 ) /8 to the relative correction coefficient. The resulting tail is Eq. (9) , with
c N ( β ) = K β ( N − 1 ) − K β ( N ) − K β ( 1 ) 144 − 5 β 2 ( N − 1 ) ( N − 2 ) 32 − 3 β ( N − 1 ) 8 = − 21 β 2 N 2 − 60 β 2 N + 54 β N + 39 β 2 − 54 β + 20 96 . (82) \begin{aligned}
c_N^{(\beta)}=\frac{K_\beta(N-1)-K_\beta(N)-K_\beta(1)}{144}-\frac{5\beta^2(N-1)(N-2)}{32}-\frac{3\beta(N-1)}{8} \\
&=-\frac{21\beta^2N^2-60\beta^2N+54\beta N+39\beta^2-54\beta+20}{96}. \tag*{(82)}
\end{aligned} c N ( β ) = 144 K β ( N − 1 ) − K β ( N ) − K β ( 1 ) − 32 5 β 2 ( N − 1 ) ( N − 2 ) − 8 3 β ( N − 1 ) = − 96 21 β 2 N 2 − 60 β 2 N + 54 β N + 39 β 2 − 54 β + 20 . (82)
As an immediate verification, for β = 2 \beta=2 β = 2 the general formula reproduces both the prefactor 2 3 N − 3 / ( N − 1 ) ! 2^{3N-3}/(N-1)! 2 3 N − 3 / ( N − 1 )! and the subleading correction − ( 21 N 2 − 33 N + 17 ) / 24 -(21N^2-33N+17)/24 − ( 21 N 2 − 33 N + 17 ) /24 obtained in [46 ] , while for N = 1 N=1 N = 1 all three symmetry classes reduce to Halperin’s result, including the − 5 / 24 -5/24 − 5/24 correction.
Physical origin of the prefactor. The contour calculation has a complementary interpretation in terms of rare optimal fluctuations of the random potential, which accounts for the leading prefactor without invoking the integrable structure of τ N ( β ) \tau_N^{(\beta)} τ N ( β ) .
For a single channel, the least costly potential supporting a level at E = − k 2 E=-k^2 E = − k 2 is the reflectionless Pöschl–Teller well
w ∗ ( x ) = − 2 k 2 sech 2 ( k x ) . w_*(x)=-2k^2\operatorname{sech}^2(kx). w ∗ ( x ) = − 2 k 2 sech 2 ( k x ) .
Its action is precisely Halperin’s S = 4 3 ( − E ) 3 / 2 / T S=\frac{4}{3}(-E)^{3/2}/T S = 3 4 ( − E ) 3/2 / T in Eq. (9) . The fluctuation determinant around this well reduces to a telescoping product equal to 1 5 \frac{1}{5} 5 1 , and the resulting prefactor reproduces the scalar answer exactly. The scalar calculation therefore fixes the normalization of the optimal-fluctuation method.
For N N N channels, the optimal well remains a single-channel fluctuation, but its orientation in channel space is arbitrary:
W ∗ ( x ) = w ∗ ( x ) u u † . W_*(x)=w_*(x)uu^\dagger. W ∗ ( x ) = w ∗ ( x ) u u † .
Here u ∈ F N u\in\mathbb{F}^N u ∈ F N is a unit vector, and multiplication by a unit of F \mathbb{F} F leaves W ∗ W_* W ∗ unchanged. The family of equivalent wells is therefore R P N − 1 \mathbb{R}P^{N-1} R P N − 1 , C P N − 1 \mathbb{C}P^{N-1} C P N − 1 , or H P N − 1 \mathbb{H}P^{N-1} H P N − 1 for β = 1 , 2 , 4 \beta=1,2,4 β = 1 , 2 , 4 , respectively. Its real dimension is β ( N − 1 ) \beta(N-1) β ( N − 1 ) .
These orientations are zero modes of the instanton. Each contributes a collective-coordinate factor 2 S / π \sqrt{2S/\pi} 2 S / π , together with a factor 3 \sqrt{3} 3 from coupling to the empty channels. The latter comes from level repulsion and a second telescoping fluctuation determinant, equal to 1 3 \frac{1}{3} 3 1 . Integrating over the projective space of orientations then gives exactly the leading prefactor in Eq. (9) .
The leading multichannel tail thus has a simple physical interpretation: Halperin’s one-channel rare well can point in any direction in channel space. The dimension and volume of this orientation space account for the symmetry-class dependence of the leading prefactor, while the tunnelling action remains that of the scalar problem.
Large N N N : planar limit and the edge
The finite-channel solution also determines both the collective spectral edge at N = ∞ N=\infty N = ∞ and its finite-N N N rounding at large N N N . Here we outline the common structure of the three symmetry classes; the symmetry-dependent edge profiles are discussed in the following section.
We take N → ∞ N\to\infty N → ∞ at fixed J J J , with T = J 2 / ( 2 N ) T=J^2/(2N) T = J 2 / ( 2 N ) . At leading order, the saddle-point equation is unchanged by multiplying the logarithm of the weight by a constant. The relation
∣ Δ ∣ β e − ∑ j U E j / T = ( ∣ Δ ∣ 2 e − ∑ j U E j / ( β T / 2 ) ) β / 2 \lvert\Delta\rvert^\beta e^{-\sum_j U_{E_j}/T}=\left(\lvert\Delta\rvert^2e^{-\sum_j U_{E_j}/(\beta T/2)}\right)^{\beta/2} ∣ Δ ∣ β e − ∑ j U E j / T = ( ∣ Δ ∣ 2 e − ∑ j U E j / ( β T /2 ) ) β /2
therefore identifies the planar saddle with that of the unitary problem at temperature β T / 2 \beta T/2 β T /2 . Equivalently, the unitary result applies with J 2 → β J 2 / 2 J^2 \to\beta J^2/2 J 2 → β J 2 /2 . Defining s β = ( β J 2 ) 1 / 3 s_\beta= (\beta J^2)^{1/3} s β = ( β J 2 ) 1/3 we obtain the limiting level density from the curve identified for β = 2 \beta= 2 β = 2 in [46 ] by replacing s J s_J s J with s β s_\beta s β . Its lower edge is consequently situated at the threshold energy E β ∗ = − 3 4 s β 2 E^*_\beta= -\frac{3}{4}s_\beta^2 E β ∗ = − 4 3 s β 2 .
The spectral edge marks the point where two solutions of the planar saddle-point equation merge. At this point the equation has a double root, and the ordinary large-N N N expansion ceases to be uniform. Finite-N N N effects must therefore be retained to describe how the sharp limiting edge is rounded.
In the unitary class, the paper [46 ] resolves this rounding through the recurrence relations of the orthogonal polynomials. The resulting critical window has width of order N − 4 / 5 N^{-4/5} N − 4/5 . The same energy scale governs the other two symmetry classes, as discussed below:
E − E β ∗ = O ( s β 2 N − 4 / 5 ) . E-E^*_\beta=O\left(s_\beta^2N^{-4/5}\right). E − E β ∗ = O ( s β 2 N − 4/5 ) .
The symmetry class thus changes the edge position and the detailed rounding profile, but not the exponent 4 / 5 4/5 4/5 .
At finite N N N , the density remains nonzero below E β ∗ E^*_\beta E β ∗ . Applying the single-eigenvalue activation argument of Ref. [41 ] , as we do below, gives to leading exponential accuracy log ρ N ≃ − A β ( E ) / T \log\rho_N\simeq-A_\beta(E)/T log ρ N ≃ − A β ( E ) / T , where A β A_\beta A β is the barrier in the effective potential created by the metastable planar distribution.
Let ρ p l \rho_{\mathrm{pl}} ρ pl denote the normalized planar density of the Riccati eigenvalues [46 , 41 ] , supported on [ c − d , c + d ] [c-d,c+d] [ c − d , c + d ] with s J s_J s J replaced by s β s_\beta s β , where
E = − c 2 − d 2 2 , c d 2 = β N T = s β 3 2 , c > s β 2 , d > 0. E=-c^2-\frac{d^2}{2},\qquad cd^2=\beta NT=\frac{s_\beta^3}{2},\qquad c>\frac{s_\beta}{2},\qquad d>0. E = − c 2 − 2 d 2 , c d 2 = β N T = 2 s β 3 , c > 2 s β , d > 0.
The effective potential for a single eigenvalue x x x escaping from the metastable planar distribution is, up to an additive constant,
V e f f ( x ) = U E ( x ) − β N T ∫ c − d c + d ρ p l ( λ ) log ∣ x − λ ∣ d λ , V_{\mathrm{eff}}(x)=U_E(x)-\beta NT\int_{c-d}^{c+d}\rho_{\mathrm{pl}}(\lambda)\log\lvert x-\lambda\rvert\,\mathrm{d}\lambda, V eff ( x ) = U E ( x ) − β N T ∫ c − d c + d ρ pl ( λ ) log ∣ x − λ ∣ d λ ,
with U E ( x ) = E x + x 3 / 3 U_E(x)=Ex+x^3/3 U E ( x ) = E x + x 3 /3 defined in Eq. (3) . To the left of the support, the planar saddle gives
V e f f ′ ( x ) = x 2 + E − β N T ∫ c − d c + d ρ p l ( λ ) x − λ d λ = − ( x + c ) ( x − c ) 2 − d 2 , x < c − d , V'_{\mathrm{eff}}(x)=x^2+E-\beta NT\int_{c-d}^{c+d}\frac{\rho_{\mathrm{pl}}(\lambda)}{x-\lambda}\,\mathrm{d}\lambda=-(x+c)\sqrt{(x-c)^2-d^2},\qquad x<c-d, V eff ′ ( x ) = x 2 + E − β N T ∫ c − d c + d x − λ ρ pl ( λ ) d λ = − ( x + c ) ( x − c ) 2 − d 2 , x < c − d ,
where the square root on the right is positive. For x < c − d x<c-d x < c − d , V e f f ′ ( x ) > 0 V'_{\mathrm{eff}}(x)>0 V eff ′ ( x ) > 0 on ( − c , c − d ) (-c,c-d) ( − c , c − d ) and V e f f ′ ( x ) < 0 V'_{\mathrm{eff}}(x)<0 V eff ′ ( x ) < 0 for x < − c x<-c x < − c , so x = − c x=-c x = − c is a local maximum (the escape saddle) separating the metastable support [ c − d , c + d ] [c-d,c+d] [ c − d , c + d ] from the runaway region x → − ∞ x\to-\infty x → − ∞ . An eigenvalue escaping from the lower endpoint c − d c-d c − d must therefore surmount the potential barrier
A β ( E ) = V e f f ( − c ) − V e f f ( c − d ) = ∫ c − d − c V e f f ′ ( x ) d x , A_\beta(E)=V_{\mathrm{eff}}(-c)-V_{\mathrm{eff}}(c-d)=\int_{c-d}^{-c}V'_{\mathrm{eff}}(x)\,\mathrm{d}x, A β ( E ) = V eff ( − c ) − V eff ( c − d ) = ∫ c − d − c V eff ′ ( x ) d x ,
which gives
A β ( E ) = ∫ − c c − d ( x + c ) ( x − c ) 2 − d 2 d x = ∫ d 2 c ( 2 c − z ) z 2 − d 2 d z = 2 c 2 + d 2 3 4 c 2 − d 2 − c d 2 log ( 2 c + 4 c 2 − d 2 d ) . \begin{aligned}
A_\beta(E)&=\int_{-c}^{c-d}(x+c)\sqrt{(x-c)^2-d^2}\,\mathrm{d}x=\int_d^{2c}(2c-z)\sqrt{z^2-d^2}\,\mathrm{d}z\\
&=\frac{2c^2+d^2}{3}\sqrt{4c^2-d^2}-cd^2\log\left(\frac{2c+\sqrt{4c^2-d^2}}{d}\right).
\end{aligned} A β ( E ) = ∫ − c c − d ( x + c ) ( x − c ) 2 − d 2 d x = ∫ d 2 c ( 2 c − z ) z 2 − d 2 d z = 3 2 c 2 + d 2 4 c 2 − d 2 − c d 2 log ( d 2 c + 4 c 2 − d 2 ) .
At the spectral edge, c = s β / 2 c=s_\beta/2 c = s β /2 and d = s β d=s_\beta d = s β : the escape saddle meets the lower endpoint of the support and the barrier vanishes. Writing δ = E β ∗ − E > 0 \delta=E^*_\beta-E>0 δ = E β ∗ − E > 0 , we find
c = s β 2 + δ 3 + O ( δ s β ) , d = s β − δ 3 + O ( δ s β ) , c=\frac{s_\beta}{2}+\sqrt{\frac{\delta}{3}}+O\left(\frac{\delta}{s_\beta}\right),\qquad d=s_\beta-\sqrt{\frac{\delta}{3}}+O\left(\frac{\delta}{s_\beta}\right), c = 2 s β + 3 δ + O ( s β δ ) , d = s β − 3 δ + O ( s β δ ) ,
and hence
A β ( E ) ∼ 4 5 2 3 1 / 4 s β 1 / 2 δ 5 / 4 . \mathcal{A}_{\beta}(E) \sim\frac{4}{5}\sqrt{2}3^{1/4}s_{\beta}^{1/2}\delta^{5/4}. A β ( E ) ∼ 5 4 2 3 1/4 s β 1/2 δ 5/4 .
The near-edge activation law applies when δ ≪ s β 2 \delta\ll s_{\beta}^{2} δ ≪ s β 2 but A β ( E ) / T ≫ 1 \mathcal{A}_{\beta}(E)/T \gg1 A β ( E ) / T ≫ 1 , outside the finite-N N N rounding region.
Deep in the forbidden region, put k = − E k=\sqrt{-E} k = − E . For k ≫ s β k \gg s_{\beta} k ≫ s β , we have c ∼ k c\sim k c ∼ k and d / c → 0 d/c\to0 d / c → 0 . Rescaling z = c t z=ct z = c t gives
A β ( E ) c 3 = ∫ d / c 2 ( 2 − t ) t 2 − ( d / c ) 2 d t ⟶ ∫ 0 2 ( 2 − t ) t d t = 4 3 . \frac{\mathcal{A}_{\beta}(E)}{c^{3}}=\int_{d/c}^{2}(2-t)\sqrt{t^{2}-(d/c)^{2}}\,dt\longrightarrow\int_{0}^{2}(2-t)t\,dt=\frac{4}{3}. c 3 A β ( E ) = ∫ d / c 2 ( 2 − t ) t 2 − ( d / c ) 2 d t ⟶ ∫ 0 2 ( 2 − t ) t d t = 3 4 .
Thus A β ( E ) ∼ 4 3 ( − E ) 3 / 2 \mathcal{A}_{\beta}(E)\sim\frac{4}{3}(-E)^{3/2} A β ( E ) ∼ 3 4 ( − E ) 3/2 , recovering the leading Lifshitz exponent 4 3 ( − E ) 3 / 2 / T \frac{4}{3}(-E)^{3/2}/T 3 4 ( − E ) 3/2 / T of § 9 . The near-edge activation law and the deep Lifshitz tail are therefore two limits of the same barrier; the finite-N N N prefactor is determined separately in § 9 .
Although the planar limit and the critical exponent follow from this common saddle-point picture, the full edge profile depends on the symmetry class. For β = 2 \beta=2 β = 2 , it is governed by the complex tritronquée solution of Painlevé I [46 ] specified in Eq. (106) , rather than the solution that is real on the real axis. For β = 1 , 4 \beta=1,4 β = 1 , 4 , the finite-N N N expansion also contains non-orientable contributions.
The corresponding edge profiles can nevertheless be obtained using the unitary solution. For β = 4 \beta=4 β = 4 , § 11 gives an exact finite-N N N representation through the unitary orthogonal polynomials and an additional three-term recursion, whose edge asymptotics supplies the correction to the density profile. The analogous construction for β = 1 \beta=1 β = 1 and its critical limit are described in Subsecs. 11D and 11E. Those sections distinguish the exact finite-N N N identities from the large-N N N scaling arguments, which are given at the physics level.
The β = 4 \beta=4 β = 4 and β = 1 \beta=1 β = 1 edges: exact reduction to β = 2 \beta=2 β = 2 orthogonal polynomials
For β = 2 \beta=2 β = 2 the large-N N N edge follows from the string equations for the recurrence coefficients of polynomials orthogonal on Γ − \Gamma_{-} Γ − [46 ] . This section shows that for β = 4 \beta=4 β = 4 the partition function is expressed exactly, at every finite N N N , through the recurrence coefficients of the same polynomials, via one additional three-term recursion. It is the starting point for the β = 4 \beta=4 β = 4 analogue of the Painlevé I edge. Note that essentially the same structure is operative behind the Tracy–Widom relation between the β = 4 \beta=4 β = 4 and β = 2 \beta=2 β = 2 soft-edge behaviour [86 , 88 ] . Subsections 11B and 11C give the exact β = 4 \beta=4 β = 4 reduction. Subsection 11D presents the analogous exact reduction for β = 1 \beta=1 β = 1 and even N N N , proved in Appendix § 21 . Subsection 11E derives the double-scaling forms from these exact reductions at the physics level. It also specifies the relevant Painlevé I solution and the selected solution of the auxiliary equation in each class. The selections follow from analyticity in the upper half energy plane and the β \beta β -dependence of the barrier; for β = 1 \beta=1 β = 1 the finite-N N N counterpart, the minimal-solution condition for X 0 X_0 X 0 , is left as a conjecture.
Orthogonal polynomials and basic identities
Let π n ( λ ) \pi_n(\lambda) π n ( λ ) be the monic polynomials orthogonal on Γ − \Gamma_{-} Γ − with the weight w ( λ ) = e − U E ( λ ) / T w(\lambda)=e^{-U_E(\lambda)/T} w ( λ ) = e − U E ( λ ) / T ,
∫ Γ − π m π n w d λ = h n δ m n , λ π n = π n + 1 + d n π n + r n π n − 1 , r n = h n h n − 1 . \int_{\Gamma_{-}}\pi_m\pi_n w\,d\lambda=h_n\delta_{mn},\qquad
\lambda\pi_n=\pi_{n+1}+d_n\pi_n+r_n\pi_{n-1},\qquad
r_n=\frac{h_n}{h_{n-1}}. ∫ Γ − π m π n w d λ = h n δ mn , λ π n = π n + 1 + d n π n + r n π n − 1 , r n = h n − 1 h n .
These are the same polynomials as used in [46 ] and we briefly outline the procedure developed there, omitting the detail. Integration by parts along Γ − \Gamma_{-} Γ − gives the “string equations”
d n 2 + r n + r n + 1 + E = 0 , r n ( d n + d n − 1 ) = n T , (83) d_n^2+r_n+r_{n+1}+E=0,\qquad r_n(d_n+d_{n-1})=nT, \tag*{(83)} d n 2 + r n + r n + 1 + E = 0 , r n ( d n + d n − 1 ) = n T , (83)
and the Toda relations ∂ θ log h n = d n \partial_{\theta}\log h_n=d_n ∂ θ log h n = d n , ∂ θ d n = r n + 1 − r n \partial_{\theta}d_n=r_{n+1}-r_n ∂ θ d n = r n + 1 − r n , where θ = − E / T \theta=-E/T θ = − E / T is the coefficient of λ \lambda λ in − U E / T -U_E/T − U E / T . We write Z Z Z for holomorphic partition functions on Γ − \Gamma_{-} Γ − , to distinguish them from the positive normalization Z N ( E , η ) \mathcal{Z}_N(E,\eta) Z N ( E , η ) . For the unitary β = 2 \beta=2 β = 2 case Z N ( 2 ) ∝ ∏ n < M h n Z_N^{(2)}\propto\prod_{n<M}h_n Z N ( 2 ) ∝ ∏ n < M h n and hence Ξ M ( 2 ) = ∑ n < M d n \Xi_M^{(2)}=\sum_{n<M}d_n Ξ M ( 2 ) = ∑ n < M d n .
Exact identity for Z N ( 4 ) Z_N^{(4)} Z N ( 4 )
Exact reduction. Up to an E E E -independent constant Z N ( 4 ) ∝ ∏ j = 0 N − 1 h 2 j + 1 ⋅ Q N Z_N^{(4)}\propto\prod_{j=0}^{N-1}h_{2j+1}\cdot Q_N Z N ( 4 ) ∝ ∏ j = 0 N − 1 h 2 j + 1 ⋅ Q N , with Q m Q_m Q m satisfying the recursion
Q m + 1 = ( d 2 m + d 2 m + 1 ) Q m − r 2 m Q m − 1 , Q 0 = 1 , Q 1 = d 0 + d 1 . (84) Q_{m+1}=(d_{2m}+d_{2m+1})Q_m-r_{2m}Q_{m-1},\qquad Q_0=1,\quad Q_1=d_0+d_1. \tag*{(84)} Q m + 1 = ( d 2 m + d 2 m + 1 ) Q m − r 2 m Q m − 1 , Q 0 = 1 , Q 1 = d 0 + d 1 . (84)
Equivalently, Z N ( 4 ) / Z 2 N ( 2 ) ∝ ∏ j < N r 2 j + 1 1 / 2 Q N Z_N^{(4)}/\sqrt{Z_{2N}^{(2)}}\propto\prod_{j<N}r_{2j+1}^{1/2}Q_N Z N ( 4 ) / Z 2 N ( 2 ) ∝ ∏ j < N r 2 j + 1 1/2 Q N .
Proof. By de Bruijn’s formula [19 ] , Z N ( 4 ) ∝ Pf [ ⟨ λ j , λ k ⟩ 4 ] 0 ≤ j , k < 2 N Z_N^{(4)}\propto\operatorname{Pf}\left[\langle\lambda^j,\lambda^k\rangle_4\right]_{0\leq j,k<2N} Z N ( 4 ) ∝ Pf [ ⟨ λ j , λ k ⟩ 4 ] 0 ≤ j , k < 2 N , where ⟨ f , g ⟩ 4 = ∫ Γ − ( f g ′ − f ′ g ) w d λ \langle f,g\rangle_4=\int_{\Gamma_-}(fg'-f'g)w\,\mathrm{d}\lambda ⟨ f , g ⟩ 4 = ∫ Γ − ( f g ′ − f ′ g ) w d λ . Since the orthogonal polynomials π j \pi_j π j are monic, replacing λ j \lambda^j λ j by π j \pi_j π j is a unitriangular change of basis. Its determinant is one, so the Pfaffian is unchanged. We may therefore work with the skew matrix a j k : = ⟨ π j , π k ⟩ 4 a_{jk}:=\langle\pi_j,\pi_k\rangle_4 a j k := ⟨ π j , π k ⟩ 4 .
Using w ′ / w = − ( E + λ 2 ) / T w'/w=-(E+\lambda^2)/T w ′ / w = − ( E + λ 2 ) / T and integrating by parts along Γ − \Gamma_- Γ − , we obtain
a j k = − 2 ∫ Γ − π j ′ π k w d λ + 1 T Φ j k , Φ j k = ∫ Γ − π j π k ( E + λ 2 ) w d λ . a_{jk}=-2\int_{\Gamma_-}\pi'_j\pi_k\,w\,\mathrm{d}\lambda+\frac{1}{T}\Phi_{jk},\qquad\Phi_{jk}=\int_{\Gamma_-}\pi_j\pi_k(E+\lambda^2)w\,\mathrm{d}\lambda. a j k = − 2 ∫ Γ − π j ′ π k w d λ + T 1 Φ j k , Φ j k = ∫ Γ − π j π k ( E + λ 2 ) w d λ .
The boundary terms vanish because the weight decays at both ends of the contour. For k > j k>j k > j , the first integral vanishes by orthogonality, since deg π j ′ < j < k \deg\pi'_j<j<k deg π j ′ < j < k . Thus a j k = Φ j k / T a_{jk}=\Phi_{jk}/T a j k = Φ j k / T for k > j k>j k > j . Since Φ j k = Φ k j \Phi_{jk}=\Phi_{kj} Φ j k = Φ k j and a j k = − a k j a_{jk}=-a_{kj} a j k = − a k j , this gives
a j k = sgn ( k − j ) T Φ j k , a j j = 0. a_{jk}=\frac{\operatorname{sgn}(k-j)}{T}\Phi_{jk},\qquad a_{jj}=0. a j k = T sgn ( k − j ) Φ j k , a j j = 0.
The three-term recurrence λ π j = π j + 1 + d j π j + r j π j − 1 \lambda\pi_j=\pi_{j+1}+d_j\pi_j+r_j\pi_{j-1} λ π j = π j + 1 + d j π j + r j π j − 1 shows that
λ 2 π j = π j + 2 + ( d j + d j + 1 ) π j + 1 + terms of degree at most j . \lambda^2\pi_j=\pi_{j+2}+(d_j+d_{j+1})\pi_{j+1}+\text{terms of degree at most }j. λ 2 π j = π j + 2 + ( d j + d j + 1 ) π j + 1 + terms of degree at most j .
Orthogonality consequently gives
Φ j , j + 1 = h j + 1 ( d j + d j + 1 ) , Φ j , j + 2 = h j + 2 , Φ j k = 0 for ∣ j − k ∣ ≥ 3. \Phi_{j,j+1}=h_{j+1}(d_j+d_{j+1}),\qquad\Phi_{j,j+2}=h_{j+2},\qquad\Phi_{jk}=0\quad\text{for }|j-k|\geq3. Φ j , j + 1 = h j + 1 ( d j + d j + 1 ) , Φ j , j + 2 = h j + 2 , Φ j k = 0 for ∣ j − k ∣ ≥ 3.
Hence the skew matrix a j k a_{jk} a j k is five-diagonal. The finite bandwidth is a direct consequence of the quadratic logarithmic derivative w ′ / w w'/w w ′ / w . Indeed, for k > j k>j k > j , integration by parts reduces a j k a_{jk} a j k to the overlap of π k \pi_k π k with ( E + λ 2 ) π j (E+\lambda^2)\pi_j ( E + λ 2 ) π j . This is a polynomial of degree j + 2 j+2 j + 2 , so orthogonality gives a j k = 0 a_{jk}=0 a j k = 0 whenever k > j + 2 k>j+2 k > j + 2 . Skew-symmetry gives the corresponding restriction below the diagonal, leaving only the first two off-diagonals on either side. More generally, a polynomial logarithmic derivative of degree d d d restricts the skew matrix to ∣ j − k ∣ ≤ d |j-k|\leq d ∣ j − k ∣ ≤ d . For a non-polynomial logarithmic derivative, this degree argument no longer applies, and the skew matrix need not have finite bandwidth.
Let now consider P m = Pf [ a j k ] 0 ≤ j , k < 2 m P_m=\operatorname{Pf}[a_{jk}]_{0\leq j,k<2m} P m = Pf [ a j k ] 0 ≤ j , k < 2 m , P 0 = 1 P_0=1 P 0 = 1 . In the expansion of P m + 1 P_{m+1} P m + 1 along its last row, the index 2 m + 1 2m+1 2 m + 1 can be paired only with 2 m 2m 2 m or 2 m − 1 2m-1 2 m − 1 . Pairing it with 2 m 2m 2 m leaves P m P_m P m . Pairing it with 2 m − 1 2m-1 2 m − 1 leaves the index 2 m 2m 2 m , which must then be paired with 2 m − 2 2m-2 2 m − 2 , leaving P m − 1 P_{m-1} P m − 1 . The Pfaffian signs therefore give
P m + 1 = a 2 m , 2 m + 1 P m − a 2 m − 1 , 2 m + 1 a 2 m − 2 , 2 m P m − 1 , m ≥ 1. P_{m+1}=a_{2m,2m+1}P_m-a_{2m-1,2m+1}a_{2m-2,2m}P_{m-1},\qquad m\geq1. P m + 1 = a 2 m , 2 m + 1 P m − a 2 m − 1 , 2 m + 1 a 2 m − 2 , 2 m P m − 1 , m ≥ 1.
Substituting the matrix entries yields
P m + 1 = h 2 m + 1 T ( d 2 m + d 2 m + 1 ) P m − h 2 m + 1 h 2 m T 2 P m − 1 . (85) P_{m+1}=\frac{h_{2m+1}}{T}(d_{2m}+d_{2m+1})P_m-\frac{h_{2m+1}h_{2m}}{T^2}P_{m-1}. \tag*{(85)} P m + 1 = T h 2 m + 1 ( d 2 m + d 2 m + 1 ) P m − T 2 h 2 m + 1 h 2 m P m − 1 . (85)
We now factor out the odd-indexed norms by writing
P m = [ ∏ j = 0 m − 1 h 2 j + 1 T ] Q m . P_m=\left[\prod_{j=0}^{m-1}\frac{h_{2j+1}}{T}\right]Q_m. P m = [ j = 0 ∏ m − 1 T h 2 j + 1 ] Q m .
Dividing the preceding relation by ∏ j = 0 m ( h 2 j + 1 / T ) \prod_{j=0}^{m}(h_{2j+1}/T) ∏ j = 0 m ( h 2 j + 1 / T ) gives
Q m + 1 = ( d 2 m + d 2 m + 1 ) Q m − h 2 m h 2 m − 1 Q m − 1 . Q_{m+1}=(d_{2m}+d_{2m+1})Q_m-\frac{h_{2m}}{h_{2m-1}}Q_{m-1}. Q m + 1 = ( d 2 m + d 2 m + 1 ) Q m − h 2 m − 1 h 2 m Q m − 1 .
Since h 2 m / h 2 m − 1 = r 2 m h_{2m}/h_{2m-1}=r_{2m} h 2 m / h 2 m − 1 = r 2 m , we obtain
Q m + 1 = ( d 2 m + d 2 m + 1 ) Q m − r 2 m Q m − 1 , Q 0 = 1 , Q 1 = d 0 + d 1 . Q_{m+1}=(d_{2m}+d_{2m+1})Q_m-r_{2m}Q_{m-1},\qquad Q_0=1,\qquad Q_1=d_0+d_1. Q m + 1 = ( d 2 m + d 2 m + 1 ) Q m − r 2 m Q m − 1 , Q 0 = 1 , Q 1 = d 0 + d 1 .
This is precisely the recursion in (84) . Finally,
Z N ( 4 ) ∝ P N = T − N [ ∏ j = 0 N − 1 h 2 j + 1 ] Q N . Z_N^{(4)}\propto P_N=T^{-N}\left[\prod_{j=0}^{N-1}h_{2j+1}\right]Q_N. Z N ( 4 ) ∝ P N = T − N [ j = 0 ∏ N − 1 h 2 j + 1 ] Q N .
At fixed T T T , the factor T − N T^{-N} T − N is independent of E E E and is absorbed into the proportionality constant, giving the stated partition-function identity.
With Ξ = − T ∂ E log Z \Xi=-T\partial_E\log Z Ξ = − T ∂ E log Z and the Toda relation − T ∂ E log h n = d n -T\partial_E\log h_n=d_n − T ∂ E log h n = d n , Eq. (84) gives
Ξ N ( 4 ) = ∑ j = 0 N − 1 d 2 j + 1 − T ∂ E log Q N . (86) \Xi_N^{(4)}=\sum_{j=0}^{N-1}d_{2j+1}-T\partial_E\log Q_N. \tag*{(86)} Ξ N ( 4 ) = j = 0 ∑ N − 1 d 2 j + 1 − T ∂ E log Q N . (86)
Here − T ∂ E = ∂ θ -T\partial_E=\partial_\theta − T ∂ E = ∂ θ acts on Q N Q_N Q N through ∂ θ d n = r n + 1 − r n \partial_\theta d_n=r_{n+1}-r_n ∂ θ d n = r n + 1 − r n and ∂ θ r n = r n ( d n − d n − 1 ) \partial_\theta r_n=r_n(d_n-d_{n-1}) ∂ θ r n = r n ( d n − d n − 1 ) . It can therefore be evaluated recursively alongside Q m Q_m Q m . It is instructive to check N = 1 N=1 N = 1 . Eq. (85) gives d 1 + r 2 / ( d 0 + d 1 ) d_1+r_2/(d_0+d_1) d 1 + r 2 / ( d 0 + d 1 ) , while the first two string equations give r 2 = d 0 2 − d 1 2 r_2=d_0^2-d_1^2 r 2 = d 0 2 − d 1 2 . Thus the expression reduces to d 0 = Ξ 1 d_0=\Xi_1 d 0 = Ξ 1 , as it must: for one particle, the partition function is independent of the Dyson index.
Comparing now with the β = 2 \beta=2 β = 2 relation Ξ 2 N ( 2 ) = ∑ n < 2 N d n \Xi_{2N}^{(2)}=\sum_{n<2N}d_n Ξ 2 N ( 2 ) = ∑ n < 2 N d n we get
Ξ N ( 4 ) = 1 2 Ξ 2 N ( 2 ) + 1 2 ∑ j < N ( d 2 j + 1 − d 2 j ) − T ∂ E log Q N . (87) \Xi_N^{(4)}=\frac{1}{2}\Xi_{2N}^{(2)}+\frac{1}{2}\sum_{j<N}(d_{2j+1}-d_{2j})-T\partial_E\log Q_N. \tag*{(87)} Ξ N ( 4 ) = 2 1 Ξ 2 N ( 2 ) + 2 1 j < N ∑ ( d 2 j + 1 − d 2 j ) − T ∂ E log Q N . (87)
The β = 4 \beta=4 β = 4 exponent is therefore half the β = 2 \beta=2 β = 2 exponent of the ensemble with twice as many particles, with the same weight and the same T T T , plus corrections of two types: (i) an even–odd difference, which reduces at leading order to a boundary term when the d n d_n d n admit a slowly varying smooth profile and (ii) the logarithmic derivative of the solution of a three-term recursion driven by the β = 2 \beta=2 β = 2 data. The first term dominates in the large-N N N limit and determines the planar values s 4 = ( 4 J ) 1 / 3 s_4=(4J)^{1/3} s 4 = ( 4 J ) 1/3 and E 4 ∗ = − 3 4 s 4 2 E_4^*=-\frac{3}{4}s_4^2 E 4 ∗ = − 4 3 s 4 2 , in agreement with § 10 .
The recursion captures the part specific to the Dyson index β = 4 \beta=4 β = 4 and reveals the mechanism by which the Dyson-index-dependent correction can survive in the double-scaling edge limit. In the planar regime, d n → d d_n\to d d n → d and r n → r r_n\to r r n → r , and the recursion has characteristic roots μ ± = d ± d 2 − r \mu_\pm=d\pm\sqrt{d^2-r} μ ± = d ± d 2 − r . At the critical point ( d ∗ , r ∗ ) = ( s / 2 , s 2 / 4 ) (d_*,r_*)=(s/2,s^2/4) ( d ∗ , r ∗ ) = ( s /2 , s 2 /4 ) , with s = s 4 s=s_4 s = s 4 , the two characteristic roots coalesce. After factoring out the common exponential growth, the leading recurrence becomes a second difference. In the double-scaling window, this second difference balances the departures of the recurrence coefficients from their critical values, so the recursion retains a nontrivial continuum limit. Thus the separated-root planar approximation breaks down precisely at criticality, and the recursion must be retained in the edge-scaling analysis. This continuum limit is derived in § 11 E.
β = 1 \beta= 1 β = 1 , even N N N : an exact reduction of the same type
For β = 1 \beta= 1 β = 1 the role of the unitary orthogonal polynomials is played by those of the squared weight e − 2 U E ( λ ) / T e^{-2U_E(\lambda)/T} e − 2 U E ( λ ) / T , i.e. of the unitary ensemble at temperature T / 2 T/2 T /2 . To distinguish the integration variable from the weight, set
w = − λ / T 1 / 3 , s = − E / T 2 / 3 , ρ s ( w ) = e w 3 / 3 − s w . w=-\lambda/T^{1/3},\qquad s=-E/T^{2/3},\qquad\rho_s(w)=e^{w^3/3-sw}. w = − λ / T 1/3 , s = − E / T 2/3 , ρ s ( w ) = e w 3 /3 − s w .
Let p n ( w ) p_n(w) p n ( w ) be the monic polynomials orthogonal on the outgoing contour C − C_- C − with respect to ρ s 2 \rho_s^2 ρ s 2 :
∫ C − p j ( w ) p k ( w ) ρ s ( w ) 2 d w = h ~ j δ j k , w p n = p n + 1 + d ~ n p n + r ~ n p n − 1 , r ~ n = h ~ n h ~ n − 1 . \int_{C_-}p_j(w)p_k(w)\rho_s(w)^2\,dw=\widetilde{h}_j\delta_{jk},\qquad wp_n=p_{n+1}+\widetilde{d}_np_n+\widetilde{r}_np_{n-1},\qquad\widetilde{r}_n=\frac{\widetilde{h}_n}{\widetilde{h}_{n-1}}. ∫ C − p j ( w ) p k ( w ) ρ s ( w ) 2 d w = h j δ j k , w p n = p n + 1 + d n p n + r n p n − 1 , r n = h n − 1 h n .
All tilded quantities below refer to this dimensionless variable. The planar limit of this unitary ensemble with N N N particles coincides with that of the β = 1 \beta=1 β = 1 ensemble with N N N particles, reproducing s 1 = ( J 2 ) 1 / 3 s_1=(J^2)^{1/3} s 1 = ( J 2 ) 1/3 and E 1 ∗ = − 3 4 s 1 2 E^*_1=-\frac{3}{4}s_1^2 E 1 ∗ = − 4 3 s 1 2 of § 10 .
For N = 2 m ≥ 2 N=2m\ge2 N = 2 m ≥ 2 , the exact reduction, proved in Appendix § 21 , is
τ 2 m ( 1 ) ( s ) = ( − 2 ) m [ ∏ i = 0 m − 2 h ~ 2 i + 1 ( s ) ] X m − 1 ( s ) , (88) \tau^{(1)}_{2m}(s)=(-2)^m\left[\prod_{i=0}^{m-2}\widetilde{h}_{2i+1}(s)\right]X_{m-1}(s), \tag*{(88)} τ 2 m ( 1 ) ( s ) = ( − 2 ) m [ i = 0 ∏ m − 2 h 2 i + 1 ( s ) ] X m − 1 ( s ) , (88)
X k + 1 = ( d ~ 2 k + d ~ 2 k + 1 ) X k − r ~ 2 k X k − 1 , k ≥ 1. X_{k+1}=(\widetilde{d}_{2k}+\widetilde{d}_{2k+1})X_k-\widetilde{r}_{2k}X_{k-1},\qquad k\ge1. X k + 1 = ( d 2 k + d 2 k + 1 ) X k − r 2 k X k − 1 , k ≥ 1.
Here the product is empty for m = 1 m=1 m = 1 and is then set equal to unity. The recursion starts from
X 0 ( s ) = i π 3 + π 2 ∫ 0 2 2 / 3 s [ Bi ( x ) + i Ai ( x ) ] d x , X 1 = ( d ~ 0 + d ~ 1 ) X 0 − h ~ 0 . X_0(s)=\frac{i\pi}{3}+\frac{\pi}{2}\int_0^{2^{2/3}s}[\operatorname{Bi}(x)+i\operatorname{Ai}(x)]\,dx,\qquad X_1=(\widetilde{d}_0+\widetilde{d}_1)X_0-\widetilde{h}_0. X 0 ( s ) = 3 iπ + 2 π ∫ 0 2 2/3 s [ Bi ( x ) + i Ai ( x )] d x , X 1 = ( d 0 + d 1 ) X 0 − h 0 .
We use τ 2 m ( 1 ) \tau^{(1)}_{2m} τ 2 m ( 1 ) to denote explicitly the holomorphic ordered integral of Eq. (4) . The factor ( − 2 ) m (-2)^m ( − 2 ) m is fixed by that contour and normalization convention and drops out of logarithmic derivatives. Since ∂ s log h ~ n = − 2 d ~ n \partial_s\log\widetilde{h}_n=-2\widetilde{d}_n ∂ s log h n = − 2 d n , the characteristic exponent is
Ξ 2 m ( 1 ) / T 1 / 3 = − 2 ∑ i = 0 m − 2 d ~ 2 i + 1 + ∂ s log X m − 1 . (89) \Xi^{(1)}_{2m}/T^{1/3}=-2\sum_{i=0}^{m-2}\widetilde{d}_{2i+1}+\partial_s\log X_{m-1}. \tag*{(89)} Ξ 2 m ( 1 ) / T 1/3 = − 2 i = 0 ∑ m − 2 d 2 i + 1 + ∂ s log X m − 1 . (89)
The derivative can be evaluated recursively using the Toda flows of the recurrence coefficients of the squared weight,
∂ s d ~ n = − 2 ( r ~ n + 1 − r ~ n ) , ∂ s r ~ n = − 2 r ~ n ( d ~ n − d ~ n − 1 ) . \partial_s\widetilde{d}_n=-2(\widetilde{r}_{n+1}-\widetilde{r}_n),\qquad\partial_s\widetilde{r}_n=-2\widetilde{r}_n(\widetilde{d}_n-\widetilde{d}_{n-1}). ∂ s d n = − 2 ( r n + 1 − r n ) , ∂ s r n = − 2 r n ( d n − d n − 1 ) .
The sequence { X k } k ≥ 0 \{X_k\}_{k\ge0} { X k } k ≥ 0 obeys the same homogeneous three-term recursion as { Q k } k ≥ 0 \{Q_k\}_{k\ge0} { Q k } k ≥ 0 in Eq. (84) , now with coefficients built from the recurrence coefficients of the squared weight, but with different initial values. In the present contour normalization,
h ~ 0 ( s ) = − π 2 1 / 3 [ Bi ( 2 2 / 3 s ) + i Ai ( 2 2 / 3 s ) ] , ∂ s X 0 = − h ~ 0 . \widetilde{h}_0(s)=-\frac{\pi}{2^{1/3}}[\operatorname{Bi}(2^{2/3}s)+i\operatorname{Ai}(2^{2/3}s)],\qquad\partial_sX_0=-\widetilde{h}_0. h 0 ( s ) = − 2 1/3 π [ Bi ( 2 2/3 s ) + i Ai ( 2 2/3 s )] , ∂ s X 0 = − h 0 .
For real s s s , the initial function can equivalently be written as
X 0 ( s ) = − ∫ − ∞ s h ~ 0 ( t ) d t , X 0 ( s ) ⟶ 0 as s → − ∞ , X_0(s)=-\int_{-\infty}^{s}\widetilde{h}_0(t)\,dt,\qquad X_0(s)\longrightarrow0\quad\text{as }s\to-\infty, X 0 ( s ) = − ∫ − ∞ s h 0 ( t ) d t , X 0 ( s ) ⟶ 0 as s → − ∞ ,
where the integral is understood in the oscillatory improper sense. The proof has a direct finite-dimensional form. For polynomials f , g f,g f , g , introduce the ordered-contour skew product
[ f , g ] 1 : = ∫ x ≺ y [ f ( x ) g ( y ) − g ( x ) f ( y ) ] ρ s ( x ) ρ s ( y ) d x d y . [f,g]_1:=\int_{x\prec y}[f(x)g(y)-g(x)f(y)]\rho_s(x)\rho_s(y)\,dx\,dy. [ f , g ] 1 := ∫ x ≺ y [ f ( x ) g ( y ) − g ( x ) f ( y )] ρ s ( x ) ρ s ( y ) d x d y .
where x ≺ y x \prec y x ≺ y means that x x x precedes y y y along the oriented contour C − C_- C − . Thus [ f , g ] 1 = − [ g , f ] 1 [f,g]_1=-[g,f]_1 [ f , g ] 1 = − [ g , f ] 1 ; the brackets denote a bilinear form, not a commutator. Define also the weighted derivative
D f = f ′ + ( w 2 − s ) f = ρ s − 1 ( f ρ s ) ′ . \mathcal{D}f=f'+(w^2-s)f=\rho_s^{-1}(f\rho_s)'. D f = f ′ + ( w 2 − s ) f = ρ s − 1 ( f ρ s ) ′ .
Integration by parts gives
[ D f , g ] 1 = 2 ∫ C − f ( w ) g ( w ) ρ s ( w ) 2 d w . [\mathcal{D}f,g]_1=2\int_{C_-}f(w)g(w)\rho_s(w)^2\,\mathrm{d}w. [ D f , g ] 1 = 2 ∫ C − f ( w ) g ( w ) ρ s ( w ) 2 d w .
with vanishing endpoint terms because the weight decays at both ends of the contour. Replacing the monic polynomial basis by
p 0 , p 1 , D p 0 , D p 1 , … p_0,\quad p_1,\quad\mathcal{D}p_0,\quad\mathcal{D}p_1,\quad\ldots p 0 , p 1 , D p 0 , D p 1 , …
is a unitriangular transformation and therefore preserves the Pfaffian. In this basis the skew matrix is five-diagonal: multiplication by w 2 − s w^2-s w 2 − s raises the polynomial degree by at most two, and orthogonality eliminates all more distant entries. Expansion of its Pfaffian gives Eq. (88) ; the first two rows supply the exceptional initial term − h ~ 0 -\tilde{h}_0 − h ~ 0 in X 1 X_1 X 1 . Finally, direct evaluation of the two-particle ordered integral at s = 0 s=0 s = 0 fixes X 0 ( 0 ) = i π / 3 X_0(0)=i\pi/3 X 0 ( 0 ) = iπ /3 . Appendix § 21 gives the matrix entries, the Pfaffian signs and the contour calculation explicitly.
Double scaling for β = 4 \beta=4 β = 4 and β = 1 \beta=1 β = 1 : derivation and solution selection
We now derive the double-scaling limits of the β = 4 \beta=4 β = 4 and β = 1 \beta=1 β = 1 edges from the exact finite-N N N reductions (84) and (88) . The derivation is at the physics level: we take continuum limits of the string equations and of the three-term recursions, keep the leading order in N − 1 / 5 N^{-1/5} N − 1/5 , and fix the relevant solutions by matching. The steps are as follows.
A common form. Both reductions lead to the same kind of object: a unitary ensemble of M M M particles on Γ − \Gamma_- Γ − , in the variable λ \lambda λ , at an effective temperature T e f f T_{\rm eff} T eff , described by its recurrence coefficients ( d n , r n ) (d_n,r_n) ( d n , r n ) . For β = 4 \beta=4 β = 4 the weight is e − U E ( λ ) / T e^{-U_E(\lambda)/T} e − U E ( λ ) / T itself, so T e f f = T T_{\rm eff}=T T eff = T , the number of particles is doubled, M = 2 N M=2N M = 2 N , and the recurrence coefficients are those of § 11 B. For β = 1 \beta=1 β = 1 and even N N N the weight is squared, e − 2 U E ( λ ) / T e^{-2U_E(\lambda)/T} e − 2 U E ( λ ) / T , which halves the temperature, T e f f = T / 2 T_{\rm eff}=T/2 T eff = T /2 , while the number of particles stays M = N M=N M = N . Its recurrence coefficients were computed in § 11 D in the dimensionless variable w w w ; since λ = − T 1 / 3 w \lambda=-T^{1/3}w λ = − T 1/3 w , they translate into d n = − T 1 / 3 d ~ n d_n=-T^{1/3}\tilde{d}_n d n = − T 1/3 d ~ n and r n = T 2 / 3 r ~ n r_n=T^{2/3}\tilde{r}_n r n = T 2/3 r ~ n in the variable λ \lambda λ .
In the rest of this subsection ( d n , r n ) (d_n,r_n) ( d n , r n ) denote whichever of the two sets is relevant. In both cases the string equations combined with the Toda flow take the form
d n 2 + r n + r n + 1 + E = 0 , r n ( d n + d n − 1 ) = n T e f f , ∂ E d n = − r n + 1 − r n T e f f . (90) d_n^2+r_n+r_{n+1}+E=0,\qquad r_n(d_n+d_{n-1})=nT_{\rm eff},\qquad\partial_E d_n=-\frac{r_{n+1}-r_n}{T_{\rm eff}}. \tag*{(90)} d n 2 + r n + r n + 1 + E = 0 , r n ( d n + d n − 1 ) = n T eff , ∂ E d n = − T eff r n + 1 − r n . (90)
The rescaling X k p h = ( − T 1 / 3 ) k X k X_k^{\mathrm{ph}}=(-T^{1/3})^kX_k X k ph = ( − T 1/3 ) k X k brings the β = 1 \beta=1 β = 1 recursion (88) to the same form as the β = 4 \beta=4 β = 4 one:
R m + 1 = ( d 2 m + d 2 m + 1 ) R m − r 2 m R m − 1 , R = Q ( β = 4 ) , R = X p h ( β = 1 ) . (91) R_{m+1}=(d_{2m}+d_{2m+1})R_m-r_{2m}R_{m-1},\qquad R=Q\;(\beta=4),\qquad R=X^{\mathrm{ph}}\;(\beta=1). \tag*{(91)} R m + 1 = ( d 2 m + d 2 m + 1 ) R m − r 2 m R m − 1 , R = Q ( β = 4 ) , R = X ph ( β = 1 ) . (91)
The characteristic exponents (85) and (89) read, in physical units,
Ξ N ( 4 ) = ∑ j = 0 N − 1 d 2 j + 1 − T ∂ E log Q N , Ξ 2 m ( 1 ) = 2 ∑ i = 0 m − 2 d 2 i + 1 − T ∂ E log X m − 1 p h . (92) \Xi_N^{(4)}=\sum_{j=0}^{N-1}d_{2j+1}-T\partial_E\log Q_N,\qquad\Xi_{2m}^{(1)}=2\sum_{i=0}^{m-2}d_{2i+1}-T\partial_E\log X_{m-1}^{\mathrm{ph}}. \tag*{(92)} Ξ N ( 4 ) = j = 0 ∑ N − 1 d 2 j + 1 − T ∂ E log Q N , Ξ 2 m ( 1 ) = 2 i = 0 ∑ m − 2 d 2 i + 1 − T ∂ E log X m − 1 ph . (92)
Note that the last terms carry the physical temperature T T T , not T e f f T_{\mathrm{eff}} T eff . With ρ = − π − 1 Im ∂ E Ξ \rho=-\pi^{-1}\operatorname{Im}\partial_E\Xi ρ = − π − 1 Im ∂ E Ξ and the last relation in (90) , the densities of states are, exactly at finite N N N ,
ρ N ( 4 ) = 1 π T Im ∑ j = 0 N − 1 ( r 2 j + 2 − r 2 j + 1 ) + T π Im ∂ E 2 log Q N , \rho_N^{(4)}=\frac{1}{\pi T}\operatorname{Im}\sum_{j=0}^{N-1}(r_{2j+2}-r_{2j+1})+\frac{T}{\pi}\operatorname{Im}\partial_E^2\log Q_N, ρ N ( 4 ) = π T 1 Im j = 0 ∑ N − 1 ( r 2 j + 2 − r 2 j + 1 ) + π T Im ∂ E 2 log Q N ,
ρ 2 m ( 1 ) = 4 π T Im ∑ i = 0 m − 2 ( r 2 i + 2 − r 2 i + 1 ) + T π Im ∂ E 2 log X m − 1 p h . (93) \rho_{2m}^{(1)}=\frac{4}{\pi T}\operatorname{Im}\sum_{i=0}^{m-2}(r_{2i+2}-r_{2i+1})+\frac{T}{\pi}\operatorname{Im}\partial_E^2\log X_{m-1}^{\mathrm{ph}}. \tag*{(93)} ρ 2 m ( 1 ) = π T 4 Im i = 0 ∑ m − 2 ( r 2 i + 2 − r 2 i + 1 ) + π T Im ∂ E 2 log X m − 1 ph . (93)
For the unitary ensemble itself, Ξ M ( 2 ) = ∑ n < M d n \Xi_M^{(2)}=\sum_{n<M}d_n Ξ M ( 2 ) = ∑ n < M d n and r 0 = 0 r_0=0 r 0 = 0 give the exact identity
ρ M ( 2 ) ∣ T e f f = Im r M π T e f f . (94) \left.\rho_M^{(2)}\right|_{T_{\mathrm{eff}}}=\frac{\operatorname{Im}r_M}{\pi T_{\mathrm{eff}}}. \tag*{(94)} ρ M ( 2 ) T eff = π T eff Im r M . (94)
The critical point
In the planar regime (M → ∞ M\to\infty M → ∞ , i.e. N → ∞ N\to\infty N → ∞ ), the recurrence coefficients vary slowly near n = M n=M n = M , d n ≃ d d_n\simeq d d n ≃ d and r n ≃ r r_n\simeq r r n ≃ r , and the string equations at n = M n=M n = M give the single cubic equation d 3 + E d + M T e f f = 0 d^3+Ed+MT_{\mathrm{eff}}=0 d 3 + E d + M T eff = 0 , or equivalently E = − d 2 − M T e f f / d E=-d^2-MT_{\mathrm{eff}}/d E = − d 2 − M T eff / d . Two positive real branches of this cubic merge at the maximum of − d 2 − M T e f f / d -d^2-MT_{\mathrm{eff}}/d − d 2 − M T eff / d for d > 0 d>0 d > 0 :
d ∗ = s 2 , r ∗ = d ∗ 2 = s 2 4 , E ∗ = − 3 d ∗ 2 = − 3 4 s 2 , s 3 = 4 M T e f f . (95) d_*=\frac{s}{2},\qquad r_*=d_*^2=\frac{s^2}{4},\qquad E^*=-3d_*^2=-\frac{3}{4}s^2,\qquad s^3=4MT_{\mathrm{eff}}. \tag*{(95)} d ∗ = 2 s , r ∗ = d ∗ 2 = 4 s 2 , E ∗ = − 3 d ∗ 2 = − 4 3 s 2 , s 3 = 4 M T eff . (95)
This reproduces s β = ( β J 2 ) 1 / 3 s_\beta=(\beta J^2)^{1/3} s β = ( β J 2 ) 1/3 of Sec. 10: for β = 2 \beta=2 β = 2 we have 4 N T = 2 J 2 4NT=2J^2 4 N T = 2 J 2 while for β = 4 \beta=4 β = 4 it is 4 ⋅ 2 N ⋅ T = 4 J 2 4\cdot2N\cdot T=4J^2 4 ⋅ 2 N ⋅ T = 4 J 2 and for β = 1 \beta=1 β = 1 it is 4 N ⋅ T / 2 = J 2 4N\cdot T/2=J^2 4 N ⋅ T /2 = J 2 . The physical branch on the forbidden side E < E ∗ E<E^* E < E ∗ is the one continuing to the Lifshitz saddle, d ≃ − E d\simeq\sqrt{-E} d ≃ − E as E → − ∞ E\to-\infty E → − ∞ . On it d > d ∗ d>d_* d > d ∗ , hence d 2 − r = d 2 − M T e f f / ( 2 d ) > 0 d^2-r=d^2-MT_{\mathrm{eff}}/(2d)>0 d 2 − r = d 2 − M T eff / ( 2 d ) > 0 . Thus the characteristic roots μ ± = d ± d 2 − r \mu_\pm=d\pm\sqrt{d^2-r} μ ± = d ± d 2 − r of the planar recursion (91) (i.e. the roots of μ 2 − 2 d μ + r = 0 \mu^2-2d\mu+r=0 μ 2 − 2 d μ + r = 0 ) are real and distinct on the forbidden side, and coalesce at E = E ∗ E=E^* E = E ∗ .
Continuum string equations and Painlevé I
At large but finite N N N (and hence large M M M ), resolving the rounding of the spectral edge requires going beyond the leading planar approximation and retaining finite-M M M corrections in the double-scaling window around the critical point (95) . To this end, put n = M + ν n=M+\nu n = M + ν , Δ E = E − E ∗ \Delta E=E-E^* Δ E = E − E ∗ , and write d n = d ∗ + b ( n ) d_n=d_*+b(n) d n = d ∗ + b ( n ) and r n = d ∗ 2 + ϱ ( n ) r_n=d_*^2+\varrho(n) r n = d ∗ 2 + ϱ ( n ) , with b b b and ϱ \varrho ϱ smooth in n n n . In the edge window we assume the scalings b ∼ M − 2 / 5 b\sim M^{-2/5} b ∼ M − 2/5 , ∂ n ∼ M − 1 / 5 \partial_n\sim M^{-1/5} ∂ n ∼ M − 1/5 and Δ E , ν T e f f ∼ M − 4 / 5 \Delta E,\nu T_{\mathrm{eff}}\sim M^{-4/5} Δ E , ν T eff ∼ M − 4/5 , which are confirmed by the result below, and expand the string equations (90) in powers of M − 1 / 5 M^{-1/5} M − 1/5 . At order M − 2 / 5 M^{-2/5} M − 2/5 both equations give ϱ = − d ∗ b \varrho=-d_*b ϱ = − d ∗ b , and at order M − 3 / 5 M^{-3/5} M − 3/5 both give the correction ϱ = − d ∗ b + 1 2 d ∗ b ′ \varrho=-d_*b+\frac{1}{2}d_*b' ϱ = − d ∗ b + 2 1 d ∗ b ′ , where primes denote ∂ n \partial_n ∂ n . At order M − 4 / 5 M^{-4/5} M − 4/5 one finds b 2 + 2 ϱ 2 + Δ E = 0 b^2+2\varrho_2+\Delta E=0 b 2 + 2 ϱ 2 + Δ E = 0 and 1 2 d ∗ 2 b ′ ′ + 2 d ∗ ϱ 2 − 2 d ∗ b 2 = ν T e f f \frac{1}{2}d_*^2b''+2d_*\varrho_2-2d_*b^2=\nu T_{\mathrm{eff}} 2 1 d ∗ 2 b ′′ + 2 d ∗ ϱ 2 − 2 d ∗ b 2 = ν T eff , where ϱ 2 \varrho_2 ϱ 2 is the part of ϱ \varrho ϱ of this order. Now eliminating ϱ 2 \varrho_2 ϱ 2 yields
b ′ ′ = 6 d ∗ b 2 + 2 d ∗ 2 ( ν T e f f + d ∗ Δ E ) . (96) b''=\frac{6}{d_*}b^2+\frac{2}{d_*^2}\left(\nu T_{\mathrm{eff}}+d_*\Delta E\right). \tag*{(96)} b ′′ = d ∗ 6 b 2 + d ∗ 2 2 ( ν T eff + d ∗ Δ E ) . (96)
Its planar (b ′ ′ = 0 b''=0 b ′′ = 0 ) solution b 2 = − ( Δ E + ν T e f f / d ∗ ) / 3 b^2=-(\Delta E+\nu T_{\mathrm{eff}}/d_*)/3 b 2 = − ( Δ E + ν T eff / d ∗ ) /3 is the expansion of the planar branch near E ∗ E^* E ∗ . The substitution
b = d ∗ κ y ( t ) , t = a ( Δ E + ν T e f f d ∗ ) , κ = 2 4 / 5 M − 2 / 5 , a = 2 7 / 5 M 4 / 5 s 2 , b=d_*\kappa y(t),\qquad t=a\left(\Delta E+\frac{\nu T_{\mathrm{eff}}}{d_*}\right),\qquad\kappa=2^{4/5}M^{-2/5},\qquad a=\frac{2^{7/5}M^{4/5}}{s^2}, b = d ∗ κ y ( t ) , t = a ( Δ E + d ∗ ν T eff ) , κ = 2 4/5 M − 2/5 , a = s 2 2 7/5 M 4/5 ,
turns (96) into
y ′ ′ ( t ) = 6 y 2 + t . (97) y''(t)=6y^2+t. \tag*{(97)} y ′′ ( t ) = 6 y 2 + t . (97)
The constants are fixed by κ d ∗ 2 = a 2 T e f f 2 \kappa d_*^2=a^2T_{\mathrm{eff}}^2 κ d ∗ 2 = a 2 T eff 2 and κ a 3 T e f f 2 = 2 \kappa a^3T_{\mathrm{eff}}^2=2 κ a 3 T eff 2 = 2 , together with (95) . At ν = 0 \nu=0 ν = 0 this gives the scaling variables
E = E β ∗ + s β 2 N e f f − 4 / 5 X ′ , t = 2 7 / 5 X ′ , N e f f = M , (98) E=E_\beta^*+s_\beta^2N_{\mathrm{eff}}^{-4/5}X',\qquad t=2^{7/5}X',\qquad N_{\mathrm{eff}}=M, \tag*{(98)} E = E β ∗ + s β 2 N eff − 4/5 X ′ , t = 2 7/5 X ′ , N eff = M , (98)
with N e f f = 2 N N_{\mathrm{eff}}=2N N eff = 2 N for β = 4 \beta=4 β = 4 and N e f f = N N_{\mathrm{eff}}=N N eff = N for β = 1 \beta=1 β = 1 . Moreover ∂ t / ∂ n = 2 2 / 5 M − 1 / 5 \partial t/\partial n=2^{2/5}M^{-1/5} ∂ t / ∂ n = 2 2/5 M − 1/5 , and
d n = d ∗ [ 1 + 2 4 / 5 M − 2 / 5 y ( t ) ] , r n = d ∗ 2 [ 1 − 2 4 / 5 M − 2 / 5 y ( t ) ] , (99) d_n=d_*\left[1+2^{4/5}M^{-2/5}y(t)\right],\qquad r_n=d_*^2\left[1-2^{4/5}M^{-2/5}y(t)\right], \tag*{(99)} d n = d ∗ [ 1 + 2 4/5 M − 2/5 y ( t ) ] , r n = d ∗ 2 [ 1 − 2 4/5 M − 2/5 y ( t ) ] , (99)
up to relative corrections O ( M − 1 / 5 ) O(M^{-1/5}) O ( M − 1/5 ) . Combining (99) with (94) and (95) gives the unitary edge density [46 ] in the present conventions:
ρ M ( 2 ) ∣ T e f f ≃ − 2 4 / 5 π M 3 / 5 s Im y ( t ) . (100) \rho_M^{(2)}\big|_{T_{\mathrm{eff}}}\simeq-\frac{2^{4/5}}{\pi}\frac{M^{3/5}}{s}\operatorname{Im}y(t). \tag*{(100)} ρ M ( 2 ) T eff ≃ − π 2 4/5 s M 3/5 Im y ( t ) . (100)
The auxiliary equation
Put R m = d ∗ m q m R_m=d_*^mq_m R m = d ∗ m q m in (91) . Then
q m + 1 − 2 q m + q m − 1 = [ d 2 m + d 2 m + 1 d ∗ − 2 ] q m − [ r 2 m d ∗ 2 − 1 ] q m − 1 = 3 b d ∗ q m [ 1 + O ( M − 1 / 5 ) ] . q_{m+1}-2q_m+q_{m-1}=\left[\frac{d_{2m}+d_{2m+1}}{d_*}-2\right]q_m-\left[\frac{r_{2m}}{d_*^2}-1\right]q_{m-1}=\frac{3b}{d_*}q_m\left[1+O(M^{-1/5})\right]. q m + 1 − 2 q m + q m − 1 = [ d ∗ d 2 m + d 2 m + 1 − 2 ] q m − [ d ∗ 2 r 2 m − 1 ] q m − 1 = d ∗ 3 b q m [ 1 + O ( M − 1/5 ) ] .
Here we used ϱ = − d ∗ b + O ( ϵ 3 / 2 ) \varrho=-d_*b+O(\epsilon^{3/2}) ϱ = − d ∗ b + O ( ϵ 3/2 ) . The terms containing b ′ b' b ′ and q m − q m − 1 q_m-q_{m-1} q m − q m − 1 are smaller by a factor M − 1 / 5 M^{-1/5} M − 1/5 . One step in m m m is two steps in n n n , i.e. Δ t = 2 7 / 5 M − 1 / 5 \Delta t=2^{7/5}M^{-1/5} Δ t = 2 7/5 M − 1/5 . The left-hand side therefore becomes 2 14 / 5 M − 2 / 5 q ′ ′ ( t ) 2^{14/5}M^{-2/5}q''(t) 2 14/5 M − 2/5 q ′′ ( t ) and the right-hand side 3 ⋅ 2 4 / 5 M − 2 / 5 y q 3\cdot2^{4/5}M^{-2/5}yq 3 ⋅ 2 4/5 M − 2/5 y q . Hence
q ′ ′ ( t ) = 3 4 y ( t ) q ( t ) , (101) q''(t)=\frac{3}{4}y(t)q(t), \tag*{(101)} q ′′ ( t ) = 4 3 y ( t ) q ( t ) , (101)
for both ensembles.
For ∣ t ∣ ≫ 1 |t|\gg1 ∣ t ∣ ≫ 1 the function y y y varies slowly on the scale of the local wavelength, ∣ y ′ ∣ ≪ ∣ y ∣ 3 / 2 |y'|\ll|y|^{3/2} ∣ y ′ ∣ ≪ ∣ y ∣ 3/2 , since ∣ y ∣ ≃ ∣ t ∣ / 6 |y|\simeq\sqrt{|t|/6} ∣ y ∣ ≃ ∣ t ∣/6 on both sides of the window. There the solutions of (101) take the WKB form q ± ≃ ( 3 y / 4 ) − 1 / 4 exp [ ± ∫ t 3 y / 4 d t ′ ] q_\pm\simeq(3y/4)^{-1/4}\exp\left[\pm\int^t\sqrt{3y/4}\,\mathrm{d}t'\right] q ± ≃ ( 3 y /4 ) − 1/4 exp [ ± ∫ t 3 y /4 d t ′ ] . Inside the window, ∣ t ∣ = O ( 1 ) |t|=O(1) ∣ t ∣ = O ( 1 ) , q q q is given only by the exact solution of (101) ; the WKB forms serve to label solutions by their behaviour as t → ± ∞ t\to\pm\infty t → ± ∞ and to match them onto the planar regime. In that regime they are the continuations of the planar roots μ ± \mu_\pm μ ± . Indeed, μ ± / d ∗ ≃ 1 ± 3 b / d ∗ \mu_\pm/d_*\simeq1\pm\sqrt{3b/d_*} μ ± / d ∗ ≃ 1 ± 3 b / d ∗ , and 3 b / d ∗ / Δ t = 3 y / 4 \sqrt{3b/d_*}/\Delta t=\sqrt{3y/4} 3 b / d ∗ /Δ t = 3 y /4 . In particular, the larger characteristic root μ + \mu_+ μ + , which dominates the growth of R m R_m R m over μ − \mu_- μ − under forward iteration in m m m , corresponds to the solution growing with t t t . For the densities, ∂ E = ( ∂ t / ∂ E ) ∂ t \partial_E=(\partial t/\partial E)\partial_t ∂ E = ( ∂ t / ∂ E ) ∂ t with ∂ t / ∂ E = 2 7 / 5 M 4 / 5 / s 2 \partial t/\partial E=2^{7/5}M^{4/5}/s^2 ∂ t / ∂ E = 2 7/5 M 4/5 / s 2 .
The last terms in (93) are then ( T / T e f f ) ( 2 4 / 5 / π ) ( M 3 / 5 / s ) Im ( log q ) ′ ′ (T/T_{\mathrm{eff}})(2^{4/5}/\pi)(M^{3/5}/s)\operatorname{Im}(\log q)'' ( T / T eff ) ( 2 4/5 / π ) ( M 3/5 / s ) Im ( log q ) ′′ , and ( log q ) ′ ′ = 3 4 y − v 2 (\log q)''=\frac{3}{4}y-v^2 ( log q ) ′′ = 4 3 y − v 2 with v = q ′ / q v=q'/q v = q ′ / q . To estimate the sums in (93) , note that, for even K K K and r 0 = 0 r_0=0 r 0 = 0 ,
∑ j = 0 K / 2 − 1 ( r 2 j + 2 − r 2 j + 1 ) = 1 2 r K + 1 2 ∑ j = 0 K / 2 − 1 ( r 2 j + 2 − 2 r 2 j + 1 + r 2 j ) . \sum_{j=0}^{K/2-1}(r_{2j+2}-r_{2j+1})=\frac{1}{2}r_K+\frac{1}{2}\sum_{j=0}^{K/2-1}(r_{2j+2}-2r_{2j+1}+r_{2j}). j = 0 ∑ K /2 − 1 ( r 2 j + 2 − r 2 j + 1 ) = 2 1 r K + 2 1 j = 0 ∑ K /2 − 1 ( r 2 j + 2 − 2 r 2 j + 1 + r 2 j ) .
Thus the deviation from 1 2 r K \frac{1}{2}r_K 2 1 r K is a sum of second differences. By (99) , the critical part of r n r_n r n has magnitude O ( s 2 M − 2 / 5 ) O(s^2M^{-2/5}) O ( s 2 M − 2/5 ) and varies over O ( M 1 / 5 ) O(M^{1/5}) O ( M 1/5 ) indices. Each second difference is therefore O ( s 2 M − 4 / 5 ) O(s^{2}M^{-4/5}) O ( s 2 M − 4/5 ) , and their sum over the critical region (of width O ( M 1 / 5 ) O(M^{1/5}) O ( M 1/5 ) ) is O ( s 2 M − 3 / 5 ) O(s^{2}M^{-3/5}) O ( s 2 M − 3/5 ) . Matching to the forbidden-side solution, whose imaginary part decays away from this region, then gives
Im ∑ j = 0 K / 2 − 1 ( r 2 j + 2 − r 2 j + 1 ) = 1 2 Im r K + O ( s 2 M − 3 / 5 ) , \operatorname{Im}\sum_{j=0}^{K/2-1}(r_{2j+2}-r_{2j+1})=\frac{1}{2}\operatorname{Im}r_K+O(s^{2}M^{-3/5}), Im j = 0 ∑ K /2 − 1 ( r 2 j + 2 − r 2 j + 1 ) = 2 1 Im r K + O ( s 2 M − 3/5 ) ,
where we used that 1 2 Im r K = O ( s 2 M − 2 / 5 ) \frac{1}{2}\operatorname{Im}r_K=O(s^{2}M^{-2/5}) 2 1 Im r K = O ( s 2 M − 2/5 ) in the critical window (99) . Here K = M K=M K = M for β = 4 \beta=4 β = 4 and K = M − 2 K=M-2 K = M − 2 for β = 1 \beta=1 β = 1 . In the latter case, replacing r K r_K r K by r M r_M r M introduces an error of the same order, since a fixed index shift changes t t t by O ( M − 1 / 5 ) O(M^{-1/5}) O ( M − 1/5 ) . Compared with the leading critical contribution O ( s 2 M − 2 / 5 ) O(s^{2}M^{-2/5}) O ( s 2 M − 2/5 ) , these errors are smaller by M − 1 / 5 M^{-1/5} M − 1/5 .
Two further approximations are within the stated O ( N − 1 / 5 ) O(N^{-1/5}) O ( N − 1/5 ) relative accuracy. First, for β = 1 \beta=1 β = 1 the final index 2 m − 2 = N − 2 2m-2=N-2 2 m − 2 = N − 2 differs from N N N by a shift of t t t of order M − 1 / 5 M^{-1/5} M − 1/5 . Second, any smooth E E E -dependence of log R \log R log R not captured by the scaling function contributes to ρ \rho ρ a term of order one; since the leading edge density is of order N 3 / 5 N^{3/5} N 3/5 (see Eqs. (102) and (103) ), this is a relative correction of order N − 3 / 5 N^{-3/5} N − 3/5 , smaller than the O ( N − 1 / 5 ) O(N^{-1/5}) O ( N − 1/5 ) corrections already neglected.
Writing v 4 = q ′ / q v_4=q'/q v 4 = q ′ / q and v 1 = x ′ / x v_1=x'/x v 1 = x ′ / x , with q q q and x x x the solutions selected below, and using (100) , we obtain
ρ N ( 4 ) ≃ N ′ 3 / 5 s 4 2 4 / 5 π Im [ 1 4 y − v 4 2 ] , N ′ = 2 N , (102) \rho_N^{(4)}\simeq\frac{N^{\prime3/5}}{s_4}\frac{2^{4/5}}{\pi}\operatorname{Im}\left[\frac{1}{4}y-v_4^2\right],\qquad N'=2N, \tag*{(102)} ρ N ( 4 ) ≃ s 4 N ′3/5 π 2 4/5 Im [ 4 1 y − v 4 2 ] , N ′ = 2 N , (102)
and, for even N N N ,
ρ N ( 1 ) ≃ N 3 / 5 s 1 2 9 / 5 π Im [ 1 4 y − v 1 2 ] . (103) \rho_N^{(1)}\simeq\frac{N^{3/5}}{s_1}\frac{2^{9/5}}{\pi}\operatorname{Im}\left[\frac{1}{4}y-v_1^2\right]. \tag*{(103)} ρ N ( 1 ) ≃ s 1 N 3/5 π 2 9/5 Im [ 4 1 y − v 1 2 ] . (103)
Both hold up to relative corrections O ( N − 1 / 5 ) O(N^{-1/5}) O ( N − 1/5 ) . Equivalently,
ρ N ( 4 ) ≃ 1 2 ρ 2 N ( 2 ) + N ′ 3 / 5 s 4 2 4 / 5 π Im ( log q ) ′ ′ , ρ N ( 1 ) ≃ ρ N ( 2 ) ∣ T / 2 + N 3 / 5 s 1 2 9 / 5 π Im ( log x ) ′ ′ . \rho_N^{(4)}\simeq\frac{1}{2}\rho_{2N}^{(2)}+\frac{N^{\prime3/5}}{s_4}\frac{2^{4/5}}{\pi}\operatorname{Im}(\log q)'',
\qquad
\rho_N^{(1)}\simeq\left.\rho_N^{(2)}\right|_{T/2}+\frac{N^{3/5}}{s_1}\frac{2^{9/5}}{\pi}\operatorname{Im}(\log x)''. ρ N ( 4 ) ≃ 2 1 ρ 2 N ( 2 ) + s 4 N ′3/5 π 2 4/5 Im ( log q ) ′′ , ρ N ( 1 ) ≃ ρ N ( 2 ) T /2 + s 1 N 3/5 π 2 9/5 Im ( log x ) ′′ .
In (103) the factor 2 9 / 5 = 2 ⋅ 2 4 / 5 2^{9/5}=2\cdot2^{4/5} 2 9/5 = 2 ⋅ 2 4/5 reflects T / T e f f = 2 T/T_{\rm eff}=2 T / T eff = 2 . The logarithmic derivatives satisfy v β = 3 4 y − v β 2 v_\beta=\frac{3}{4}y-v_\beta^2 v β = 4 3 y − v β 2 .
The physical solution of Painlevé I
Generic solutions of Painlevé I have poles in every direction of the complex t t t -plane; two requirements single out the physical one. First, by the zero-free region (61) Ξ N \Xi_N Ξ N has no poles for Im E ≥ 0 \operatorname{Im}E\geq0 Im E ≥ 0 . Since t = a ( Δ E + ν T e f f / d ∗ ) t=a(\Delta E+\nu T_{\rm eff}/d_*) t = a ( Δ E + ν T eff / d ∗ ) with a > 0 a>0 a > 0 , and y y y enters ∂ θ Ξ M ( 2 ) = r M \partial_\theta\Xi_M^{(2)}=r_M ∂ θ Ξ M ( 2 ) = r M linearly by (99) , y y y has no poles in the closed upper half t t t -plane. Second, for ∣ t ∣ → ∞ |t|\to\infty ∣ t ∣ → ∞ in that half-plane the planar branch, continued from the Lifshitz saddle (b > 0 b>0 b > 0 ) through Im E > 0 \operatorname{Im}E>0 Im E > 0 , must be recovered:
y ( t ) ≃ − t / 6 , 0 ≤ arg t ≤ π , (104) y(t)\simeq\sqrt{-t/6},\qquad0\leq\arg t\leq\pi, \tag*{(104)} y ( t ) ≃ − t /6 , 0 ≤ arg t ≤ π , (104)
with the principal branch; in particular y ≃ − i t / 6 y\simeq-\mathrm{i}\sqrt{t/6} y ≃ − i t /6 for t → + ∞ t\to+\infty t → + ∞ , which gives a positive density in (100) .
These conditions fix the solution uniquely. Linearizing about y 0 = − t / 6 y_0=\sqrt{-t/6} y 0 = − t /6 , δ y ′ ′ = 12 y 0 δ y \delta y''=12y_0\delta y δ y ′′ = 12 y 0 δ y , gives the two WKB solutions
δ y ± ≃ ( − t ) − 1 / 8 exp [ ± 4 5 2 3 / 2 3 1 / 4 ( − t ) 5 / 4 ] . (105) \delta y_\pm\simeq(-t)^{-1/8}\exp\left[\pm\frac{4}{5}2^{3/2}3^{1/4}(-t)^{5/4}\right]. \tag*{(105)} δ y ± ≃ ( − t ) − 1/8 exp [ ± 5 4 2 3/2 3 1/4 ( − t ) 5/4 ] . (105)
The two solutions exchange growth and decay across the Stokes ray arg t = 3 π / 5 \arg t=3\pi/5 arg t = 3 π /5 , so a solution obeying (104) on both sides of this direction can contain neither, and both integration constants are fixed. (About − − t / 6 -\sqrt{-t/6} − − t /6 the corresponding linearized WKB solutions oscillate instead, which is why the real tronquée cannot describe the activated tail.) With ω = e 2 π i / 5 \omega=e^{2\pi\mathrm{i}/5} ω = e 2 π i /5 and y r y_r y r the real tronquée, which behaves as y r ∼ − − t / 6 y_r\sim-\sqrt{-t/6} y r ∼ − − t /6 for ∣ arg ( − t ) ∣ < 4 π / 5 |\arg(-t)|<4\pi/5 ∣ arg ( − t ) ∣ < 4 π /5 , the solution is
y ( t ) = y 1 ( t ) : = ω 2 y r ( ω t ) , (106) y(t)=y_1(t):=\omega^2y_r(\omega t), \tag*{(106)} y ( t ) = y 1 ( t ) := ω 2 y r ( ω t ) , (106)
which behaves as − t / 6 \sqrt{-t/6} − t /6 for − π / 5 < arg t < 7 π / 5 -\pi/5 < \arg t < 7\pi/5 − π /5 < arg t < 7 π /5 (Kapaev’s y 0 y_0 y 0 [59 ] ). By the proof of Dubrovin’s conjecture [16 ] , see also [58 ] , its poles lie in − 3 π / 5 < arg t < − π / 5 -3\pi/5 < \arg t < -\pi/5 − 3 π /5 < arg t < − π /5 , i.e. at Im E < 0 \operatorname{Im} \mathcal{E} < 0 Im E < 0 ; it is regular on the real axis, and analyticity holds for all finite t t t . The conjugate y − 1 ( t ) = y 1 ( t ˉ ) ‾ y_{-1}(t)=\overline{y_1(\bar{t})} y − 1 ( t ) = y 1 ( t ˉ ) belongs to E − i 0 \mathcal{E}-i0 E − i 0 . The same y 1 y_1 y 1 follows from mapping the outgoing contour to the critical cubic model [46 , 8 , 9 , 2 ] .
On the forbidden side y ± 1 y_{\pm1} y ± 1 share the real series about + − t / 6 +\sqrt{-t/6} + − t /6 and differ, like Bi ± i Ai \operatorname{Bi}\pm i\operatorname{Ai} Bi ± i Ai , by the exponentially small jump computed by Kapaev [59 ] :
Im y 1 ( t ) ≃ − 2 − 19 / 8 3 − 1 / 8 π ∣ t ∣ − 1 / 8 e − 4 5 2 3 / 4 3 1 / 4 ∣ t ∣ 5 / 4 , t → − ∞ . (107) \operatorname{Im}y_1(t)\simeq-\frac{2^{-19/8}3^{-1/8}}{\sqrt{\pi}}|t|^{-1/8}e^{-\frac{4}{5}2^{3/4}3^{1/4}|t|^{5/4}},\qquad t\to-\infty. \tag*{(107)} Im y 1 ( t ) ≃ − π 2 − 19/8 3 − 1/8 ∣ t ∣ − 1/8 e − 5 4 2 3/4 3 1/4 ∣ t ∣ 5/4 , t → − ∞. (107)
Substituting (98) into the exponent of (107) gives 4 5 2 3 1 / 4 s 1 / 2 ( E ∗ − E ) 5 / 4 / T e f f \frac{4}{5}\sqrt{2}3^{1/4}s^{1/2}(E^*-E)^{5/4}/T_{\mathrm{eff}} 5 4 2 3 1/4 s 1/2 ( E ∗ − E ) 5/4 / T eff , which is precisely the near-edge single-eigenvalue barrier of Sec. 10 at temperature T e f f T_{\mathrm{eff}} T eff .
Selection for β = 4 \beta=4 β = 4
The recursion (91) for Q Q Q starts at m = 0 m=0 m = 0 from Q 0 = 1 Q_0=1 Q 0 = 1 and Q 1 = d 0 + d 1 Q_1=d_0+d_1 Q 1 = d 0 + d 1 . For small n n n one has E < E ∗ ( n ) E<E^*(n) E < E ∗ ( n ) , where E ∗ ( n ) = − 3 4 ( 4 n T e f f ) 2 / 3 E^*(n)=-\frac{3}{4}(4nT_{\mathrm{eff}})^{2/3} E ∗ ( n ) = − 4 3 ( 4 n T eff ) 2/3 , so the early steps lie on the forbidden side of the local criticality, where the roots μ ± \mu_\pm μ ± are real and separated. Under forward iteration in m m m , the component of Q m Q_m Q m along the larger characteristic root μ + \mu_+ μ + is amplified exponentially relative to μ − \mu_- μ − for generic initial values, unless these are fine-tuned to suppress it. Nothing in the polynomial initial values of Q Q Q does so, and Q m Q_m Q m therefore follows μ + \mu_+ μ + up to exponentially small corrections. In the window this is the solution growing with t t t , i.e. the solution of (101) decaying as t → − ∞ t\to-\infty t → − ∞ :
The consequences are the following.
Forbidden side. Here v 4 = 3 y / 4 + … v_4=\sqrt{3y/4}+\ldots v 4 = 3 y /4 + … , where the dots denote real terms plus terms proportional to Im y \operatorname{Im}y Im y . Hence Im [ 1 4 y − v 4 2 ] \operatorname{Im}[\frac{1}{4}y-v_4^2] Im [ 4 1 y − v 4 2 ] is exponentially small at the rate (105) , reproducing the near-edge single-eigenvalue barrier of Sec. 10 at temperature T T T with s = s 4 s=s_4 s = s 4 , as required.
Allowed side. q q q is purely q + q_+ q + near arg t = 3 π / 5 \arg t=3\pi/5 arg t = 3 π /5 , and q + q_+ q + dominates for arg t < 3 π / 5 \arg t<3\pi/5 arg t < 3 π /5 , so q q q grows as t → + ∞ t\to+\infty t → + ∞ . Then v 4 2 = 3 4 y + O ( t − 1 ) v_4^2=\frac{3}{4}y+O(t^{-1}) v 4 2 = 4 3 y + O ( t − 1 ) , and (102) reduces to 1 2 ρ 2 N ( 2 ) \frac{1}{2}\rho_{2N}^{(2)} 2 1 ρ 2 N ( 2 ) , the planar β = 4 \beta=4 β = 4 density.
Selection for β = 1 \beta=1 β = 1
The sequence X p h X^{\mathrm{ph}} X ph obeys the same recursion; with generic initial values it would follow μ + \mu_+ μ + like Q Q Q . Two conditions show that it does not.
(a) Analyticity. The zero-free region holds for every β \beta β , so the term T ∂ E log X m − 1 p h T\partial_E\log X_{m-1}^{\mathrm{ph}} T ∂ E log X m − 1 ph in (92) has no poles, and x x x no zeros, for Im t > 0 \operatorname{Im}t>0 Im t > 0 . In the WKB forms q ± ≃ ( 3 y / 4 ) − 1 / 4 e ± W q_\pm\simeq(3y/4)^{-1/4}e^{\pm W} q ± ≃ ( 3 y /4 ) − 1/4 e ± W , W ′ = 3 y / 4 W'=\sqrt{3y/4} W ′ = 3 y /4 , dominance changes at the same direction arg t = 3 π / 5 \arg t=3\pi/5 arg t = 3 π /5 as in Par. 5, and a solution containing both q ± q_\pm q ± there has a sequence of zeros along it, as Ai + c Bi \operatorname{Ai}+c\operatorname{Bi} Ai + c Bi has along a ray. Hence x x x is purely q + q_+ q + near that direction, i.e. decays as t → − ∞ t\to-\infty t → − ∞ , or purely q − q_- q − , i.e. decays as t → + ∞ t\to+\infty t → + ∞ .
(b) The barrier. The activation exponent at fixed T T T does not depend on β \beta β (Secs. 9 and 10), while the recursion is driven by the unitary ensemble at T e f f = T / 2 T_{\mathrm{eff}}=T/2 T eff = T /2 , so Im y \operatorname{Im}y Im y carries twice the required exponent. The first candidate, purely q + q_+ q + on the negative axis, inherits this exponent and is excluded. The second, continued to the negative axis, acquires across arg t = π \arg t=\pi arg t = π an admixture of q + q_+ q + with a complex coefficient, which contributes to Im [ 1 4 y − v 1 2 ] \operatorname{Im}[\frac{1}{4}y-v_1^2] Im [ 4 1 y − v 1 2 ] at order exp [ − 2 ∫ 0 t 3 y / 4 ] \exp[-2\int_0^t\sqrt{3y/4}] exp [ − 2 ∫ 0 t 3 y /4 ] . Since 2 3 / 4 = 1 2 12 2\sqrt{3/4}=\frac{1}{2}\sqrt{12} 2 3/4 = 2 1 12 , this is half of (105) , the barrier at temperature T T T with s = s 1 s=s_1 s = s 1 . Hence x x x decays as t → + ∞ t\to+\infty t → + ∞ . For β = 4 \beta=4 β = 4 (T e f f = T T_{\mathrm{eff}}=T T eff = T ) the same argument excludes the second candidate, in agreement with Par. 6.
Checks. For t → + ∞ t\to+\infty t → + ∞ , log q ≃ + ∫ t 3 y / 4 \log q\simeq+\int^t\sqrt{3y/4} log q ≃ + ∫ t 3 y /4 and log x ≃ − ∫ t 3 y / 4 \log x\simeq-\int^t\sqrt{3y/4} log x ≃ − ∫ t 3 y /4 . Since the unitary parts of log Z ( 4 ) \log Z^{(4)} log Z ( 4 ) and log Z ( 1 ) \log Z^{(1)} log Z ( 1 ) are oriented, log q \log q log q and log x \log x log x carry the leading projective-plane terms, and these are opposite in the variables ( N e f f , T e f f ) (N_{\mathrm{eff}},T_{\mathrm{eff}}) ( N eff , T eff ) , as the β ↔ 4 / β \beta\leftrightarrow4/\beta β ↔ 4/ β duality requires [72 ] ; cf. the N 2 N^2 N 2 terms ± 9 4 N e f f − 1 \pm\frac{9}{4}N_{\mathrm{eff}}^{-1} ± 4 9 N eff − 1 of K β ( N ) K_\beta(N) K β ( N ) in (78) . At finite N N N the initial values X 0 , X 1 X_0,X_1 X 0 , X 1 must make X p h X^{\mathrm{ph}} X ph the minimal solution of the recursion on the forbidden side; we leave the exact finite-N N N verification of this minimal-solution condition for the Airy primitive X 0 X_0 X 0 (which vanishes as s → − ∞ s\to-\infty s → − ∞ ) as a conjecture.
Summary
Both edges are thus built from solutions of the same Schrödinger operator − ∂ t 2 + 3 4 y ( t ) -\partial_t^2 + \frac{3}{4}y(t) − ∂ t 2 + 4 3 y ( t ) , selected at opposite ends: decaying at t → − ∞ t \to-\infty t → − ∞ for β = 4 \beta= 4 β = 4 , and at t → + ∞ t \to+\infty t → + ∞ for β = 1 \beta= 1 β = 1 . In both cases y y y is the complex tritronquée (106) . For β = 2 \beta= 2 β = 2 the edge profile is given by y y y alone; for β = 1 , 4 \beta= 1,4 β = 1 , 4 the selected solution of the auxiliary equation adds a contribution of the same order.
On the allowed side both scaling forms match their planar limits. This leading-order matching is insensitive to the selection, since v β 2 ≃ 3 4 y v_\beta^2 \simeq\frac{3}{4}y v β 2 ≃ 4 3 y for either solution; the selection shows up only at subleading order. On the forbidden side both reproduce the single-eigenvalue barrier at temperature T T T , with the respective s β s_\beta s β .
The construction is analogous to the way the Tracy–Widom distributions F 1 F_1 F 1 and F 4 F_4 F 4 are built from the same Painlevé II data [86 ] , although the driving equation and the selection problem here are different.
Discussion and outlook
The central problem of this paper is the spectral behaviour of an N N N -channel Schrödinger operator with matrix-valued white-noise disorder. The stationary Riccati distribution provides a common description of this problem in all three Dyson classes. Its normalization yields the density of states and the sum of Lyapunov exponents, while its asymptotic regimes connect the rare fluctuations responsible for the Lifshitz tail to the collective behaviour near the large-N N N band edge. The appearance of outgoing Kontsevich integrals is the exact mathematical structure underlying these spectral results.
Exact spectral description. At complex energy, the Riccati matrix has an explicit stationary probability distribution for β = 1 , 2 , 4 \beta= 1,2,4 β = 1 , 2 , 4 . This distribution is not of Gibbs type. Its stationarity follows from a drift of the form ( D + A ) ∇ Φ (D + A)\nabla\Phi ( D + A ) ∇Φ , with constant antisymmetric A A A . The relevant trace identities remain valid for non-commuting matrices and in all three Dyson classes. At real energy, the same construction gives a Gibbs/anti-Gibbs two-matrix representation, extending Halperin’s scalar formula to the multichannel problem.
The outgoing integrals τ N ( β ) \tau_N^{(\beta)} τ N ( β ) arise as the holomorphic factors of the normalization. They therefore give closed expressions for the spectral observables and explain why the unitary result of Ref. [46 ] extends to the orthogonal and symplectic classes. For β = 4 \beta= 4 β = 4 , and for β = 1 \beta= 1 β = 1 with even N N N , the finite-N N N reductions also identify the additional recursions that distinguish their spectral behaviour from that of the underlying β = 2 \beta= 2 β = 2 problem.
Rare fluctuations and the Lifshitz tail. The forbidden-region density is governed by a fluctuation that binds a state in one direction in channel space. The calculation in § 9 determines both the exponential barrier and the N N N - and β \beta β -dependent prefactor, together with its first correction. The channel dependence has a geometric origin: the direction selected by the fluctuation belongs to R P N − 1 \mathbb{R}\mathbb{P}^{N-1} R P N − 1 , C P N − 1 \mathbb{C}\mathbb{P}^{N-1} C P N − 1 or H P N − 1 \mathbb{H}\mathbb{P}^{N-1} H P N − 1 , according to the Dyson class. This geometry enters the prefactor, while the tunnelling configuration determines the exponential suppression.
These tail results concern the Gaussian white-noise model. They should not be transferred without qualification to disorder with bounded support: sufficiently deep in the forbidden region, the optimal fluctuation probes the microscopic distribution of the disorder.
The critical edge. The rounded planar edge limit at large N N N describes a different regime from the fixed-N N N deep Lifshitz tail. In the critical window, the recurrence coefficients of the underlying β = 2 \beta= 2 β = 2 orthogonal polynomials are governed by the complex Painlevé I tritronquée (106) . For β = 1 , 4 \beta= 1,4 β = 1 , 4 , this does not by itself determine the density: the additional finite-N N N recursions retain a nontrivial continuum limit and yield q ′ ′ ( t ) = 3 4 y ( t ) q ( t ) q''(t) = \frac{3}{4}y(t)q(t) q ′′ ( t ) = 4 3 y ( t ) q ( t ) . The two classes select different solutions of this equation, and their logarithmic derivatives supply the corrections to the unitary scaling function.
For β = 4 \beta= 4 β = 4 , analyticity in the upper half energy plane and the near-edge single-eigenvalue barrier at temperature T T T , as well as the finite-N N N recursion, select the solution decaying as t → − ∞ t \to-\infty t → − ∞ .
For β = 1 \beta= 1 β = 1 (even N N N ), analyticity in the upper half energy plane and the near-edge single-eigenvalue barrier at temperature T T T select the solution decaying as t → + ∞ t \to+\infty t → + ∞ , in agreement with the leading projective-plane contribution to the free energy.
The resulting edge forms match the planar densities on the allowed side and the single-eigenvalue barrier on the forbidden side at the level of the exponential decay rates.
A universality conjecture. We expect the edge scaling functions obtained here to extend beyond the Gaussian white-noise operator. The natural candidate family consists of quasi-one-dimensional disordered systems whose bulk limit is described by the one-dimensional Efetov–Wegner nonlinear σ \sigma σ -model in the corresponding Dyson class. Such a reduction was established for random band matrices in the large-bandwidth limit by Fyodorov and Mirlin [36 , 37 ] ; for recent rigorous progress on the 1D random-band-matrix localization transition, see also [50 ] . We refer to this family here as the 1D Efetov–Wegner class and conjecture that, at a regular band edge and after matching the model-dependent scales, its members exhibit the density scaling functions obtained in this paper.
The conjecture is supported by the random-band-matrix edge theory of Sodin [83 ] and of Liu and Zou [67 , 68 ] (see also Ref. [82 ] for early insights). For β = 2 \beta= 2 β = 2 the correspondence was proposed in Ref. [46 ] . For β = 1 \beta= 1 β = 1 the comparison is even more constraining, since the metric factors are inherited from the unitary case: both the projective-plane coefficient and the regularized first moment of the band-matrix edge intensity follow without free parameters, and both agree with the band-matrix diagram expansion (Appendix § 22 ). An analogous edge universality for chiral random band matrices will be developed in [40 ] . As a critical-window phenomenon, this universality governs the scaling regime near a regular band edge, while deeper in the forbidden region the asymptotic Lifshitz tail naturally reflects the microscopic disorder distribution.
Beyond the holomorphic sector. The spectral results above use the normalization and holomorphic observables of the stationary law. The full distribution also provides a starting point for non-holomorphic scattering observables. Appendix § 23 collects the boundary-response identities, their relation to time delays, and the high-energy connection with absorbing-waveguide theory, and outlines the programme of Paper II [44 ] .
Open problems
Mathematical completion. Three steps are argued here at the physics level. The first is the uniqueness of the stationary distribution, both at η > 0 \eta> 0 η > 0 and for the real-energy dynamics continued through the Riccati poles. At real energy this uniqueness is also what identifies Eq. (63) with the current-carrying stationary state for β = 1 , 4 \beta= 1, 4 β = 1 , 4 ; for β = 2 \beta= 2 β = 2 that identification follows independently from Appendix § 20 and the fermionic solution of Ref. [46 ] . The second is the decomposition (57) of the integration surface behind Result 2. The third is a rigorous Riemann–Hilbert double-scaling analysis of the string equations and the β = 1 , 4 \beta= 1, 4 β = 1 , 4 auxiliary recursions in § 11 (including the solution-selection argument and the minimal-solution condition for X 0 X_0 X 0 ).
The Y Y Y -ensemble at large N N N . Eq. (50) is an explicit positive ensemble with a Schur-type interaction. Its planar limit and edge would give an alternative, manifestly positive route to the β = 1 , 4 \beta=1,4 β = 1 , 4 edge functions of Sec. 11.
Finite length L L L . The stationary distribution gives neither the L L L -dependent counting statistics nor first-escape rates, and in particular does not determine the banded-matrix analogue of the Tracy–Widom edge distribution at finite L L L . For β = 2 \beta=2 β = 2 , the L L L -dependent counting statistics and first-escape rates are accessible via the fermionic structure of Ref. [46 ] (though the finite-L L L edge distribution itself was not derived there); for β = 1 , 4 \beta=1,4 β = 1 , 4 , even the L L L -dependent formulation remains open.
Non-orientable topology. Do the β = 1 , 4 \beta=1,4 β = 1 , 4 outgoing Kontsevich integrals with non-degenerate thresholds generate intersection-type invariants of moduli spaces of non-orientable surfaces, e.g. within refined topological recursion [13 , 48 ] ? A related concrete problem is a proof of the twisted-gluing conjecture of Appendix E2, i.e. of the asymptotic irrelevance of the side constraint for the loop count of randomly twisted triangle gluings.
Acknowledgements
The author thanks Pierre Le Doussal for the collaboration on Ref. [46 ] , on which this work builds, and Aurellien Grabsch for encouraging interest in the results and methods of the paper. Prof. Sergei Chmutov is gratefully acknowledged for useful correspondence about statistics of randomly twisted triangle gluings.
The research reported in this paper was supported by UKRI grant UKRI1015, “Non-Hermitian Random Matrices: Theory and Applications.”
The author acknowledges assistance from GPT-6.1 Sol (OpenAI) with organizing the systematics of results transfer from N = 1 N=1 N = 1 to N ≥ 1 N\geq1 N ≥ 1 (in particular, in § 15 B)-§ 15 C)) and writing the Appendices. Claude Opus 5 (Anthropic) was used for designing and executing numerical computations, for help with analysing various asymptotics of Kontsevich integrals, and for critical reading. Finally, Gemini 4 Hydrogen (Google) was used for improving the clarity of the presentation and performing consistency checks. All scientific arguments, results, and conclusions remain the responsibility of the author.
Algebraic equivalence of the matrix Halperin distribution and the fermionic stationary state
We establish the equivalence at β = 2 \beta=2 β = 2 directly, without invoking uniqueness of the stationary matrix diffusion. Throughout, T > 0 T>0 T > 0 , E ∈ R E\in\mathbb{R} E ∈ R , and
U ( x ) = E x + x 3 3 , r = N − 1 , Δ ( x ) = ∏ i < j ( x j − x i ) . (108) U(x)=Ex+\frac{x^3}{3},\qquad r=N-1,\qquad\Delta(x)=\prod_{i<j}(x_j-x_i). \tag*{(108)} U ( x ) = E x + 3 x 3 , r = N − 1 , Δ ( x ) = i < j ∏ ( x j − x i ) . (108)
Flat matrix measure means the product of the diagonal differentials and the real and imaginary off-diagonal differentials. Haar measure on U ( N ) U(N) U ( N ) is normalized to one. Constants independent of the eigenvalues are absorbed into normalization constants.
Statement and conventions
Consider the normalized Hermitian matrix density
P m a t ( Λ ) = 1 Z e − Tr U ( Λ ) / T ∫ M < Λ e Tr U ( M ) / T d M . (109) P_{\mathrm{mat}}(\Lambda)=\frac{1}{\mathcal{Z}}e^{-\operatorname{Tr}U(\Lambda)/T}\int_{M<\Lambda}e^{\operatorname{Tr}U(M)/T}\,dM. \tag*{(109)} P mat ( Λ ) = Z 1 e − Tr U ( Λ ) / T ∫ M < Λ e Tr U ( M ) / T d M . (109)
The matrix inequality B < A B < A B < A means that A − B A - B A − B is positive definite. The integral converges for fixed Λ \Lambda Λ : putting M = Λ − Y M = \Lambda- Y M = Λ − Y , Y > 0 Y > 0 Y > 0 , gives
Tr U ( Λ − Y ) = Tr U ( Λ ) − Tr [ ( E I + Λ 2 ) Y ] + Tr ( Λ Y 2 ) − 1 3 Tr Y 3 . (110) \operatorname{Tr}U(\Lambda-Y)=\operatorname{Tr}U(\Lambda)-\operatorname{Tr}\left[(EI+\Lambda^2)Y\right]+\operatorname{Tr}(\Lambda Y^2)-\frac{1}{3}\operatorname{Tr}Y^3. \tag*{(110)} Tr U ( Λ − Y ) = Tr U ( Λ ) − Tr [ ( E I + Λ 2 ) Y ] + Tr ( Λ Y 2 ) − 3 1 Tr Y 3 . (110)
The cubic term dominates uniformly at infinity on the positive cone.
The fermionic construction uses the adjoint pair
L N = T ∂ x 2 + ( E + x 2 ) ∂ x + ( N + 1 ) x , L N † = T ∂ x 2 − ( E + x 2 ) ∂ x + ( N − 1 ) x . (111) L_N=T\partial_x^2+(E+x^2)\partial_x+(N+1)x,\qquad L_N^\dagger=T\partial_x^2-(E+x^2)\partial_x+(N-1)x. \tag*{(111)} L N = T ∂ x 2 + ( E + x 2 ) ∂ x + ( N + 1 ) x , L N † = T ∂ x 2 − ( E + x 2 ) ∂ x + ( N − 1 ) x . (111)
The adjoint operator L N † L_N^\dagger L N † preserves P r = span { 1 , x , … , x r } \mathcal{P}_r=\operatorname{span}\{1,x,\ldots,x^r\} P r = span { 1 , x , … , x r } . For a polynomial eigenbasis p a p_a p a , write
L N † p a = ϵ a p a , L N v a = ϵ a v a . (112) L_N^\dagger p_a=\epsilon_a p_a,\qquad L_N v_a=\epsilon_a v_a. \tag*{(112)} L N † p a = ϵ a p a , L N v a = ϵ a v a . (112)
The physical right modes are those regular through the Riccati winding; equivalently, their Lagrange currents may be normalized to
J [ p a , v a ] = T ( p a v a ′ − p a ′ v a ) + ( E + x 2 ) p a v a = 1. (113) J[p_a,v_a]=T(p_a v_a'-p_a'v_a)+(E+x^2)p_a v_a=1. \tag*{(113)} J [ p a , v a ] = T ( p a v a ′ − p a ′ v a ) + ( E + x 2 ) p a v a = 1. (113)
In the paper [46 ] these modes are written
v a ( x ) = p a ( x ) T e − U ( x ) / T FP ∫ − ∞ x e U ( t ) / T p a ( t ) 2 d t . (114) v_a(x)=\frac{p_a(x)}{T}e^{-U(x)/T}\operatorname{FP}\int_{-\infty}^{x}\frac{e^{U(t)/T}}{p_a(t)^2}\,\mathrm{d}t. \tag*{(114)} v a ( x ) = T p a ( x ) e − U ( x ) / T FP ∫ − ∞ x p a ( t ) 2 e U ( t ) / T d t . (114)
The finite part is interpreted by continuation of the regular solution across polynomial zeros. Our construction below has no such apparent poles. In fact, with p a ( x ) = c a x r + ⋯ p_a(x)=c_a x^r+\cdots p a ( x ) = c a x r + ⋯ (r = N − 1 r=N-1 r = N − 1 ), the normalized right mode takes the manifestly regular form
v a ( x ) = ( − 1 ) r c a 2 r ! T r + 1 e − U ( x ) / T ∫ − ∞ x ( x − t ) r p a ( − t ) e U ( t ) / T d t , (115) v_a(x)=\frac{(-1)^r}{c_a^2r!T^{r+1}}e^{-U(x)/T}\int_{-\infty}^{x}(x-t)^r p_a(-t)e^{U(t)/T}\,\mathrm{d}t, \tag*{(115)} v a ( x ) = c a 2 r ! T r + 1 ( − 1 ) r e − U ( x ) / T ∫ − ∞ x ( x − t ) r p a ( − t ) e U ( t ) / T d t , (115)
which is derived in Eq. (134) below.
Proposition 3. The ordered eigenvalue density induced by (A2) is the normalized fermionic density
P o r d ( λ ) = 1 N det [ p a ( λ i ) ] a , i = 1 N det [ v a ( λ i ) ] a , i = 1 N , λ 1 < ⋯ < λ N . (116) P_{\mathrm{ord}}(\lambda)=\frac{1}{\mathcal{N}}\det[p_a(\lambda_i)]_{a,i=1}^{N}\det[v_a(\lambda_i)]_{a,i=1}^{N},\qquad\lambda_1<\cdots<\lambda_N. \tag*{(116)} P ord ( λ ) = N 1 det [ p a ( λ i ) ] a , i = 1 N det [ v a ( λ i ) ] a , i = 1 N , λ 1 < ⋯ < λ N . (116) At parameter values where a polynomial eigenbasis degenerates, the identity holds by continuation, or in a generalized polynomial basis.
Angular integration and a polynomial ensemble
We use the matrix Heaviside integral of Kieburg, Kuijlaars and Stivigny, Theorem 2.3 of Ref. [60 ] . For Hermitian matrices A A A , B B B with distinct spectra a a a , b b b ,
∫ U ( N ) 1 { A − V B V † > 0 } d V = c N , 0 det [ ( a i − b j ) + N − 1 ] i , j = 1 N Δ ( a ) Δ ( b ) , c N , 0 = ∏ j = 0 N − 1 ( N − 1 j ) − 1 . (117) \int_{U(N)}\mathbf{1}_{\{A-VBV^\dagger>0\}}\,\mathrm{d}V=c_{N,0}\frac{\det\left[(a_i-b_j)_+^{N-1}\right]_{i,j=1}^{N}}{\Delta(a)\Delta(b)},\qquad c_{N,0}=\prod_{j=0}^{N-1}\binom{N-1}{j}^{-1}. \tag*{(117)} ∫ U ( N ) 1 { A − V B V † > 0 } d V = c N , 0 Δ ( a ) Δ ( b ) det [ ( a i − b j ) + N − 1 ] i , j = 1 N , c N , 0 = j = 0 ∏ N − 1 ( j N − 1 ) − 1 . (117)
Here ( s ) + 0 (s)_+^0 ( s ) + 0 means 1 { s > 0 } \mathbf{1}_{\{s>0\}} 1 { s > 0 } when N = 1 N=1 N = 1 . Repeated eigenvalues are treated by limits.
Fix Λ = diag ( λ 1 , … , λ N ) \Lambda=\operatorname{diag}(\lambda_1,\ldots,\lambda_N) Λ = diag ( λ 1 , … , λ N ) and diagonalize M = V diag ( m ) V † M=V\operatorname{diag}(m)V^\dagger M = V diag ( m ) V † . The Weyl integration formula and (A10) give
∫ M < Λ e Tr U ( M ) / T d M ∝ 1 Δ ( λ ) ∫ R N Δ ( m ) det [ ( λ i − m j ) + N − 1 ] i , j = 1 N ∏ j = 1 N e U ( m j ) / T d m j . (118) \int_{M<\Lambda}e^{\operatorname{Tr}U(M)/T}\,\mathrm{d}M\propto\frac{1}{\Delta(\lambda)}\int_{\mathbb{R}^N}\Delta(m)\det\left[(\lambda_i-m_j)_+^{N-1}\right]_{i,j=1}^{N}\prod_{j=1}^{N}e^{U(m_j)/T}\,\mathrm{d}m_j. \tag*{(118)} ∫ M < Λ e Tr U ( M ) / T d M ∝ Δ ( λ ) 1 ∫ R N Δ ( m ) det [ ( λ i − m j ) + N − 1 ] i , j = 1 N j = 1 ∏ N e U ( m j ) / T d m j . (118)
Integration over unordered m m m introduces only a constant N ! N! N ! compared with the ordered convention. Although e U ( m ) / T e^{U(m)/T} e U ( m ) / T grows at + ∞ +\infty + ∞ , the truncated-power determinant vanishes when any m j > max i λ i m_j > \max_i \lambda_i m j > max i λ i ; at − ∞ -\infty − ∞ there is cubic exponential decay. Thus all integrals used here are convergent.
Using Δ ( m ) = det [ m k j ] k = 0 , … , r ; j = 1 , … , N \Delta(m)=\det[m_k^j]_{k=0,\ldots,r;\ j=1,\ldots,N} Δ ( m ) = det [ m k j ] k = 0 , … , r ; j = 1 , … , N , Andreief’s identity reduces (A11) to a determinant of scalar integrals. Including the eigenvalue Jacobian Δ ( λ ) 2 \Delta(\lambda)^2 Δ ( λ ) 2 of Λ \Lambda Λ , we obtain
P o r d ( λ ) ∝ Δ ( λ ) det [ g k ( λ i ) ] i = 1 , … , N ; k = 0 , … , r (119) P_{\mathrm{ord}}(\lambda) \propto\Delta(\lambda)\det[g_k(\lambda_i)]_{i=1,\ldots,N;\ k=0,\ldots,r} \tag*{(119)} P ord ( λ ) ∝ Δ ( λ ) det [ g k ( λ i ) ] i = 1 , … , N ; k = 0 , … , r (119)
where
g k ( x ) = e − U ( x ) / T ∫ − ∞ x ( x − t ) r t k e U ( t ) / T d t . (120) g_k(x)=e^{-U(x)/T}\int_{-\infty}^{x}(x-t)^r t^k e^{U(t)/T}\,\mathrm{d}t. \tag*{(120)} g k ( x ) = e − U ( x ) / T ∫ − ∞ x ( x − t ) r t k e U ( t ) / T d t . (120)
Consequently, only the identification of the right-function space remains.
An explicit intertwiner
Define a linear map on P r \mathcal{P}_r P r by
( K p ) ( x ) = e − U ( x ) / T ∫ − ∞ x ( x − t ) r p ( − t ) e U ( t ) / T d t . (121) (Kp)(x)=e^{-U(x)/T}\int_{-\infty}^{x}(x-t)^r p(-t)e^{U(t)/T}\,\mathrm{d}t. \tag*{(121)} ( K p ) ( x ) = e − U ( x ) / T ∫ − ∞ x ( x − t ) r p ( − t ) e U ( t ) / T d t . (121)
Thus K [ x k ] = h k = ( − 1 ) k g k K[x^k]=h_k=(-1)^k g_k K [ x k ] = h k = ( − 1 ) k g k . We claim the exact identity
L N K p = K L N † p , p ∈ P r . (122) \mathcal{L}_N Kp=K\mathcal{L}_N^\dagger p,\qquad p\in\mathcal{P}_r. \tag*{(122)} L N K p = K L N † p , p ∈ P r . (122)
For completeness we prove it in the monomial basis. For r ≥ 1 r\ge1 r ≥ 1 , introduce
I s , k ( x ) = ∫ − ∞ x ( x − t ) s t k e U ( t ) / T d t . (123) I_{s,k}(x)=\int_{-\infty}^{x}(x-t)^s t^k e^{U(t)/T}\,\mathrm{d}t. \tag*{(123)} I s , k ( x ) = ∫ − ∞ x ( x − t ) s t k e U ( t ) / T d t . (123)
The gauge identity
L N ( e − U / T F ) = e − U / T [ T F ′ ′ − ( E + x 2 ) F ′ + r x F ′ ] (124) \mathcal{L}_N(e^{-U/T}F)=e^{-U/T}\left[TF''-(E+x^2)F'+rxF'\right] \tag*{(124)} L N ( e − U / T F ) = e − U / T [ T F ′′ − ( E + x 2 ) F ′ + r x F ′ ] (124)
first yields, with endpoint terms included,
L N g k = r e − U / T [ T k I r − 1 , k − 1 − I r , k + 1 ] . (125) \mathcal{L}_N g_k=re^{-U/T}\left[TkI_{r-1,k-1}-I_{r,k+1}\right]. \tag*{(125)} L N g k = r e − U / T [ T k I r − 1 , k − 1 − I r , k + 1 ] . (125)
Indeed, integration of ∂ t [ ( x − t ) r − 1 t k e U ( t ) / T ] \partial_t[(x-t)^{r-1}t^k e^{U(t)/T}] ∂ t [( x − t ) r − 1 t k e U ( t ) / T ] replaces the term containing E + t 2 E+t^2 E + t 2 ; when r = 1 r=1 r = 1 its endpoint contribution cancels the endpoint contribution from F ′ ′ F'' F ′′ in (A17). For k ≥ 1 k\ge1 k ≥ 1 , integration of ∂ t [ ( x − t ) r t k − 1 e U ( t ) / T ] \partial_t[(x-t)^r t^{k-1}e^{U(t)/T}] ∂ t [( x − t ) r t k − 1 e U ( t ) / T ] gives
T r I r − 1 , k − 1 = I r , k + 1 + E I r , k − 1 + T ( k − 1 ) I r , k − 2 . (126) TrI_{r-1,k-1}=I_{r,k+1}+EI_{r,k-1}+T(k-1)I_{r,k-2}. \tag*{(126)} T r I r − 1 , k − 1 = I r , k + 1 + E I r , k − 1 + T ( k − 1 ) I r , k − 2 . (126)
For k = 0 k=0 k = 0 , the first term on the right of (A18) is absent. Hence, for 0 ≤ k ≤ r 0\le k\le r 0 ≤ k ≤ r ,
L N g k = ( k − r ) g k + 1 + E k g k − 1 + T k ( k − 1 ) g k − 2 , (127) \mathcal{L}_N g_k=(k-r)g_{k+1}+Ek g_{k-1}+Tk(k-1)g_{k-2}, \tag*{(127)} L N g k = ( k − r ) g k + 1 + E k g k − 1 + T k ( k − 1 ) g k − 2 , (127)
where terms with zero coefficient are omitted. In particular, no g r + 1 g_{r+1} g r + 1 is needed. Multiplying by ( − 1 ) k (-1)^k ( − 1 ) k gives
L N h k = ( r − k ) h k + 1 − E k h k − 1 + T k ( k − 1 ) h k − 2 , (128) \mathcal{L}_N h_k=(r-k)h_{k+1}-Ek h_{k-1}+Tk(k-1)h_{k-2}, \tag*{(128)} L N h k = ( r − k ) h k + 1 − E k h k − 1 + T k ( k − 1 ) h k − 2 , (128)
which is exactly the image under K K K of
L N † x k = ( r − k ) x k + 1 − E k x k − 1 + T k ( k − 1 ) x k − 2 . (129) \mathcal{L}_N^\dagger x^k=(r-k)x^{k+1}-Ekx^{k-1}+Tk(k-1)x^{k-2}. \tag*{(129)} L N † x k = ( r − k ) x k + 1 − E k x k − 1 + T k ( k − 1 ) x k − 2 . (129)
For N = 1 N=1 N = 1 , K 1 = e − U / T ∫ − ∞ x e U ( t ) / T d t K1=e^{-U/T}\int_{-\infty}^{x}e^{U(t)/T}\,\mathrm{d}t K 1 = e − U / T ∫ − ∞ x e U ( t ) / T d t and direct differentiation gives L 1 K 1 = 0 \mathcal{L}_1K1=0 L 1 K 1 = 0 . This completes the proof of (A15) for every N N N .
Identification of the physical modes
Let p a ( x ) = c a x r + ⋯ p_a(x)=c_a x^r+\cdots p a ( x ) = c a x r + ⋯ be a nonzero polynomial eigenfunction. Its degree is exactly r r r : otherwise the raising term in (A4) would contradict the eigenvalue equation. By (A15),
w a = K p a , L N w a = ϵ a w a . (130) w_a=\mathcal{K}p_a,\qquad\mathcal{L}_N w_a=\epsilon_a w_a. \tag*{(130)} w a = K p a , L N w a = ϵ a w a . (130)
Hence each left eigenfunction p a p_a p a is carried into a right eigenfunction w a w_a w a with the same eigenvalue. Since the integrand in (A14) is smooth and decays like e t 3 / 3 T e^{t^3/3T} e t 3 /3 T as t → − ∞ t\to-\infty t → − ∞ , w a w_a w a is regular on the whole real line, with no apparent poles at the zeros of p a p_a p a . It remains to check that w a w_a w a is an admissible right mode, i.e. that it carries a nonzero Lagrange current; this follows from its behaviour at the two real infinities.
Endpoint asymptotics of (A14), with x − t x-t x − t of order T / x 2 T/x^2 T / x 2 , give at both real infinities
w a ( x ) = ( − 1 ) r c a r ! T r + 1 x − r − 2 ( 1 + O ( x − 1 ) ) . (131) w_a(x)=(-1)^r c_a r!T^{r+1}x^{-r-2}\left(1+O(x^{-1})\right). \tag*{(131)} w a ( x ) = ( − 1 ) r c a r ! T r + 1 x − r − 2 ( 1 + O ( x − 1 ) ) . (131)
At + ∞ +\infty + ∞ , portions of the integral away from the endpoint are exponentially subleading. The analogous expansion holds at − ∞ -\infty − ∞ . These expansions can be differentiated and extended to arbitrary algebraic order.
The Lagrange identity is
∂ x J [ p , w ] = p L N w − w L N † p . (132) \partial_x J[p,w]=p\mathcal{L}_Nw-w\mathcal{L}_N^\dagger p. \tag*{(132)} ∂ x J [ p , w ] = p L N w − w L N † p . (132)
Thus J [ p a , w a ] J[p_a,w_a] J [ p a , w a ] is constant, and (A24) evaluates it explicitly:
J [ p a , w a ] = ( − 1 ) r c a 2 r ! T r + 1 ≠ 0. (133) J[p_a,w_a]=(-1)^r c_a^2r!T^{r+1}\ne0. \tag*{(133)} J [ p a , w a ] = ( − 1 ) r c a 2 r ! T r + 1 = 0. (133)
Define
v ^ a = K p a ( − 1 ) r c a 2 r ! T r + 1 . (134) \widehat{v}_a=\frac{\mathcal{K}p_a}{(-1)^r c_a^2r!T^{r+1}}. \tag*{(134)} v a = ( − 1 ) r c a 2 r ! T r + 1 K p a . (134)
Then J [ p a , v ^ a ] = 1 J[p_a,\widehat{v}_a]=1 J [ p a , v a ] = 1 . On any interval free of zeros of p a p_a p a , solving this first-order equation gives precisely (A7), up to addition of a constant multiple of e − U / T p a e^{-U/T}p_a e − U / T p a . That homogeneous term grows exponentially at − ∞ -\infty − ∞ , whereas (A24) is algebraic. Its coefficient is therefore zero for the physical solution selected at − ∞ -\infty − ∞ . Below the lowest real zero of p a p_a p a the integral in (A7) is an ordinary one, and v ^ a \widehat{v}_a v a coincides with the right mode of [46 ] identically. Beyond that zero, (A14) provides the regular continuation, so the finite-part prescription of [46 ] is no longer needed: it is replaced by the ordinary convergent integral (A14).
This argument is a scalar differential-equation identification, not an appeal to uniqueness of the stationary diffusion. It also shows what the integral representation (A14) adds to the form (A7) which appeared in [46 ] . The two describe the same modes, but the integrand in (A14) is regular, so polynomial zeros cause no difficulty, and (A14) supplies the regular meaning of the finite part in (A7) at those zeros. Moreover, K \mathcal{K} K is linear in p p p : a single operator carries the whole of P r \mathcal{P}_r P r onto the space of right modes, with K [ x k ] = h k \mathcal{K}[x^k]=h_k K [ x k ] = h k . This is what turns the determinant of right modes into the polynomial-ensemble form used below without diagonalizing L N † \mathcal{L}_N^\dagger L N † , and it keeps that form well defined at parameter values where diagonalization fails. The intertwining relation (A15) shows directly that each left eigenfunction is carried into a right eigenfunction with the same eigenvalue, and the algebraic tails (A24), hence the value (A26) of the current, follow from the endpoint expansion of a single integral. The form in [46 ] , in turn, displays the unit Lagrange current (A6) mode by mode.
Equality of the normalized densities
Write p a ( x ) = ∑ k = 0 r C a k x k p_a(x)=\sum_{k=0}^r C_{ak}x^k p a ( x ) = ∑ k = 0 r C ak x k . For a polynomial eigenbasis C C C is invertible, so
det [ p a ( λ i ) ] = det C Δ ( λ ) , det [ ( K p a ) ( λ i ) ] = det C det [ h k ( λ i ) ] k , i . (135) \det[p_a(\lambda_i)]=\det C\,\Delta(\lambda),\qquad
\det[(\mathcal{K}p_a)(\lambda_i)]=\det C\,\det[h_k(\lambda_i)]_{k,i}. \tag*{(135)} det [ p a ( λ i )] = det C Δ ( λ ) , det [( K p a ) ( λ i )] = det C det [ h k ( λ i ) ] k , i . (135)
Replacing g k g_k g k by h k = ( − 1 ) k g k h_k = (-1)^k g_k h k = ( − 1 ) k g k , and then K p a \mathcal{K}_{pa} K p a by the current-normalized modes (A27), changes determinants only by λ \lambda λ -independent constants. Equations (A12) and (A27) therefore prove (A9) up to an overall constant, which is fixed to unity by normalization on the same ordered chamber. At exceptional parameter values where diagonalization of the finite polynomial operator fails, (A12) remains well-defined and equality extends from generic parameters by continuity. Since both matrix distributions are conjugation invariant, equality of their joint eigenvalue densities implies equality of their full matrix probability densities (with the unordered symmetric density obtained by dividing by N ! N! N ! ). For N = 1 N=1 N = 1 , the construction reduces immediately to P ( x ) ∝ e − U ( x ) / T ∫ − ∞ x e U ( t ) / T d t P(x) \propto e^{-U(x)/T}\int_{-\infty}^{x}e^{U(t)/T}\,dt P ( x ) ∝ e − U ( x ) / T ∫ − ∞ x e U ( t ) / T d t , the scalar Halperin density.
Exact reduction of the even-N N N orthogonal ensemble to unitary recurrence coefficients
We give a direct proof of the finite-N N N reduction for β = 1 \beta=1 β = 1 and even N N N . The result uses the unitary orthogonal polynomials for the squared one-particle weight and one explicitly determined Airy primitive. The central step is a finite change of polynomial basis which turns the ordered-integral skew matrix into a five-diagonal matrix. Its Pfaffian then obeys the same three-term recursion as in the symplectic calculation, with different initial values.
Conventions and statement of the result
We work in the dimensionless variable w = − λ / T 1 / 3 w=-\lambda/T^{1/3} w = − λ / T 1/3 and set
s = − E / T 2 / 3 , ρ s ( w ) = exp ( w 3 / 3 − s w ) . (136) s=-E/T^{2/3},\qquad\rho_s(w)=\exp(w^3/3-sw). \tag*{(136)} s = − E / T 2/3 , ρ s ( w ) = exp ( w 3 /3 − s w ) . (136)
The outgoing contour C − C_- C − is oriented from e i π / 3 ∞ e^{i\pi/3}\infty e iπ /3 ∞ to − ∞ -\infty − ∞ and may be taken as the two straight rays joined at the origin. The symbol ≺ \prec ≺ refers to order along this oriented contour. To keep the normalization unambiguous, we use the holomorphic ordered integral
τ N ( 1 ) ( s ) = ∫ w 1 ≺ ⋯ ≺ w N ∏ i < j ( w j − w i ) ∏ j = 1 N ρ s ( w j ) d w j . (137) \tau_N^{(1)}(s)=\int_{w_1\prec\cdots\prec w_N}\prod_{i<j}(w_j-w_i)\prod_{j=1}^{N}\rho_s(w_j)\,dw_j. \tag*{(137)} τ N ( 1 ) ( s ) = ∫ w 1 ≺ ⋯ ≺ w N i < j ∏ ( w j − w i ) j = 1 ∏ N ρ s ( w j ) d w j . (137)
This is the holomorphic partition function entering the characteristic exponent, not the real normalization proportional to its squared modulus.
Let p n ( w ) = w n + ⋯ p_n(w)=w^n+\cdots p n ( w ) = w n + ⋯ be the monic polynomials satisfying the bilinear, without complex conjugation, orthogonality relations
( p j , p k ) 2 : = ∫ C − p j ( w ) p k ( w ) ρ s ( w ) 2 d w = h ~ j δ j k . (138) (p_j,p_k)_2:=\int_{C_-}p_j(w)p_k(w)\rho_s(w)^2\,dw=\widetilde h_j\delta_{jk}. \tag*{(138)} ( p j , p k ) 2 := ∫ C − p j ( w ) p k ( w ) ρ s ( w ) 2 d w = h j δ j k . (138)
w p n = p n + 1 + d ~ n p n + r ~ n p n − 1 , r ~ n = h ~ n h ~ n − 1 , r ~ 0 = 0. (139) wp_n=p_{n+1}+\widetilde d_n p_n+\widetilde r_n p_{n-1},\qquad\widetilde r_n=\frac{\widetilde h_n}{\widetilde h_{n-1}},\qquad\widetilde r_0=0. \tag*{(139)} w p n = p n + 1 + d n p n + r n p n − 1 , r n = h n − 1 h n , r 0 = 0. (139)
These are precisely the unitary recurrence coefficients at temperature T / 2 T/2 T /2 before the dimensionless change of variables. In this appendix all tilded quantities refer to the w w w variable. We abbreviate
B n = d ~ n + d ~ n + 1 . (140) B_n=\widetilde d_n+\widetilde d_{n+1}. \tag*{(140)} B n = d n + d n + 1 . (140)
Exact identity. For N = 2 m ≥ 2 N=2m\ge2 N = 2 m ≥ 2 , with an empty product interpreted as one,
τ 2 m ( 1 ) ( s ) = ( − 2 ) m [ ∏ j = 0 m − 2 h ~ 2 j + 1 ( s ) ] X m − 1 ( s ) . (141) \tau_{2m}^{(1)}(s)=(-2)^m\left[\prod_{j=0}^{m-2}\widetilde h_{2j+1}(s)\right]X_{m-1}(s). \tag*{(141)} τ 2 m ( 1 ) ( s ) = ( − 2 ) m [ j = 0 ∏ m − 2 h 2 j + 1 ( s ) ] X m − 1 ( s ) . (141)
The sequence X k X_k X k is specified by
X 0 ( s ) = i π 3 + π 2 ∫ 0 2 2 / 3 s [ Bi ( u ) + i Ai ( u ) ] d u , (142) X_0(s)=\frac{\mathrm{i}\pi}{3}+\frac{\pi}{2}\int_{0}^{2^{2/3}s}[\operatorname{Bi}(u)+\mathrm{i}\operatorname{Ai}(u)]\,\mathrm{d}u, \tag*{(142)} X 0 ( s ) = 3 i π + 2 π ∫ 0 2 2/3 s [ Bi ( u ) + i Ai ( u )] d u , (142)
X 1 = B 0 X 0 − h ~ 0 , (143) X_1=B_0X_0-\widetilde h_0, \tag*{(143)} X 1 = B 0 X 0 − h 0 , (143)
X k + 1 = B 2 k X k − r ~ 2 k X k − 1 , k ≥ 1. (144) X_{k+1}=B_{2k}X_k-\widetilde r_{2k}X_{k-1},\qquad k\ge1. \tag*{(144)} X k + 1 = B 2 k X k − r 2 k X k − 1 , k ≥ 1. (144)
The integral in (B7) is path independent for complex s s s , since its integrand is entire. The factor ( − 2 ) m (-2)^m ( − 2 ) m in (B6) is fixed by the orientation and ordered-integral convention in (B2). Dropping this s s s -independent factor gives the proportionality formula in the main text.
The corresponding characteristic exponent is
Ξ 2 m ( 1 ) ( E ) T 1 / 3 = − 2 ∑ j = 0 m − 2 d ~ 2 j + 1 ( s ) + ∂ s X m − 1 ( s ) X m − 1 ( s ) , s = − E / T 2 / 3 . (145) \frac{\Xi_{2m}^{(1)}(E)}{T^{1/3}}=-2\sum_{j=0}^{m-2}\widetilde d_{2j+1}(s)+\frac{\partial_sX_{m-1}(s)}{X_{m-1}(s)},\qquad s=-E/T^{2/3}. \tag*{(145)} T 1/3 Ξ 2 m ( 1 ) ( E ) = − 2 j = 0 ∑ m − 2 d 2 j + 1 ( s ) + X m − 1 ( s ) ∂ s X m − 1 ( s ) , s = − E / T 2/3 . (145)
We first prove the partition-function identity and then derive (B10).
The orthogonal-polynomial argument is made initially where the finitely many required Hankel determinants are nonzero. Because the contour integrals are entire in s s s , the complete identity extends across isolated singularities of the recurrence representation by analytic continuation. The logarithmic derivative is understood away from zeros of τ 2 m ( 1 ) \tau_{2m}^{(1)} τ 2 m ( 1 ) .
Ordered integrals and the integration-by-parts identity
For polynomials f , g f,g f , g , define the skew form
[ f , g ] 1 = ∫ x < y [ f ( x ) g ( y ) − g ( x ) f ( y ) ] ρ s ( x ) ρ s ( y ) d x d y . (146) [f,g]_1=\int_{x<y}[f(x)g(y)-g(x)f(y)]\rho_s(x)\rho_s(y)\,\mathrm{d}x\,\mathrm{d}y. \tag*{(146)} [ f , g ] 1 = ∫ x < y [ f ( x ) g ( y ) − g ( x ) f ( y )] ρ s ( x ) ρ s ( y ) d x d y . (146)
In particular, [ f , g ] 1 = − [ g , f ] 1 [f,g]_1=-[g,f]_1 [ f , g ] 1 = − [ g , f ] 1 . The ordered version of de Bruijn’s formula [19 ] gives
τ 2 m ( 1 ) = Pf ( [ p j , p k ] 1 ) 0 ≤ j , k < 2 m . (147) \tau_{2m}^{(1)}=\operatorname{Pf}\left([p_j,p_k]_1\right)_{0\le j,k<2m}. \tag*{(147)} τ 2 m ( 1 ) = Pf ( [ p j , p k ] 1 ) 0 ≤ j , k < 2 m . (147)
There is no factorial in this convention: the determinant of the monic polynomials evaluated at the ordered variables is ∏ i < j ( w j − w i ) \prod_{i<j}(w_j-w_i) ∏ i < j ( w j − w i ) . Replacing the monomials by any other monic basis of successive degrees multiplies that determinant, and the associated Pfaffian, by one.
Introduce the weighted derivative
D f = f ′ + ( w 2 − s ) f = ρ s − 1 ( f ρ s ) ′ . (148) Df=f'+(w^2-s)f=\rho_s^{-1}(f\rho_s)'. \tag*{(148)} D f = f ′ + ( w 2 − s ) f = ρ s − 1 ( f ρ s ) ′ . (148)
Its essential property is
[ D f , g ] 1 = 2 ( f , g ) 2 . (149) [Df,g]_1=2(f,g)_2. \tag*{(149)} [ D f , g ] 1 = 2 ( f , g ) 2 . (149)
To verify both the sign and the boundary terms, write
F f ( y ) = ∫ C s t a r t y f ( x ) ρ s ( x ) d x , M f = ∫ C − f ( x ) ρ s ( x ) d x . F_f(y)=\int_{C_{\mathrm{start}}}^{y}f(x)\rho_s(x)\,\mathrm{d}x,\qquad M_f=\int_{C_-}f(x)\rho_s(x)\,\mathrm{d}x. F f ( y ) = ∫ C start y f ( x ) ρ s ( x ) d x , M f = ∫ C − f ( x ) ρ s ( x ) d x .
Splitting the ordered double integral in (B11) then gives
[ f , g ] 1 = ∫ C − [ 2 F f ( y ) − M f ] g ( y ) ρ s ( y ) d y . (150) [f,g]_1=\int_{C_-}[2F_f(y)-M_f]g(y)\rho_s(y)\,\mathrm{d}y. \tag*{(150)} [ f , g ] 1 = ∫ C − [ 2 F f ( y ) − M f ] g ( y ) ρ s ( y ) d y . (150)
Since a polynomial times ρ s \rho_s ρ s vanishes at both endpoints, F D f ( y ) = f ( y ) ρ s ( y ) F_{Df}(y)=f(y)\rho_s(y) F D f ( y ) = f ( y ) ρ s ( y ) and M D f = 0 M_{Df}=0 M D f = 0 . Substitution proves (B14). All these integrals converge absolutely on the two rays, locally uniformly in s s s . If integration is performed separately on the rays, the contributions at their common endpoint cancel. In particular, there is no extra term at the corner of the contour. Skew-symmetry also gives
[ f , D g ] 1 = − 2 ( f , g ) 2 . (151) [f,Dg]_1=-2(f,g)_2. \tag*{(151)} [ f , D g ] 1 = − 2 ( f , g ) 2 . (151)
A finite basis change and the five-diagonal matrix
Use the following single sequence of monic polynomials:
q 0 = p 0 = 1 , q 1 = p 1 , q n + 2 = D p n ( n ≥ 0 ) . (152) q_{0}=p_{0}=1,\qquad q_{1}=p_{1},\qquad q_{n+2}=\mathcal{D}p_{n}\quad(n\geq0). \tag*{(152)} q 0 = p 0 = 1 , q 1 = p 1 , q n + 2 = D p n ( n ≥ 0 ) . (152)
The degree of D p n \mathcal{D}p_{n} D p n is n + 2 n+2 n + 2 , with leading coefficient one. Thus for every finite 2 m 2m 2 m the transformation from p 0 , … , p 2 m − 1 p_{0},\ldots,p_{2m-1} p 0 , … , p 2 m − 1 to q 0 , … , q 2 m − 1 q_{0},\ldots,q_{2m-1} q 0 , … , q 2 m − 1 is unitriangular. It is a finite transformation: its last member uses only D p 2 m − 3 \mathcal{D}p_{2m-3} D p 2 m − 3 , whose degree is 2 m − 1 2m-1 2 m − 1 . Consequently, with
C j k = [ q j , q k ] 1 , P m = Pf [ C j k ] 0 ≤ j , k < 2 m , P 0 = 1 , C_{jk}=[q_{j},q_{k}]_{1},\qquad P_{m}=\operatorname{Pf}[C_{jk}]_{0\leq j,k<2m},\qquad P_{0}=1, C j k = [ q j , q k ] 1 , P m = Pf [ C j k ] 0 ≤ j , k < 2 m , P 0 = 1 ,
we have P m = τ 2 m ( 1 ) P_{m}=\tau_{2m}^{(1)} P m = τ 2 m ( 1 ) .
Set
A ( s ) = [ 1 , w ] 1 = [ p 0 , p 1 ] 1 . (153) A(s)=[1,w]_{1}=[p_{0},p_{1}]_{1}. \tag*{(153)} A ( s ) = [ 1 , w ] 1 = [ p 0 , p 1 ] 1 . (153)
The first two rows follow immediately from (B16):
C 01 = A , C 0 , n + 2 = − 2 h ~ 0 δ n 0 , C 1 , n + 2 = − 2 h ~ 1 δ n 1 . (154) C_{01}=A,\qquad C_{0,n+2}=-2\widetilde{h}_{0}\delta_{n0},\qquad C_{1,n+2}=-2\widetilde{h}_{1}\delta_{n1}. \tag*{(154)} C 01 = A , C 0 , n + 2 = − 2 h 0 δ n 0 , C 1 , n + 2 = − 2 h 1 δ n 1 . (154)
For the remaining entries, (B14) gives
C i + 2 , j + 2 = [ D p i , D p j ] 1 = 2 ( p i , D p j ) 2 = − 2 ( D p i , p j ) 2 . (155) C_{i+2,j+2}=[\mathcal{D}p_{i},\mathcal{D}p_{j}]_{1}=2(p_{i},\mathcal{D}p_{j})_{2}=-2(\mathcal{D}p_{i},p_{j})_{2}. \tag*{(155)} C i + 2 , j + 2 = [ D p i , D p j ] 1 = 2 ( p i , D p j ) 2 = − 2 ( D p i , p j ) 2 . (155)
The last equality follows directly by integrating ( p i p j ρ s 2 ) ′ (p_{i}p_{j}\rho s^{2})' ( p i p j ρ s 2 ) ′ :
( p i , D p j ) 2 + ( D p i , p j ) 2 = 0. (p_{i},\mathcal{D}p_{j})_{2}+(\mathcal{D}p_{i},p_{j})_{2}=0. ( p i , D p j ) 2 + ( D p i , p j ) 2 = 0.
It holds on the complex contour because the orthogonality is bilinear and the endpoint terms vanish.
Applying (B4) twice yields
D p i = p i + 2 + B i p i + 1 + a polynomial of degree at most i . (156) \mathcal{D}p_{i}=p_{i+2}+B_{i}p_{i+1}+\text{a polynomial of degree at most }i. \tag*{(156)} D p i = p i + 2 + B i p i + 1 + a polynomial of degree at most i . (156)
For j > i j>i j > i , orthogonality therefore determines every entry of (B20):
C i + 2 , j + 2 = { − 2 h ~ i + 1 B i , j = i + 1 , − 2 h ~ i + 2 , j = i + 2 , 0 , j ≥ i + 3. (157) C_{i+2,j+2}=
\begin{cases}
-2\widetilde{h}_{i+1}B_{i}, & j=i+1,\\
-2\widetilde{h}_{i+2}, & j=i+2,\\
0, & j\geq i+3.
\end{cases} \tag*{(157)} C i + 2 , j + 2 = ⎩ ⎨ ⎧ − 2 h i + 1 B i , − 2 h i + 2 , 0 , j = i + 1 , j = i + 2 , j ≥ i + 3. (157)
The lower triangle is fixed by skew-symmetry. Equivalently, the only nonzero upper-triangular entries are
C 01 = A , C 12 = 0 , C_{01}=A,\qquad C_{12}=0, C 01 = A , C 12 = 0 ,
C k , k + 1 = − 2 h ~ k − 1 B k − 2 ( k ≥ 2 ) , C k , k + 2 = − 2 h ~ k ( k ≥ 0 ) . (158) C_{k,k+1}=-2\widetilde{h}_{k-1}B_{k-2}\quad(k\geq2),\qquad C_{k,k+2}=-2\widetilde{h}_{k}\quad(k\geq0). \tag*{(158)} C k , k + 1 = − 2 h k − 1 B k − 2 ( k ≥ 2 ) , C k , k + 2 = − 2 h k ( k ≥ 0 ) . (158)
This is the required five-diagonal matrix. Its exceptional first rows are explicit in (B19); they will produce the inhomogeneous term in the initial value X 1 X_{1} X 1 .
Pfaffian expansion and the three-term recursion
For clarity, consider first any skew matrix with C j k = 0 C_{jk}=0 C j k = 0 for ∣ j − k ∣ > 2 \lvert j-k\rvert>2 ∣ j − k ∣ > 2 . In the last-row expansion of its 2 m × 2 m 2m\times2m 2 m × 2 m Pfaffian, the final index 2 m − 1 2m-1 2 m − 1 can be paired only with 2 m − 2 2m-2 2 m − 2 or 2 m − 3 2m-3 2 m − 3 . The first choice leaves the leading ( 2 m − 2 ) × ( 2 m − 2 ) (2m-2)\times(2m-2) ( 2 m − 2 ) × ( 2 m − 2 ) Pfaffian. In the second choice the remaining index 2 m − 2 2m-2 2 m − 2 must pair with 2 m − 4 2m-4 2 m − 4 , leaving the leading ( 2 m − 4 ) × ( 2 m − 4 ) (2m-4)\times(2m-4) ( 2 m − 4 ) × ( 2 m − 4 ) Pfaffian. The two pairings ( 2 m − 4 , 2 m − 2 ) (2m-4,2m-2) ( 2 m − 4 , 2 m − 2 ) and ( 2 m − 3 , 2 m − 1 ) (2m-3,2m-1) ( 2 m − 3 , 2 m − 1 ) cross in the ordered index list and contribute with a minus sign. Hence
P m = C 2 m − 2 , 2 m − 1 P m − 1 − C 2 m − 4 , 2 m − 2 C 2 m − 3 , 2 m − 1 P m − 2 , m ≥ 2. (159) P_m=C_{2m-2,2m-1}P_{m-1}-C_{2m-4,2m-2}C_{2m-3,2m-1}P_{m-2},\qquad m\ge2. \tag*{(159)} P m = C 2 m − 2 , 2 m − 1 P m − 1 − C 2 m − 4 , 2 m − 2 C 2 m − 3 , 2 m − 1 P m − 2 , m ≥ 2. (159)
For m = 2 m=2 m = 2 this is simply Pf C = C 01 C 23 − C 02 C 13 + C 03 C 12 \operatorname{Pf}C=C_{01}C_{23}-C_{02}C_{13}+C_{03}C_{12} Pf C = C 01 C 23 − C 02 C 13 + C 03 C 12 , with C 03 = 0 C_{03}=0 C 03 = 0 .
Inserting (B23) gives
P m = − 2 h ~ 2 m − 3 B 2 m − 4 P m − 1 − 4 h ~ 2 m − 4 h ~ 2 m − 3 P m − 2 , P 0 = 1 , P 1 = A . (160) P_m=-2\widetilde{h}_{2m-3}B_{2m-4}P_{m-1}-4\widetilde{h}_{2m-4}\widetilde{h}_{2m-3}P_{m-2},\qquad P_0=1,\qquad P_1=A. \tag*{(160)} P m = − 2 h 2 m − 3 B 2 m − 4 P m − 1 − 4 h 2 m − 4 h 2 m − 3 P m − 2 , P 0 = 1 , P 1 = A . (160)
In particular,
P 2 = − 2 h ~ 1 B 0 A − 4 h ~ 0 h ~ 1 . P_2=-2\widetilde{h}_1B_0A-4\widetilde{h}_0\widetilde{h}_1. P 2 = − 2 h 1 B 0 A − 4 h 0 h 1 .
Define X 0 = − A / 2 X_0=-A/2 X 0 = − A /2 and, for m ≥ 1 m\ge1 m ≥ 1 , factor out the norms by
P m = ( − 2 ) m [ ∏ j = 0 m − 2 h ~ 2 j + 1 ] X m − 1 . (161) P_m=(-2)^m\left[\prod_{j=0}^{m-2}\widetilde{h}_{2j+1}\right]X_{m-1}. \tag*{(161)} P m = ( − 2 ) m [ j = 0 ∏ m − 2 h 2 j + 1 ] X m − 1 . (161)
The displayed formula for P 2 P_2 P 2 gives X 1 = B 0 X 0 − h ~ 0 X_1=B_0X_0-\widetilde{h}_0 X 1 = B 0 X 0 − h 0 . For m ≥ 3 m\ge3 m ≥ 3 , division of (B25) by the factor in (B26) gives
X m − 1 = B 2 m − 4 X m − 2 − h ~ 2 m − 4 h ~ 2 m − 5 X m − 3 . X_{m-1}=B_{2m-4}X_{m-2}-\frac{\widetilde{h}_{2m-4}}{\widetilde{h}_{2m-5}}X_{m-3}. X m − 1 = B 2 m − 4 X m − 2 − h 2 m − 5 h 2 m − 4 X m − 3 .
Putting k = m − 2 k=m-2 k = m − 2 proves (B9). Thus the partition-function reduction is already established, apart from the explicit evaluation of X 0 X_0 X 0 .
The contour fixes the Airy primitive
First differentiate A = [ 1 , w ] 1 A=[1,w]_1 A = [ 1 , w ] 1 with respect to s s s . Each derivative of a weight brings down minus its integration variable. Therefore
A ′ ( s ) = − [ w , w ] 1 − [ 1 , w 2 ] 1 = − [ 1 , w 2 ] 1 . (162) A'(s)=-[w,w]_1-[1,w^2]_1=-[1,w^2]_1. \tag*{(162)} A ′ ( s ) = − [ w , w ] 1 − [ 1 , w 2 ] 1 = − [ 1 , w 2 ] 1 . (162)
Since D 1 = w 2 − s \mathcal{D}_1=w^2-s D 1 = w 2 − s , (B16) gives [ 1 , w 2 ] 1 = − 2 h ~ 0 [1,w^2]_1=-2\widetilde{h}_0 [ 1 , w 2 ] 1 = − 2 h 0 . Hence
A ′ = 2 h ~ 0 , X 0 ′ = − h ~ 0 . (163) A'=2\widetilde{h}_0,\qquad X'_0=-\widetilde{h}_0. \tag*{(163)} A ′ = 2 h 0 , X 0 ′ = − h 0 . (163)
Rescaling the squared-weight integral by u = 2 1 / 3 w u=2^{1/3}w u = 2 1/3 w and using the outgoing Airy contour gives
h ~ 0 ( s ) = − π 2 1 / 3 [ Bi ( 2 2 / 3 s ) + i Ai ( 2 2 / 3 s ) ] . (164) \widetilde{h}_0(s)=-\frac{\pi}{2^{1/3}}\left[\operatorname{Bi}(2^{2/3}s)+i\operatorname{Ai}(2^{2/3}s)\right]. \tag*{(164)} h 0 ( s ) = − 2 1/3 π [ Bi ( 2 2/3 s ) + i Ai ( 2 2/3 s ) ] . (164)
The minus sign is fixed by the orientation from e i π / 3 ∞ e^{i\pi/3}\infty e iπ /3 ∞ to − ∞ -\infty − ∞ ; the usual Airy integral representations give this same convention [74 Sec. 9.5] . Equation (B28) now determines X 0 X_0 X 0 up to one constant. We compute that constant from the ordered integral itself, rather than impose a normalization afterwards.
At s = 0 s=0 s = 0 put a = e i π / 3 a=e^{i\pi/3} a = e iπ /3 and b = − 1 b=-1 b = − 1 . The two parts of the contour are w = a t w=at w = a t with t t t decreasing from infinity to zero and w = b t w=bt w = b t with t t t increasing from zero to infinity. In
A ( 0 ) = ∫ x < y ( y − x ) e ( x 3 + y 3 ) / 3 d x d y , A(0)=\int_{x<y}(y-x)e^{(x^3+y^3)/3}\,dx\,dy, A ( 0 ) = ∫ x < y ( y − x ) e ( x 3 + y 3 ) /3 d x d y ,
the contributions with both variables on the same ray cancel: both have cubic phase − t 3 / 3 -t^3/3 − t 3 /3 , while reversing their radial order reverses the factor y − x y-x y − x . For the remaining contribution, x x x is on the incoming ray and y y y is on the outgoing ray. Define
J ℓ = ∫ 0 ∞ t ℓ e − t 3 / 3 d t = 3 ( ℓ − 2 ) / 3 Γ ( ℓ + 1 3 ) . J_\ell=\int_0^\infty t^\ell e^{-t^3/3}\,dt=3^{(\ell-2)/3}\Gamma\left(\frac{\ell+1}{3}\right). J ℓ = ∫ 0 ∞ t ℓ e − t 3 /3 d t = 3 ( ℓ − 2 ) /3 Γ ( 3 ℓ + 1 ) .
Including the incoming-ray orientation, the cross contribution is
A ( 0 ) = − a b ∫ 0 ∞ ∫ 0 ∞ ( b v − a u ) e − ( u 3 + v 3 ) / 3 d u d v = a b ( a − b ) J 0 J 1 = − i 3 Γ ( 1 / 3 ) Γ ( 2 / 3 ) 3 = − 2 i π 3 . (165) A(0)=-ab\int_{0}^{\infty}\int_{0}^{\infty}(bv-au)e^{-(u^{3}+v^{3})/3}\,du\,dv
=ab(a-b)J_{0}J_{1}=-i\sqrt{3}\frac{\Gamma(1/3)\Gamma(2/3)}{3}=-\frac{2i\pi}{3}. \tag*{(165)} A ( 0 ) = − ab ∫ 0 ∞ ∫ 0 ∞ ( b v − a u ) e − ( u 3 + v 3 ) /3 d u d v = ab ( a − b ) J 0 J 1 = − i 3 3 Γ ( 1/3 ) Γ ( 2/3 ) = − 3 2 iπ . (165)
Thus X 0 ( 0 ) = i π / 3 X_{0}(0)=i\pi/3 X 0 ( 0 ) = iπ /3 . Integration of (B28), using (B29), proves (B7), including its additive constant.
For real s s s , the standard oscillatory integrals
∫ − ∞ 0 Ai ( u ) d u = 2 3 , ∫ − ∞ 0 Bi ( u ) d u = 0 \int_{-\infty}^{0}\operatorname{Ai}(u)\,du=\frac{2}{3},\qquad\int_{-\infty}^{0}\operatorname{Bi}(u)\,du=0 ∫ − ∞ 0 Ai ( u ) d u = 3 2 , ∫ − ∞ 0 Bi ( u ) d u = 0
[74 Sec. 9.10(iv)] give the equivalent form
X 0 ( s ) = − ∫ − ∞ s h ~ 0 ( t ) d t , lim s → − ∞ X 0 ( s ) = 0. (166) X_{0}(s)=-\int_{-\infty}^{s}\widetilde{h}_{0}(t)\,dt,\qquad\lim_{s\to-\infty}X_{0}(s)=0. \tag*{(166)} X 0 ( s ) = − ∫ − ∞ s h 0 ( t ) d t , s → − ∞ lim X 0 ( s ) = 0. (166)
The improper integral is conditionally convergent. Its oscillatory Airy asymptotics give X 0 ( s ) = O ( ∣ s ∣ − 3 / 4 ) X_{0}(s)=O(|s|^{-3/4}) X 0 ( s ) = O ( ∣ s ∣ − 3/4 ) as s → − ∞ s\to-\infty s → − ∞ . The finite-base-point formula (B7) specifies its entire continuation for complex s s s .
Characteristic exponent from the unitary recurrence coefficients
Differentiating the norm in (B3) gives
∂ s h ~ n = 2 ∫ p n ( ∂ s p n ) ρ s 2 d w − 2 ∫ w p n 2 ρ s 2 d w . \partial_{s}\widetilde{h}_{n}=2\int p_{n}(\partial_{s}p_{n})\rho_{s}^{2}\,dw-2\int wp_{n}^{2}\rho_{s}^{2}\,dw. ∂ s h n = 2 ∫ p n ( ∂ s p n ) ρ s 2 d w − 2 ∫ w p n 2 ρ s 2 d w .
The first integral vanishes because ∂ s p n \partial_{s}p_{n} ∂ s p n has degree at most n − 1 n-1 n − 1 ; the recurrence (B4) evaluates the second. Consequently,
∂ s log h ~ n = − 2 d ~ n . (167) \partial_{s}\log\widetilde{h}_{n}=-2\widetilde{d}_{n}. \tag*{(167)} ∂ s log h n = − 2 d n . (167)
Using Ξ 2 m ( 1 ) / T 1 / 3 = ∂ s log τ ~ 2 m ( 1 ) \Xi_{2m}^{(1)}/T^{1/3}=\partial_{s}\log\widetilde{\tau}_{2m}^{(1)} Ξ 2 m ( 1 ) / T 1/3 = ∂ s log τ 2 m ( 1 ) in (B6) now proves (B10).
For completeness, all recurrence coefficients needed in this formula can be generated from (B29) by the unitary string equations in the present dimensionless convention:
d ~ n 2 + r ~ n + r ~ n + 1 = s , \widetilde{d}_{n}^{2}+\widetilde{r}_{n}+\widetilde{r}_{n+1}=s, d n 2 + r n + r n + 1 = s ,
r ~ n ( d ~ n + d ~ n − 1 ) = − n 2 , n ≥ 1. (168) \widetilde{r}_{n}(\widetilde{d}_{n}+\widetilde{d}_{n-1})=-\frac{n}{2},\qquad n\geq1. \tag*{(168)} r n ( d n + d n − 1 ) = − 2 n , n ≥ 1. (168)
The first follows by integrating ( p n 2 ρ s 2 ) ′ (p_{n}^{2}\rho_{s}^{2})' ( p n 2 ρ s 2 ) ′ : the derivative of p n p_{n} p n is orthogonal to p n p_{n} p n , and ( ρ s 2 ) ′ = 2 ( w 2 − s ) ρ s 2 (\rho_{s}^{2})'=2(w^{2}-s)\rho_{s}^{2} ( ρ s 2 ) ′ = 2 ( w 2 − s ) ρ s 2 . For the second, integration of ( p n p n − 1 ρ s 2 ) ′ (p_{n}p_{n-1}\rho_{s}^{2})' ( p n p n − 1 ρ s 2 ) ′ gives n h ~ n − 1 + 2 h ~ n ( d ~ n + d ~ n − 1 ) = 0 n\widetilde{h}_{n-1}+2\widetilde{h}_{n}(\widetilde{d}_{n}+\widetilde{d}_{n-1})=0 n h n − 1 + 2 h n ( d n + d n − 1 ) = 0 . In particular,
d ~ 0 = − 1 2 ∂ s log h ~ 0 , r ~ 1 = s − d ~ 0 2 , h ~ n = r ~ n h ~ n − 1 . \widetilde{d}_{0}=-\frac{1}{2}\partial_{s}\log\widetilde{h}_{0},\qquad\widetilde{r}_{1}=s-\widetilde{d}_{0}^{2},\qquad\widetilde{h}_{n}=\widetilde{r}_{n}\widetilde{h}_{n-1}. d 0 = − 2 1 ∂ s log h 0 , r 1 = s − d 0 2 , h n = r n h n − 1 .
The negative sign in the second string equation reflects the weight e 2 w 3 / 3 − 2 s w e^{2w^{3}/3-2sw} e 2 w 3 /3 − 2 s w and the change from λ \lambda λ to w w w .
The associated Toda relations are
∂ s d ~ n = − 2 ( r ~ n + 1 − r ~ n ) , ∂ s r ~ n = − 2 r ~ n ( d ~ n − d ~ n − 1 ) . (169) \partial_{s}\widetilde{d}_{n}=-2(\widetilde{r}_{n+1}-\widetilde{r}_{n}),\qquad\partial_{s}\widetilde{r}_{n}=-2\widetilde{r}_{n}(\widetilde{d}_{n}-\widetilde{d}_{n-1}). \tag*{(169)} ∂ s d n = − 2 ( r n + 1 − r n ) , ∂ s r n = − 2 r n ( d n − d n − 1 ) . (169)
For example, differentiation of the orthogonality relations gives ∂ s p n = 2 r ~ n p n − 1 \partial_{s}p_{n}=2\widetilde{r}_{n}p_{n-1} ∂ s p n = 2 r n p n − 1 ; comparison of the next-to-leading coefficients in (B4) gives the first equation in (B34), and (B32) gives the second. Thus the derivative in (B10) can be evaluated along with X k X_{k} X k , without numerical differentiation. Writing a dot for ∂ s \partial_{s} ∂ s , one uses
X ˙ 0 = − h ~ 0 , X ˙ 1 = B ˙ 0 X 0 + ( d ~ 0 − d ~ 1 ) h ~ 0 , \dot{X}_{0}=-\widetilde{h}_{0},\qquad\dot{X}_{1}=\dot{B}_{0}X_{0}+(\widetilde{d}_{0}-\widetilde{d}_{1})\widetilde{h}_{0}, X ˙ 0 = − h 0 , X ˙ 1 = B ˙ 0 X 0 + ( d 0 − d 1 ) h 0 ,
X ˙ k + 1 = B ˙ 2 k X k + B 2 k X ˙ k − r ~ ˙ 2 k X k − 1 − r ~ 2 k X ˙ k − 1 , k ≥ 1. (170) \dot{X}_{k+1}=\dot{B}_{2k}X_{k}+B_{2k}\dot{X}_{k}-\dot{\widetilde{r}}_{2k}X_{k-1}-\widetilde{r}_{2k}\dot{X}_{k-1},\qquad k\geq1. \tag*{(170)} X ˙ k + 1 = B ˙ 2 k X k + B 2 k X ˙ k − r ˙ 2 k X k − 1 − r 2 k X ˙ k − 1 , k ≥ 1. (170)
Low-dimensional checks. The first three even sizes read
τ 2 ( 1 ) = − 2 X 0 , τ 4 ( 1 ) = 4 h ~ 1 ( B 0 X 0 − h ~ 0 ) , \tau^{(1)}_2=-2X_0,\qquad\tau^{(1)}_4=4\widetilde{h}_1(B_0X_0-\widetilde{h}_0), τ 2 ( 1 ) = − 2 X 0 , τ 4 ( 1 ) = 4 h 1 ( B 0 X 0 − h 0 ) ,
τ 6 ( 1 ) = − 8 h ~ 1 h ~ 3 [ ( B 2 B 0 − r ~ 2 ) X 0 − B 2 h ~ 0 ] . (171) \tau^{(1)}_6=-8\widetilde{h}_1\widetilde{h}_3\left[(B_2B_0-\widetilde{r}_2)X_0-B_2\widetilde{h}_0\right]. \tag*{(171)} τ 6 ( 1 ) = − 8 h 1 h 3 [ ( B 2 B 0 − r 2 ) X 0 − B 2 h 0 ] . (171)
These follow both from direct expansion of the corresponding Pfaffians and from the recursion. At N = 2 N=2 N = 2 , Ξ 2 ( 1 ) / T 1 / 3 = − h ~ 0 / X 0 \Xi^{(1)}_2/T^{1/3}=-\widetilde{h}_0/X_0 Ξ 2 ( 1 ) / T 1/3 = − h 0 / X 0 , which also follows directly from A ′ = 2 h ~ 0 A'=2\widetilde{h}_0 A ′ = 2 h 0 and τ 2 ( 1 ) = A \tau^{(1)}_2=A τ 2 ( 1 ) = A .
Relation to the symplectic reduction. At fixed s s s , let Q k Q_k Q k and R k R_k R k satisfy the homogeneous recursion (B9), with
Q 0 = 1 , Q 1 = B 0 , R 0 = 0 , R 1 = 1. Q_0=1,\qquad Q_1=B_0,\qquad R_0=0,\qquad R_1=1. Q 0 = 1 , Q 1 = B 0 , R 0 = 0 , R 1 = 1.
Then the exact solution is
X k = Q k X 0 − h ~ 0 R k . (172) X_k=Q_kX_0-\widetilde{h}_0R_k. \tag*{(172)} X k = Q k X 0 − h 0 R k . (172)
The sequence Q k Q_k Q k is the same one that appears in the symplectic reduction when built from these unitary recurrence coefficients. The orthogonal ensemble selects a different linear combination through the ordered-contour primitive X 0 X_0 X 0 and the term − h ~ 0 -\widetilde{h}_0 − h 0 in X 1 X_1 X 1 . Thus the extra β = 1 \beta=1 β = 1 information is confined to explicit scalar initial values, while all coefficients of the finite-N N N recursion come from the β = 2 \beta=2 β = 2 problem at temperature T / 2 T/2 T /2 . This completes the exact reduction for even N N N . Its double-scaling limit, including the selection of the Painlevé I solution, is carried out in Sec. 11E. Odd N N N , which requires a bordered Pfaffian, remains open.
The β = 1 \beta=1 β = 1 edge and real symmetric random band matrices
In Ref. [46 ] the unitary edge scaling function was compared with the localized spectral edge of complex Hermitian random band matrices. We retain the rescaling constants fixed there and extend the proposed correspondence to the other symmetry classes. The common semicircle term is required by normalization. We make only one nontrivial comparison here: the coefficient of the first symmetry-dependent correction inside the band, evaluated for the real symmetric ensemble.
The band-matrix edge. Consider M × M M\times M M × M real symmetric periodic band matrices. For u < v u<v u < v with 0 < ∣ u − v ∣ M ≤ W 0<|u-v|_M\le W 0 < ∣ u − v ∣ M ≤ W , the entries H u v = H v u = ± 1 H_{uv}=H_{vu}=\pm1 H uv = H v u = ± 1 are independent random signs with equal probabilities; all remaining entries, including the diagonal, vanish. Let λ α \lambda_\alpha λ α denote the eigenvalues of H / 2 W H/\sqrt{2W} H / 2 W , whose limiting semicircle has edges at ± 2 \pm2 ± 2 . Define the edge coordinate ξ α = ( 2 − λ α ) W 4 / 5 \xi_\alpha=(2-\lambda_\alpha)W^{4/5} ξ α = ( 2 − λ α ) W 4/5 and the rescaled cumulative eigenvalue count
σ 1 ( ξ ) = lim W 6 / 5 M ⟨ # { α : ξ α < ξ } ⟩ , 1 ≪ W ≪ M 5 / 6 . (173) \sigma_1(\xi)=\lim\frac{W^{6/5}}{M}\left\langle\#\{\alpha:\xi_\alpha<\xi\}\right\rangle,\qquad1\ll W\ll M^{5/6}. \tag*{(173)} σ 1 ( ξ ) = lim M W 6/5 ⟨ # { α : ξ α < ξ } ⟩ , 1 ≪ W ≪ M 5/6 . (173)
The brackets in Eq. (C1) denote expectation over the independent random signs defining the band-matrix ensemble. Positive ξ \xi ξ points into the band, while negative ξ \xi ξ describes eigenvalues beyond the limiting edge. Sodin proved the existence of this limiting counting function for the random-sign model and obtained
σ 1 ( ξ ) = 2 3 π ξ 3 / 2 + O ( ξ ) , ξ → + ∞ . \sigma_1(\xi)=\frac{2}{3\pi}\xi^{3/2}+O(\xi),\qquad\xi\to+\infty. σ 1 ( ξ ) = 3 π 2 ξ 3/2 + O ( ξ ) , ξ → + ∞.
The leading coefficient is fixed by the semicircle normalization [83 ] . To prove corrections to this leading behaviour, one uses averaged traces of Chebyshev polynomials U n U_n U n of the second kind [83 , 67 ] . Their degree n n n scales with the bandwidth so that
τ = n W 2 / 5 > 0 \tau=\frac{n}{W^{2/5}}>0 τ = W 2/5 n > 0
remains fixed. Indeed, near the upper edge,
1 n + 1 U n ( λ 2 ) ⟶ sin ( τ ξ ) τ ξ , λ = 2 − ξ W − 4 / 5 . \frac{1}{n+1}U_n\left(\frac{\lambda}{2}\right) \longrightarrow\frac{\sin(\tau\sqrt{\xi})}{\tau\sqrt{\xi}}, \qquad\lambda=2-\xi W^{-4/5}. n + 1 1 U n ( 2 λ ) ⟶ τ ξ sin ( τ ξ ) , λ = 2 − ξ W − 4/5 .
For even degrees, the two edges contribute equally, giving the transform
ϕ 1 ( τ ) = 2 ∫ − ∞ ∞ sin ( τ ξ ) τ ξ d σ 1 ( ξ ) . (174) \phi_1(\tau)=2\int_{-\infty}^{\infty}\frac{\sin(\tau\sqrt{\xi})}{\tau\sqrt{\xi}}\,d\sigma_1(\xi). \tag*{(174)} ϕ 1 ( τ ) = 2 ∫ − ∞ ∞ τ ξ sin ( τ ξ ) d σ 1 ( ξ ) . (174)
For ξ < 0 \xi<0 ξ < 0 , the kernel is understood by analytic continuation as sinh ( τ − ξ ) / ( τ − ξ ) \sinh(\tau\sqrt{-\xi})/(\tau\sqrt{-\xi}) sinh ( τ − ξ ) / ( τ − ξ ) . The integral uses the regularization of Ref. [46 ] , which removes the leading semicircle contribution. This reflects the exact Chebyshev orthogonality relation
∫ − 2 2 4 − λ 2 2 π U n ( λ / 2 ) d λ = 0 , n ≥ 1. \int_{-2}^{2}\frac{\sqrt{4-\lambda^2}}{2\pi}U_n(\lambda/2)\,d\lambda=0,\qquad n\geq1. ∫ − 2 2 2 π 4 − λ 2 U n ( λ /2 ) d λ = 0 , n ≥ 1.
In the walk representation of these traces, τ \tau τ is the total walk length measured in units of W 2 / 5 W^{2/5} W 2/5 ; diagram edge lengths such as α \alpha α use the same units. The Chebyshev expansion cancels immediate backtracking contributions. For random signs, the remaining walks may traverse an edge twice in the same direction, producing non-orientable diagrams. The leading such diagram has no unitary counterpart and supplies the single symmetry-dependent correction compared below.
The orthogonal scaling function. For β = 1 \beta=1 β = 1 and even N N N we write E = E 1 ∗ + s 1 2 N − 4 / 5 X E=E_1^*+s_1^2N^{-4/5}X E = E 1 ∗ + s 1 2 N − 4/5 X and t = 2 7 / 5 X t=2^{7/5}X t = 2 7/5 X , so that Eq. (103) reads ρ N ( 1 ) ≃ ( N 3 / 5 / s 1 ) Φ 1 ′ ( X ) \rho_N^{(1)}\simeq(N^{3/5}/s_1)\Phi_1'(X) ρ N ( 1 ) ≃ ( N 3/5 / s 1 ) Φ 1 ′ ( X ) with Φ 1 ′ ( X ) = ( 2 9 / 5 / π ) Im [ 1 4 y − v 1 2 ] \Phi_1'(X)=(2^{9/5}/\pi)\operatorname{Im}[\frac{1}{4}y-v_1^2] Φ 1 ′ ( X ) = ( 2 9/5 / π ) Im [ 4 1 y − v 1 2 ] . Since v 1 ′ = 3 4 y − v 1 2 v_1'=\frac{3}{4}y-v_1^2 v 1 ′ = 4 3 y − v 1 2 , we have 1 4 y − v 1 2 = − 1 2 y + v 1 ′ \frac{1}{4}y-v_1^2=-\frac{1}{2}y+v_1' 4 1 y − v 1 2 = − 2 1 y + v 1 ′ . The first term gives the unitary scaling function Φ ′ ( X ) = − ( 2 4 / 5 / π ) Im y 1 ( 2 7 / 5 X ) \Phi'(X)=-(2^{4/5}/\pi)\operatorname{Im}y_1(2^{7/5}X) Φ ′ ( X ) = − ( 2 4/5 / π ) Im y 1 ( 2 7/5 X ) of Ref. [46 ] , and the second is a total derivative. The cumulative intensity Φ 1 ( X ) = ∫ − ∞ X Φ 1 ′ ( Y ) d Y \Phi_1(X)=\int_{-\infty}^{X}\Phi_1'(Y)\,dY Φ 1 ( X ) = ∫ − ∞ X Φ 1 ′ ( Y ) d Y is therefore
Φ 1 ( X ) = Φ ( X ) + 2 2 / 5 π Im v 1 ( 2 7 / 5 X ) , (175) \Phi_1(X)=\Phi(X)+\frac{2^{2/5}}{\pi}\operatorname{Im}v_1(2^{7/5}X), \tag*{(175)} Φ 1 ( X ) = Φ ( X ) + π 2 2/5 Im v 1 ( 2 7/5 X ) , (175)
where we have used that v 1 ≃ − 3 y / 4 v_1\simeq-\sqrt{3y/4} v 1 ≃ − 3 y /4 is real, up to exponentially small terms, as t → − ∞ t\to-\infty t → − ∞ . Equation (C3) plays the role that the construction of F 1 F_1 F 1 from F 2 F_2 F 2 plays for the Tracy–Widom distributions.
On the allowed side y 1 ≃ − i t / 6 − 1 / ( 48 t 2 ) + … y_1\simeq-i\sqrt{t/6}-1/(48t^2)+\ldots y 1 ≃ − i t /6 − 1/ ( 48 t 2 ) + … , and the selected solution x x x decays as t → + ∞ t\to+\infty t → + ∞ , so that v 1 = − 3 y / 4 − 1 / ( 8 t ) + O ( t − 9 / 4 ) v_1=-\sqrt{3y/4}-1/(8t)+O(t^{-9/4}) v 1 = − 3 y /4 − 1/ ( 8 t ) + O ( t − 9/4 ) , with the branch Re 3 y / 4 > 0 \operatorname{Re}\sqrt{3y/4}>0 Re 3 y /4 > 0 . Hence Im v 1 = 1 4 6 1 / 4 t 1 / 4 + O ( t − 9 / 4 ) \operatorname{Im}v_1=\frac{1}{4}6^{1/4}t^{1/4}+O(t^{-9/4}) Im v 1 = 4 1 6 1/4 t 1/4 + O ( t − 9/4 ) , and with the unitary expansion Φ ( X ) = A + X 3 / 2 + a Φ X − 7 / 2 + … \Phi(X)=A_+X^{3/2}+a_\Phi X^{-7/2}+\ldots Φ ( X ) = A + X 3/2 + a Φ X − 7/2 + … of Ref. [46 ] ,
Φ 1 ( X ) = A + X 3 / 2 + 3 1 / 4 2 π X 1 / 4 + O ( X − 9 / 4 ) , A + = 4 3 π 3 , X → + ∞ . (176) \Phi_1(X)=A_+X^{3/2}+\frac{3^{1/4}}{2\pi}X^{1/4}+O(X^{-9/4}),\qquad A_+=\frac{4}{3\pi\sqrt{3}},\qquad X\to+\infty. \tag*{(176)} Φ 1 ( X ) = A + X 3/2 + 2 π 3 1/4 X 1/4 + O ( X − 9/4 ) , A + = 3 π 3 4 , X → + ∞. (176)
The term X 1 / 4 X^{1/4} X 1/4 is the projective-plane contribution. Its sign is fixed by the selection of Par. 7 of Sec. 11E; for β = 4 \beta=4 β = 4 it is reversed.
Rescaling fixed by the unitary case. We conjecture Φ 1 ( c ξ ) = A σ 1 ( ξ ) \Phi_1(c\xi)=A\sigma_1(\xi) Φ 1 ( c ξ ) = A σ 1 ( ξ ) , using
c = ( 3 / 64 ) 1 / 5 , A = 48 − 1 / 5 , c=(3/64)^{1/5},\qquad A=48^{-1/5}, c = ( 3/64 ) 1/5 , A = 4 8 − 1/5 ,
as fixed in Ref. [46 ] . With the same variance profile, bandwidth convention, and spectral normalization, the orientable contributions retain the unitary normalization. No additional constants are fitted in the orthogonal comparison. The leading term 2 ξ 3 / 2 / ( 3 π ) 2\xi^{3/2}/(3\pi) 2 ξ 3/2 / ( 3 π ) is required by the common semicircle law and is not an independent test of the correspondence. Equation (C4) then predicts, without any free parameter,
σ 1 ( ξ ) = 2 3 π ξ 3 / 2 + 6 2 π ξ 1 / 4 + O ( ξ − 9 / 4 ) , ξ → + ∞ . (177) \sigma_1(\xi)=\frac{2}{3\pi}\xi^{3/2}+\frac{\sqrt{6}}{2\pi}\xi^{1/4}+O(\xi^{-9/4}),\qquad\xi\to+\infty. \tag*{(177)} σ 1 ( ξ ) = 3 π 2 ξ 3/2 + 2 π 6 ξ 1/4 + O ( ξ − 9/4 ) , ξ → + ∞. (177)
The single nontrivial check: the projective-plane term. The leading non-orientable diagram consists of a marked tail attached to a single loop, traversed twice in the same direction; see Figure 2 . The non-backtracking condition excludes the opposite orientation of the second passage, so this diagram has no unitary counterpart. The diagram has one independent spatial loop. For the uniform band profile, a step measured in units of the bandwidth has variance ς 2 = ∫ − 1 1 u 2 d u / 2 = 1 / 3 \varsigma^2 = \int_{-1}^{1} u^2\,\mathrm{d}u/2 = 1/3 ς 2 = ∫ − 1 1 u 2 d u /2 = 1/3 . The continuum return probability for a loop of length α \alpha α is therefore
Figure 2. The leading non-orientable walk diagram. Starting at the marked point, the walk follows the tail, goes twice around the same loop in the same direction, and returns along the tail. Reversing direction on the second circuit would immediately retrace the last edge of the first circuit, violating the non-backtracking condition. The loop has length α \alpha α and is traversed twice, giving 2 α ≤ τ 2\alpha\le\tau 2 α ≤ τ . The drawing shows the underlying graph and walk; its ribbon thickening gives the projective-plane contribution.
p α ( 0 ) = 1 2 π ς 2 α . p_\alpha(0) = \frac{1}{\sqrt{2\pi\varsigma^2\alpha}}. p α ( 0 ) = 2 π ς 2 α 1 .
Although the walk traverses the loop twice, its spatial closure imposes only one return condition, giving one factor p α ( 0 ) p_\alpha(0) p α ( 0 ) . The two passages use total length 2 α 2\alpha 2 α ; the remaining length is assigned to the outward and return passages along the marked tail. After eliminating the tail length, the admissible range is 0 < α < τ / 2 0 < \alpha< \tau/2 0 < α < τ /2 . With the normalized-trace factor 1 / τ 1/\tau 1/ τ , the diagram contributes
ϕ 1 , p r o j ( τ ) = 1 τ ∫ 0 τ / 2 p α ( 0 ) d α = 1 τ 2 π ς 2 2 τ 2 = 3 π τ − 1 / 2 . (178) \phi_{1,\mathrm{proj}}(\tau) = \frac{1}{\tau}\int_{0}^{\tau/2} p_\alpha(0)\,\mathrm{d}\alpha= \frac{1}{\tau\sqrt{2\pi\varsigma^2}}\,2\sqrt{\frac{\tau}{2}} = \sqrt{\frac{3}{\pi}}\,\tau^{-1/2}. \tag*{(178)} ϕ 1 , proj ( τ ) = τ 1 ∫ 0 τ /2 p α ( 0 ) d α = τ 2 π ς 2 1 2 2 τ = π 3 τ − 1/2 . (178)
The diagram evaluated above is Sodin’s s = 1 s=1 s = 1 diagram, with its continuum weight given by Eq. (4.6) of Ref. [83 ] . A term B ξ 1 / 4 B\xi^{1/4} B ξ 1/4 in σ 1 \sigma_1 σ 1 contributes B 2 π τ − 1 / 2 B\sqrt{2\pi}\,\tau^{-1/2} B 2 π τ − 1/2 to Eq. (174) . Thus the result 3 / π τ − 1 / 2 \sqrt{3/\pi}\,\tau^{-1/2} 3/ π τ − 1/2 agrees precisely with the Riccati prediction B = 6 / ( 2 π ) B=\sqrt{6}/(2\pi) B = 6 / ( 2 π ) in Eq. (177) , at the level of the transformed edge asymptotics. The corresponding ξ 1 / 4 \xi^{1/4} ξ 1/4 correction to the edge counting function was conjectured by Sodin [83 Remark 7.2] .
Symplectic prediction and status. For quaternion self-dual band matrices with independent Haar-random unit-quaternion entries in the band, count each Kramers pair once. Keeping the unitary rescaling in the doubled-size convention of § 11 E, the corresponding prediction in the natural band coordinate ξ = ( 2 − λ ) W 4 / 5 \xi=(2-\lambda)W^{4/5} ξ = ( 2 − λ ) W 4/5 is
σ 4 ( ξ ) = 2 3 π ξ 3 / 2 − 6 4 π ξ 1 / 4 + O ( ξ − 9 / 4 ) , ξ → + ∞ . (179) \sigma_4(\xi) = \frac{2}{3\pi}\xi^{3/2} - \frac{\sqrt{6}}{4\pi}\xi^{1/4} + O(\xi^{-9/4}), \qquad\xi\to+\infty. \tag*{(179)} σ 4 ( ξ ) = 3 π 2 ξ 3/2 − 4 π 6 ξ 1/4 + O ( ξ − 9/4 ) , ξ → + ∞. (179)
The reversal of the first correction reflects the different selection of the auxiliary solution for β = 4 \beta=4 β = 4 . This is a prediction; we do not perform an additional quaternion band-matrix comparison here.
On the Riccati side the arguments are at the physics level, and on the band-matrix side the single-loop contribution follows from the continuum counting formulas of Refs. [83 , 67 ] , while the passage from the transform to the asymptotics of σ 1 \sigma_1 σ 1 itself is the step that Sodin could establish only in a weak sense. The correspondence Φ 1 ( c ξ ) = A σ 1 ( ξ ) \Phi_1(c\xi)=A\sigma_1(\xi) Φ 1 ( c ξ ) = A σ 1 ( ξ ) remains a conjecture, now supported by the first symmetry-dependent allowed-side coefficient, with no adjustment of the unitary rescaling constants.
Boundary admittance, absorption and scattering time delays
This appendix collects the scattering interpretation of the Riccati variable without interrupting the spectral derivation in the main text. It distinguishes the internal norm for a prescribed boundary amplitude, the dwell time referred to incident flux, and the Wigner–Smith delay. It also records the high-energy connection with absorbing-waveguide statistics and outlines the programme of Paper II [44 ] . Throughout the boundary-response derivations, the sample occupies 0 < x < L 0 < x < L 0 < x < L , with a Dirichlet boundary at x = 0 x = 0 x = 0 and a lead, when present, attached at x = L x = L x = L ; the positive coordinate direction points from the Dirichlet end towards the lead. Thus T = − ∂ E Z \mathcal{T} = -\partial_E Z T = − ∂ E Z , consistently with the main text. The semi-infinite sample is obtained by taking L → ∞ L \to\infty L → ∞ .
Boundary response, absorbed flux and internal norm
Replacing E E E by E = E + i η \mathcal{E} = E + i\eta E = E + i η , with η > 0 \eta> 0 η > 0 , is equivalent to adding the absorbing potential − i η -i\eta − i η to W W W . Consider the sample 0 < x < L 0 < x < L 0 < x < L , with a Dirichlet condition at x = 0 x = 0 x = 0 , and prescribe the wave amplitude at the other end:
ψ ( 0 ) = 0 , ψ ( L ) = b . \psi(0) = 0,\qquad\psi(L) = b. ψ ( 0 ) = 0 , ψ ( L ) = b .
If Ψ \Psi Ψ is the matrix solution satisfying Ψ ( 0 ) = 0 \Psi(0) = 0 Ψ ( 0 ) = 0 and Ψ ′ ( 0 ) = I \Psi'(0) = I Ψ ′ ( 0 ) = I , then
ψ ( x ) = F ( x ; E ) b , F ( x ; E ) : = Ψ ( x ; E ) Ψ ( L ; E ) − 1 . \psi(x) = F(x;\mathcal{E})b,\qquad F(x;\mathcal{E}) := \Psi(x;\mathcal{E})\Psi(L;\mathcal{E})^{-1}. ψ ( x ) = F ( x ; E ) b , F ( x ; E ) := Ψ ( x ; E ) Ψ ( L ; E ) − 1 .
In particular,
ψ ′ ( L ) = Z ( L ; E ) b . \psi'(L) = Z(L;\mathcal{E})b. ψ ′ ( L ) = Z ( L ; E ) b .
Thus Z Z Z maps the boundary amplitude to the boundary slope. This is its Dirichlet-to-Neumann interpretation; with the appropriate physical normalization it is an input admittance.
The quantity measured by the imaginary part of Z Z Z follows directly from flux balance. Define
j ( x ) = Im ( ψ † ψ ′ ) , j(x) = \operatorname{Im}(\psi^\dagger\psi'), j ( x ) = Im ( ψ † ψ ′ ) ,
positive towards increasing x x x . Since ψ ′ ′ = ( W − E ) ψ \psi'' = (W - \mathcal{E})\psi ψ ′′ = ( W − E ) ψ and W W W is Hermitian,
j ′ ( x ) = − η ψ † ψ . j'(x) = -\eta\psi^\dagger\psi. j ′ ( x ) = − η ψ † ψ .
The Dirichlet boundary condition ψ ( 0 ) = 0 \psi(0) = 0 ψ ( 0 ) = 0 implies that no probability current flows through the boundary at x = 0 x = 0 x = 0 , so j ( 0 ) = 0 j(0) = 0 j ( 0 ) = 0 . Writing Z = A − i P Z = A - iP Z = A − i P , with A A A and P P P Hermitian, we therefore obtain
− j ( L ) = b † P b = η ∫ 0 L ψ † ( x ) ψ ( x ) d x . -j(L) = b^\dagger P b = \eta\int_0^L \psi^\dagger(x)\psi(x)\,\mathrm{d}x. − j ( L ) = b † P b = η ∫ 0 L ψ † ( x ) ψ ( x ) d x .
The minus sign reflects the geometry: flux injected at x = L x = L x = L travels towards decreasing x x x . Since this identity holds for every boundary vector b b b , it gives the matrix identity
P η = ∫ 0 L F † ( x ; E ) F ( x ; E ) d x . \frac{P}{\eta} = \int_0^L F^\dagger(x;\mathcal{E})F(x;\mathcal{E})\,\mathrm{d}x. η P = ∫ 0 L F † ( x ; E ) F ( x ; E ) d x .
Thus P / η P/\eta P / η is exactly the internal-norm matrix for a prescribed boundary amplitude, at any η > 0 \eta> 0 η > 0 . Its positivity expresses the positivity of absorption.
At real energy, away from the Dirichlet eigenvalues, the same internal norm is obtained by differentiating the boundary response with respect to energy. The Wronskian identity gives
T ( E ) : = − ∂ E Z ( L ; E ) = ∫ 0 L F ( x ; E ) † F ( x ; E ) d x . \mathcal{T}(E) := -\partial_E Z(L;E) = \int_0^L F(x;E)^\dagger F(x;E)\,\mathrm{d}x. T ( E ) := − ∂ E Z ( L ; E ) = ∫ 0 L F ( x ; E ) † F ( x ; E ) d x .
Consequently,
lim η → 0 + P ( E , η ) η = T ( E ) . \lim_{\eta\to0^+}\frac{P(E,\eta)}{\eta} = \mathcal{T}(E). η → 0 + lim η P ( E , η ) = T ( E ) .
The distinction at finite absorption is simple: P / η P/\eta P / η measures the norm of the solution at E + i η E + i\eta E + i η , whereas T \mathcal{T} T measures that of the real-energy solution. They coincide only in the zero-absorption limit (see § 6 ).
Scalar local and integrated currents. For N = 1 N=1 N = 1 , Z = ψ ′ / ψ = T 1 / 3 ( u − i v ) Z=\psi'/\psi=T^{1/3}(u-iv) Z = ψ ′ / ψ = T 1/3 ( u − i v ) . In the current normalization used in the scalar warm-up,
j ( x ) = ℏ m ∣ ψ ( x ) ∣ 2 Im Z ( x ) = − ℏ m T 1 / 3 ∣ ψ ( x ) ∣ 2 v ( x ) . j(x)=\frac{\hbar}{m}|\psi(x)|^2\operatorname{Im}Z(x)=-\frac{\hbar}{m}T^{1/3}|\psi(x)|^2v(x). j ( x ) = m ℏ ∣ ψ ( x ) ∣ 2 Im Z ( x ) = − m ℏ T 1/3 ∣ ψ ( x ) ∣ 2 v ( x ) .
The local intensity appears because ψ ′ = Z ψ \psi'=Z\psi ψ ′ = Z ψ . The integrated Wronskian identity gives the equivalent flux-balance expression
∣ ψ ( x ) ∣ 2 Im Z ( x ) = − η ∫ 0 x ∣ ψ ( s ) ∣ 2 d s . |\psi(x)|^2\operatorname{Im}Z(x)=-\eta\int_0^x|\psi(s)|^2\,\mathrm{d}s. ∣ ψ ( x ) ∣ 2 Im Z ( x ) = − η ∫ 0 x ∣ ψ ( s ) ∣ 2 d s .
The explicit factor ℏ / m \hbar/m ℏ/ m only restores the conventional current normalization; E E E and η \eta η retain the rescaled Schrödinger convention of the main text.
Ideal-lead matching and the absorption deficit
To connect this boundary response to scattering, attach an ideal lead at x = L x=L x = L . For a common channel wavenumber q > 0 q>0 q > 0 , write the lead wave as
ψ l e a d ( x ) = e − i q ( x − L ) a i n + e i q ( x − L ) a o u t , a o u t = r a i n . \psi_{\mathrm{lead}}(x)=e^{-iq(x-L)}a_{\mathrm{in}}+e^{iq(x-L)}a_{\mathrm{out}},\qquad a_{\mathrm{out}}=r\,a_{\mathrm{in}}. ψ lead ( x ) = e − i q ( x − L ) a in + e i q ( x − L ) a out , a out = r a in .
Matching the amplitude and slope at x = L x=L x = L gives
r = − ( Z − i q I ) − 1 ( Z + i q I ) , b = ( I + r ) a i n . r=-(Z-iqI)^{-1}(Z+iqI),\qquad b=(I+r)a_{\mathrm{in}}. r = − ( Z − i q I ) − 1 ( Z + i q I ) , b = ( I + r ) a in .
Thus fixing the incident amplitude is different from fixing b b b : reflection changes the amplitude actually applied to the sample. Flux balance now reads
I − r † r = 1 q ( I + r ) † P ( I + r ) . I-r^\dagger r=\frac{1}{q}(I+r)^\dagger P(I+r). I − r † r = q 1 ( I + r ) † P ( I + r ) .
This is the scattering form of the same absorption identity. The factors I + r I+r I + r convert the prescribed incident amplitude into the boundary amplitude.
A closely related convention for absorbing scattering systems is to write
I − S Γ † S Γ = Γ Q Γ , Γ = 2 η . I-S_\Gamma^\dagger S_\Gamma=\Gamma Q_\Gamma,\qquad\Gamma=2\eta. I − S Γ † S Γ = Γ Q Γ , Γ = 2 η .
Here Q Γ Q_\Gamma Q Γ is the finite-absorption delay matrix [80 , 38 , 12 ] . For the lead construction above, the corresponding absorption-deficit matrix is
I − r † r 2 η = 1 2 q ( I + r ) † P η ( I + r ) . \frac{I-r^\dagger r}{2\eta}=\frac{1}{2q}(I+r)^\dagger\frac{P}{\eta}(I+r). 2 η I − r † r = 2 q 1 ( I + r ) † η P ( I + r ) .
This formula makes explicit both the internal norm and the conversion to incident-flux normalization.
Dwell time and the Wigner–Smith matrix
We retain the same coordinate convention: the sample occupies 0 < x < L 0<x<L 0 < x < L , the Dirichlet boundary is at x = 0 x=0 x = 0 , and the lead is attached at x = L x=L x = L . In particular, T = − ∂ E Z ( L ; E ) T=-\partial_E Z(L;E) T = − ∂ E Z ( L ; E ) , as in the main text. All formulas in this subsection refer to real energy and a lossless sample. The potential and the Dirichlet boundary condition are independent of energy. We use units ℏ = 2 m = 1 \hbar=2m=1 ℏ = 2 m = 1 .
With F ( x ; E ) = Ψ ( x ; E ) Ψ ( L ; E ) − 1 F(x;E)=\Psi(x;E)\Psi(L;E)^{-1} F ( x ; E ) = Ψ ( x ; E ) Ψ ( L ; E ) − 1 , the scattering state is
ψ ( x ) = F ( x ; E ) ( I + r ) a i n . \psi(x)=F(x;E)(I+r)a_{\mathrm{in}}. ψ ( x ) = F ( x ; E ) ( I + r ) a in .
The integrated probability in the sample is therefore
∫ 0 L ψ † ψ d x = a i n † ( I + r ) † T ( I + r ) a i n . \int_{0}^{L}\psi^{\dagger}\psi\,dx=a_{\mathrm{in}}^{\dagger}(I+r)^{\dagger}\mathcal{T}(I+r)a_{\mathrm{in}}. ∫ 0 L ψ † ψ d x = a in † ( I + r ) † T ( I + r ) a in .
Since the magnitude of the incident flux is 2 q a i n † a i n 2q\,a_{\mathrm{in}}^{\dagger}a_{\mathrm{in}} 2 q a in † a in , the dwell-time matrix is
Q d w e l l = 1 2 q ( I + r ) † T ( I + r ) . (180) Q_{\mathrm{dwell}}=\frac{1}{2q}(I+r)^{\dagger}\mathcal{T}(I+r). \tag*{(180)} Q dwell = 2 q 1 ( I + r ) † T ( I + r ) . (180)
It is positive and converts the internal norm for a prescribed boundary amplitude into a time referred to incident-flux normalization.
To verify the energy-derivative identity used here, differentiate − F ′ ′ + W F = E F -F^{\prime\prime}+WF=EF − F ′′ + W F = E F at real energy. The Wronskian relation becomes
∂ x [ F † ∂ E F ′ − F ′ † ∂ E F ] = − F † F . \partial_x\left[F^{\dagger}\partial_EF^{\prime}-F^{\prime\dagger}\partial_EF\right]=-F^{\dagger}F. ∂ x [ F † ∂ E F ′ − F ′ † ∂ E F ] = − F † F .
At x = 0 x=0 x = 0 the bracket vanishes because F ( 0 ; E ) = 0 F(0;E)=0 F ( 0 ; E ) = 0 for all E E E . At x = L x=L x = L , F ( L ; E ) = I F(L;E)=I F ( L ; E ) = I and ∂ E F ( L ; E ) = 0 \partial_EF(L;E)=0 ∂ E F ( L ; E ) = 0 , so the bracket is ∂ E Z ( L ; E ) \partial_EZ(L;E) ∂ E Z ( L ; E ) . Hence
− ∂ E Z ( L ; E ) = ∫ 0 L F ( x ; E ) † F ( x ; E ) d x = T . -\partial_EZ(L;E)=\int_{0}^{L}F(x;E)^{\dagger}F(x;E)\,dx=\mathcal{T}. − ∂ E Z ( L ; E ) = ∫ 0 L F ( x ; E ) † F ( x ; E ) d x = T .
No change of boundary orientation is involved.
The Wigner–Smith matrix differs from the dwell-time matrix when the lead wavenumber depends on energy. In the present convention, boundary matching gives
Z ( I + r ) = i q ( r − I ) , r = − ( Z − i q I ) − 1 ( Z + i q I ) . Z(I+r)=iq(r-I),\qquad r=-(Z-iqI)^{-1}(Z+iqI). Z ( I + r ) = i q ( r − I ) , r = − ( Z − i q I ) − 1 ( Z + i q I ) .
At real energy Z = Z † Z=Z^{\dagger} Z = Z † , so r † r = I r^{\dagger}r=I r † r = I . Using ∂ E Z = − T \partial_EZ=-\mathcal{T} ∂ E Z = − T , differentiation gives
( Z − i q I ) ∂ E r = i ( ∂ E q ) ( r − I ) + T ( I + r ) . (Z-iqI)\partial_Er=i(\partial_Eq)(r-I)+\mathcal{T}(I+r). ( Z − i q I ) ∂ E r = i ( ∂ E q ) ( r − I ) + T ( I + r ) .
The Cayley representation implies
r † ( Z − i q I ) − 1 = − ( I + r ) † 2 i q . r^{\dagger}(Z-iqI)^{-1}=-\frac{(I+r)^{\dagger}}{2iq}. r † ( Z − i q I ) − 1 = − 2 i q ( I + r ) † .
Consequently, without assuming that Z Z Z commutes with T \mathcal{T} T ,
Q W S : = − i r † ∂ E r = 1 2 q ( I + r ) † T ( I + r ) + i ∂ E q 2 q ( I + r ) † ( r − I ) = Q d w e l l + i ∂ E q 2 q ( r − r † ) . \begin{aligned}
Q_{\mathrm{WS}}&:=-ir^{\dagger}\partial_Er=\frac{1}{2q}(I+r)^{\dagger}\mathcal{T}(I+r)+\frac{i\partial_Eq}{2q}(I+r)^{\dagger}(r-I)\\
&=Q_{\mathrm{dwell}}+\frac{i\partial_Eq}{2q}(r-r^{\dagger}).
\end{aligned} Q WS := − i r † ∂ E r = 2 q 1 ( I + r ) † T ( I + r ) + 2 q i ∂ E q ( I + r ) † ( r − I ) = Q dwell + 2 q i ∂ E q ( r − r † ) .
where the last equality uses unitarity. Thus
Q W S = Q d w e l l + i ∂ E q 2 q ( r − r † ) . (181) Q_{\mathrm{WS}}=Q_{\mathrm{dwell}}+\frac{i\partial_Eq}{2q}(r-r^{\dagger}). \tag*{(181)} Q WS = Q dwell + 2 q i ∂ E q ( r − r † ) . (181)
The additional term is the self-interference contribution associated with the incoming and reflected waves in the lead and need not be positive [53 , 89 ] . For a free lead, E = q 2 E=q^{2} E = q 2 and ∂ E q = 1 / ( 2 q ) \partial_Eq=1/(2q) ∂ E q = 1/ ( 2 q ) , giving
Q W S − Q d w e l l = i 4 q 2 ( r − r † ) . Q_{\mathrm{WS}}-Q_{\mathrm{dwell}}=\frac{i}{4q^{2}}(r-r^{\dagger}). Q WS − Q dwell = 4 q 2 i ( r − r † ) .
In the scalar case r = e i ϕ ( E ) r=e^{i\phi(E)} r = e i ϕ ( E ) , this reads
∂ E ϕ = Q d w e l l − sin ϕ 2 q 2 . \partial_E\phi=Q_{\mathrm{dwell}}-\frac{\sin\phi}{2q^{2}}. ∂ E ϕ = Q dwell − 2 q 2 sin ϕ .
If the lead wavenumber is held fixed when differentiating with respect to the sample energy, the correction vanishes. The zero-absorption limit of the absorption-deficit matrix then agrees directly with the Wigner–Smith matrix. With a dispersive physical lead, the contact term above must be retained.
Waveguide geometry and the high-energy reflection limit
The connection between random waveguide scattering and hyperbolic geometry goes back to the classical paper of Gertsenshtein and Vasil’ev [49 ] , an early treatment of one-dimensional wave localization. A multichannel extension of such framework to Siegel’s circle was studied later in [91 ] , with closely related recent work [71 ] constructing a Dyson-type stochastic evolution on the Siegel half-space. Here it is also appropriate to mention the extremely successful characterization of transport properties (transmission, conductance and other scattering characteristics) of N N N -channel system in the framework of the so-called DMPK [21 , 70 ] approach, with many exact results obtained over the last three decades, see e.g. [4 , 11 , 10 , 5 , 77 ] .
Microscopic versus effective diffusion. The invariant volume element also appears in the classical description of random waveguide scattering by Gertsenshtein and Vasil’ev [49 ] , whose evolution is generated by the hyperbolic Laplacian, with absorption represented by an additional drift. The microscopic Riccati dynamics (19) is not itself isotropic hyperbolic diffusion: the noise acts only on u u u , while the equation for v v v contains no Brownian increment. This distinction does not preclude an isotropic effective description after rapid-phase averaging in a suitable high-energy limit. Indeed, the high-energy scaling of the stationary distribution suggests a connection with the classical reflection statistics of absorbing waveguides [10 , 5 ] . The scaling limit below makes this connection explicit; its detailed scattering interpretation will be developed in Paper II [44 ] .
High-energy stationary limit. Consider E → ∞ E \to\infty E → ∞ with η E / T > 0 \eta\sqrt{E}/T > 0 η E / T > 0 fixed. The stationary law (49) suggests an isotropic limiting description in reflection coordinates. Indeed, setting E = k 2 E = k^2 E = k 2 , H = A / k H = A/k H = A / k , Q = P / k Q = P/k Q = P / k , and c = η k / T c = \eta k/T c = η k / T , its exponent has, as k → ∞ k \to\infty k → ∞ with c > 0 c > 0 c > 0 fixed, the limiting form
− c Tr [ ( H 2 + I ) Q − 1 + Q ] . -c\operatorname{Tr}\left[(H^2+I)Q^{-1}+Q\right]. − c Tr [ ( H 2 + I ) Q − 1 + Q ] .
The corresponding candidate limiting distribution is therefore
d P ∞ ( H , Q ) ∝ exp { − c Tr [ ( H 2 + I ) Q − 1 + Q ] } d μ ( H , Q ) , \mathrm{d}\mathbb{P}_{\infty}(H,Q) \propto\exp\left\{-c\operatorname{Tr}\left[(H^2+I)Q^{-1}+Q\right]\right\}\mathrm{d}\mu(H,Q), d P ∞ ( H , Q ) ∝ exp { − c Tr [ ( H 2 + I ) Q − 1 + Q ] } d μ ( H , Q ) ,
where d μ \mathrm{d}\mu d μ is the invariant volume of the matrix half-space. Under the Cayley transformation to reflection coordinates, this weight becomes radial and leads formally to
p ∞ ( λ ) ∝ ∣ Δ ( λ ) ∣ β exp [ − 4 c ∑ j = 1 N λ j ] , λ j = R j 1 − R j ≥ 0 , p_{\infty}(\lambda) \propto|\Delta(\lambda)|^{\beta}\exp\left[-4c\sum_{j=1}^{N}\lambda_j\right], \qquad\lambda_j=\frac{R_j}{1-R_j}\ge0, p ∞ ( λ ) ∝ ∣Δ ( λ ) ∣ β exp [ − 4 c j = 1 ∑ N λ j ] , λ j = 1 − R j R j ≥ 0 ,
where R j R_j R j are the reflection eigenvalues. This is the Laguerre form of the stationary reflection-eigenvalue distribution previously obtained for semi-infinite absorbing waveguides in the isotropic scattering description [10 , 5 ] , with N = 1 N=1 N = 1 case considered earlier in [34 ] . The high-energy limit of our microscopic admittance distribution thus provides a direct connection with this classical result. The matching of scattering and absorption conventions, and a detailed treatment of the limiting reflection statistics, will be given in Paper II [44 ] .
Scattering statistics and the programme of Paper II
Unlike the complex contour weight, which reproduces holomorphic averages, the full stationary admittance law can be used to evaluate non-holomorphic functions of Z Z Z and Z † Z^\dagger Z † . With the lead convention specified above, these include reflection eigenvalues, absorbed-flux observables, and resonance-counting functions such as log ∣ det r ∣ 2 \log|\det r|^2 log ∣ det r ∣ 2 . Time delays additionally require the real-energy limit and the distinction between the internal norm and the physical Wigner–Smith matrix established above. Related scalar scattering problems were studied long ago in [34 ] and more recently in Refs. [43 , 45 ] , and a σ \sigma σ -model treatment of reflection resonances is given in Ref. [39 ] .
The following programme concerns Paper II [44 ] ; the scattering-statistics calculations listed here are not derived in Paper I. Paper II will use the full stationary law for a semi-infinite absorbing wire attached to an ideal lead, in all three Dyson classes.
Admittance and absorbing DMPK. At high energy, the stationary admittance distribution reduces to the Laguerre-β \beta β ensemble of absorbing DMPK theory (see Appendix D4). The exact distribution also determines the corrections to this limit: a parity symmetry cancels the first-order contribution to the DMPK marginal, so its leading correction is of order ( k ℓ ) − 2 (k\ell)^{-2} ( k ℓ ) − 2 , where k k k is the wave number and ℓ \ell ℓ the mean free path.
Reflection resonances. Through a Poincaré–Lelong counting identity, the distribution gives the density of resonances for all three classes. This includes Ward identities that make narrow resonances β \beta β -independent, closed-form σ \sigma σ -model-type hard-edge densities for β = 1 , 4 \beta=1,4 β = 1 , 4 , and the broad-resonance tail.
Time delays. The zero-absorption law gives the joint statistics of the boundary admittance and its energy derivative, which will be used to determine the reflection and Wigner–Smith time-delay statistics (incorporating the dispersive-lead correction in Eq. (181) ).
The band edge. In the window E ∼ T 2 / 3 E\sim T^{2/3} E ∼ T 2/3 , where there is no DMPK regime, the narrow-resonance coefficient is a density of states. For N = 1 N=1 N = 1 it is exactly Halperin’s ν 1 \nu_1 ν 1 ; for general N N N it is conjecturally the ν N \nu_N ν N of Eq. (7) .
Experiment. A concrete experimental realization using disordered, lossy L C LC L C ladders—where ohmic dissipation plays the role of the uniform absorption η \eta η —will be developed in Paper II. Measuring the complex input impedance as frequency and loss are varied provides direct access to the disorder statistics of the boundary admittance Z Z Z , allowing quantitative comparison with the stationary law (3) and its predicted crossover with absorption.
The finite-length problem in the unitary class, including the full finite-N N N solution of the absorbing DMPK equation, will be treated separately [47 ] .
Intersection numbers and the outgoing contour
For β = 2 \beta=2 β = 2 , the formal expansion of the normalized Kontsevich integral generates intersection numbers on moduli spaces of curves [90 , 62 ] . These numbers are geometric coefficients obtained by integrating products of the classes ψ i \psi_i ψ i ; ψ i \psi_i ψ i is the first Chern class of the cotangent line at the i i i -th marked point. They are also the correlation functions of two-dimensional topological gravity [90 ] .
Our wire problem supplies an outgoing contour prescription for this integral. Admissible contours with the same dominant-saddle contribution can share a formal perturbative expansion while differing by exponentially small terms. The outgoing prescription fixes these additional saddle contributions and hence selects a nonperturbative completion of the formal series. The connection developed below is that the tunnelling contribution responsible for the Lifshitz tail also controls the large-order growth of the perturbative coefficients; these coefficients, in turn, count gluings of triangles.
Perturbative content. For β = 2 \beta=2 β = 2 , shifting to a saddle and removing the saddle exponential and Gaussian normalization identifies the perturbative expansion with the scalar-external-field specialization of the Witten–Kontsevich model. Its cubic vertices generate trivalent ribbon graphs with a factor N N N per face [7 , 64 ] . The coefficients therefore give weighted sums of intersection numbers, rather than resolving the individual numbers encoded by the full generating function with a general external matrix.
For β = 1 , 4 \beta= 1, 4 β = 1 , 4 , the analogous expansion also contains non-orientable graphs [51 , 72 ] . We do not establish here an intersection-number interpretation of the non-orientable coefficients of these outgoing integrals. Their explicit map expansions provide benchmarks for possible comparisons with refined (b b b -deformed) topological recursion [13 , 48 ] .
Large-order growth and random surface gluings
Here large order means k → ∞ k \to\infty k → ∞ at fixed N N N . Each coefficient sums over maps of different topologies; it is not a fixed-genus coefficient. For a connected orientable trivalent map with 2 k 2k 2 k vertices and F F F faces, Euler’s relation gives k = 2 g − 2 + F k = 2g - 2 + F k = 2 g − 2 + F . The partition function includes disconnected maps, whereas its logarithm selects connected ones.
The formal series. Consider the Lifshitz tail regime s → + ∞ s \to+\infty s → + ∞ of § 9 , and put all N N N integration variables near the dominant saddle point, w j = − s + u j w_j = -\sqrt{s} + u_j w j = − s + u j . Since φ ( − s + u ) = φ ( − s ) − s u 2 + u 3 / 3 \varphi(-\sqrt{s} + u) = \varphi(-\sqrt{s}) - \sqrt{s}\,u^2 + u^3/3 φ ( − s + u ) = φ ( − s ) − s u 2 + u 3 /3 exactly, the perturbative part of τ N ( β ) \tau_N^{(\beta)} τ N ( β ) , divided by its leading term, is a formal average over a Gaussian β \beta β -ensemble. The exponential is expanded before taking Gaussian moments; the unexpanded real Gaussian integral with a cubic exponential is not convergent:
P N ( ζ ) = ⟨ e p 3 / 3 ⟩ G , β = ∑ k ≥ 0 a k ( N ) ζ − k , ζ = 2 3 s 3 / 2 , a k ( N ) = ⟨ p 3 2 k ⟩ G , β ∣ σ = 1 g k ( 2 k ) ! 12 k . (182) P_N(\zeta) = \left\langle e^{p_3/3}\right\rangle_{\mathrm{G},\beta} = \sum_{k\geq0} a_k(N)\,\zeta^{-k}, \qquad\zeta= \frac{2}{3}s^{3/2}, \qquad a_k(N) = \frac{\left\langle p_3^{2k}\right\rangle_{\mathrm{G},\beta}\big|_{\sigma=1}}{g^k(2k)!\,12^k}. \tag*{(182)} P N ( ζ ) = ⟨ e p 3 /3 ⟩ G , β = k ≥ 0 ∑ a k ( N ) ζ − k , ζ = 3 2 s 3/2 , a k ( N ) = g k ( 2 k )! 1 2 k ⟨ p 3 2 k ⟩ G , β σ = 1 . (182)
Here p 3 = ∑ j u j 3 p_3 = \sum_j u_j^3 p 3 = ∑ j u j 3 , and the Gaussian variance parameter is σ = 1 / ( 2 s ) \sigma= 1/(2\sqrt{s}) σ = 1/ ( 2 s ) , so that σ 3 k = ( 12 ζ ) − k \sigma^{3k} = (12\zeta)^{-k} σ 3 k = ( 12 ζ ) − k . The even-moment formula makes a k a_k a k positive for all three symmetry classes, including β = 4 \beta= 4 β = 4 , whose graph representation has signed weights. The obstruction to Borel summation on the positive s s s -axis is the positive-axis singularity associated with the tunnelling saddle described next.
Large-order behaviour. At fixed N N N , τ N ( β ) \tau_N^{(\beta)} τ N ( β ) is a finite-dimensional exponential integral, and the large-order behaviour of its expansion is controlled by the other saddle points. Their relative actions, in units of ζ \zeta ζ , are 2 ℓ 2\ell 2 ℓ for the configuration with ℓ \ell ℓ eigenvalues moved to w + w_+ w + . The nearest, ℓ = 1 \ell= 1 ℓ = 1 , is the one-eigenvalue tunnelling configuration of § 9 . Because the series is not summable directly on the positive s s s -axis, it can be resummed by approaching this axis from either side. The two results differ by i P N τ N ( 1 ) / τ N ( 0 ) iP_N\tau_N^{(1)}/\tau_N^{(0)} i P N τ N ( 1 ) / τ N ( 0 ) , the full one-eigenvalue contribution; on the axis itself it enters only with the factor 1 2 \frac{1}{2} 2 1 found in § 9 . The standard dispersion argument [66 , 6 , 69 ] then gives, with b = β ( N − 1 ) 2 b = \frac{\beta(N-1)}{2} b = 2 β ( N − 1 ) ,
a k ( N ) ≃ 1 2 π N Γ ( 1 + β / 2 ) Γ ( 1 + N β / 2 ) b 2 2 − k [ Γ ( k + b ) + γ N ( β ) Γ ( k + b − 1 ) + O ( Γ ( k + b − 2 ) ) ] . (183) a_k(N) \simeq\frac{1}{2\pi}\frac{N\Gamma(1+\beta/2)}{\Gamma(1+N\beta/2)}\,b^2\,2^{-k}\left[\Gamma(k+b)+\gamma_N^{(\beta)}\Gamma(k+b-1)+O\bigl(\Gamma(k+b-2)\bigr)\right]. \tag*{(183)} a k ( N ) ≃ 2 π 1 Γ ( 1 + N β /2 ) N Γ ( 1 + β /2 ) b 2 2 − k [ Γ ( k + b ) + γ N ( β ) Γ ( k + b − 1 ) + O ( Γ ( k + b − 2 ) ) ] . (183)
The coefficient of the first correction is
γ N ( β ) = 4 3 κ N + K β ( N ) 108 = ( N − 2 ) ( 4 β 2 N 2 − 43 β 2 N + 18 β N + 39 β 2 − 18 β + 20 ) 144 . (184) \gamma_N^{(\beta)} = \frac{4}{3}\kappa_N + \frac{K_\beta(N)}{108} = \frac{(N-2)\left(4\beta^2N^2 - 43\beta^2N + 18\beta N + 39\beta^2 - 18\beta+ 20\right)}{144}. \tag*{(184)} γ N ( β ) = 3 4 κ N + 108 K β ( N ) = 144 ( N − 2 ) ( 4 β 2 N 2 − 43 β 2 N + 18 β N + 39 β 2 − 18 β + 20 ) . (184)
It combines two contributions:
the relative correction κ N \kappa_N κ N to the instanton ratio τ N ( 1 ) / τ N ( 0 ) \tau_N^{(1)}/\tau_N^{(0)} τ N ( 1 ) / τ N ( 0 ) (§ 9 );
the first perturbative coefficient a 1 ( N ) = K β ( N ) / 216 a_1(N) = K_\beta(N)/216 a 1 ( N ) = K β ( N ) /216 , obtained from Eq. (76) at the saddle w − w_- w − .
It is κ N \kappa_N κ N , not c N ( β ) c_N^{(\beta)} c N ( β ) , that enters here: c N ( β ) c_N^{(\beta)} c N ( β ) contains in addition the term − 3 β ( N − 1 ) / 8 -3\beta(N-1)/8 − 3 β ( N − 1 ) /8 , which comes from differentiating ν N \nu_N ν N .
This dispersion argument assumes that the nearest Borel singularity is the one-eigenvalue tunnelling configuration, with the Stokes constant i i i whose half, i / 2 i/2 i /2 , appears in § 9 . The resulting large-order asymptotics is therefore stated at the same physics level as that saddle analysis. For N = 1 N=1 N = 1 , Eq. (183) gives
a k ≃ 2 − k 2 π [ Γ ( k ) − 5 36 Γ ( k − 1 ) ] . a_k \simeq\frac{2^{-k}}{2\pi}\left[\Gamma(k)-\frac{5}{36}\Gamma(k-1)\right]. a k ≃ 2 π 2 − k [ Γ ( k ) − 36 5 Γ ( k − 1 ) ] .
in agreement with the known Airy coefficients, since 5 36 = 2 u 1 \frac{5}{36}=2u_1 36 5 = 2 u 1 with u 1 = 5 72 u_1=\frac{5}{72} u 1 = 72 5 . For N = 2 N=2 N = 2 , Eq. (184) gives γ 2 ( β ) = 0 \gamma_2^{(\beta)}=0 γ 2 ( β ) = 0 : the first subleading Gamma term vanishes.
The Lifshitz tail thus determines the large-order growth of the generating functions of β \beta β -weighted trivalent maps:
The large-order growth is controlled by a single-eigenvalue tunnelling saddle. The only remaining freedom is the direction in channel space along which it does so, a unit vector up to a unit of F \mathbb{F} F . These directions form the projective space R P N − 1 \mathbb{R}P^{N-1} R P N − 1 , C P N − 1 \mathbb{C}P^{N-1} C P N − 1 or H P N − 1 \mathbb{H}P^{N-1} H P N − 1 , of real dimension β ( N − 1 ) \beta(N-1) β ( N − 1 ) .
The number of colours N N N enters the leading large-order constant only through the dimension and the volume of this space of directions. In the Fubini–Study normalization of § 9 , the constant in Eq. (183) relative to N = 1 N=1 N = 1 is Vol ( F P N − 1 ) ( 6 / π ) β ( N − 1 ) / 2 \operatorname{Vol}(\mathbb{F}P^{N-1})(6/\pi)^{\beta(N-1)/2} Vol ( F P N − 1 ) ( 6/ π ) β ( N − 1 ) /2 .
What is counted. Dually, a trivalent map with 2 k 2k 2 k vertices is a surface glued from n = 2 k n=2k n = 2 k triangles. Pair the 3 n 3n 3 n sides of the triangles at random and glue each pair.
For β = 2 \beta=2 β = 2 every pair is glued so that the orientations match, and the surface is orientable.
For β = 1 \beta=1 β = 1 each pair is glued straight or with a flip, with probability 1 2 \frac{1}{2} 2 1 each, so non-orientable surfaces (projective planes, Klein bottles, … \ldots … ) occur. The equal weights follow from the real-symmetric covariance of Sec. 2.
Let F F F be the number of vertices of the glued surface, i.e. of faces of the trivalent map; its Euler characteristic is F − k F-k F − k , so, for a connected orientable surface at fixed k k k , fewer vertices means higher genus. Each vertex carries a matrix index, so the weight N F N^F N F is the number of ways to colour the vertices with N N N colours. By Wick’s theorem, exactly, for β = 1 , 2 \beta=1,2 β = 1 , 2 ,
a k ( N ) a k ( 1 ) = E [ N F ] (185) \frac{a_k(N)}{a_k(1)}=\mathbb{E}\left[N^F\right] \tag*{(185)} a k ( 1 ) a k ( N ) = E [ N F ] (185)
over random gluings of 2 k 2k 2 k triangles (possibly disconnected).
For β = 4 \beta=4 β = 4 the same twisted gluings enter with signed weights, through the β = 1 ↔ 4 \beta=1\leftrightarrow4 β = 1 ↔ 4 duality [72 ] :
a k ( β = 4 ) ( N ) a k ( 1 ) = ( − 2 ) k E [ ( − 2 N ) F ] β = 1 g l u i n g s . (186) \frac{a_k^{(\beta=4)}(N)}{a_k(1)}=(-2)^k\mathbb{E}\left[(-2N)^F\right]_{\beta=1\ \mathrm{gluings}}. \tag*{(186)} a k ( 1 ) a k ( β = 4 ) ( N ) = ( − 2 ) k E [ ( − 2 N ) F ] β = 1 gluings . (186)
Two facts fix this form:
the duality acts on a graph with 2 k 2k 2 k vertices and 3 k 3k 3 k edges by rescalings, which produce a factor c k c^k c k ;
c = − 2 c=-2 c = − 2 follows from the case k = 1 k=1 k = 1 , since K 4 ( N ) = − 2 K 1 ( − 2 N ) K_4(N)=-2K_1(-2N) K 4 ( N ) = − 2 K 1 ( − 2 N ) in Eq. (78) .
The Ewens comparison and twisted-gluing conjecture
With θ = β / 2 \theta=\beta/2 θ = β /2 and n = 2 k n=2k n = 2 k , Eqs. (E2) and (E4) give, for β = 1 , 2 \beta=1,2 β = 1 , 2 and every integer N ≥ 1 N\ge1 N ≥ 1 ,
E [ N F ] = Γ ( θ ) Γ ( N θ ) 6 θ ( N − 1 ) Γ ( k + θ ( N − 1 ) ) Γ ( k ) [ 1 + γ N ( β ) + 5 36 k + O ( k − 2 ) ] ≃ Γ ( θ ) Γ ( N θ ) ( 3 n ) θ ( N − 1 ) . (187) \mathbb{E}[N^F]=\frac{\Gamma(\theta)}{\Gamma(N\theta)}6^{\theta(N-1)}\frac{\Gamma(k+\theta(N-1))}{\Gamma(k)}\left[1+\frac{\gamma_N^{(\beta)}+\frac{5}{36}}{k}+O(k^{-2})\right]\simeq\frac{\Gamma(\theta)}{\Gamma(N\theta)}(3n)^{\theta(N-1)}. \tag*{(187)} E [ N F ] = Γ ( N θ ) Γ ( θ ) 6 θ ( N − 1 ) Γ ( k ) Γ ( k + θ ( N − 1 )) [ 1 + k γ N ( β ) + 36 5 + O ( k − 2 ) ] ≃ Γ ( N θ ) Γ ( θ ) ( 3 n ) θ ( N − 1 ) . (187)
The leading form coincides with the large-n n n asymptotics of the generating function, at z = N z=N z = N , of the number of cycles of a random permutation of 3 n 3n 3 n elements under the Ewens distribution with parameter θ \theta θ [28 ] . That generating function is ∏ j = 0 3 n − 1 ( z θ + j ) / ( θ + j ) \prod_{j=0}^{3n-1}(z\theta+j)/(\theta+j) ∏ j = 0 3 n − 1 ( z θ + j ) / ( θ + j ) . The constant Γ ( θ ) / Γ ( N θ ) \Gamma(\theta)/\Gamma(N\theta) Γ ( θ ) /Γ ( N θ ) is the Ewens normalization; in the instanton language it is Vol ( F P N − 1 ) / π β ( N − 1 ) / 2 \operatorname{Vol}(FP^{N-1})/\pi^{\beta(N-1)/2} Vol ( F P N − 1 ) / π β ( N − 1 ) /2 .
The change from θ = 1 \theta=1 θ = 1 for orientable gluings to θ = 1 2 \theta=\frac{1}{2} θ = 2 1 for gluings with random flips has a combinatorial interpretation. For orientable gluings, the orientation induces a permutation of the 3 n 3n 3 n sides whose cycles correspond to vertices. For gluings with random flips, a useful description instead uses two perfect matchings of the 6 n 6n 6 n side-endpoints: the fixed matching α \alpha α that joins the endpoints meeting at each triangle corner, and the matching μ g l u e \mu_{\mathrm{glue}} μ glue induced by the gluing. Each vertex of the glued surface is one loop in their union.
The matching μ g l u e \mu_{\mathrm{glue}} μ glue is constrained by the pairing of whole sides. If it were uniformly random among all ( 6 n − 1 ) ! ! (6n-1)!! ( 6 n − 1 )!! matchings, the resulting loops would form a random cyclation in the sense of Pippenger [79 ] . Their number would then have the exact generating function
∏ j = 0 3 n − 1 z + 2 j 1 + 2 j , \prod_{j=0}^{3n-1}\frac{z+2j}{1+2j}, j = 0 ∏ 3 n − 1 1 + 2 j z + 2 j ,
equal to that of the cycle count of an Ewens permutation with parameter 1 2 \frac{1}{2} 2 1 .
This comparison motivates the conjecture below. Equation (E6) itself concerns fixed integer z = N ≥ 1 z=N\ge1 z = N ≥ 1 ; extending its asymptotics to non-integer z z z requires additional justification.
For β = 4 \beta=4 β = 4 , Eq. (E6) with θ = 2 \theta=2 θ = 2 holds for the signed sums (E5), whose interpretation as a probability measure on gluings is not evident.
Equation (E6) also predicts the 1 / n 1/n 1/ n corrections. For N = 2 N=2 N = 2 it gives
E [ 2 F ] = 6 k + 5 6 + O ( k − 1 ) \mathbb{E}[2^F]=6k+\frac{5}{6}+O(k^{-1}) E [ 2 F ] = 6 k + 6 5 + O ( k − 1 ) for orientable gluings, against 3 n + 1 = 6 k + 1 3n+1=6k+1 3 n + 1 = 6 k + 1 for a uniform permutation;
E [ 2 F ] = 6 π k [ 1 + 1 72 k + O ( k − 2 ) ] \mathbb{E}[2^F]=\sqrt{6\pi k}\left[1+\frac{1}{72k}+O(k^{-2})\right] E [ 2 F ] = 6 π k [ 1 + 72 k 1 + O ( k − 2 ) ] for twisted gluings, against 6 π k [ 1 + 1 48 k + O ( k − 2 ) ] \sqrt{6\pi k}\left[1+\frac{1}{48k}+O(k^{-2})\right] 6 π k [ 1 + 48 k 1 + O ( k − 2 ) ] for two uniform matchings.
Gluings therefore agree with the uniform models only to leading order.
Relation to known results. For β = 2 \beta=2 β = 2 , the distribution of the vertex count has been studied rigorously. Pippenger and Schleich conjectured E F = log ( 3 n ) + γ \mathbb{E}F=\log(3n)+\gamma E F = log ( 3 n ) + γ and Var F = log ( 3 n ) + γ − π 2 / 6 \operatorname{Var}F=\log(3n)+\gamma-\pi^2/6 Var F = log ( 3 n ) + γ − π 2 /6 from simulations, and proved E F = log n + O ( 1 ) \mathbb{E}F=\log n+O(1) E F = log n + O ( 1 ) [78 ] . Their conjectures were confirmed in certain cases in Ref. [31 ] . Chmutov and Pittel encoded the glued surface by a permutation of the sides, whose cycles are the vertices, and showed that it is asymptotically uniform on the alternating group or on its complement, depending on parity [14 ] . They deduced a local central limit theorem for the number of vertices and a Gaussian genus with variance 1 4 log ( 3 n ) \frac{1}{4}\log(3n) 4 1 log ( 3 n ) for triangles.
For a uniform permutation of 3 n 3n 3 n elements, conditioning on parity does not change the moment E [ N F ] \mathbb{E}[N^F] E [ N F ] at an integer 1 ≤ N < 3 n 1\le N<3n 1 ≤ N < 3 n : the signed generating polynomial ∑ π sgn ( π ) z K ( π ) = z ( z − 1 ) ⋯ ( z − 3 n + 1 ) \sum_{\pi}\operatorname{sgn}(\pi)z^{K(\pi)}=z(z-1)\cdots(z-3n+1) ∑ π sgn ( π ) z K ( π ) = z ( z − 1 ) ⋯ ( z − 3 n + 1 ) vanishes there. The uniform model therefore gives
Γ ( 3 n + N ) Γ ( N ) Γ ( 3 n + 1 ) ≃ ( 3 n ) N − 1 Γ ( N ) , (188) \frac{\Gamma(3n+N)}{\Gamma(N)\Gamma(3n+1)}\simeq\frac{(3n)^{N-1}}{\Gamma(N)}, \tag*{(188)} Γ ( N ) Γ ( 3 n + 1 ) Γ ( 3 n + N ) ≃ Γ ( N ) ( 3 n ) N − 1 , (188)
precisely the leading expression predicted by Eq. (187) at θ = 1 \theta= 1 θ = 1 . This agreement is consistent with asymptotic uniformity, but is not by itself an independent proof of the instanton prefactor: transferring exponential moments from the uniform model to the gluing model requires additional tail control, beyond convergence in total variation.
For β = 1 \beta= 1 β = 1 we have not found the corresponding statement in the literature, and we formulate it as a conjecture.
Conjecture (twisted gluings). Let n n n be even. Glue n n n triangles by a uniformly random pairing of their 3 n 3n 3 n sides, each pair glued straight or with a flip with probability 1 2 \frac{1}{2} 2 1 , and let F n F_n F n be the number of vertices of the resulting surface. Then
E [ z F n ] = Γ ( 1 / 2 ) Γ ( z / 2 ) ( 3 n ) ( z − 1 ) / 2 ( 1 + o ( 1 ) ) , \mathbb{E}\left[z^{F_n}\right] = \frac{\Gamma(1/2)}{\Gamma(z/2)}(3n)^{(z-1)/2}(1+o(1)), E [ z F n ] = Γ ( z /2 ) Γ ( 1/2 ) ( 3 n ) ( z − 1 ) /2 ( 1 + o ( 1 )) , uniformly for z z z in a complex neighbourhood of z = 1 z = 1 z = 1 , and pointwise for z > 0 z > 0 z > 0 . Thus the conjecture predicts the same leading probability-generating-function asymptotics as the cycle count of an Ewens( 1 2 ) (\frac{1}{2}) ( 2 1 ) random permutation of 3 n 3n 3 n elements.
For integer z = N ≥ 1 z = N \ge1 z = N ≥ 1 the leading asymptotics, together with its 1 / n 1/n 1/ n correction (E6), follows from the spectral calculation at the physics level; the conjectural part is the extension to non-integer z z z . If the asymptotic form holds uniformly for z z z near 1 1 1 , a standard result on generating functions of this type [55 ] yields E F n = 1 2 log ( 3 n ) + 1 2 ( γ + 2 log 2 ) + o ( 1 ) \mathbb{E}F_n = \frac{1}{2}\log(3n) + \frac{1}{2}(\gamma+ 2\log2) + o(1) E F n = 2 1 log ( 3 n ) + 2 1 ( γ + 2 log 2 ) + o ( 1 ) and Var F n = 1 2 log ( 3 n ) + 1 2 ( γ + 2 log 2 ) − π 2 8 + o ( 1 ) \operatorname{Var} F_n = \frac{1}{2}\log(3n) + \frac{1}{2}(\gamma+ 2\log2) - \frac{\pi^2}{8} + o(1) Var F n = 2 1 log ( 3 n ) + 2 1 ( γ + 2 log 2 ) − 8 π 2 + o ( 1 ) , together with a central limit theorem for F n F_n F n . These are exactly the values for a uniformly random second matching, for which the mean and variance are ∑ j = 1 3 n 1 2 j − 1 \sum_{j=1}^{3n}\frac{1}{2j-1} ∑ j = 1 3 n 2 j − 1 1 and ∑ j = 1 3 n [ 1 2 j − 1 − 1 ( 2 j − 1 ) 2 ] \sum_{j=1}^{3n}\left[\frac{1}{2j-1}-\frac{1}{(2j-1)^2}\right] ∑ j = 1 3 n [ 2 j − 1 1 − ( 2 j − 1 ) 2 1 ] .
The conjecture asserts that the side-pairing constraint on μ g l u e \mu_{\mathrm{glue}} μ glue becomes irrelevant for the leading loop-count asymptotics. It is an analogue, at the level of these generating functions, of the orientable near-uniformity results. The constraint remains visible in the 1 / n 1/n 1/ n corrections displayed above. There is no corresponding orientability-imposed parity constraint on the vertex count in the twisted model.
This is a concrete prediction for randomly twisted gluings. A direct combinatorial derivation, possibly using the methods of Ref. [14 ] , would test the spectral large-order argument independently. Comparisons with refined recursion should first identify the same map weights and specialization. Likewise, comparison with large-genus intersection-number asymptotics [29 ] requires matching the sum over marked points and the scalar external-field specialization.
Non-degenerate thresholds and the external matrix
When each transverse mode i i i has its own threshold ϵ i \epsilon_i ϵ i , the scalar energy E E E is replaced by the diagonal matrix E = E I − diag ( ϵ 1 , … , ϵ N ) E = EI - \operatorname{diag}(\epsilon_1,\ldots,\epsilon_N) E = E I − diag ( ϵ 1 , … , ϵ N ) , which no longer commutes with Z Z Z . The local identities in § 14 and § 15 extend directly: the equation for P P P is independent of E E E , the drift identities use only ∇ X Tr ( E X ) = E \nabla_X\operatorname{Tr}(EX) = E ∇ X Tr ( E X ) = E , and the factorization enters only through Tr ( E Y ) \operatorname{Tr}(EY) Tr ( E Y ) . Continuing the outgoing integration-surface argument from equal thresholds at fixed n > 0 n > 0 n > 0 gives the stationary distribution (3) with Tr U E ( X ) = Tr ( E X ) + 1 3 Tr X 3 \operatorname{Tr}U E(X) = \operatorname{Tr}(EX) + \frac{1}{3}\operatorname{Tr}X^3 Tr U E ( X ) = Tr ( E X ) + 3 1 Tr X 3 and Z N ∝ ∫ N ( E + i X ) 1 2 \mathcal{Z}_N \propto\int_N(E + iX)^{\frac{1}{2}} Z N ∝ ∫ N ( E + i X ) 2 1 .
For β = 2 \beta= 2 β = 2 , J N \mathcal{J}_N J N reduces to Kontsevich’s determinant J N ∝ det [ A j − 1 ( i ) ( s i ) ] / Δ ( s ) \mathcal{J}_N \propto\det[A_{j-1}^{(i)}(s_i)]/\Delta(s) J N ∝ det [ A j − 1 ( i ) ( s i )] /Δ ( s ) with s i = − ( E − ϵ i + i η ) / T 2 / 3 s_i = -(E-\epsilon_i+i\eta)/T^{2/3} s i = − ( E − ϵ i + i η ) / T 2/3 , and shifting W = − Λ + U W = -\Lambda+ U W = − Λ + U with Λ = diag ( s 1 , … , s N ) \Lambda= \operatorname{diag}(\sqrt{s_1},\ldots,\sqrt{s_N}) Λ = diag ( s 1 , … , s N ) recovers the external-field Kontsevich model with Miwa times proportional to Tr Λ − 2 j + 1 \operatorname{Tr}\Lambda^{-2j+1} Tr Λ − 2 j + 1 . What is lost compared with equal thresholds is everything relying on rotational invariance, which Tr ( E Y ) \operatorname{Tr}(EY) Tr ( E Y ) breaks: the eigenvalue form (50) and the β = 1 , 4 \beta= 1,4 β = 1 , 4 reductions of § 7 and § 11 . Physically, differentiating the integrated density of states with respect to ϵ i \epsilon_i ϵ i gives the channel-resolved density of states ρ i = − ∂ ϵ i ν N \rho_i = -\partial_{\epsilon_i}\nu_N ρ i = − ∂ ϵ i ν N by the Hellmann–Feynman theorem. For β = 1 , 4 \beta= 1,4 β = 1 , 4 , the corresponding external-field integrals generate non-orientable graphs [72 , 30 ] , now equipped with the outgoing contour prescription and a positive stationary normalization; we leave any detailed comparison with unoriented gravity models [84 , 87 , 20 ] open.