Introduction
The total bandwidth
For a rational parameter α = p / q \alpha=p/q α = p / q , the spectrum of the critical Harper operator consists of finitely many bands. Their number increases with the denominator q q q , and their total length tends to zero. Thouless’s conjecture predicts the precise rate: the denominator times the total length tends to a universal constant involving Catalan’s constant. We prove this prediction uniformly in the numerator.
We use the normalization
( H α , ϑ u ) n = u n + 1 + u n − 1 + 2 cos ( 2 π α n + ϑ ) u n , u ∈ ℓ 2 ( Z ) . (1) (H_{\alpha,\vartheta}u)_n=u_{n+1}+u_{n-1}+2\cos(2\pi\alpha n+\vartheta)u_n,\qquad u\in\ell^2(\mathbb Z). \tag*{(1)} ( H α , ϑ u ) n = u n + 1 + u n − 1 + 2 cos ( 2 π α n + ϑ ) u n , u ∈ ℓ 2 ( Z ) . (1)
The coefficients of u n − 1 u_{n-1} u n − 1 and u n + 1 u_{n+1} u n + 1 are one, and the coefficient of the cosine is two. The real parameters are α \alpha α and ϑ \vartheta ϑ . For relatively prime integers 1 ≤ p < q 1\le p<q 1 ≤ p < q , define
S p , q = ⋃ ϑ ∈ R / 2 π Z σ ( H p / q , ϑ ) . (2) S_{p,q}=\bigcup_{\vartheta\in\mathbb R/2\pi\mathbb Z}\sigma(H_{p/q,\vartheta}). \tag*{(2)} S p , q = ϑ ∈ R /2 π Z ⋃ σ ( H p / q , ϑ ) . (2)
Here σ ( H ) \sigma(H) σ ( H ) denotes the spectrum of H H H . The union over ϑ \vartheta ϑ is part of the definition. Its Lebesgue measure ∣ S p , q ∣ |S_{p,q}| ∣ S p , q ∣ is the total bandwidth considered here.
Theorem 1.1 (Thouless’s bandwidth asymptotic). For the normalization (1) ,
lim q → ∞ sup 1 ≤ p < q ( p , q ) = 1 ∣ q ∣ S p , q ∣ − 32 G π ∣ = 0 , G = ∑ j = 0 ∞ ( − 1 ) j ( 2 j + 1 ) 2 . (3) \lim_{q\to\infty}\sup_{\substack{1\le p<q\\(p,q)=1}}\left|q|S_{p,q}|-\frac{32G}{\pi}\right|=0,\qquad G=\sum_{j=0}^{\infty}\frac{(-1)^j}{(2j+1)^2}. \tag*{(3)} q → ∞ lim 1 ≤ p < q ( p , q ) = 1 sup q ∣ S p , q ∣ − π 32 G = 0 , G = j = 0 ∑ ∞ ( 2 j + 1 ) 2 ( − 1 ) j . (3) Here G G G is Catalan’s constant.
The constant is 32 G / π = 9.329948928 … 32G/\pi=9.329948928\ldots 32 G / π = 9.329948928 … . In particular, the result holds along every sequence of reduced rational parameters approaching an irrational parameter, without any assumption on its continued fraction. The uniform statement (3) is stronger than this usual formulation.
Historical background
The bandwidth question grew out of the spectral theory of the Harper equation. Harper introduced the difference equation in his study of electrons on a lattice in a uniform magnetic field [8 ] . For a rational parameter, Chambers separated the dependence of the finite determinant on the two phase variables [5 ] . This relation reduces the union of the spectra over the phases to the inverse image of an interval under a single polynomial. Hofstadter’s calculations subsequently displayed the hierarchy of bands and gaps as the rational parameter varies [10 ] . The total bandwidth extracts one numerical quantity from this intricate set: the sum of the lengths of all its bands.
Thouless’s 1983 paper [23 ] studied this quantity both analytically and numerically. At the critical coupling, his computations suggested decay proportional to the inverse denominator, with a leading constant close to 9.33 over a broad range of numerators. His later analysis of the numerator-one regime identified the candidate 32 G / π 32G/\pi 32 G / π [24 ] . Here the rational parameter tends to zero as 1 / q 1/q 1/ q , so the argument concerns a particular family of long-period operators. Thouless and Tan investigated corrections to this scaling [25 ] ; the slow corrections in the numerator-one case help explain the difficulty of extrapolating the constant numerically. Their comparison with numerator two also emphasized that the approach to the limit can depend strongly on arithmetic features of the rational parameter.
Avron, van Mouche and Simon established important exact identities and estimates for spectral measures and rational approximation [3 , 4 ] . Their analysis distinguishes the union of the spectra over the phase from their intersection and gives, in particular, precise results away from the critical coupling. This distinction matters for the bandwidth problem, whose rational spectrum is the union. Building on their intersection-spectrum identity, Last and Wilkinson proved the exact sum rule [18 ]
∑ P p , q ( E ) = 0 1 ∣ P p , q ′ ( E ) ∣ = 1 q . \sum_{P_{p,q}(E)=0} \frac{1}{\lvert P'_{p,q}(E)\rvert} = \frac{1}{q}. P p , q ( E ) = 0 ∑ ∣ P p , q ′ ( E )∣ 1 = q 1 .
Here P p , q P_{p,q} P p , q denotes the Chambers polynomial, defined in (4) below. It fixes the central value exactly and yielded the lower bound ∣ S p , q ∣ ≥ 4 / q \lvert S_{p,q}\rvert\ge4/q ∣ S p , q ∣ ≥ 4/ q . The higher Taylor coefficients of the coarea density, which are needed to recover the leading constant from this local information, are additional quantities.
Last then proved the general bounds [16 ]
2 ( 5 + 1 ) q < ∣ S p , q ∣ < 8 e q . \frac{2(\sqrt{5}+1)}{q} < \lvert S_{p,q}\rvert< \frac{8e}{q}. q 2 ( 5 + 1 ) < ∣ S p , q ∣ < q 8 e .
These bounds established the correct order of magnitude uniformly in the reduced numerator. The upper bound was also central to his proof of zero spectral measure at the critical coupling for a full-measure class of irrational parameters. The determination of the leading constant required finer information. Helffer and Kerdelhue supplied rigorous semiclassical proofs of the Thouless asymptotic for 1 / q 1/q 1/ q and for 2 / q 2/q 2/ q with q q q odd [9 ] . Their proof analyzes the spectrum near the saddle energy of the symbol cos x + cos ξ \cos x+\cos\xi cos x + cos ξ , where the principal contribution to the bandwidth is concentrated. Their operator has half the energy scale used here, so their constant 16 G / π 16G/\pi 16 G / π is the same constant after normalization. In both theorems the numerator is fixed while the denominator tends to infinity.
Krasovsky obtained exact formulas and asymptotic information for the derivative of the discriminant at the central energy [14 ] . For arbitrary fixed numerator and increasing odd denominator, his expansion gives detailed information about the central band and the clustering of nearby bands. It concerns the discriminant at one energy, whereas the bandwidth involves all inverse branches over the full interval [ − 4 , 4 ] [-4,4] [ − 4 , 4 ] . Ouvry and Wu later proposed extensions of the Thouless formula to higher moments for numerator one [19 ] . Their moments are alternating sums of powers of band-edge energies, and their asymptotic analysis has numerical support; the paper explicitly leaves a firmer proof of the higher-moment extension open. These moments differ from the derivatives of the coarea density considered here.
A substantial improvement of the general upper bound was obtained by Jitomirskaya, Konstantinov and Krasovsky [12 ] :
∣ S p , q ∣ ≤ 4 π q . |S_{p,q}| \le\frac{4\pi}{q}. ∣ S p , q ∣ ≤ q 4 π .
Their proof uses a different gauge, a corresponding Chambers-type formula and matrix eigenvalue inequalities. Figueras and Puig’s recent computer-assisted work on open gaps also examines bandwidth scaling and stronger finite-denominator inequalities [7 ] . Its validated finite computations give useful evidence while leaving the asymptotic law as a conjecture. An asymptotic statement by itself does not establish those stronger finite-denominator inequalities.
The rational bandwidth problem is related to, but more quantitative than, the vanishing measure of the irrational critical spectrum. Avila and Krikorian established the measure formula for every irrational parameter [2 ] ; Avila and Jitomirskaya proved the Cantor-spectrum assertion of the Ten Martini Problem [1 ] . Neither assertion specifies the leading rate of decay of rational bandwidths. The bandwidth conjecture is often formulated along rational sequences approaching a fixed irrational parameter. The formulation in Theorem 1.1 is uniform over the reduced numerators. It therefore also encompasses the fixed-numerator regimes described above. The contact theorem supplies the local information needed for this uniform conclusion, and the integral estimate in Section 2 transfers that information to the total bandwidth.
A local question behind the total bandwidth
Chambers’ formula gives a monic real polynomial P p , q P_{p,q} P p , q of degree q q q such that
S p , q = P p , q − 1 ( [ − 4 , 4 ] ) . (4) S_{p,q}=P_{p,q}^{-1}([-4,4]). \tag*{(4)} S p , q = P p , q − 1 ([ − 4 , 4 ]) . (4)
We construct this polynomial explicitly in Section 3. Its zeros are simple and real. A critical value means a number P p , q ( E c ) P_{p,q}(E_c) P p , q ( E c ) with P p , q ′ ( E c ) = 0 P'_{p,q}(E_c)=0 P p , q ′ ( E c ) = 0 . All such values lie outside ( − 4 , 4 ) (-4,4) ( − 4 , 4 ) :
P p , q ′ ( E c ) = 0 ⟹ ∣ P p , q ( E c ) ∣ ≥ 4. P'_{p,q}(E_c)=0 \quad\Longrightarrow\quad|P_{p,q}(E_c)|\ge4. P p , q ′ ( E c ) = 0 ⟹ ∣ P p , q ( E c ) ∣ ≥ 4.
A proof of this basic, well-known fact is given in Subsection 3.3. For ∣ t ∣ < 4 |t|<4 ∣ t ∣ < 4 , set
ρ p , q ( t ) = ∑ E ∈ R P p , q ( E ) = t 1 ∣ P p , q ′ ( E ) ∣ . (5) \rho_{p,q}(t)=\sum_{\substack{E\in\mathbb{R}\\P_{p,q}(E)=t}}\frac{1}{|P'_{p,q}(E)|}. \tag*{(5)} ρ p , q ( t ) = E ∈ R P p , q ( E ) = t ∑ ∣ P p , q ′ ( E ) ∣ 1 . (5)
The variable E E E is the original spectral variable; t = P p , q ( E ) t=P_{p,q}(E) t = P p , q ( E ) is a polynomial level. Thus ρ p , q ( t ) d t \rho_{p,q}(t)\,dt ρ p , q ( t ) d t is the total length on the E E E -axis lying over a small interval of levels of length d t dt d t . The change-of-variables formula gives
∣ S p , q ∣ = ∫ − 4 4 ρ p , q ( t ) d t . (6) |S_{p,q}|=\int_{-4}^{4}\rho_{p,q}(t)\,dt. \tag*{(6)} ∣ S p , q ∣ = ∫ − 4 4 ρ p , q ( t ) d t . (6)
The Last–Wilkinson identity [18 ] is
q ρ p , q ( 0 ) = q ∑ P p , q ( E ) = 0 1 ∣ P p , q ′ ( E ) ∣ = 1. (7) q\rho_{p,q}(0)=q\sum_{\substack{P_{p,q}(E)=0}}\frac{1}{|P'_{p,q}(E)|}=1. \tag*{(7)} q ρ p , q ( 0 ) = q P p , q ( E ) = 0 ∑ ∣ P p , q ′ ( E ) ∣ 1 = 1. (7)
It is natural to seek a local coarea law before taking the integral in (6) : does the normalized density approach one function, independently of the rational parameter? For the bandwidth it is enough to answer the symmetric version of this question. Put
R p , q ( t ) = q 2 ( ρ p , q ( t ) + ρ p , q ( − t ) ) . (8) R_{p,q}(t)=\frac{q}{2}\left(\rho_{p,q}(t)+\rho_{p,q}(-t)\right). \tag*{(8)} R p , q ( t ) = 2 q ( ρ p , q ( t ) + ρ p , q ( − t ) ) . (8)
Then R p , q R_{p,q} R p , q is even, R p , q ( 0 ) = 1 R_{p,q}(0)=1 R p , q ( 0 ) = 1 , and q ∣ S p , q ∣ = ∫ − 4 4 R p , q ( t ) d t q|S_{p,q}|=\int_{-4}^{4}R_{p,q}(t)\,dt q ∣ S p , q ∣ = ∫ − 4 4 R p , q ( t ) d t . Symmetrization avoids imposing an unnecessary parity condition on the original density.
A useful guide is the case of a fixed numerator equal to one. Its limiting matrix calculation, carried out below, reduces to the angular average of ( 1 − t 2 cos 2 ψ / 16 ) − 1 / 2 (1-t^2\cos^2\psi/16)^{-1/2} ( 1 − t 2 cos 2 ψ /16 ) − 1/2 over − π / 2 ≤ ψ ≤ π / 2 -\pi/2\le\psi\le\pi/2 − π /2 ≤ ψ ≤ π /2 .
This suggests looking for the same function at every rational parameter. The Last–Wilkinson normalization (7) already fixes its value at zero. The candidate limiting profile is
G ( t ) = ∑ m = 0 ∞ ( 2 m m ) 2 256 m t 2 m = 2 F 1 ( 1 2 , 1 2 ; 1 ; t 2 16 ) , ∣ t ∣ < 4. (9) \mathcal{G}(t)=\sum_{m=0}^{\infty}\frac{\binom{2m}{m}^{2}}{256^{m}}t^{2m}={}_2F_1\left(\frac{1}{2},\frac{1}{2};1;\frac{t^2}{16}\right),\qquad|t|<4. \tag*{(9)} G ( t ) = m = 0 ∑ ∞ 25 6 m ( m 2 m ) 2 t 2 m = 2 F 1 ( 2 1 , 2 1 ; 1 ; 16 t 2 ) , ∣ t ∣ < 4. (9)
The symbol 2 F 1 {}_2F_1 2 F 1 denotes the Gauss hypergeometric function, defined for ∣ z ∣ < 1 |z|<1 ∣ z ∣ < 1 by
2 F 1 ( a , b ; c ; z ) = ∑ m = 0 ∞ ( a ) m ( b ) m ( c ) m m ! z m , c ∉ { 0 , − 1 , − 2 , … } . {}_2F_1(a,b;c;z)=\sum_{m=0}^{\infty}\frac{(a)_m(b)_m}{(c)_m m!}z^m,\qquad c\notin\{0,-1,-2,\ldots\}. 2 F 1 ( a , b ; c ; z ) = m = 0 ∑ ∞ ( c ) m m ! ( a ) m ( b ) m z m , c ∈ / { 0 , − 1 , − 2 , … } .
Here ( a ) 0 = 1 (a)_0=1 ( a ) 0 = 1 and ( a ) m = a ( a + 1 ) ⋯ ( a + m − 1 ) (a)_m=a(a+1)\cdots(a+m-1) ( a ) m = a ( a + 1 ) ⋯ ( a + m − 1 ) for $m\geq 1; this is the rising factorial. The subscripts indicate two upper parameters a , b a,b a , b and one lower parameter c c c . The square-lattice representation of G \mathcal{G} G makes the proposed universality plausible:
G ( t ) = 1 ( 2 π ) 2 ∫ 0 2 π ∫ 0 2 π d x d y 1 − t 16 ( 2 cos x + 2 cos y ) . (10) \mathcal{G}(t)=\frac{1}{(2\pi)^2}\int_0^{2\pi}\int_0^{2\pi}\frac{dx\,dy}{1-\frac{t}{16}(2\cos x+2\cos y)}. \tag*{(10)} G ( t ) = ( 2 π ) 2 1 ∫ 0 2 π ∫ 0 2 π 1 − 16 t ( 2 cos x + 2 cos y ) d x d y . (10)
Indeed, expansion of this integral counts closed square-lattice walks of even length; their number at length 2 m 2m 2 m is ( 2 m m ) 2 \binom{2m}{m}^{2} ( m 2 m ) 2 . This explains the proposed coefficients, but does not prove that the Harper inverse-slope density has them. Establishing that identification is the main task of the paper.
Rather than attempt a direct limit of all inverse branches on the whole interval, we first examine derivatives at the central level t = 0 t=0 t = 0 . Each derivative can be computed by finite matrix algebra; the positive-real estimate below will recover the local function and justify integration up to the endpoints. We call ρ p , q ( 2 m ) ( 0 ) \rho_{p,q}^{(2m)}(0) ρ p , q ( 2 m ) ( 0 ) the even central contacts ; these are simply the even Taylor derivatives of the coarea density.
Theorem 1.2 (The universal coarea limit). The functions R p , q R_{p,q} R p , q extend holomorphically to ∣ t ∣ < 4 |t|<4 ∣ t ∣ < 4 , and for every 0 < r < 4 0<r<4 0 < r < 4 ,
lim q → ∞ sup 1 ≤ p < q ( p , q ) = 1 sup ∣ t ∣ ≤ r ∣ R p , q ( t ) − G ( t ) ∣ = 0. (11) \lim_{q\to\infty}\sup_{\substack{1\leq p<q\\(p,q)=1}}\sup_{|t|\leq r}|R_{p,q}(t)-\mathcal{G}(t)|=0. \tag*{(11)} q → ∞ lim 1 ≤ p < q ( p , q ) = 1 sup ∣ t ∣ ≤ r sup ∣ R p , q ( t ) − G ( t ) ∣ = 0. (11) Equivalently, the even central contacts have the following limits. For every fixed integer $m\geq 1,
lim q → ∞ sup 1 ≤ p < q ( p , q ) = 1 ∣ q ρ p , q ( 2 m ) ( 0 ) − ( 2 m ) ! ( 2 m m ) 2 256 m ∣ = 0. (12) \lim_{q\to\infty}\sup_{\substack{1\leq p<q\\(p,q)=1}}\left|q\rho_{p,q}^{(2m)}(0)-\frac{(2m)!\binom{2m}{m}^{2}}{256^{m}}\right|=0. \tag*{(12)} q → ∞ lim 1 ≤ p < q ( p , q ) = 1 sup q ρ p , q ( 2 m ) ( 0 ) − 25 6 m ( 2 m )! ( m 2 m ) 2 = 0. (12)
Theorem 1.2 is a corollary of the stronger result in Proposition 8.4, which also gives a uniform error estimate. The coefficient formulation (12) is what the matrix and period calculations prove. Section 2 shows why it is equivalent to (11) and suffices for the bandwidth theorem.
Motivation and guiding ideas
I first encountered Thouless’s bandwidth conjecture in 2021, while reading Last’s survey [17 Section 3.1] in the volume Sturm–Liouville Theory: Past and Present . I immediately suspected a connection with Wu’s counting of square-lattice spanning trees [27 ] , because Catalan’s constant appears in both problems. The matrix-tree theorem suggested considering the square-lattice logarithmic potential
Φ ( t ) = 1 ( 2 π ) 2 ∫ 0 2 π ∫ 0 2 π log ∣ t − 2 cos x − 2 cos y ∣ d x d y . \Phi(t)=\frac{1}{(2\pi)^2}\int_0^{2\pi}\int_0^{2\pi}\log|t-2\cos x-2\cos y|\,dx\,dy. Φ ( t ) = ( 2 π ) 2 1 ∫ 0 2 π ∫ 0 2 π log ∣ t − 2 cos x − 2 cos y ∣ d x d y .
If τ n \tau_n τ n denotes the number of spanning trees of the periodic n × n n\times n n × n square grid, then
lim n → ∞ log τ n n 2 = Φ ( 4 ) = 4 G π . \lim_{n\to\infty}\frac{\log\tau_n}{n^2}=\Phi(4)=\frac{4G}{\pi}. n → ∞ lim n 2 log τ n = Φ ( 4 ) = π 4 G .
Thouless’s proposed bandwidth constant is eight times this quantity.
The heat-kernel and determinant results of Chinta, Jorgenson, and Karlsson [6 ] , with which I was already familiar, gave a further reason to look for a local spectral law. Their main theorem in dimension two gives a refined asymptotic for the logarithm of the product of the nonzero Laplacian eigenvalues on discrete tori; for square grids its leading term per vertex is again 4 G / π 4G/\pi 4 G / π . This suggested a local prediction: the normalized coarea density should tend to 4 ∣ Φ ′ ( t ) ∣ 4|\Phi'(t)| 4∣ Φ ′ ( t ) ∣ , with its value at zero interpreted continuously. The proof below uses the symmetric version of this prediction, which already determines the total bandwidth.
This potential gives exactly the candidate profile G \mathcal{G} G in (9) :
G ( t ) = 4 ∣ Φ ′ ( t ) ∣ , 0 < ∣ t ∣ < 4. \mathcal{G}(t)=4|\Phi'(t)|,\qquad0<|t|<4. G ( t ) = 4∣ Φ ′ ( t ) ∣ , 0 < ∣ t ∣ < 4.
On this interval, expressing Φ ′ \Phi' Φ ′ as the average of ( t − 2 cos x − 2 cos y ) − 1 (t-2\cos x-2\cos y)^{-1} ( t − 2 cos x − 2 cos y ) − 1 requires a Cauchy principal value. Here is an elementary calculation explaining the identity. The two-to-one torus map ( u , v ) ↦ ( u + v , u − v ) (u,v)\mapsto(u+v,u-v) ( u , v ) ↦ ( u + v , u − v ) has Jacobian of absolute value two, so it preserves normalized torus averages and replaces 2 cos x + 2 cos y 2\cos x+2\cos y 2 cos x + 2 cos y by 4 cos u cos v 4\cos u\cos v 4 cos u cos v . The one-dimensional logarithmic-potential formula, applied first in u u u , gives, for 0 < t < 4 0<t<4 0 < t < 4 ,
Φ ( t ) = 2 π ∫ arccos ( t / 4 ) π / 2 arcosh ( t 4 cos v ) d v . \Phi(t)=\frac{2}{\pi}\int_{\arccos(t/4)}^{\pi/2}\operatorname{arcosh}\left(\frac{t}{4\cos v}\right)\,dv. Φ ( t ) = π 2 ∫ a r c c o s ( t /4 ) π /2 arcosh ( 4 cos v t ) d v .
The constant term is zero because the average of log ( 2 ∣ cos u ∣ ) \log(2|\cos u|) log ( 2∣ cos u ∣ ) is zero; in particular, Φ ( 0 ) = 0 \Phi(0)=0 Φ ( 0 ) = 0 . Here arcosh is the inverse hyperbolic cosine. For 0 < b < t 0<b<t 0 < b < t , use arcosh ( t / b ) = ∫ b t ( r 2 − b 2 ) − 1 / 2 d r \operatorname{arcosh}(t/b)=\int_b^t(r^2-b^2)^{-1/2}\,dr arcosh ( t / b ) = ∫ b t ( r 2 − b 2 ) − 1/2 d r . The integrand is nonnegative, allowing the integrations to be interchanged. Substituting cos v = ( r / 4 ) sin θ \cos v=(r/4)\sin\theta cos v = ( r /4 ) sin θ then gives
Φ ′ ( t ) = 1 2 π ∫ 0 π / 2 d θ 1 − ( t 2 / 16 ) sin 2 θ = 1 4 G ( t ) , 0 < t < 4. \Phi'(t)=\frac{1}{2\pi}\int_0^{\pi/2}\frac{d\theta}{\sqrt{1-(t^2/16)\sin^2\theta}}=\frac{1}{4}\mathcal{G}(t),\qquad0<t<4. Φ ′ ( t ) = 2 π 1 ∫ 0 π /2 1 − ( t 2 /16 ) sin 2 θ d θ = 4 1 G ( t ) , 0 < t < 4.
Expanding this last integrand gives the series in (9) . Evenness of Φ \Phi Φ gives the formula for negative t t t . At zero, Φ \Phi Φ has one-sided derivatives Φ ′ ( 0 ± ) = ± 1 / 4 \Phi'(0\pm)=\pm1/4 Φ ′ ( 0 ± ) = ± 1/4 ; thus G ( 0 ) = 1 \mathcal{G}(0)=1 G ( 0 ) = 1 is the continuous extension of 4 ∣ Φ ′ ∣ 4|\Phi'| 4∣ Φ ′ ∣ . In particular, the candidate profile has the required mass
∫ − 4 4 G ( t ) d t = 8 Φ ( 4 ) = 32 G π . \int_{-4}^{4}\mathcal{G}(t)\,dt=8\Phi(4)=\frac{32G}{\pi}. ∫ − 4 4 G ( t ) d t = 8Φ ( 4 ) = π 32 G .
The determinant and closed-walk viewpoints involve the same lattice data. On a periodic square grid, the Laplacian is four times the identity minus the adjacency matrix. The matrix-tree theorem uses the product of its nonzero eigenvalues, while traces of powers of the adjacency matrix count closed walks. Logarithmic determinants and resolvents connect these descriptions: differentiation of a logarithmic determinant with respect to a scalar spectral shift gives a trace of a resolvent. The candidate coarea profile is a diagonal square-lattice resolvent, whose Taylor coefficients are scaled counts of closed walks from a fixed lattice point.
A second motivation was to generalize the Last–Wilkinson identity [18 ] . The exact value q ρ p , ρ q ( 0 ) = 1 q_{\rho_p,\rho_q}(0)=1 q ρ p , ρ q ( 0 ) = 1 is the zeroth-order central contact. Seeking analogous limits for the higher even derivatives at the same central level gives a concrete route to the whole local profile. We encode these contacts in one generating function and determine its coefficients from a differential equation. Here the equation is obtained by identifying the generating function with a period of compact genus-one curves. An elementary positive-real estimate controls the endpoint contributions and recovers the bandwidth limit from the contact limits.
Why the local law determines the total bandwidth
Taylor coefficients determine the density near zero, but an integral over ( − 4 , 4 ) (-4,4) ( − 4 , 4 ) also needs control near its endpoints. The useful finite fact is that the analytically continued inverse-slope sum has positive real part. After symmetrization it is even and normalized at zero. A disk map and Schwarz’s lemma then give the uniform, integrable bound
0 < R ρ p , ρ q ( t ) ≤ 4 16 − t 2 . 0<R_{\rho_p,\rho_q}(t)\leq\frac{4}{\sqrt{16-t^2}}. 0 < R ρ p , ρ q ( t ) ≤ 16 − t 2 4 .
The same complex-analytic estimate gives local boundedness, so convergence of the contacts implies local uniform convergence of the density. Thus one positivity argument handles both the interior and the endpoints; Section 2 gives its proof.
For the optional lower-order calculations it is convenient to use
F p , q ( z ) = R p , q ( 4 z ) = ∑ m ≥ 0 a m ( p , q ) z m , a m ( p , q ) = 16 m q ρ p , q ( 2 m ) ( 0 ) ( 2 m ) ! . (13) F_{p,q}(z)=R_{p,q}(4\sqrt{z})=\sum_{m\geq0}a_m(p,q)z^m,\qquad a_m(p,q)=\frac{16^m q\rho_{p,q}^{(2m)}(0)}{(2m)!}. \tag*{(13)} F p , q ( z ) = R p , q ( 4 z ) = m ≥ 0 ∑ a m ( p , q ) z m , a m ( p , q ) = ( 2 m )! 1 6 m q ρ p , q ( 2 m ) ( 0 ) . (13)
Evenness makes this independent of the choice of square root. The bound just obtained reads
0 ≤ F p , q ( u 2 ) ≤ 1 1 − u 2 , 0 ≤ u < 1. (14) 0\leq F_{p,q}(u^2)\leq\frac{1}{\sqrt{1-u^2}},\qquad0\leq u<1. \tag*{(14)} 0 ≤ F p , q ( u 2 ) ≤ 1 − u 2 1 , 0 ≤ u < 1. (14)
Since q ∣ S p , q ∣ = 8 ∫ 0 1 F p , q ( u 2 ) d u q|S_{p,q}|=8\int_0^1F_{p,q}(u^2)\,\mathrm{d}u q ∣ S p , q ∣ = 8 ∫ 0 1 F p , q ( u 2 ) d u , it permits passage to the limiting integral, which equals 32 G / π 32G/\pi 32 G / π . Appendix C retains a stronger positivity theorem: both F p , q F_{p,q} F p , q and F p , q 2 F_{p,q}^2 F p , q 2 are Stieltjes transforms of probability measures on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Their moment information is used in the optional bounds below.
Section 4 is optional and may be skipped completely. It gives elementary calculations of the zeroth, second, and fourth contacts, placing each bandwidth consequence beside its contact. Together with the finite positivity theorem, the Last–Wilkinson identity for the zeroth contact yields 8 ≤ q ∣ S p , q ∣ ≤ 4 π 8\leq q|S_{p,q}|\leq4\pi 8 ≤ q ∣ S p , q ∣ ≤ 4 π . The fourth contact improves these bounds, for all sufficiently large q q q , to 9.083091 ≤ q ∣ S p , q ∣ ≤ 10.076410 9.083091\leq q|S_{p,q}|\leq10.076410 9.083091 ≤ q ∣ S p , q ∣ ≤ 10.076410 . The complete sums of reciprocal even powers of sines appearing in these calculations are also special values of spectral zeta functions of cycle graphs. At higher orders, however, several coupled indices and ordering restrictions produce cancellations that complete trigonometric sums do not determine. These difficulties motivate treating all contact orders together through one generating function and one differential equation.
The proof through one normalized period
The main task is to identify one analytic function, rather than to evaluate each contact separately. The proof has four stages.
An exact finite formula. Section 3 expresses q ρ p , q ( t ) q\rho_{p,q}(t) q ρ p , q ( t ) as a matrix entry and moves it to a cut of a tridiagonal matrix. The signs of the inverse derivatives are retained by a commuting involution. This is the algebraic link between the spectral density and the matrices used in the limit.
A justified fixed-numerator limit. For a fixed odd numerator p p p , Appendix A proves convergence to an angular average G p ( t ) G_p(t) G p ( t ) of entries of a 2 p × 2 p 2p\times2p 2 p × 2 p matrix function. Its statement appears in the main text as Proposition 5.1. Strong operator convergence alone would not justify the entries sampled by the moving cut. The appendix first subtracts an exact singular contribution, proves local L 2 L^2 L 2 convergence of the remaining rows, and cancels the one-sided Fourier terms before computing the normalization. § 5 then turns the average into a scalar integral by commuting matrices, determinant differentiation, and Jensen’s formula.
The same period equation for every odd numerator. The scalar determinant defines level curves. Its explicit boundary polynomials determine a compact family, its singular levels, and a fixed differential form. Coarea identifies G p G_p G p with a closed period, including the normalization inherited from the matrix entries. § 6 verifies the compact family, and § 7 determines its period equation.
A period is an integral of a differential form over a closed curve. In an algebraic family, parameter derivatives of differentials belong, modulo exact forms, to a finite-dimensional space. A linear relation among these derivatives becomes a differential equation after integration: exact forms integrate to zero. For genus one this space has dimension two; see [26 ] , § 1.2 and [11 ] , Chapter 10. In the present family, explicit nodal-annulus calculations determine the poles and zeros of the Wronskian of two periods. They fix the differential equation, whose normalized analytic solution is G G G . Its coefficient recurrence then gives all contacts at once. The passage from the physical integral to this particular normalized period is verified before using the equation.
From fixed numerators to all rational parameters. § 8 uses the inverse of p p p modulo q q q to express each contact as a convergent sum over closed paths with Cauchy kernels. Uniform tail bounds give continuity at irrational parameters and one-sided limits at rational parameters. The
exact rational Fourier identity then controls rational point values. Only odd fixed numerators are needed: if b / d ∈ ( 0 , 1 ) b/d \in(0,1) b / d ∈ ( 0 , 1 ) is reduced with odd d d d , choose arbitrarily large positive integers q q q , a a a with d a = b q + 1 da=bq+1 d a = b q + 1 . The inverse of d d d modulo q q q is a a a , and
a q = b d + 1 d q . \frac{a}{q}=\frac{b}{d}+\frac{1}{dq}. q a = d b + d q 1 .
Thus fixed odd numerator d d d determines a one-sided limit at b / d b/d b / d in the inverse parameter. These rationals are dense. The tail bounds therefore extend the fixed-numerator evaluations to all numerators, with both denominator parities. Proposition 8.4 supplies the uniform error estimate; Section 2 then concludes the bandwidth theorem.
For the convenience of readers unfamiliar with algebraic geometry, we give the constructions in explicit coordinates. Polynomial factorization, point blowups, and intersections on a product of projective lines determine the compact family, at the level of Chapters 1–2 of [20 ] . The additional Riemann-surface arguments—holomorphic differentials, closed periods, and integration in nodal annuli—are explained where used. Experts in algebraic geometry may readily see shorter formulations of some arguments in § 6 and § 7 ; we keep the coordinate and period calculations explicit for readers with less background. The singular matrix limit in Appendix A is logically independent of the period evaluation.
Notation
The integers p p p , q q q are relatively prime unless stated otherwise, and G G G always denotes Catalan’s constant. The symbols I I I and e j e_j e j denote the identity matrix and a standard column coordinate vector, with the dimension and the starting index specified where they are used. For a matrix X X X , X T X^{T} X T is its transpose and X ∗ X^{*} X ∗ its conjugate transpose; [ X , Y ] = X Y − Y X [X,Y]=XY-YX [ X , Y ] = X Y − Y X . We write [ t n ] f ( t ) [t^{n}]f(t) [ t n ] f ( t ) for the coefficient of t n t^{n} t n in a Taylor or formal power series. The fractional part is { x } = x − ⌊ x ⌋ \{x\}=x-\lfloor x\rfloor { x } = x − ⌊ x ⌋ , and [ r ] q [r]_q [ r ] q is the least nonnegative residue of an integer r r r modulo q q q . The inverse of p p p modulo q q q , when denoted by p ˉ \bar{p} p ˉ , is chosen in { 1 , … , q − 1 } \{1,\ldots,q-1\} { 1 , … , q − 1 } . All changes in the roles of a fixed numerator and a varying denominator are stated explicitly. The notation adj X \operatorname{adj} X adj X means the classical adjugate matrix, and Tr X \operatorname{Tr} X Tr X its trace. We write PV \operatorname{PV} PV for a Cauchy principal value, sech x = 1 / cosh x \operatorname{sech} x=1/\cosh x sech x = 1/ cosh x , and csch x = 1 / sinh x \operatorname{csch} x=1/\sinh x csch x = 1/ sinh x . An indicator 1 A \mathbf{1}_{A} 1 A is one when condition A A A holds and zero otherwise.
Reading the proof
Section 2 derives the bandwidth theorem from the contacts. Section 4 is optional and may be skipped completely. Sections 3 and 5 give the matrix reduction; Sections 6–7 evaluate the resulting integral. Section 8 proves numerator uniformity. Appendix A justifies the analytic limit, and Appendix B verifies the commuting identity using two-by-two blocks. Appendix C contains the stronger positivity result used only for the optional bounds.
Dependency diagram for the two main theorems
The arrows below mean that a result is used in the next box. All proposition and theorem numbers are links to their statements. The optional elementary section is outside this chain.
Figure 2. Exact finite generator. Proposition 3.2: q p = X 11 q_p = X_{11} q p = X 11 and the cut identity. Fixed odd numerator reduction. Proposition 5.1. Appendix A: reciprocal identity (A.15); principal algebra Lemma A.2; analytic-row Lemma A.3; scalar-shift Lemma A.4; cut normalization (A.30). Independent evaluation of G p G_p G p . Proposition 5.2 and (5.16): the scalar integral. Proposition 6.5: genus one; Proposition 7.2: integral equals a closed period; Theorem 7.4: period ODE. These prove Theorem 5.3. Evaluated fixed-numerator contacts. Corollary 8.1 combines Proposition 5.1 and Theorem 5.3. From fixed numerators to uniform rational contacts. Lemma 8.2 and Proposition 8.3: exact identity at every rational. Closed-path bound (8.5): continuity, one-sided limits, and rational point error. Odd-denominator density then proves Proposition 8.4. Theorem 1.2. The even density, equivalently all its central contacts, converges uniformly in the numerator. Theorem 1.1. Coefficient limits + Lemma 2.1 + Proposition 2.2. Positive real part and evenness supply the endpoint bound; the limiting integral equals 32 G / π 32G/\pi 32 G / π .
Independent reductions. Appendix A proves the fixed-numerator reduction without evaluating G p G_p G p . Its reciprocal identity, the rational transfer, and Lemma 2.1 use only finite matrices, Fourier series, and the polynomial properties established in Section 3.
Figure 1. The universal normalized even coarea profile (left) and the monotone limiting bandwidth sums B M = 8 ∑ m = 0 M ( 2 m ) 2 / [ 16 m ( 2 m + 1 ) ] B_M = 8\sum_{m=0}^{M}(2^m)^2/[16^m(2m+1)] B M = 8 ∑ m = 0 M ( 2 m ) 2 / [ 1 6 m ( 2 m + 1 )] (right). The horizontal variable t t t is the polynomial level P p , q ( E ) P_{p,q}(E) P p , q ( E ) . The profile is plotted away from its logarithmic endpoints. The uniform endpoint estimate is needed to infer the full bandwidth from convergence near t = 0 t=0 t = 0 .
We first prove the bandwidth theorem assuming only the coefficient limits (12) . The limit theorem itself is proved later. Three finite-polynomial facts suffice here: P p , q P_{p,q} P p , q has simple real zeros, no critical value in ( − 4 , 4 ) (-4,4) ( − 4 , 4 ) , and q ρ p , q ( 0 ) = 1 q\rho_{p,q}(0)=1 q ρ p , q ( 0 ) = 1 . Section 3 proves all three directly.
There are two issues: Taylor coefficients must determine the density throughout ( − 4 , 4 ) (-4,4) ( − 4 , 4 ) , and mass must not accumulate at its endpoints. One elementary positive-real estimate resolves both. The stronger Stieltjes representations used for the optional lower-order bounds are proved separately in Appendix C.
A positive-real estimate
Put
Ω = C ∖ ( ( − ∞ , − 4 ] ∪ [ 4 , ∞ ) ) . \Omega= \mathbb{C}\setminus\left((-\infty,-4]\cup[4,\infty)\right). Ω = C ∖ ( ( − ∞ , − 4 ] ∪ [ 4 , ∞ ) ) .
For a polynomial level z ∈ Ω z\in\Omega z ∈ Ω , absolute values in the real inverse-slope sum are replaced by fixed signs of its inverse branches. The following proof explains that continuation before estimating it.
Lemma 2.1 (An integrable bound from positivity). Let P P P be a monic polynomial of degree q ≥ 2 q\ge2 q ≥ 2 with simple real zeros and no critical value in ( − 4 , 4 ) (-4,4) ( − 4 , 4 ) . Let ρ ( t ) = ∑ P ( E ) = t ∣ P ′ ( E ) ∣ − 1 \rho(t)=\sum_{P(E)=t}|P'(E)|^{-1} ρ ( t ) = ∑ P ( E ) = t ∣ P ′ ( E ) ∣ − 1 for real ∣ t ∣ < 4 |t|<4 ∣ t ∣ < 4 , and suppose q ρ ( 0 ) = 1 q\rho(0)=1 q ρ ( 0 ) = 1 . The function R ( t ) = q ( ρ ( t ) + ρ ( − t ) ) / 2 R(t)=q(\rho(t)+\rho(-t))/2 R ( t ) = q ( ρ ( t ) + ρ ( − t )) /2 extends to an even holomorphic function on Ω \Omega Ω , with positive real part and R ( 0 ) = 1 R(0)=1 R ( 0 ) = 1 . Set
ϕ ( z ) = z 4 + 16 − z 2 , 16 − 0 2 = 4. \phi(z)=\frac{z}{4+\sqrt{16-z^2}},\qquad\sqrt{16-0^2}=4. ϕ ( z ) = 4 + 16 − z 2 z , 16 − 0 2 = 4. Then
∣ R ( z ) ∣ ≤ 1 + ∣ ϕ ( z ) ∣ 2 1 − ∣ ϕ ( z ) ∣ 2 ( z ∈ Ω ) , 0 < R ( t ) ≤ 4 16 − t 2 ( − 4 < t < 4 ) . (15) |R(z)|\le\frac{1+|\phi(z)|^2}{1-|\phi(z)|^2}\quad(z\in\Omega),\qquad0<R(t)\le\frac{4}{\sqrt{16-t^2}}\quad(-4<t<4). \tag*{(15)} ∣ R ( z ) ∣ ≤ 1 − ∣ ϕ ( z ) ∣ 2 1 + ∣ ϕ ( z ) ∣ 2 ( z ∈ Ω ) , 0 < R ( t ) ≤ 16 − t 2 4 ( − 4 < t < 4 ) . (15) In particular these normalized functions form a locally bounded family on Ω \Omega Ω , independently of P P P and q q q .
Proof. Write the zeros as e 1 < ⋯ < e q e_1<\cdots<e_q e 1 < ⋯ < e q and put σ j = sgn P ′ ( e j ) \sigma_j=\operatorname{sgn}P'(e_j) σ j = sgn P ′ ( e j ) . Rolle’s theorem gives all q − 1 q-1 q − 1 critical points: one in each ( e j , e j + 1 ) (e_j,e_{j+1}) ( e j , e j + 1 ) . They are real, and the hypothesis places their values on the removed rays. Since Ω \Omega Ω is simply connected and contains no critical value, the inverse branches E j ( z ) E_j(z) E j ( z ) with E j ( 0 ) = e j E_j(0)=e_j E j ( 0 ) = e j continue holomorphically there. For real ∣ t ∣ < 4 |t|<4 ∣ t ∣ < 4 the branches remain real and simple: a departure from the real axis or a collision would require a critical value. Their derivatives therefore retain their initial signs. Hence
ρ ( z ) = ∑ j = 1 q σ j E j ′ ( z ) \rho(z)=\sum_{j=1}^{q}\sigma_j E_j'(z) ρ ( z ) = j = 1 ∑ q σ j E j ′ ( z )
is the required analytic continuation of the positive real density.
If Im z > 0 \operatorname{Im} z > 0 Im z > 0 , the roots E j ( z ) E_j(z) E j ( z ) with σ j = 1 \sigma_j = 1 σ j = 1 are exactly the roots in the upper E E E -half-plane. Indeed, this is true at z = i ε z = i\varepsilon z = i ε because E j ( i ε ) = e j + i ε / P ′ ( e j ) + O ( ε 2 ) E_j(i\varepsilon) = e_j + i\varepsilon/P'(e_j) + O(\varepsilon^2) E j ( i ε ) = e j + i ε / P ′ ( e j ) + O ( ε 2 ) , and no root can cross the real axis while z z z is nonreal. The sum of all inverse derivatives is zero: the sum of the q q q roots of P ( E ) − z P(E) - z P ( E ) − z is independent of z z z because q ≥ 2 q \ge2 q ≥ 2 . Closing a contour in the upper E E E -half-plane therefore gives
ρ ( z ) = 2 ∑ Im E j ( z ) > 0 1 P ′ ( E j ( z ) ) = 1 π i ∫ R d E P ( E ) − z . (16) \rho(z) = 2 \sum_{\operatorname{Im} E_j(z)>0} \frac{1}{P'(E_j(z))} = \frac{1}{\pi i} \int_{\mathbb{R}} \frac{dE}{P(E)-z}. \tag*{(16)} ρ ( z ) = 2 Im E j ( z ) > 0 ∑ P ′ ( E j ( z )) 1 = π i 1 ∫ R P ( E ) − z d E . (16)
The integral is absolutely convergent, and the large semicircle vanishes because its integrand is O ( ∣ E ∣ − q ) O(|E|^{-q}) O ( ∣ E ∣ − q ) . If z = x + i y z = x + iy z = x + i y , y > 0 y > 0 y > 0 , taking real parts yields
Re ρ ( x + i y ) = 1 π ∫ R y d E ( P ( E ) − x ) 2 + y 2 > 0. \operatorname{Re} \rho(x+iy) = \frac{1}{\pi} \int_{\mathbb{R}} \frac{y\,dE}{(P(E)-x)^2+y^2} > 0. Re ρ ( x + i y ) = π 1 ∫ R ( P ( E ) − x ) 2 + y 2 y d E > 0.
Reflection gives the same positivity in the lower half-plane, and the original inverse-slope formula gives it on ( − 4 , 4 ) (-4,4) ( − 4 , 4 ) . Thus R R R has positive real part throughout Ω \Omega Ω .
The map w = ϕ ( z ) w = \phi(z) w = ϕ ( z ) sends Ω \Omega Ω conformally onto the unit disk, with inverse z = 8 w / ( 1 + w 2 ) z = 8w/(1+w^2) z = 8 w / ( 1 + w 2 ) . One can verify the boundary directly: w = e i θ w=e^{i\theta} w = e i θ maps to z = 4 / cos θ z=4/\cos\theta z = 4/ cos θ , the two removed real rays; the inverse is one-to-one in the disk, since its two possible preimages have product one. Moreover,
Im 8 w 1 + w 2 = 8 Im w ( 1 − ∣ w ∣ 2 ) ∣ 1 + w 2 ∣ 2 , \operatorname{Im} \frac{8w}{1+w^2} = \frac{8\operatorname{Im} w(1-|w|^2)}{|1+w^2|^2}, Im 1 + w 2 8 w = ∣1 + w 2 ∣ 2 8 Im w ( 1 − ∣ w ∣ 2 ) ,
so a nonreal point of the disk maps into the corresponding half-plane, while a real point maps into ( − 4 , 4 ) (-4,4) ( − 4 , 4 ) . Its derivative is nonzero in the disk, and the branch with ϕ ( 0 ) = 0 \phi(0)=0 ϕ ( 0 ) = 0 is its inverse. Define
H ( w ) = R ( 8 w / ( 1 + w 2 ) ) − 1 R ( 8 w / ( 1 + w 2 ) ) + 1 . H(w) = \frac{R(8w/(1+w^2))-1}{R(8w/(1+w^2))+1}. H ( w ) = R ( 8 w / ( 1 + w 2 )) + 1 R ( 8 w / ( 1 + w 2 )) − 1 .
Positive real part gives ∣ H ∣ < 1 |H|<1 ∣ H ∣ < 1 . The function is even and H ( 0 ) = 0 H(0)=0 H ( 0 ) = 0 , so its zero at zero has order at least two. Applying Schwarz’s lemma twice gives ∣ H ( w ) ∣ ≤ ∣ w ∣ 2 |H(w)|\le|w|^2 ∣ H ( w ) ∣ ≤ ∣ w ∣ 2 . Since R = ( 1 + H ) / ( 1 − H ) R=(1+H)/(1-H) R = ( 1 + H ) / ( 1 − H ) , this proves the first bound in (15) . For real ∣ t ∣ < 4 |t|<4 ∣ t ∣ < 4 , the identity
1 + ϕ ( t ) 2 1 − ϕ ( t ) 2 = 4 16 − t 2 \frac{1+\phi(t)^2}{1-\phi(t)^2} = \frac{4}{\sqrt{16-t^2}} 1 − ϕ ( t ) 2 1 + ϕ ( t ) 2 = 16 − t 2 4
gives the second. Compact subsets of Ω \Omega Ω have ∣ ϕ ∣ |\phi| ∣ ϕ ∣ bounded strictly below one, proving local boundedness. □
Local convergence, endpoint control, and the constant
Apply the lemma to P = P p , q P=P_{p,q} P = P p , q . In the notation (13) , its real bound becomes
0 ≤ F p , q ( u 2 ) ≤ ( 1 − u 2 ) − 1 / 2 , 8 ∫ 1 − δ 1 F p , q ( u 2 ) d u ≤ 16 δ ( 0 < δ ≤ 1 ) . (17) 0 \le F_{p,q}(u^2) \le(1-u^2)^{-1/2}, \qquad8\int_{1-\delta}^{1} F_{p,q}(u^2)\,du \le16\sqrt{\delta} \qquad(0<\delta\le1). \tag*{(17)} 0 ≤ F p , q ( u 2 ) ≤ ( 1 − u 2 ) − 1/2 , 8 ∫ 1 − δ 1 F p , q ( u 2 ) d u ≤ 16 δ ( 0 < δ ≤ 1 ) . (17)
The second estimate uses 1 − u 2 ≥ 1 − u 1-u^2\ge1-u 1 − u 2 ≥ 1 − u . It is uniform in all reduced p / q p/q p / q , before any contact limit is taken.
Proposition 2.2 (Recovering the local law and the bandwidth). Assume the coefficient convergence (12) . Then R p , q ( t ) → G ( t ) R_{p,q}(t)\to G(t) R p , q ( t ) → G ( t ) locally uniformly for complex ∣ t ∣ < 4 |t|<4 ∣ t ∣ < 4 , uniformly in the reduced numerator. Moreover, Theorem 1.1 holds.
Proof. The bound (15) gives a normal family on ∣ t ∣ < 4 |t|<4 ∣ t ∣ < 4 : every sequence has a subsequence converging uniformly on compact subsets to a holomorphic function. For any sequence with q → ∞ q\to\infty q → ∞ , Cauchy’s coefficient formula and (12) force every such limit to have the Taylor series (9) ; its constant term is one by the Last–Wilkinson identity. The identity theorem therefore makes every subsequential limit equal to G G G throughout the disk. This also proves numerator uniformity. Otherwise one could choose reduced numerators, denominators tending to infinity, and points in a fixed compact disk where the error stays bounded away from zero; a convergent subsequence would contradict that choice. Conversely, local uniform convergence on complex disks implies the contact limits by Cauchy’s derivative formula. This proves the equivalence stated in Theorem 1.2 .
For the integration, write
β m = ( 2 m m ) 2 16 m , F ∗ ( z ) = ∑ m ≥ 0 β m z m = G ( 4 z ) . \beta_m = \frac{\binom{2m}{m}^2}{16^m}, \qquad F_*(z) = \sum_{m\ge0} \beta_m z^m = G(4\sqrt{z}). β m = 1 6 m ( m 2 m ) 2 , F ∗ ( z ) = m ≥ 0 ∑ β m z m = G ( 4 z ) .
The factor 16 m 16^m 1 6 m here comes from t = 4 z t=4\sqrt{z} t = 4 z : these are the coefficients in z z z , rather than the Taylor coefficients in t t t . The exact change of variables is
q ∣ S p , q ∣ = ∫ − 4 4 R p , q ( t ) d t = 8 ∫ 0 1 F p , q ( u 2 ) d u . (18) q|S_{p,q}|=\int_{-4}^{4}R_{p,q}(t)\,dt=8\int_0^1F_{p,q}(u^2)\,du. \tag*{(18)} q ∣ S p , q ∣ = ∫ − 4 4 R p , q ( t ) d t = 8 ∫ 0 1 F p , q ( u 2 ) d u . (18)
For fixed δ > 0 \delta>0 δ > 0 the integrands converge uniformly on [ 0 , 1 − δ ] [0,1-\delta] [ 0 , 1 − δ ] . On the remaining interval (17) bounds their integrals by 16 δ 16\sqrt{\delta} 16 δ . The same estimate holds for F ∗ F_* F ∗ by taking the pointwise limit in the majorant. Sending q → ∞ q\to\infty q → ∞ and then δ ↓ 0 \delta\downarrow0 δ ↓ 0 gives, uniformly in the numerator,
q ∣ S p , q ∣ ⟶ 8 ∫ 0 1 F ∗ ( u 2 ) d u . (19) q|S_{p,q}|\longrightarrow8\int_0^1F_*(u^2)\,du. \tag*{(19)} q ∣ S p , q ∣ ⟶ 8 ∫ 0 1 F ∗ ( u 2 ) d u . (19)
Equivalently, one may apply dominated convergence with majorant ( 1 − u 2 ) − 1 / 2 (1-u^2)^{-1/2} ( 1 − u 2 ) − 1/2 .
It remains to evaluate the scalar integral. Define the complete elliptic integral with modulus k k k by
K ( k ) = ∫ 0 π / 2 d θ 1 − k 2 sin 2 θ , 0 ≤ k < 1. K(k)=\int_0^{\pi/2}\frac{d\theta}{\sqrt{1-k^2\sin^2\theta}},\qquad0\le k<1. K ( k ) = ∫ 0 π /2 1 − k 2 sin 2 θ d θ , 0 ≤ k < 1.
Expanding the integrand by the binomial theorem and integrating sin 2 m θ \sin^{2m}\theta sin 2 m θ gives F ∗ ( u 2 ) = 2 K ( u ) / π F_*(u^2)=2K(u)/\pi F ∗ ( u 2 ) = 2 K ( u ) / π . To evaluate its integral, first fix 0 < r < 1 0<r<1 0 < r < 1 . The integrand defining K K K is continuous on [ 0 , r ] × [ 0 , π / 2 ] [0,r]\times[0,\pi/2] [ 0 , r ] × [ 0 , π /2 ] , so changing the order of these ordinary integrals gives
∫ 0 r K ( u ) d u = ∫ 0 π / 2 arcsin ( r sin θ ) sin θ d θ . \int_0^rK(u)\,du=\int_0^{\pi/2}\frac{\arcsin(r\sin\theta)}{\sin\theta}\,d\theta. ∫ 0 r K ( u ) d u = ∫ 0 π /2 sin θ arcsin ( r sin θ ) d θ .
The quotient at θ = 0 \theta=0 θ = 0 is defined by continuity and equals r r r . As r ↑ 1 r\uparrow1 r ↑ 1 , the integrands on the right converge uniformly:
0 ≤ θ − arcsin ( r sin θ ) sin θ = ∫ r 1 d u 1 − u 2 sin 2 θ ≤ ∫ r 1 d u 1 − u 2 = arccos r . 0\le\frac{\theta-\arcsin(r\sin\theta)}{\sin\theta}=\int_r^1\frac{du}{\sqrt{1-u^2\sin^2\theta}}\le\int_r^1\frac{du}{\sqrt{1-u^2}}=\arccos r. 0 ≤ sin θ θ − arcsin ( r sin θ ) = ∫ r 1 1 − u 2 sin 2 θ d u ≤ ∫ r 1 1 − u 2 d u = arccos r .
Consequently the improper integral is finite and
∫ 0 1 K ( u ) d u = ∫ 0 π / 2 θ sin θ d θ . \int_0^1K(u)\,du=\int_0^{\pi/2}\frac{\theta}{\sin\theta}\,d\theta. ∫ 0 1 K ( u ) d u = ∫ 0 π /2 sin θ θ d θ .
Since ( log tan ( θ / 2 ) ) ′ = 1 / sin θ (\log\tan(\theta/2))'=1/\sin\theta ( log tan ( θ /2 ) ) ′ = 1/ sin θ , integration by parts makes the last integral
− 2 ∫ 0 π / 4 log ( tan x ) d x = 2 ∫ 0 1 − log v 1 + v 2 d v = 2 ∑ j ≥ 0 ( − 1 ) j ( 2 j + 1 ) 2 = 2 G . -2\int_0^{\pi/4}\log(\tan x)\,dx=2\int_0^1\frac{-\log v}{1+v^2}\,dv=2\sum_{j\ge0}\frac{(-1)^j}{(2j+1)^2}=2G. − 2 ∫ 0 π /4 log ( tan x ) d x = 2 ∫ 0 1 1 + v 2 − log v d v = 2 j ≥ 0 ∑ ( 2 j + 1 ) 2 ( − 1 ) j = 2 G .
For the series equality, use the finite geometric expansion of ( 1 + v 2 ) − 1 (1+v^2)^{-1} ( 1 + v 2 ) − 1 . After retaining the terms j = 0 , … , N j=0,\ldots,N j = 0 , … , N , the absolute value of the integrated remainder is at most
2 ∫ 0 1 ( − log v ) v 2 N + 2 d v = 2 ( 2 N + 3 ) 2 , 2\int_0^1(-\log v)v^{2N+2}\,dv=\frac{2}{(2N+3)^2}, 2 ∫ 0 1 ( − log v ) v 2 N + 2 d v = ( 2 N + 3 ) 2 2 ,
which tends to zero. This also justifies the termwise integration without a convergence theorem for infinite sums. Combining these identities with (19) gives 8 ( 2 / π ) ( 2 G ) = 32 G / π 8(2/\pi)(2G)=32G/\pi 8 ( 2/ π ) ( 2 G ) = 32 G / π , as required. □
The remaining task is Theorem 1.2 . We first express the coarea density by finite matrices, obtaining an exact formula for each derivative at zero. We use a different representation from the usual cyclic tridiagonal Harper matrix, whose boundary entries contain phase factors. The real symmetric tridiagonal matrix J J J below has zero diagonal and no corner entries. Its characteristic polynomial is the same Chambers polynomial; § 3.2 proves this by an explicit unitary change of basis.
This representation makes two useful operations possible. First, J J J commutes with reversal of the coordinates, so reflection identifies the alternating signs in the inverse branches of its characteristic polynomial. Second, adding a single corner entry gives a matrix J t J_t J t with characteristic polynomial exactly P ( E ) − t P(E)-t P ( E ) − t . Thus the inverse branches become eigenvalue branches of one matrix depending linearly on t t t . The matrix B B B is chosen so that T t = ε R + t B T_t=\varepsilon R+tB T t = εR + tB commutes with J t J_t J t and continues the reflection signs along those branches. These identities lead to the inverse square root X ( t ) X(t) X ( t ) and the formula q ρ ( t ) = X ( t ) 11 q_\rho(t)=X(t)_{11} q ρ ( t ) = X ( t ) 11 .
A second formula computes the same entry at any cut along the tridiagonal matrix. This freedom is useful when taking limits: the first diagonal entry itself need not have a pointwise limit after rescaling. All constructions in this section are finite-dimensional. In particular, the simplicity of the eigenvalue branches near t = 0 t=0 t = 0 is established directly from J J J , before its identification with the Chambers polynomial. Background on symmetric tridiagonal matrices can be found in [22 Chapter 7] .
The matrices and the two identities
Let q ≥ 2 q\ge2 q ≥ 2 , ( p , q ) = 1 (p,q)=1 ( p , q ) = 1 , and choose an odd lift p ^ ≡ p ( m o d q ) \widehat p\equiv p\pmod q p ≡ p ( mod q ) (take p ^ = p \widehat p=p p = p for odd p p p and p ^ = p + q \widehat p=p+q p = p + q otherwise). Put
θ = π p ^ q , h j = 2 sin ( j θ ) , P ( E ) = det ( E I − J ) , \theta=\frac{\pi\widehat p}{q},\qquad h_j=2\sin(j\theta),\qquad P(E)=\det(EI-J), θ = q π p , h j = 2 sin ( j θ ) , P ( E ) = det ( E I − J ) ,
where J = Jac ( h 1 , … , h q − 1 ) J=\operatorname{Jac}(h_1,\ldots,h_{q-1}) J = Jac ( h 1 , … , h q − 1 ) denotes the following q × q q\times q q × q real symmetric tridiagonal matrix with zero diagonal:
J = ( 0 h 1 0 ⋯ 0 h 1 0 h 2 ⋱ ⋮ 0 h 2 0 ⋱ 0 ⋮ ⋱ ⋱ ⋱ h q − 1 0 ⋯ 0 h q − 1 0 ) (20) J=
\begin{pmatrix}
0 & h_1 & 0 & \cdots& 0 \\
h_1 & 0 & h_2 & \ddots& \vdots\\
0 & h_2 & 0 & \ddots& 0 \\
\vdots& \ddots& \ddots& \ddots& h_{q-1} \\
0 & \cdots& 0 & h_{q-1} & 0
\end{pmatrix} \tag*{(20)} J = 0 h 1 0 ⋮ 0 h 1 0 h 2 ⋱ ⋯ 0 h 2 0 ⋱ 0 ⋯ ⋱ ⋱ ⋱ h q − 1 0 ⋮ 0 h q − 1 0 (20)
For q = 2 q=2 q = 2 this means ( 0 h 1 h 1 0 ) \begin{pmatrix}0&h_1\\h_1&0\end{pmatrix} ( 0 h 1 h 1 0 ) . In this section coordinates are numbered 1 , … , q 1,\ldots,q 1 , … , q . Let R R R reverse these coordinates, so that R e j = e q + 1 − j Re_j=e_{q+1-j} R e j = e q + 1 − j , and let N = e 1 e q T N=e_1e_q^{\mathsf T} N = e 1 e q T have its only nonzero entry, equal to one, in the top right corner. Define B B B by its explicit upper-triangular pattern:
B = 1 q ( 0 h 1 − 1 0 − h 2 − 1 0 h 3 − 1 ⋯ 0 0 h 2 − 1 0 − h 3 − 1 0 ⋯ 0 0 0 h 3 − 1 0 − h 4 − 1 ⋯ ⋮ ⋮ ⋮ ⋱ ⋱ ⋱ ⋱ 0 ⋯ 0 0 0 h q − 2 − 1 0 0 ⋯ 0 0 0 0 h q − 1 − 1 0 ⋯ 0 0 0 0 0 ) (21) B=\frac{1}{q}
\begin{pmatrix}
0 & h_1^{-1} & 0 & -h_2^{-1} & 0 & h_3^{-1} & \cdots\\
0 & 0 & h_2^{-1} & 0 & -h_3^{-1} & 0 & \cdots\\
0 & 0 & 0 & h_3^{-1} & 0 & -h_4^{-1} & \cdots\\
\vdots& \vdots& \vdots& \ddots& \ddots& \ddots& \ddots\\
0 & \cdots& 0 & 0 & 0 & h_{q-2}^{-1} & 0 \\
0 & \cdots& 0 & 0 & 0 & 0 & h_{q-1}^{-1} \\
0 & \cdots& 0 & 0 & 0 & 0 & 0
\end{pmatrix} \tag*{(21)} B = q 1 0 0 0 ⋮ 0 0 0 h 1 − 1 0 0 ⋮ ⋯ ⋯ ⋯ 0 h 2 − 1 0 ⋮ 0 0 0 − h 2 − 1 0 h 3 − 1 ⋱ 0 0 0 0 − h 3 − 1 0 ⋱ 0 0 0 h 3 − 1 0 − h 4 − 1 ⋱ h q − 2 − 1 0 0 ⋯ ⋯ ⋯ ⋱ 0 h q − 1 − 1 0 (21)
More precisely, its first, third, fifth, and subsequent odd upper diagonals are, in order,
1 q ( h 1 − 1 , … , h q − 1 − 1 ) , − 1 q ( h 2 − 1 , … , h q − 2 − 1 ) , 1 q ( h 3 − 1 , … , h q − 3 − 1 ) , … \frac{1}{q}(h_1^{-1},\ldots,h_{q-1}^{-1}),\qquad
-\frac{1}{q}(h_2^{-1},\ldots,h_{q-2}^{-1}),\qquad
\frac{1}{q}(h_3^{-1},\ldots,h_{q-3}^{-1}),\qquad\ldots q 1 ( h 1 − 1 , … , h q − 1 − 1 ) , − q 1 ( h 2 − 1 , … , h q − 2 − 1 ) , q 1 ( h 3 − 1 , … , h q − 3 − 1 ) , …
The intervening upper diagonals, the main diagonal, and all lower diagonals are zero. This description defines B B B in every size. For q = 2 m q=2m q = 2 m , the last nonzero upper diagonal has the single entry ( − 1 ) m − 1 / ( q h m ) (-1)^{m-1}/(qh_m) ( − 1 ) m − 1 / ( q h m ) ; for q = 2 m + 1 q=2m+1 q = 2 m + 1 , it has the two entries ( − 1 ) m − 1 / ( q h m ) (-1)^{m-1}/(qh_m) ( − 1 ) m − 1 / ( q h m ) and ( − 1 ) m − 1 / ( q h m + 1 ) (-1)^{m-1}/(qh_{m+1}) ( − 1 ) m − 1 / ( q h m + 1 ) . For example, the odd and even cases q = 5 , 6 q=5,6 q = 5 , 6 are shown side by side:
B ∣ q = 5 = 1 5 ( 0 h 1 − 1 0 − h 2 − 1 0 0 0 h 2 − 1 0 − h 3 − 1 0 0 0 h 3 − 1 0 0 0 0 0 h 4 − 1 0 0 0 0 0 ) , B ∣ q = 6 = 1 6 ( 0 h 1 − 1 0 − h 2 − 1 0 h 3 − 1 0 0 h 2 − 1 0 − h 3 − 1 0 0 0 0 h 3 − 1 0 − h 4 − 1 0 0 0 0 h 4 − 1 0 0 0 0 0 0 h 5 − 1 0 0 0 0 0 0 ) . \left.B\right|_{q=5}=\frac{1}{5}\begin{pmatrix}0&h_1^{-1}&0&-h_2^{-1}&0\\0&0&h_2^{-1}&0&-h_3^{-1}\\0&0&0&h_3^{-1}&0\\0&0&0&0&h_4^{-1}\\0&0&0&0&0\end{pmatrix},\qquad\left.B\right|_{q=6}=\frac{1}{6}\begin{pmatrix}0&h_1^{-1}&0&-h_2^{-1}&0&h_3^{-1}\\0&0&h_2^{-1}&0&-h_3^{-1}&0\\0&0&0&h_3^{-1}&0&-h_4^{-1}\\0&0&0&0&h_4^{-1}&0\\0&0&0&0&0&h_5^{-1}\\0&0&0&0&0&0\end{pmatrix}. B ∣ q = 5 = 5 1 0 0 0 0 0 h 1 − 1 0 0 0 0 0 h 2 − 1 0 0 0 − h 2 − 1 0 h 3 − 1 0 0 0 − h 3 − 1 0 h 4 − 1 0 , B ∣ q = 6 = 6 1 0 0 0 0 0 0 h 1 − 1 0 0 0 0 0 0 h 2 − 1 0 0 0 0 − h 2 − 1 0 h 3 − 1 0 0 0 0 − h 3 − 1 0 h 4 − 1 0 0 h 3 − 1 0 − h 4 − 1 0 h 5 − 1 0 .
For q = 2 q=2 q = 2 , B = 1 2 ( 0 h 1 − 1 0 0 ) B=\frac{1}{2}\begin{pmatrix}0&h_1^{-1}\\0&0\end{pmatrix} B = 2 1 ( 0 0 h 1 − 1 0 ) . The alternating upper diagonals make the commutator in the next lemma vanish away from the two endpoints. All denominators are nonzero, because ( p ^ , q ) = 1 (\hat{p},q)=1 ( p ^ , q ) = 1 .
Persymmetry. A square matrix is persymmetric if its entries are symmetric across the anti-diagonal. For a q × q q\times q q × q matrix A A A , this means
A i j = A q + 1 − j , q + 1 − i , equivalently R A T R = A . A_{ij}=A_{q+1-j,q+1-i},\qquad\text{equivalently}\qquad RA^{T}R=A. A ij = A q + 1 − j , q + 1 − i , equivalently R A T R = A .
Here T T T is the ordinary transpose, without complex conjugation. Equivalently, both A R AR A R and R A RA R A are symmetric. Powers, convergent power series, and the inverse when it exists preserve persymmetry. The odd lift gives h q − j = h j h_{q-j}=h_j h q − j = h j , so both B B B and J J J are persymmetric. In particular B R BR B R and R B RB R B are symmetric, a fact used in the second-contact calculation below. Persymmetry alone does not imply commutation with R R R . Since J J J is also symmetric, however, R J R = J T = J RJR=J^{T}=J R J R = J T = J , hence R J = J R RJ=JR R J = J R . This is the reflection symmetry used for the eigenvectors in Proposition 3.2 .
After defining B B B , introduce the scalar sign
ε = 1 q ∏ j = 1 q − 1 h j . (22) \varepsilon=\frac{1}{q}\prod_{j=1}^{q-1}h_j. \tag*{(22)} ε = q 1 j = 1 ∏ q − 1 h j . (22)
To see that ε ∈ { 1 , − 1 } \varepsilon\in\{1,-1\} ε ∈ { 1 , − 1 } , put ζ = e 2 π i p ^ / q \zeta=e^{2\pi i\hat{p}/q} ζ = e 2 π i p ^ / q . Its powers ζ , … , ζ q − 1 \zeta,\ldots,\zeta^{q-1} ζ , … , ζ q − 1 are all the nontrivial q q q th roots of unity. Evaluating ( x q − 1 ) / ( x − 1 ) (x^q-1)/(x-1) ( x q − 1 ) / ( x − 1 ) at x = 1 x=1 x = 1 therefore gives
∏ j = 1 q − 1 ∣ h j ∣ = ∏ j = 1 q − 1 ∣ 1 − ζ j ∣ = q . \prod_{j=1}^{q-1}|h_j|=\prod_{j=1}^{q-1}|1-\zeta^j|=q. j = 1 ∏ q − 1 ∣ h j ∣ = j = 1 ∏ q − 1 ∣1 − ζ j ∣ = q .
The factors h j h_j h j are real, so their product is either q q q or − q -q − q , which proves the assertion.
Here is the reason for these choices. The corner cofactor of E I − J EI-J E I − J is q ε q\varepsilon q ε , so the perturbation
J t = J + ε t q N J_t=J+\frac{\varepsilon t}{q}N J t = J + q εt N
changes its characteristic polynomial by exactly − t -t − t . We seek a commuting matrix that starts at ε R \varepsilon R εR , because this reflection will identify the signs in the density. With the linear form T t = ε R + t B T_t=\varepsilon R+tB T t = εR + tB , expansion gives
[ T t , J t ] = t ( [ B , J ] + 1 q [ R , N ] ) + ε t 2 q [ B , N ] . [T_t,J_t]=t\left([B,J]+\frac{1}{q}[R,N]\right)+\frac{\varepsilon t^2}{q}[B,N]. [ T t , J t ] = t ( [ B , J ] + q 1 [ R , N ] ) + q ε t 2 [ B , N ] .
Since [ R , N ] = e q e 1 T − e 1 e q T [R,N]=e_qe_1^{T}-e_1e_q^{T} [ R , N ] = e q e 1 T − e 1 e q T , the desired commutation follows if
[ B , J ] = 1 q ( e 1 e 1 T − e q e q T ) , [ B , N ] = 0. [B,J]=\frac{1}{q}(e_1e_1^{T}-e_qe_q^{T}),\qquad[B,N]=0. [ B , J ] = q 1 ( e 1 e 1 T − e q e q T ) , [ B , N ] = 0.
These are the equations that motivate B B B . Its first upper diagonal is forced by the diagonal part of the first equation: the products h j B j , j + 1 h_jB_{j,j+1} h j B j , j + 1 must all equal 1 / q 1/q 1/ q . The subsequent odd upper diagonals cancel the remaining entries. The sine addition formula verifies those cancellations below. Finally, the zero first column and last row give the stronger identities B N = N B = 0 BN=NB=0 B N = N B = 0 .
Lemma 3.1 (Finite commutator identities). One has
[ B , J ] = e 1 e 1 T − e q e q T q , B N = N B = 0. (23) [B,J]=\frac{e_1e_1^{T}-e_qe_q^{T}}{q},\qquad BN=NB=0. \tag*{(23)} [ B , J ] = q e 1 e 1 T − e q e q T , B N = N B = 0. (23)
Consequently, for
J t = J + ε t q N , T t = ε R + t B , J_t = J + \frac{\varepsilon t}{q}N,\qquad T_t = \varepsilon R + tB, J t = J + q εt N , T t = εR + tB ,
[ T t , J t ] = 0 , det ( E I − J t ) = P ( E ) − t , [T_t,J_t]=0,\qquad\det(EI-J_t)=P(E)-t, [ T t , J t ] = 0 , det ( E I − J t ) = P ( E ) − t ,
T t 2 = I + ε t ( R B + B R ) + t 2 B 2 . (24) T_t^2=I+\varepsilon t(RB+BR)+t^2B^2. \tag*{(24)} T t 2 = I + εt ( R B + B R ) + t 2 B 2 . (24)
Proof. The diagonal of [ B , J ] [B,J] [ B , J ] is ( 1 , 0 , … , 0 , − 1 ) / q (1,0,\ldots,0,-1)/q ( 1 , 0 , … , 0 , − 1 ) / q . All strictly lower entries vanish, as do entries of opposite coordinate parity. For i < j i<j i < j of the same parity put k = ( i + j ) / 2 k=(i+j)/2 k = ( i + j ) /2 and a = ( j − i ) / 2 a=(j-i)/2 a = ( j − i ) /2 . The remaining entry, multiplied by q ( − 1 ) a − 1 q(-1)^{a-1} q ( − 1 ) a − 1 , is
h j − 1 + h i − 1 h k − 1 − h j + h i h k = 0 , \frac{h_{j-1}+h_{i-1}}{h_{k-1}}-\frac{h_j+h_i}{h_k}=0, h k − 1 h j − 1 + h i − 1 − h k h j + h i = 0 ,
by h k + a + h k − a = 2 cos ( a θ ) h k h_{k+a}+h_{k-a}=2\cos(a\theta)h_k h k + a + h k − a = 2 cos ( a θ ) h k ; endpoint terms are interpreted with h 0 = h q = 0 h_0=h_q=0 h 0 = h q = 0 . The first column and last row of B B B vanish, proving B N = N B = 0 BN=NB=0 B N = N B = 0 . Now [ R , N ] = e q e q T − e 1 e 1 T [R,N]=e_q e_q^T-e_1e_1^T [ R , N ] = e q e q T − e 1 e 1 T , so the two terms of degree one in [ T t , J t ] [T_t,J_t] [ T t , J t ] cancel, and the other terms vanish. The formula for T t 2 T_t^2 T t 2 is immediate.
The corner cofactor of the tridiagonal matrix gives
e q T adj ( E I − J ) e 1 = ∏ j = 1 q − 1 h j = q ε . e_q^T\operatorname{adj}(EI-J)e_1=\prod_{j=1}^{q-1}h_j=q\varepsilon. e q T adj ( E I − J ) e 1 = j = 1 ∏ q − 1 h j = q ε .
The rank-one determinant formula, valid as a polynomial identity, therefore gives det ( E I − J t ) = P ( E ) − t \det(EI-J_t)=P(E)-t det ( E I − J t ) = P ( E ) − t . □
The real symmetric tridiagonal matrix J J J has simple real eigenvalues λ 1 , … , λ q \lambda_1,\ldots,\lambda_q λ 1 , … , λ q . To see simplicity directly, the equation J u = λ u Ju=\lambda u J u = λ u gives
h 1 u ( 2 ) = λ u ( 1 ) , h k u ( k + 1 ) = λ u ( k ) − h k − 1 u ( k − 1 ) ( 2 ≤ k < q ) . h_1u(2)=\lambda u(1),\qquad h_ku(k+1)=\lambda u(k)-h_{k-1}u(k-1)\qquad(2\le k<q). h 1 u ( 2 ) = λ u ( 1 ) , h k u ( k + 1 ) = λ u ( k ) − h k − 1 u ( k − 1 ) ( 2 ≤ k < q ) .
Since every h k h_k h k is nonzero, the first coordinate determines all the others; if it is zero, then u = 0 u=0 u = 0 . Each eigenspace therefore has dimension at most one. As J J J is real symmetric, it has an orthonormal eigenbasis, so algebraic multiplicities are also one. In particular, P ′ ( λ j ) ≠ 0 P'(\lambda_j)\ne0 P ′ ( λ j ) = 0 .
The implicit function theorem applied to P ( E ) − t P(E)-t P ( E ) − t at ( E , t ) = ( λ j , 0 ) (E,t)=(\lambda_j,0) ( E , t ) = ( λ j , 0 ) now gives analytic branches λ j ( t ) \lambda_j(t) λ j ( t ) with P ( λ j ( t ) ) = t P(\lambda_j(t))=t P ( λ j ( t )) = t and λ j ( 0 ) = λ j \lambda_j(0)=\lambda_j λ j ( 0 ) = λ j . For sufficiently small ∣ t ∣ |t| ∣ t ∣ , these q q q branches are distinct and P ′ ( λ j ( t ) ) ≠ 0 P'(\lambda_j(t))\ne0 P ′ ( λ j ( t )) = 0 , by continuity. This is a local statement about the parameter t t t near zero, independent of any identification of P P P with the Chambers polynomial. Write σ j = sgn P ′ ( λ j ) \sigma_j=\operatorname{sgn}P'(\lambda_j) σ j = sgn P ′ ( λ j ) and define the oriented inverse-slope sum
ρ ( t ) = ∑ j = 1 q σ j λ j ′ ( t ) , λ j ′ ( t ) = 1 P ′ ( λ j ( t ) ) . \rho(t)=\sum_{j=1}^{q}\sigma_j\lambda_j'(t),\qquad\lambda_j'(t)=\frac{1}{P'(\lambda_j(t))}. ρ ( t ) = j = 1 ∑ q σ j λ j ′ ( t ) , λ j ′ ( t ) = P ′ ( λ j ( t )) 1 .
For real t t t near zero, the signs of P ′ ( λ j ( t ) ) P'(\lambda_j(t)) P ′ ( λ j ( t )) remain σ j \sigma_j σ j , so ρ ( t ) = ∑ P ( E ) = t ∣ P ′ ( E ) ∣ − 1 \rho(t)=\sum_{P(E)=t}|P'(E)|^{-1} ρ ( t ) = ∑ P ( E ) = t ∣ P ′ ( E ) ∣ − 1 .
We define X ( t ) X(t) X ( t ) directly from the single matrix pencil T t T_t T t . Since T 0 2 = I T_0^2=I T 0 2 = I , one has ∥ T t 2 − I ∥ < 1 \lVert T_t^2-I\rVert<1 ∥ T t 2 − I ∥ < 1 for ∣ t ∣ |t| ∣ t ∣ sufficiently small. Put
X ( t ) = ( T t 2 ) − 1 / 2 = ∑ k = 0 ∞ ( − 1 / 2 k ) ( T t 2 − I ) k . (25) X(t)=(T_t^2)^{-1/2}=\sum_{k=0}^{\infty}\binom{-1/2}{k}(T_t^2-I)^k. \tag*{(25)} X ( t ) = ( T t 2 ) − 1/2 = k = 0 ∑ ∞ ( k − 1/2 ) ( T t 2 − I ) k . (25)
Here ( α 0 ) = 1 \binom{\alpha}{0}=1 ( 0 α ) = 1 and ( α k ) = α ( α − 1 ) ⋯ ( α − k + 1 ) / k ! \binom{\alpha}{k}=\alpha(\alpha-1)\cdots(\alpha-k+1)/k! ( k α ) = α ( α − 1 ) ⋯ ( α − k + 1 ) / k ! for k ≥ 1 k\ge1 k ≥ 1 . Thus the series converges in matrix norm and fixes the branch by X ( 0 ) = I X(0)=I X ( 0 ) = I . The square root is taken at T t 2 T_t^2 T t 2 near I I I , even though T 0 = ε R T_0=\varepsilon R T 0 = εR has both signs. Immediately,
X ( t ) 2 T t 2 = I , [ X ( t ) , T t ] = [ X ( t ) , J t ] = 0 , X(t)^2T_t^2=I,\qquad[X(t),T_t]=[X(t),J_t]=0, X ( t ) 2 T t 2 = I , [ X ( t ) , T t ] = [ X ( t ) , J t ] = 0 ,
because every term of the series is a polynomial in T t 2 T_t^2 T t 2 . No eigenvectors are needed to define X X X .
Proposition 3.2 (The density and a position-independent matrix identity). For t t t in a neighborhood of zero,
q ρ ( t ) = X ( t ) 11 = e 1 T ( T t 2 ) − 1 / 2 e 1 . (26) q\rho(t)=X(t)_{11}=e_1^{T}(T_t^2)^{-1/2}e_1. \tag*{(26)} q ρ ( t ) = X ( t ) 11 = e 1 T ( T t 2 ) − 1/2 e 1 . (26) Moreover, for every cut 1 ≤ j < q 1\le j<q 1 ≤ j < q ,
X ( t ) 11 = q h j { ( X ( t ) B ) j , j + 1 − ( X ( t ) B ) j + 1 , j } . (27) X(t)_{11}=qh_j\left\{(X(t)B)_{j,j+1}-(X(t)B)_{j+1,j}\right\}. \tag*{(27)} X ( t ) 11 = q h j { ( X ( t ) B ) j , j + 1 − ( X ( t ) B ) j + 1 , j } . (27) The inverse square root is the branch defined in (25) . The right side of (27) is independent of j j j . Thus this identity holds separately for every coefficient [ t n ] X ( t ) [t^n]X(t) [ t n ] X ( t ) , in particular for the even coefficients used below.
Proof. We separate the argument into three steps.
Reflection determines the signs. Choose a normalized real eigenvector u j u_j u j with J u j = λ j u j Ju_j=\lambda_j u_j J u j = λ j u j . Commutation gives
J ( R u j ) = R ( J u j ) = λ j R u j . J(Ru_j)=R(Ju_j)=\lambda_jRu_j. J ( R u j ) = R ( J u j ) = λ j R u j . Thus R u j Ru_j R u j belongs to the same eigenspace as u j u_j u j . That space is one-dimensional, so R u j = r j u j Ru_j=r_ju_j R u j = r j u j for some scalar r j r_j r j . As R 2 = I R^2=I R 2 = I , applying R R R twice gives r j 2 = 1 r_j^2=1 r j 2 = 1 , hence r j ∈ { 1 , − 1 } r_j\in\{1,-1\} r j ∈ { 1 , − 1 } . In coordinates, u j ( q ) = r j u j ( 1 ) u_j(q)=r_ju_j(1) u j ( q ) = r j u j ( 1 ) , and u j ( 1 ) ≠ 0 u_j(1)\ne0 u j ( 1 ) = 0 by the recurrence above.
The orthonormal eigenbasis gives the resolvent expansion
( E I − J ) − 1 = ∑ k = 1 q u k u k T E − λ k , P ( E ) ≠ 0. (EI-J)^{-1}=\sum_{k=1}^{q}\frac{u_ku_k^{T}}{E-\lambda_k},\qquad P(E)\ne0. ( E I − J ) − 1 = k = 1 ∑ q E − λ k u k u k T , P ( E ) = 0. For the corner entry, the cofactor formula gives
( E I − J ) q 1 − 1 = ( − 1 ) 1 + q det M 1 q P ( E ) = ∏ k = 1 q − 1 h k P ( E ) = q ε P ( E ) , (EI-J)^{-1}_{q1}=\frac{(-1)^{1+q}\det M_{1q}}{P(E)}
=\frac{\prod_{k=1}^{q-1}h_k}{P(E)}
=\frac{q\varepsilon}{P(E)}, ( E I − J ) q 1 − 1 = P ( E ) ( − 1 ) 1 + q det M 1 q = P ( E ) ∏ k = 1 q − 1 h k = P ( E ) q ε , where M 1 q M_{1q} M 1 q is obtained from E I − J EI-J E I − J by deleting row 1 and column q q q . It is upper triangular with diagonal − h 1 , … , − h q − 1 -h_1,\ldots,-h_{q-1} − h 1 , … , − h q − 1 , so its determinant is ( − 1 ) q − 1 ∏ k = 1 q − 1 h k (-1)^{q-1}\prod_{k=1}^{q-1}h_k ( − 1 ) q − 1 ∏ k = 1 q − 1 h k ; the factor ( − 1 ) q − 1 (-1)^{q-1} ( − 1 ) q − 1 cancels the cofactor sign ( − 1 ) q + 1 (-1)^{q+1} ( − 1 ) q + 1 . Comparing this with the ( q , 1 ) (q,1) ( q , 1 ) entry of the resolvent expansion and taking the residue at E = λ j E=\lambda_j E = λ j gives
r j ∣ u j ( 1 ) ∣ 2 = q ε P ′ ( λ j ) , σ j = ε r j . (28) r_j\lvert u_j(1)\rvert^2=\frac{q\varepsilon}{P'(\lambda_j)},\qquad\sigma_j=\varepsilon r_j. \tag*{(28)} r j ∣ u j ( 1 ) ∣ 2 = P ′ ( λ j ) q ε , σ j = ε r j . (28) The sign conclusion uses ∣ u j ( 1 ) ∣ 2 > 0 \lvert u_j(1)\rvert^2>0 ∣ u j ( 1 ) ∣ 2 > 0 .
The signed derivative sum is a matrix entry. Set S t = T t X ( t ) S_t=T_tX(t) S t = T t X ( t ) . The convergent series defining X X X gives
S t 2 = I , S 0 = ε R , [ S t , J t ] = 0. S_t^2=I,\qquad S_0=\varepsilon R,\qquad[S_t,J_t]=0. S t 2 = I , S 0 = εR , [ S t , J t ] = 0. As proved before the proposition, the roots λ j ( t ) \lambda_j(t) λ j ( t ) of P ( E ) − t P(E)-t P ( E ) − t remain simple and distinct for sufficiently small ∣ t ∣ \lvert t\rvert ∣ t ∣ . By (24) , they are exactly the eigenvalues of J t J_t J t . Here simplicity at t = 0 t=0 t = 0 refers to the parameter value, not to a particular root E = 0 E=0 E = 0 ; no Chambers-polynomial result is needed.
If J t v j = λ j ( t ) v j J_tv_j=\lambda_j(t)v_j J t v j = λ j ( t ) v j , then
J t ( S t v j ) = S t ( J t v j ) = λ j ( t ) S t v j . J_t(S_tv_j)=S_t(J_tv_j)=\lambda_j(t)S_tv_j. J t ( S t v j ) = S t ( J t v j ) = λ j ( t ) S t v j . The eigenspace is one-dimensional, so S t v j = s j ( t ) v j S_tv_j=s_j(t)v_j S t v j = s j ( t ) v j . The relation S t 2 = I S_t^2=I S t 2 = I implies s j ( t ) 2 = 1 s_j(t)^2=1 s j ( t ) 2 = 1 . Choosing the eigenvector branch continuously, the scalar s j ( t ) s_j(t) s j ( t ) is continuous and therefore constant near zero. At t = 0 t=0 t = 0 , (28) gives s j ( 0 ) = ε r j = σ j s_j(0)=\varepsilon r_j=\sigma_j s j ( 0 ) = ε r j = σ j . Thus S t S_t S t acts by σ j \sigma_j σ j on the λ j ( t ) \lambda_j(t) λ j ( t ) eigenspace. This explains the notation S t = sign T t S_t=\operatorname{sign}T_t S t = sign T t used for this analytic matrix function.
For the derivative calculation, J t J_t J t is generally nonsymmetric. Choose analytic right eigenvectors v j ( t ) v_j(t) v j ( t ) , and let w j ( t ) T w_j(t)^T w j ( t ) T be the rows of the inverse eigenvector matrix. Then
J t v j = λ j v j , w j T J t = λ j w j T , w j T v k = δ j k . J_tv_j=\lambda_jv_j,\qquad w_j^TJ_t=\lambda_jw_j^T,\qquad w_j^Tv_k=\delta_{jk}. J t v j = λ j v j , w j T J t = λ j w j T , w j T v k = δ j k .
where δ j k \delta_{jk} δ j k is one when j = k j=k j = k and zero otherwise. Differentiating the right-eigenvector equation and multiplying by w j T w_j^T w j T cancels the eigenvector-derivative terms, because w j T J t = λ j w j T w_j^T J_t=\lambda_j w_j^T w j T J t = λ j w j T and w j T v j = 1 w_j^T v_j=1 w j T v j = 1 . Thus
λ j ′ = w j T J t ′ v j = Tr ( v j w j T J t ′ ) . \lambda_j'=w_j^T J_t'v_j=\operatorname{Tr}(v_jw_j^T J_t'). λ j ′ = w j T J t ′ v j = Tr ( v j w j T J t ′ ) .
Since the eigenvectors form a basis and S t v j = σ j v j S_t v_j=\sigma_jv_j S t v j = σ j v j , S t = ∑ j σ j v j w j T S_t=\sum_j\sigma_jv_jw_j^T S t = ∑ j σ j v j w j T . Consequently
ρ ( t ) = ∑ j σ j λ j ′ ( t ) = Tr ( S t J t ′ ) . \rho(t)=\sum_j\sigma_j\lambda_j'(t)=\operatorname{Tr}(S_tJ_t'). ρ ( t ) = j ∑ σ j λ j ′ ( t ) = Tr ( S t J t ′ ) .
Now J t ′ = ε N / q J_t'=\varepsilon N/q J t ′ = εN / q and Tr ( S t N ) = e q T S t e 1 \operatorname{Tr}(S_tN)=e_q^T S_t e_1 Tr ( S t N ) = e q T S t e 1 . The last row of B B B vanishes, so e q T T t = ε e 1 T e_q^T T_t=\varepsilon e_1^T e q T T t = ε e 1 T . Therefore
q ρ ( t ) = ε Tr ( S t N ) = ε e q T T t X ( t ) e 1 = ε 2 e 1 T X ( t ) e 1 = X ( t ) 11 . q\rho(t)=\varepsilon\operatorname{Tr}(S_tN)=\varepsilon e_q^T T_tX(t)e_1
=\varepsilon^2e_1^T X(t)e_1=X(t)_{11}. q ρ ( t ) = ε Tr ( S t N ) = ε e q T T t X ( t ) e 1 = ε 2 e 1 T X ( t ) e 1 = X ( t ) 11 .
This proves (26) , including q ρ ( 0 ) = 1 q\rho(0)=1 q ρ ( 0 ) = 1 .
[ K , J t ] = X ( t ) [ B , J t ] = 1 q X ( t ) ( e 1 e 1 T − e q e q T ) . [K,J_t]=X(t)[B,J_t]=\frac{1}{q}X(t)(e_1e_1^T-e_qe_q^T). [ K , J t ] = X ( t ) [ B , J t ] = q 1 X ( t ) ( e 1 e 1 T − e q e q T ) .
For 1 ≤ j < q 1\leq j<q 1 ≤ j < q , sum the first j j j diagonal entries of this equation. The right side is X ( t ) 11 / q X(t)_{11}/q X ( t ) 11 / q , because the last diagonal entry is excluded. The corner perturbation J t − J = ε t N / q J_t-J=\varepsilon tN/q J t − J = εtN / q contributes zero: K e 1 = 0 Ke_1=0 K e 1 = 0 , so K N = 0 KN=0 K N = 0 and the partial trace of N K = e 1 e q T K NK=e_1e_q^TK N K = e 1 e q T K is K q 1 = 0 K_{q1}=0 K q 1 = 0 . For the tridiagonal part, expand
∑ a = 1 j [ K , J ] a a = ∑ a = 1 j ∑ b = 1 q ( K a b J b a − J a b K b a ) . \sum_{a=1}^{j}[K,J]_{aa}=\sum_{a=1}^{j}\sum_{b=1}^{q}(K_{ab}J_{ba}-J_{ab}K_{ba}). a = 1 ∑ j [ K , J ] aa = a = 1 ∑ j b = 1 ∑ q ( K ab J ba − J ab K ba ) .
Terms with a , b ≤ j a,b\leq j a , b ≤ j cancel in pairs. The only remaining tridiagonal connection across the cut is the pair j , j + 1 j,j+1 j , j + 1 , so
h j ( K j , j + 1 − K j + 1 , j ) = X ( t ) 11 / q . h_j(K_{j,j+1}-K_{j+1,j})=X(t)_{11}/q. h j ( K j , j + 1 − K j + 1 , j ) = X ( t ) 11 / q .
Multiplying by q q q proves (27) . □
Can the formula for X ( t ) X(t) X ( t ) use one fixed matrix? The formula X ( t ) = ( T t 2 ) − 1 / 2 X(t)=(T_t^2)^{-1/2} X ( t ) = ( T t 2 ) − 1/2 uses the matrix T t T_t T t , which varies with t t t . We now ask whether X ( t ) X(t) X ( t ) can instead be written as f t ( K ) f_t(K) f t ( K ) , where f t f_t f t is a scalar function and K K K is one fixed q × q q\times q q × q matrix, independent of t t t . In general this is impossible because the coefficient matrices need not commute. Put X n = [ t n ] X ( t ) X_n=[t^n]X(t) X n = [ t n ] X ( t ) . Already for p = 1 p=1 p = 1 , q = 3 q=3 q = 3 , expansion of the series (25) gives
[ X 2 , X 4 ] = 5 a 6 128 ( E 33 − E 11 ) ≠ 0 , a = 1 3 3 , E i j = e i e j T . [X_2,X_4]=\frac{5a^6}{128}(E_{33}-E_{11})\ne0,\qquad a=\frac{1}{3\sqrt{3}},\quad E_{ij}=e_ie_j^T. [ X 2 , X 4 ] = 128 5 a 6 ( E 33 − E 11 ) = 0 , a = 3 3 1 , E ij = e i e j T .
For verification, the invariant block on coordinates 1 , 3 1,3 1 , 3 is
( X 2 ) { 1 , 3 } = a 2 8 ( 3 − 1 3 3 ) , ( X 4 ) { 1 , 3 } = a 4 64 ( 15 − 5 35 15 ) . (X_2)_{\{1,3\}}=\frac{a^2}{8}\begin{pmatrix}3&-1\\3&3\end{pmatrix},\qquad
(X_4)_{\{1,3\}}=\frac{a^4}{64}\begin{pmatrix}15&-5\\35&15\end{pmatrix}. ( X 2 ) { 1 , 3 } = 8 a 2 ( 3 3 − 1 3 ) , ( X 4 ) { 1 , 3 } = 64 a 4 ( 15 35 − 5 15 ) .
Scalar functions of a single fixed matrix commute. Hence even the even part of the full matrix family X ( t ) X(t) X ( t ) cannot in general be written as f t ( K ) f_t(K) f t ( K ) with one fixed q × q q\times q q × q matrix K K K . This concerns the full matrix family; it does not rule out a moment representation for its selected entry. The analytic simplification in Appendix A instead compares two matrix pencils by one resolvent identity, before expanding in the contact order.
Identification of P P P with the Chambers polynomial
The material in this subsection and the next is standard. Experts may skip both subsections.
Chambers’ phase-dependence formula is classical [5 ] . The sine tridiagonal representation below appears for odd denominators in [14 Section 2] ; related finite determinant formulas in both parity cases are given in [12 Section 2.1, Lemma 4] . We retain the direct derivation to fix the normalization and signs for both denominator parities.
We now connect the preceding finite matrix formula with the usual Harper matrices. The argument computes the characteristic polynomial of one Hermitian matrix in two ways. Its description by two unitary matrices gives the Chambers polynomial, whereas a change of basis turns it into the tridiagonal matrix J J J . The argument uses only finite determinants and changes of orthonormal basis.
Put ζ = e 2 π i p / q = e 2 π i p / q \zeta=e^{2\pi i p/q}=e^{2\pi i p/q} ζ = e 2 π i p / q = e 2 π i p / q and define
D = diag ( ζ , ζ 2 , … , ζ q ) , S = ( 0 0 ⋯ 0 1 1 0 ⋯ 0 0 0 1 ⋱ ⋮ ⋮ ⋮ ⋱ ⋱ 0 0 0 ⋯ 0 1 0 ) . D=\operatorname{diag}(\zeta,\zeta^{2},\ldots,\zeta^{q}),\qquad
S=\begin{pmatrix}
0&0&\cdots&0&1\\
1&0&\cdots&0&0\\
0&1&\ddots&\vdots&\vdots\\
\vdots&\ddots&\ddots&0&0\\
0&\cdots&0&1&0
\end{pmatrix}. D = diag ( ζ , ζ 2 , … , ζ q ) , S = 0 1 0 ⋮ 0 0 0 1 ⋱ ⋯ ⋯ ⋯ ⋱ ⋱ 0 0 0 ⋮ 0 1 1 0 ⋮ 0 0 .
Thus S e j = e j + 1 Se_j=e_{j+1} S e j = e j + 1 , with S e q = e 1 Se_q=e_1 S e q = e 1 , and direct multiplication gives
D S = ζ S D , D q = S q = I . DS=\zeta SD,\qquad D^q=S^q=I. D S = ζ S D , D q = S q = I .
For complex numbers a , b a,b a , b of absolute value one, consider
H ( a , b ) = a D + a − 1 D − 1 + b S + b − 1 S − 1 . H(a,b)=aD+a^{-1}D^{-1}+bS+b^{-1}S^{-1}. H ( a , b ) = a D + a − 1 D − 1 + b S + b − 1 S − 1 .
The diagonal entries are 2 cos ( arg a + 2 π p j / q ) 2\cos(\arg a+2\pi pj/q) 2 cos ( arg a + 2 π p j / q ) ; the other terms give the cyclic nearest-neighbor coefficients. This is the usual finite Harper matrix, with the two phases represented by a , b a,b a , b . Its adjoint equals itself because D − 1 = D ∗ D^{-1}=D^* D − 1 = D ∗ and S − 1 = S ∗ S^{-1}=S^* S − 1 = S ∗ . For q > 2 q>2 q > 2 , replacing e j e_j e j by b j e j b^j e_j b j e j makes the interior off-diagonal coefficients equal to one and places b q , b − q b^q,b^{-q} b q , b − q in the two corners. Thus this is also the usual representation with phases in the corner entries. The matrix sum above covers q = 2 q=2 q = 2 as well, when two contributions occupy the same entry.
We first determine the dependence of its characteristic polynomial on these phases. Write
F ( E ; a , b ) = det ( E I − H ( a , b ) ) . F(E;a,b)=\det(EI-H(a,b)). F ( E ; a , b ) = det ( E I − H ( a , b )) .
The determinant expansion makes F F F a Laurent polynomial in a , b a,b a , b , meaning a finite sum of integer powers, including negative powers. The relations between D D D and S S S give
S − 1 H ( a , b ) S = H ( ζ a , b ) , D H ( a , b ) D − 1 = H ( a , ζ b ) . S^{-1}H(a,b)S=H(\zeta a,b),\qquad DH(a,b)D^{-1}=H(a,\zeta b). S − 1 H ( a , b ) S = H ( ζ a , b ) , D H ( a , b ) D − 1 = H ( a , ζ b ) .
Since conjugation preserves determinants, F F F is invariant under either replacement a ↦ ζ a a\mapsto\zeta a a ↦ ζ a or b ↦ ζ b b\mapsto\zeta b b ↦ ζ b . Consequently a Laurent monomial a m b n a^m b^n a m b n with nonzero coefficient must satisfy ζ m = ζ n = 1 \zeta^m=\zeta^n=1 ζ m = ζ n = 1 . Since ( p , q ) = 1 (p,q)=1 ( p , q ) = 1 , ζ \zeta ζ is primitive, so q q q divides both m m m and n n n .
Each determinant term has q q q factors, and each factor contributes at most one power of a , a − 1 , b , b − 1 a,a^{-1},b,b^{-1} a , a − 1 , b , b − 1 . Thus ∣ m ∣ + ∣ n ∣ ≤ q |m|+|n|\le q ∣ m ∣ + ∣ n ∣ ≤ q . The only possible phase monomials are therefore
1 , a q , a − q , b q , b − q . 1,\qquad a^q,\qquad a^{-q},\qquad b^q,\qquad b^{-q}. 1 , a q , a − q , b q , b − q .
To obtain a q a^q a q , every factor must come from − a D -aD − a D , so its coefficient is det ( − D ) = − 1 \det(-D)=-1 det ( − D ) = − 1 . Here det D = ( − 1 ) q − 1 \det D=(-1)^{q-1} det D = ( − 1 ) q − 1 because its diagonal entries are all the q q q th roots of unity. Likewise det S = ( − 1 ) q − 1 \det S=(-1)^{q-1} det S = ( − 1 ) q − 1 because S S S is a cyclic permutation. The coefficients of a − q , b q , b − q a^{-q},b^q,b^{-q} a − q , b q , b − q are also − 1 -1 − 1 . The remaining, phase-independent part is a monic degree-q q q polynomial in E E E , namely the Chambers polynomial P p , q C h P^{\mathrm{Ch}}_{p,q} P p , q Ch . Hence
F ( E ; a , b ) = P p , q C h ( E ) − a q − a − q − b q − b − q . F(E;a,b)=P^{\mathrm{Ch}}_{p,q}(E)-a^q-a^{-q}-b^q-b^{-q}. F ( E ; a , b ) = P p , q Ch ( E ) − a q − a − q − b q − b − q .
We also need this formula for any q × q q \times q q × q unitary pair U , V U,V U , V satisfying U V = ζ V U UV=\zeta VU U V = ζ V U , U q = u I U^q=uI U q = u I , and V q = v I V^q=vI V q = v I . Choose a normalized eigenvector f 1 f_1 f 1 of U U U , say U f 1 = λ f 1 Uf_1=\lambda f_1 U f 1 = λ f 1 . Then
U ( V k f 1 ) = λ ζ k V k f 1 . U(V^k f_1)=\lambda\zeta^k V^k f_1. U ( V k f 1 ) = λ ζ k V k f 1 .
The q q q eigenvalues for 0 ≤ k < q 0 \le k < q 0 ≤ k < q are distinct. Their eigenvectors are orthogonal because U U U is unitary. Choose b b b with b q = v b^q=v b q = v , put a = λ / ζ a=\lambda/\zeta a = λ / ζ , and define an orthonormal basis by
f j = b − ( j − 1 ) V j − 1 f 1 , 1 ≤ j ≤ q . f_j=b^{-(j-1)}V^{j-1}f_1,\qquad1\le j\le q. f j = b − ( j − 1 ) V j − 1 f 1 , 1 ≤ j ≤ q .
In this basis U f j = a ζ j f j Uf_j=a\zeta^j f_j U f j = a ζ j f j and V f j = b f j + 1 Vf_j=bf_{j+1} V f j = b f j + 1 cyclically; the last equality uses V q = v I = b q I V^q=vI=b^qI V q = v I = b q I . Thus U , V U,V U , V become a D , b S aD,bS a D , b S , and a q = u a^q=u a q = u . Applying the preceding determinant formula gives
det ( E I − U − U ∗ − V − V ∗ ) = P p , q C h ( E ) − ( u + u − 1 + v + v − 1 ) . (29) \det(EI-U-U^*-V-V^*)=P^{\mathrm{Ch}}_{p,q}(E)-(u+u^{-1}+v+v^{-1}). \tag*{(29)} det ( E I − U − U ∗ − V − V ∗ ) = P p , q Ch ( E ) − ( u + u − 1 + v + v − 1 ) . (29)
Now choose U = − D S U=-DS U = − D S and V = S V=S V = S . They obey the required commutation relation and V q = I V^q=I V q = I . Since S D = ζ − 1 D S SD=\zeta^{-1}DS S D = ζ − 1 D S ,
U q = ( − 1 ) q ζ − q ( q − 1 ) / 2 I = ( − 1 ) q e − π i p ^ ( q − 1 ) I = ( − 1 ) q ( − 1 ) q − 1 I = − I , \begin{aligned}
U^q=(-1)^q\zeta^{-q(q-1)/2}I \\
&=(-1)^q e^{-\pi i\widehat p(q-1)}I=(-1)^q(-1)^{q-1}I=-I,
\end{aligned} U q = ( − 1 ) q ζ − q ( q − 1 ) /2 I = ( − 1 ) q e − π i p ( q − 1 ) I = ( − 1 ) q ( − 1 ) q − 1 I = − I ,
where the last step uses the odd lift p ^ \widehat p p . Thus u = − 1 u=-1 u = − 1 , v = 1 v=1 v = 1 , and the scalar terms in (29) cancel. For
H ∗ = U + U ∗ + V + V ∗ = ( I − D ) S + S ∗ ( I − D ∗ ) H_* = U+U^*+V+V^*=(I-D)S+S^*(I-D^*) H ∗ = U + U ∗ + V + V ∗ = ( I − D ) S + S ∗ ( I − D ∗ )
we obtain det ( E I − H ∗ ) = P p , q C h ( E ) \det(EI-H_*)=P^{\mathrm{Ch}}_{p,q}(E) det ( E I − H ∗ ) = P p , q Ch ( E ) .
To compute the same determinant in another basis, put K = ( I − D ) S K=(I-D)S K = ( I − D ) S , so H ∗ = K + K ∗ H_*=K+K^* H ∗ = K + K ∗ . The coefficient in K e q − 1 = ( 1 − ζ q ) e q Ke_{q-1}=(1-\zeta^q)e_q K e q − 1 = ( 1 − ζ q ) e q is zero. Order the coordinates instead as
f 1 = e q , f 2 = e 1 , … , f q = e q − 1 . f_1=e_q,\qquad f_2=e_1,\qquad\ldots,\qquad f_q=e_{q-1}. f 1 = e q , f 2 = e 1 , … , f q = e q − 1 .
In this basis
K f j = c j f j + 1 ( 1 ≤ j < q ) , K f q = 0 , c j = 1 − ζ j . Kf_j=c_jf_{j+1}\quad(1\le j<q),\qquad Kf_q=0,\qquad c_j=1-\zeta^j. K f j = c j f j + 1 ( 1 ≤ j < q ) , K f q = 0 , c j = 1 − ζ j .
Therefore H ∗ H_* H ∗ becomes the tridiagonal matrix
( 0 c 1 ‾ 0 ⋯ 0 c 1 0 c 2 ‾ ⋱ ⋮ 0 c 2 0 ⋱ 0 ⋮ ⋱ ⋱ ⋱ c q − 1 ‾ 0 ⋯ 0 c q − 1 0 ) . \begin{pmatrix}
0 & \overline{c_1} & 0 & \cdots& 0\\
c_1 & 0 & \overline{c_2} & \ddots& \vdots\\
0 & c_2 & 0 & \ddots& 0\\
\vdots& \ddots& \ddots& \ddots& \overline{c_{q-1}}\\
0 & \cdots& 0 & c_{q-1} & 0
\end{pmatrix}. 0 c 1 0 ⋮ 0 c 1 0 c 2 ⋱ ⋯ 0 c 2 0 ⋱ 0 ⋯ ⋱ ⋱ ⋱ c q − 1 0 ⋮ 0 c q − 1 0 .
Recall θ = π p ^ / q \theta=\pi\widehat p/q θ = π p / q and h j = 2 sin ( j θ ) h_j=2\sin(j\theta) h j = 2 sin ( j θ ) . Then
c j = 1 − e 2 i j θ = − i e i j θ h j = α j h j , α j = − i e i j θ , ∣ α j ∣ = 1. c_j=1-e^{2ij\theta}=-ie^{ij\theta}h_j=\alpha_jh_j,\qquad\alpha_j=-ie^{ij\theta},\qquad|\alpha_j|=1. c j = 1 − e 2 ij θ = − i e ij θ h j = α j h j , α j = − i e ij θ , ∣ α j ∣ = 1.
Choose γ 1 = 1 \gamma_1=1 γ 1 = 1 , γ j + 1 = α j γ j \gamma_{j+1}=\alpha_j\gamma_j γ j + 1 = α j γ j , and replace f j f_j f j by g j = γ j f j g_j=\gamma_jf_j g j = γ j f j . These vectors still form an orthonormal basis. The coefficient from g j g_j g j to g j + 1 g_{j+1} g j + 1 is
γ j γ j + 1 c j = h j . \frac{\gamma_j}{\gamma_{j+1}}c_j=h_j. γ j + 1 γ j c j = h j .
The reverse coefficient is also h j h_j h j , by Hermitian symmetry. Thus H ∗ H_* H ∗ in this basis is exactly J J J . We have proved
P ( E ) = det ( E I − J ) = det ( E I − H ∗ ) = P p , q C h ( E ) . (30) P(E)=\det(EI-J)=\det(EI-H_*)=P^{\mathrm{Ch}}_{p,q}(E). \tag*{(30)} P ( E ) = det ( E I − J ) = det ( E I − H ∗ ) = P p , q Ch ( E ) . (30)
This also accounts for every phase and sign in the identification. The primitive root ζ \zeta ζ is unchanged by replacing p p p by its odd lift; for the finite identities involving B B B , R R R , ε \varepsilon ε , we retain the stated choice
h j = 2 sin ( π p ^ j / q ) . h_j=2\sin(\pi\widehat p j/q). h j = 2 sin ( π p j / q ) .
Real inverse branches and the bandwidth
The inverse-image description of the spectrum and the corresponding location of the critical values are standard; see [15 Section 1] . We give the elementary argument needed for the bandwidth formula in our normalization.
The identification just proved extends the local real inverse branches used in Proposition 3.2 to the interval ( − 4 , 4 ) (-4,4) ( − 4 , 4 ) . This stronger statement is not needed to justify the local eigenvalue arguments in that proposition.
For any s ∈ [ − 4 , 4 ] s \in[-4,4] s ∈ [ − 4 , 4 ] , choose real ϕ , ψ \phi,\psi ϕ , ψ with 2 cos ϕ + 2 cos ψ = s 2\cos\phi+2\cos\psi=s 2 cos ϕ + 2 cos ψ = s . Set u = e i ϕ u=e^{i\phi} u = e i ϕ , v = e i ψ v=e^{i\psi} v = e i ψ , and choose unit-modulus a , b a,b a , b with a q = u a^q=u a q = u , b q = v b^q=v b q = v . Equations (29) –(30) make P ( E ) − s P(E)-s P ( E ) − s the characteristic polynomial of the Hermitian matrix H ( a , b ) H(a,b) H ( a , b ) . Hence all its roots are real.
We next show that no critical value of P P P lies in ( − 4 , 4 ) (-4,4) ( − 4 , 4 ) . The q q q simple real roots of P P P give, by Rolle’s theorem, q − 1 q-1 q − 1 distinct roots of P ′ P' P ′ , one between each consecutive pair. These exhaust its degree, so every critical point E c E_c E c is simple: P ′ ′ ( E c ) ≠ 0 P''(E_c)\ne0 P ′′ ( E c ) = 0 . Put s c = P ( E c ) s_c=P(E_c) s c = P ( E c ) . At a local minimum,
P ( E c + x ) = s c + 1 2 P ′ ′ ( E c ) x 2 + O ( x 3 ) , P ′ ′ ( E c ) > 0. P(E_c+x)=s_c+\frac{1}{2}P''(E_c)x^2+O(x^3),\qquad P''(E_c)>0. P ( E c + x ) = s c + 2 1 P ′′ ( E c ) x 2 + O ( x 3 ) , P ′′ ( E c ) > 0.
For s s s just above s c s_c s c there are two real roots near E c E_c E c ; for s s s just below it there are none. Exactly two complex roots, counted with multiplicity, remain near E c E_c E c by continuity of polynomial roots. On the lower side they form a nonreal conjugate pair. At a local maximum the two sides are reversed. If s c ∈ ( − 4 , 4 ) s_c\in(-4,4) s c ∈ ( − 4 , 4 ) , both nearby levels remain in ( − 4 , 4 ) (-4,4) ( − 4 , 4 ) , which contradicts the real-root property. Thus all critical values have absolute value at least 4 4 4 .
Consequently the roots of P ( E ) − s P(E)-s P ( E ) − s are simple for − 4 < s < 4 -4<s<4 − 4 < s < 4 , and the local branches continue throughout this interval with their derivative signs unchanged. The usual union of Harper bands is S p , q = P − 1 ( [ − 4 , 4 ] ) S_{p,q}=P^{-1}([-4,4]) S p , q = P − 1 ([ − 4 , 4 ]) , since the phase sum ranges over precisely [ − 4 , 4 ] [-4,4] [ − 4 , 4 ] . Changing variables on each monotone inverse branch gives
∣ S p , q ∣ = ∫ − 4 4 ρ ( t ) d t . |S_{p,q}|=\int_{-4}^{4}\rho(t)\,dt. ∣ S p , q ∣ = ∫ − 4 4 ρ ( t ) d t .
where the integral is understood as an improper integral at its endpoints. Critical values at 4 4 4 or − 4 -4 − 4 are allowed.
For q = 2 q=2 q = 2 , P ( E ) = E 2 − 4 P(E)=E^2-4 P ( E ) = E 2 − 4 , ε h 1 = 2 \varepsilon h_1=2 ε h 1 = 2 , and T t 2 = ( 1 + t / 4 ) I T_t^2=(1+t/4)I T t 2 = ( 1 + t /4 ) I . Thus (26) gives 2 ρ ( t ) = ( 1 + t / 4 ) − 1 / 2 2\rho(t)=(1+t/4)^{-1/2} 2 ρ ( t ) = ( 1 + t /4 ) − 1/2 , agreeing with ρ ( t ) = 1 / 4 + t \rho(t)=1/\sqrt{4+t} ρ ( t ) = 1/ 4 + t . The construction of K K K and the basis change above remain valid even though the two cyclic shift directions coincide when q = 2 q=2 q = 2 .
This optional section may be skipped completely; no subsequent proof uses it. We first combine the Last–Wilkinson identity for the zeroth contact with finite positivity to obtain the first bandwidth bounds. We then calculate the second and fourth contacts, deriving successively sharper lower and upper bounds immediately after each contact limit. The complete sums of reciprocal even powers of sines involved also occur as special values of spectral zeta functions of cycle graphs.
The zeroth contact: the Last–Wilkinson identity
The zeroth contact q p p , q ( 0 ) = 1 q p_{p,q}(0)=1 q p p , q ( 0 ) = 1 is exactly the Last–Wilkinson identity [18 ] , recalled in (7) . Proposition 3.2 recovers this known identity immediately from X ( 0 ) = I X(0)=I X ( 0 ) = I . Combined with finite positivity, it already has a bandwidth consequence. Recall F ( z ) = F p , q ( z ) = R p , q ( 4 z ) F(z)=F_{p,q}(z)=R_{p,q}(4\sqrt{z}) F ( z ) = F p , q ( z ) = R p , q ( 4 z ) and q ∣ S p , q ∣ = 8 ∫ 0 1 F ( u 2 ) d u q|S_{p,q}|=8\int_0^1F(u^2)\,du q ∣ S p , q ∣ = 8 ∫ 0 1 F ( u 2 ) d u . Theorem C.1 gives probability representations for F F F and F 2 F^2 F 2 . Their total sums alone yields F ≥ 1 F\ge1 F ≥ 1 and F 2 ≤ ( 1 − z ) − 1 F^2\le(1-z)^{-1} F 2 ≤ ( 1 − z ) − 1 for 0 ≤ z < 1 0\le z<1 0 ≤ z < 1 , so the zeroth contact already gives
8 ≤ q ∣ S p , q ∣ ≤ 4 π (31) 8\le q|S_{p,q}|\le4\pi \tag*{(31)} 8 ≤ q ∣ S p , q ∣ ≤ 4 π (31)
The lower constant improves Last’s bound 2 ( 5 + 1 ) ≈ 6.472136 2(\sqrt{5}+1)\approx6.472136 2 ( 5 + 1 ) ≈ 6.472136 [16 Lemma 1] . The upper constant recovers the bound of Jitomirskaya, Konstantinov and Krasovsky [12 Theorem 1] , which improved Last’s earlier upper constant 8 e ≈ 21.746255 8e\approx21.746255 8 e ≈ 21.746255 .
The second contact and sharper bounds
For the pencil T t = ε R + t B T_t=\varepsilon R+tB T t = εR + tB of Section 3, abbreviate the linear coefficient of T t 2 T_t^2 T t 2 by H = ε ( R B + B R ) H=\varepsilon(RB+BR) H = ε ( R B + B R ) . The binomial series gives
[ t 2 ] X ( t ) = − 1 2 B 2 + 3 8 H 2 . (32) [t^2]X(t)=-\frac{1}{2}B^2+\frac{3}{8}H^2. \tag*{(32)} [ t 2 ] X ( t ) = − 2 1 B 2 + 8 3 H 2 . (32)
Because B B B is strictly upper triangular, ( B 2 ) 11 = 0 (B^2)_{11}=0 ( B 2 ) 11 = 0 . The relation B T = R B R B^{\mathsf{T}}=RBR B T = R B R , obtained by reflecting the displayed entries of B B B and using h q − j = h j h_{q-j}=h_j h q − j = h j , makes both B R BR B R and R B RB R B symmetric, so H H H is symmetric. The first row of R B RB R B vanishes because it is the last row of B B B . Thus the first row of H H H is the reversed first row of ε B \varepsilon B εB , and
( H 2 ) 11 = ∑ j = 1 q H 1 j 2 = ∑ j = 1 q B 1 j 2 = 1 q 2 ∑ k = 1 ⌊ q / 2 ⌋ 1 h k 2 . (H^2)_{11}=\sum_{j=1}^{q}H_{1j}^2=\sum_{j=1}^{q}B_{1j}^2=\frac{1}{q^2}\sum_{k=1}^{\lfloor q/2\rfloor}\frac{1}{h_k^2}. ( H 2 ) 11 = j = 1 ∑ q H 1 j 2 = j = 1 ∑ q B 1 j 2 = q 2 1 k = 1 ∑ ⌊ q /2 ⌋ h k 2 1 .
In the last equality the nonzero entries of the first row of B B B have column 2 k 2k 2 k and absolute value 1 / ( q ∣ h k ∣ ) 1/(q|h_k|) 1/ ( q ∣ h k ∣ ) . Remembering that a second derivative is twice the coefficient of t 2 t^2 t 2 , and that h k = 2 sin ( π p ^ k / q ) h_k=2\sin(\pi\widehat{p}k/q) h k = 2 sin ( π p k / q ) , we obtain
q ρ p , q ′ ′ ( 0 ) = 3 16 q 2 ∑ k = 1 ⌊ q / 2 ⌋ csc 2 π p ^ k q . (33) q\rho_{p,q}^{\prime\prime}(0)=\frac{3}{16q^2}\sum_{k=1}^{\lfloor q/2\rfloor}\csc^2\frac{\pi\widehat{p}k}{q}. \tag*{(33)} q ρ p , q ′′ ( 0 ) = 16 q 2 3 k = 1 ∑ ⌊ q /2 ⌋ csc 2 q π p k . (33)
We use the classical complete-sum evaluation
∑ k = 1 q − 1 csc 2 ( k π / q ) = q 2 − 1 3 . \sum_{k=1}^{q-1}\csc^2(k\pi/q)=\frac{q^2-1}{3}. k = 1 ∑ q − 1 csc 2 ( k π / q ) = 3 q 2 − 1 .
For a resolvent-kernel method of evaluating such sums, see Jorgenson–Karlsson–Smajlović [13 ] . For their interpretation as special values of spectral zeta functions of cycle graphs, see Xie–Zhao–Zhao [28 ] , Section 4.
Multiplication by p ^ \widehat{p} p permutes the nonzero residues modulo q q q , so the complete sum is unchanged. Pair k k k with q − k q-k q − k . For odd q q q every term belongs to a pair; for even q q q the unpaired term at k = q / 2 k=q/2 k = q /2 is one. We have proved the following exact formula.
Proposition 4.1 (The second contact). *For every reduced p / q p/q p / q with q ≥ 2 q\ge2 q ≥ 2 ,
q ρ p , q ′ ′ ( 0 ) = { ( q 2 − 1 ) / ( 32 q 2 ) , q odd , ( q 2 + 2 ) / ( 32 q 2 ) , q even . (34) q\rho_{p,q}^{\prime\prime}(0)=
\begin{cases}
(q^2-1)/(32q^2), & q\ \text{odd},\\
(q^2+2)/(32q^2), & q\ \text{even}.
\end{cases} \tag*{(34)} q ρ p , q ′′ ( 0 ) = { ( q 2 − 1 ) / ( 32 q 2 ) , ( q 2 + 2 ) / ( 32 q 2 ) , q odd , q even . (34) In particular the second-contact limit is 1 / 32 1/32 1/32 , uniformly in p p p .
The second contact sharpens both bounds for large denominators.
Proposition 4.2 (Bounds using only the second contact). *For all sufficiently large q q q , uniformly over reduced p / q p/q p / q ,
8.885765 ≤ q ∣ S p , q ∣ ≤ 10.805152. (35) 8.885765\le q|S_{p,q}|\le10.805152. \tag*{(35)} 8.885765 ≤ q ∣ S p , q ∣ ≤ 10.805152. (35) Thus the second contact raises the initial lower bound from 8 8 8 to 8.885765 8.885765 8.885765 and lowers the upper bound from 4 π 4\pi 4 π to $10.805152.
Proof. Let ν \nu ν be the probability measure representing F 2 F^2 F 2 . Its first moment is b q = [ z ] F 2 = 16 q ρ p , q ′ ′ ( 0 ) → 1 / 2 b_q=[z]F^2=16q\rho_{p,q}^{\prime\prime}(0)\to1/2 b q = [ z ] F 2 = 16 q ρ p , q ′′ ( 0 ) → 1/2 , uniformly by Proposition 4.1. The convex function x ↦ ( 1 − z x ) x\mapsto(1-zx) x ↦ ( 1 − z x ) lies above its tangent at b q b_q b q and below its chord between zero and one. Integration against ν \nu ν gives
1 1 − b q z ≤ F ( z ) 2 ≤ 1 − ( 1 − b q ) z 1 − z . \frac{1}{1-b_qz}\le F(z)^2\le\frac{1-(1-b_q)z}{1-z}. 1 − b q z 1 ≤ F ( z ) 2 ≤ 1 − z 1 − ( 1 − b q ) z .
Take positive square roots and integrate with z = u 2 z=u^2 z = u 2 . As b q → 1 / 2 b_q\to1/2 b q → 1/2 , the lower integral tends to 2 2 π 2\sqrt{2\pi} 2 2 π ; the substitution u = sin θ u=\sin\theta u = sin θ gives the upper limiting constant
8 ∫ 0 π / 2 1 − 1 2 sin 2 θ d θ = 10.805151048 … 8\int_0^{\pi/2}\sqrt{1-\frac{1}{2}\sin^2\theta}\,d\theta=10.805151048\ldots 8 ∫ 0 π /2 1 − 2 1 sin 2 θ d θ = 10.805151048 …
The lower limiting constant is 2 2 π = 8.885765876 … 2\sqrt{2\pi}=8.885765876\ldots 2 2 π = 8.885765876 … . Both lie strictly inside the displayed bounds. The integrable majorant ( 1 − u 2 ) − 1 / 2 (1-u^2)^{-1/2} ( 1 − u 2 ) − 1/2 justifies these limiting integrals uniformly, so (35) holds for every sufficiently large q q q . For a direct check of the upper numerical bound, truncate ( 1 − z 2 / 2 ) 1 / 2 (1-z^2/2)^{1/2} ( 1 − z 2 /2 ) 1/2 after z 64 z^{64} z 64 ; all omitted coefficients are negative. Integrating at z = sin 2 θ z=\sin^2\theta z = sin 2 θ and using π < 104348 / 33215 \pi<104348/33215 π < 104348/33215 gives an upper bound less than 10.805152 10.805152 10.805152 . □ \square □
Two examples. For p = 1 , q = 3 p=1,q=3 p = 1 , q = 3 , the entries are h 1 = h 2 = 3 h_1=h_2=\sqrt{3} h 1 = h 2 = 3 , and
B = 1 3 3 ( 0 1 0 0 0 1 0 0 0 ) , P ( E ) = E 3 − 6 E . B=\frac{1}{3\sqrt{3}}\begin{pmatrix}0&1&0\\0&0&1\\0&0&0\end{pmatrix},\qquad P(E)=E^3-6E. B = 3 3 1 0 0 0 1 0 0 0 1 0 , P ( E ) = E 3 − 6 E .
Thus ( H 2 ) 11 = 1 / 27 (H^2)_{11}=1/27 ( H 2 ) 11 = 1/27 , [ t 2 ] X 11 = 1 / 72 [t^2]X_{11}=1/72 [ t 2 ] X 11 = 1/72 , and 3 ρ 1 , 3 ′ ′ ( 0 ) = 1 / 36 3\rho_{1,3}''(0)=1/36 3 ρ 1 , 3 ′′ ( 0 ) = 1/36 , as in (34) . For p = 1 , q = 4 p=1,q=4 p = 1 , q = 4 , we have ( h 1 , h 2 , h 3 ) = ( 2 , 2 , 2 ) (h_1,h_2,h_3)=(\sqrt{2},2,\sqrt{2}) ( h 1 , h 2 , h 3 ) = ( 2 , 2 , 2 ) and
B = 1 4 ( 0 1 / 2 0 − 1 / 2 0 0 1 / 2 0 0 0 0 1 / 2 0 0 0 0 ) , P ( E ) = E 4 − 8 E 2 + 4. B=\frac{1}{4}\begin{pmatrix}0&1/\sqrt{2}&0&-1/2\\0&0&1/2&0\\0&0&0&1/\sqrt{2}\\0&0&0&0\end{pmatrix},\qquad P(E)=E^4-8E^2+4. B = 4 1 0 0 0 0 1/ 2 0 0 0 0 1/2 0 0 − 1/2 0 1/ 2 0 , P ( E ) = E 4 − 8 E 2 + 4.
Now ( H 2 ) 11 = 3 / 64 (H^2)_{11}=3/64 ( H 2 ) 11 = 3/64 , [ t 2 ] X 11 = 9 / 512 [t^2]X_{11}=9/512 [ t 2 ] X 11 = 9/512 , and 4 ρ 1 , 4 ′ ′ ( 0 ) = 9 / 256 4\rho_{1,4}''(0)=9/256 4 ρ 1 , 4 ′′ ( 0 ) = 9/256 .
These calculations can also be checked directly from the polynomials, without using B B B . For a simple real inverse branch P ( E ( t ) ) = t P(E(t))=t P ( E ( t )) = t , differentiate E ′ = 1 / P ′ ( E ) E'=1/P'(E) E ′ = 1/ P ′ ( E ) twice and retain its fixed orientation. Summing gives the useful check
ρ ′ ′ ( 0 ) = ∑ P ( λ ) = 0 3 P ′ ′ ( λ ) 2 − P ′ ′ ′ ( λ ) P ′ ( λ ) ∣ P ′ ( λ ) ∣ 5 . \rho''(0)=\sum_{P(\lambda)=0}\frac{3P''(\lambda)^2-P'''(\lambda)P'(\lambda)}{|P'(\lambda)|^5}. ρ ′′ ( 0 ) = P ( λ ) = 0 ∑ ∣ P ′ ( λ ) ∣ 5 3 P ′′ ( λ ) 2 − P ′′′ ( λ ) P ′ ( λ ) .
For E 3 − 6 E E^3-6E E 3 − 6 E the roots are 0 , ± 6 0,\pm\sqrt{6} 0 , ± 6 ; their contributions are 1 / 216 , 1 / 432 , 1 / 432 1/216,1/432,1/432 1/216 , 1/432 , 1/432 . For E 4 − 8 E 2 + 4 E^4-8E^2+4 E 4 − 8 E 2 + 4 the roots are ± 4 ± 2 3 \pm\sqrt{4\pm2\sqrt{3}} ± 4 ± 2 3 ; substitution gives ρ ′ ′ ( 0 ) = 9 / 1024 \rho''(0)=9/1024 ρ ′′ ( 0 ) = 9/1024 . Both checks agree with the matrix computation.
The fourth contact and further bounds
Proposition 4.3 (The fourth contact). Uniformly over all reduced p / q p/q p / q ,
q ρ p , q ( 4 ) ( 0 ) = 27 2048 + O ( log 2 ( q ) q ) . (36) q\rho_{p,q}^{(4)}(0)=\frac{27}{2048}+O\left(\frac{\log^2(q)}{q}\right). \tag*{(36)} q ρ p , q ( 4 ) ( 0 ) = 2048 27 + O ( q log 2 ( q ) ) . (36)
Corollary 4.4 (Bounds using contacts through order four). For all sufficiently large q q q , uniformly over reduced p / q p/q p / q ,
9.083091 ≤ q ∣ S p , q ∣ ≤ 10.076410. (37) 9.083091\le q|S_{p,q}|\le10.076410. \tag*{(37)} 9.083091 ≤ q ∣ S p , q ∣ ≤ 10.076410. (37)
The fourth contact raises the preceding lower bound from 8.885765 8.885765 8.885765 to $9.083091 and lowers the upper bound from 10.805152 10.805152 10.805152 to $10.076410. These explicit bounds improve both sides of the classical estimates using only contacts through order four. The all-contact theorem will identify the sharp value 32 G / π ≈ 9.329949 32G/\pi\approx9.329949 32 G / π ≈ 9.329949 .
Proof of the bandwidth bounds. The limiting comparison integrals are
L 4 = 8 ∫ 0 1 1 − 3 u 2 / 16 1 − 11 u 2 / 16 d u = 9.083091914 … , L_4=8\int_0^1\sqrt{\frac{1-3u^2/16}{1-11u^2/16}}\,du=9.083091914\ldots, L 4 = 8 ∫ 0 1 1 − 11 u 2 /16 1 − 3 u 2 /16 d u = 9.083091914 … ,
U 4 = 8 ∫ 0 π / 2 1 − 13 sin 2 θ / 16 1 − 5 sin 2 θ / 16 d θ = 10.076409963 … . (38) U_4=8\int_0^{\pi/2}\sqrt{\frac{1-13\sin^2\theta/16}{1-5\sin^2\theta/16}}\,d\theta=10.076409963\ldots. \tag*{(38)} U 4 = 8 ∫ 0 π /2 1 − 5 sin 2 θ /16 1 − 13 sin 2 θ /16 d θ = 10.076409963 … . (38)
Write F ( z ) = 1 + a 1 z + a 2 z 2 + ⋯ F(z)=1+a_1z+a_2z^2+\cdots F ( z ) = 1 + a 1 z + a 2 z 2 + ⋯ . The second and fourth contacts give a 1 → 1 / 4 a_1\to1/4 a 1 → 1/4 , a 2 → 9 / 64 a_2\to9/64 a 2 → 9/64 , uniformly. Thus the first two moments of the probability measure ν \nu ν for F 2 F^2 F 2 satisfy
b = ∫ x d ν = 2 a 1 ⟶ 1 2 , c = ∫ x 2 d ν = 2 a 2 + a 1 2 ⟶ 11 32 . b=\int x\,d\nu=2a_1\longrightarrow\frac{1}{2},\qquad c=\int x^2\,d\nu=2a_2+a_1^2\longrightarrow\frac{11}{32}. b = ∫ x d ν = 2 a 1 ⟶ 2 1 , c = ∫ x 2 d ν = 2 a 2 + a 1 2 ⟶ 32 11 .
For large q q q , put r = c / b r=c/b r = c / b and s = ( b − c ) / ( 1 − b ) s=(b-c)/(1-b) s = ( b − c ) / ( 1 − b ) . The moment inequalities b 2 ≤ c ≤ b b^2\le c\le b b 2 ≤ c ≤ b give 0 ≤ s ≤ b ≤ r ≤ 1 0\le s\le b\le r\le1 0 ≤ s ≤ b ≤ r ≤ 1 . Apply Jensen’s inequality to the probability measures x d ν / b x\,d\nu/b x d ν / b and ( 1 − x ) d ν / ( 1 − b ) (1-x)\,d\nu/(1-b) ( 1 − x ) d ν / ( 1 − b ) , respectively. It gives
1 + b z 1 − r z ≤ F ( z ) 2 ≤ 1 − ( 1 + s − b ) z ( 1 − z ) ( 1 − s z ) , 0 ≤ z < 1. 1+\frac{bz}{1-rz}\le F(z)^2\le\frac{1-(1+s-b)z}{(1-z)(1-sz)},\qquad0\le z<1. 1 + 1 − r z b z ≤ F ( z ) 2 ≤ ( 1 − z ) ( 1 − sz ) 1 − ( 1 + s − b ) z , 0 ≤ z < 1.
At the limiting moments these are ( 1 − 3 z / 16 ) / ( 1 − 11 z / 16 ) (1-3z/16)/(1-11z/16) ( 1 − 3 z /16 ) / ( 1 − 11 z /16 ) and ( 1 − 13 z / 16 ) / ( ( 1 − z ) ( 1 − 5 z / 16 ) ) (1-13z/16)/((1-z)(1-5z/16)) ( 1 − 13 z /16 ) / (( 1 − z ) ( 1 − 5 z /16 )) . Take positive square roots and integrate the bandwidth formula. Both bounding integrands are dominated by ( 1 − u 2 ) − 1 / 2 (1-u^2)^{-1/2} ( 1 − u 2 ) − 1/2 ; their integrals therefore converge uniformly as the moment pair converges. In the upper integral use u = sin θ u=\sin\theta u = sin θ . Since L 4 > 9.083091 L_4>9.083091 L 4 > 9.083091 and U 4 < 10.076410 U_4<10.076410 U 4 < 10.076410 , these uniform limiting inequalities give (37) . The decimal comparisons can be certified by finite binomial expansions: for f ( z ) = ( 1 − a z ) / ( 1 − b z ) f(z)=\sqrt{(1-az)/(1-bz)} f ( z ) = ( 1 − a z ) / ( 1 − b z ) ,
f ′ ( z ) = 1 2 ( b − a ) ( 1 − a z ) − 1 / 2 ( 1 − b z ) − 3 / 2 . f'(z)=\frac{1}{2}(b-a)(1-az)^{-1/2}(1-bz)^{-3/2}. f ′ ( z ) = 2 1 ( b − a ) ( 1 − a z ) − 1/2 ( 1 − b z ) − 3/2 .
Its nonconstant coefficients have the sign of b − a b-a b − a . Truncating f f f through degree 64 gives a lower bound for L 4 L_4 L 4 greater than 9.083091 and an upper bound for U 4 U_4 U 4 less than 10.076410. These are rational comparisons after using ∫ 0 π / 2 sin 2 k θ d θ = π 4 k ( 2 k k ) \int_{0}^{\pi/2}\sin^{2k}\theta\,d\theta=\frac{\pi}{4^k}\binom{2k}{k} ∫ 0 π /2 sin 2 k θ d θ = 4 k π ( k 2 k ) and π < 104348 / 33215 \pi<104348/33215 π < 104348/33215 . □
Proof sketch for odd denominators. Take q = 2 M + 1 q=2M+1 q = 2 M + 1 . The calculation has three steps: write the matrix coefficient as sine sums, complete these sums to a full period, and evaluate the correction caused by the original index restrictions. This correction has a nonzero limit; it cannot be treated as an error.
From matrices to sine sums. Put θ = π p / q \theta=\pi p/q θ = π p / q and β k = ( 2 q sin ( k θ ) ) − 1 \beta_k=(2q\sin(k\theta))^{-1} β k = ( 2 q sin ( k θ ) ) − 1 . For 0 ≤ d ≤ M 0\le d\le M 0 ≤ d ≤ M , set
U d = ∑ k = d + 1 M β k β k + d , V d = ∑ k = 1 d β k β k + d , W d = U d − V d . U_d=\sum_{k=d+1}^{M}\beta_k\beta_{k+d},\qquad V_d=\sum_{k=1}^{d}\beta_k\beta_{k+d},\qquad W_d=U_d-V_d. U d = k = d + 1 ∑ M β k β k + d , V d = k = 1 ∑ d β k β k + d , W d = U d − V d .
with empty sums equal to zero. Expanding the generator to fourth order and multiplying the displayed entries of B B B gives
K p , q : = 16 15 q ρ p , q ( 4 ) ( 0 ) = 7 U 0 2 + ∑ d = 1 M ( 8 U d 2 − W d 2 ) . K_{p,q}:=\frac{16}{15}q\rho_{p,q}^{(4)}(0)=7U_0^2+\sum_{d=1}^{M}(8U_d^2-W_d^2). K p , q := 15 16 q ρ p , q ( 4 ) ( 0 ) = 7 U 0 2 + d = 1 ∑ M ( 8 U d 2 − W d 2 ) .
The identity csc x csc ( x + y ) = ( cot x − cot ( x + y ) ) / sin y \csc x\csc(x+y)=(\cot x-\cot(x+y))/\sin y csc x csc ( x + y ) = ( cot x − cot ( x + y )) / sin y turns U d , V d U_d,V_d U d , V d into partial cotangent sums.
Complete the sums, keeping the correction. Define, for 1 ≤ d ≤ M 1\le d\le M 1 ≤ d ≤ M ,
a d = 1 8 q 2 sin 2 ( d θ ) cos ( d θ ) , r d = − 1 4 q 2 cos ( d θ ) . a_d=\frac{1}{8q^2\sin^2(d\theta)\cos(d\theta)},\qquad r_d=-\frac{1}{4q^2\cos(d\theta)}. a d = 8 q 2 sin 2 ( d θ ) cos ( d θ ) 1 , r d = − 4 q 2 cos ( d θ ) 1 .
In the cotangent expressions, reflection to the complementary index range replaces ( U d , W d ) (U_d,W_d) ( U d , W d ) by ( U d + a d , W d + 4 a d + r d ) (U_d+a_d,W_d+4a_d+r_d) ( U d + a d , W d + 4 a d + r d ) . Averaging the corresponding quadratic expressions completes the sums to a full period. Subtracting that average from the original expression gives the exact split K p , q = C q + E p , q + R p , q K_{p,q}=C_q+\mathcal{E}_{p,q}+R_{p,q} K p , q = C q + E p , q + R p , q below. Here C q C_q C q is the average. After finite Fourier inversion, its completed numerators contain only 1 , ζ d , ζ 2 d 1,\zeta^d,\zeta^{2d} 1 , ζ d , ζ 2 d , where ζ q = 1 \zeta^q=1 ζ q = 1 . The nine pairings of these terms are evaluated by the finite Fourier transform of the complete cosecant-square sum. Their sum reduces to
C q = 3 128 [ 2 q 2 ∑ j = 1 q − 1 csc 2 π j q − 1 q 4 ∑ j = 1 q − 1 csc 4 π j q ] = 29 1920 + O ( q − 2 ) . C_q=\frac{3}{128}\left[\frac{2}{q^2}\sum_{j=1}^{q-1}\csc^2\frac{\pi j}{q}-\frac{1}{q^4}\sum_{j=1}^{q-1}\csc^4\frac{\pi j}{q}\right]=\frac{29}{1920}+O(q^{-2}). C q = 128 3 [ q 2 2 j = 1 ∑ q − 1 csc 2 q π j − q 4 1 j = 1 ∑ q − 1 csc 4 q π j ] = 1920 29 + O ( q − 2 ) .
Here the two complete evaluations are
∑ j = 1 q − 1 csc 2 π j q = q 2 − 1 3 , ∑ j = 1 q − 1 csc 4 π j q = ( q 2 − 1 ) ( q 2 + 11 ) 45 . \sum_{j=1}^{q-1}\csc^2\frac{\pi j}{q}=\frac{q^2-1}{3},\qquad\sum_{j=1}^{q-1}\csc^4\frac{\pi j}{q}=\frac{(q^2-1)(q^2+11)}{45}. j = 1 ∑ q − 1 csc 2 q π j = 3 q 2 − 1 , j = 1 ∑ q − 1 csc 4 q π j = 45 ( q 2 − 1 ) ( q 2 + 11 ) .
Again these sums are related to spectral zeta values of graphs; see [13 , 28 ] . Thus the completed contribution involves the first two integer values of the cycle spectral zeta function. The terms added at the ends of the original ranges produce E p , q \mathcal{E}_{p,q} E p , q ; the remaining terms form R p , q \mathcal{R}_{p,q} R p , q .
3. Use symmetry before estimating. Let ⟨ n ⟩ q \langle n\rangle_q ⟨ n ⟩ q be the representative of n n n modulo q q q in { − M , … , M } \{-M,\ldots,M\} { − M , … , M } , and put D = { ⟨ p d ⟩ q : 1 ≤ d ≤ M } \mathcal{D}=\{\langle pd\rangle_q:1\le d\le M\} D = {⟨ p d ⟩ q : 1 ≤ d ≤ M } . This set contains one element from each pair { j , − j } \{j,-j\} { j , − j } . Write s j = sin ( π j / q ) s_j=\sin(\pi j/q) s j = sin ( π j / q ) , c j = cos ( π j / q ) c_j=\cos(\pi j/q) c j = cos ( π j / q ) , and f ( D , K ) = ( s D 2 c D s K s D + K ) − 1 f(D,K)=(s_D^2c_Ds_Ks_{D+K})^{-1} f ( D , K ) = ( s D 2 c D s K s D + K ) − 1 . The correction has the explicit form
E p , q = 1 16 q 4 [ ∑ D ∈ D 1 s D 4 c D 2 − 2 ∑ D , K ∈ D f ( D , K ) ] . \mathcal{E}_{p,q}=\frac{1}{16q^4}\left[\sum_{D\in\mathcal{D}}\frac{1}{s_D^4c_D^2}-2\sum_{D,K\in\mathcal{D}}f(D,K)\right]. E p , q = 16 q 4 1 D ∈ D ∑ s D 4 c D 2 1 − 2 D , K ∈ D ∑ f ( D , K ) .
Its double sum looks dependent on p p p , but exchanging its indices gives
f ( D , K ) + f ( K , D ) = 1 s D 2 s K 2 + 1 s D c D s K c K . f(D,K)+f(K,D)=\frac{1}{s_D^2s_K^2}+\frac{1}{s_Dc_Ds_Kc_K}. f ( D , K ) + f ( K , D ) = s D 2 s K 2 1 + s D c D s K c K 1 .
Thus 2 ∑ f = A q 2 + Λ 2 2\sum f=A_q^2+\Lambda^2 2 ∑ f = A q 2 + Λ 2 , where A q = ∑ D ∈ D s D − 2 = ( q 2 − 1 ) / 6 A_q=\sum_{D\in\mathcal{D}}s_D^{-2}=(q^2-1)/6 A q = ∑ D ∈ D s D − 2 = ( q 2 − 1 ) /6 and Λ = ∑ D ∈ D ( s D c D ) − 1 \Lambda=\sum_{D\in\mathcal{D}}(s_Dc_D)^{-1} Λ = ∑ D ∈ D ( s D c D ) − 1 . The remaining diagonal sum is ( q 2 − 1 ) ( q 2 + 71 ) / 90 (q^2-1)(q^2+71)/90 ( q 2 − 1 ) ( q 2 + 71 ) /90 . Substitution gives the nonzero limiting correction:
E p , q = − ( q 2 − 1 ) ( q 2 − 49 ) 960 q 4 − Λ 2 16 q 4 = − 1 960 + O ( q − 2 log 2 q ) . \mathcal{E}_{p,q}=-\frac{(q^2-1)(q^2-49)}{960q^4}-\frac{\Lambda^2}{16q^4}=-\frac{1}{960}+O(q^{-2}\log^2q). E p , q = − 960 q 4 ( q 2 − 1 ) ( q 2 − 49 ) − 16 q 4 Λ 2 = − 960 1 + O ( q − 2 log 2 q ) .
Indeed ∣ Λ ∣ = O ( q log q ) |\Lambda|=O(q\log q) ∣Λ∣ = O ( q log q ) , uniformly, by summing ∣ s j c j ∣ − 1 |s_jc_j|^{-1} ∣ s j c j ∣ − 1 over the complete set of opposite pairs.
For completeness, the remainder estimate uses no further cancellation. Expanding the difference of the quadratic expressions gives
R p , q = ∑ d = 1 M r d ( W d + 4 a d ) + 1 2 ∑ d = 1 M r d 2 . \mathcal{R}_{p,q}=\sum_{d=1}^{M}r_d(W_d+4a_d)+\frac{1}{2}\sum_{d=1}^{M}r_d^2. R p , q = d = 1 ∑ M r d ( W d + 4 a d ) + 2 1 d = 1 ∑ M r d 2 .
The Euclidean norms satisfy ∥ r ∥ 2 = O ( q − 1 ) \lVert r\rVert_2=O(q^{-1}) ∥ r ∥ 2 = O ( q − 1 ) and ∥ a ∥ 2 = O ( 1 ) \lVert a\rVert_2=O(1) ∥ a ∥ 2 = O ( 1 ) . The W d W_d W d are, up to alternating signs, coordinates of H 2 e 1 H^2e_1 H 2 e 1 , with H = ε ( R B + B R ) H=\varepsilon(RB+BR) H = ε ( R B + B R ) as above. The row and column sums of the absolute values of B B B are O ( log q ) O(\log q) O ( log q ) , uniformly in p p p , by permutation of the nonzero residue classes. Thus ∥ W ∥ 2 ≤ ∥ H ∥ 2 = O ( log 2 q ) \lVert W\rVert_2\le\lVert H\rVert^2=O(\log^2q) ∥ W ∥ 2 ≤ ∥ H ∥ 2 = O ( log 2 q ) . Cauchy–Schwarz yields R p , q = O ( q − 1 log 2 q ) \mathcal{R}_{p,q}=O(q^{-1}\log^2q) R p , q = O ( q − 1 log 2 q ) . Combining the three parts and multiplying by 15 / 16 15/16 15/16 proves
q ρ p , q ( 4 ) ( 0 ) = 15 16 ( 29 1920 − 1 960 ) + O ( log 2 ( q ) q ) = 27 2048 + O ( log 2 ( q ) q ) . q\rho_{p,q}^{(4)}(0)=\frac{15}{16}\left(\frac{29}{1920}-\frac{1}{960}\right)+O\left(\frac{\log^2(q)}{q}\right)=\frac{27}{2048}+O\left(\frac{\log^2(q)}{q}\right). q ρ p , q ( 4 ) ( 0 ) = 16 15 ( 1920 29 − 960 1 ) + O ( q log 2 ( q ) ) = 2048 27 + O ( q log 2 ( q ) ) .
We omit the separate even-denominator calculation from this sketch. Proposition 8.4 below also gives a complete proof of (36) for both parities.
Difficulties at higher contact orders. At sixth and higher orders, longer matrix products introduce several coupled summation indices and overlapping ordering constraints. Complete cosecant identities evaluate unrestricted sums, but do not evaluate the new ordered corrections. The fourth contact already contains a restricted correction with a nonzero limit, so higher-order terms cannot simply be discarded without a cancellation argument. The two-index symmetry used above gives no immediate corresponding identity at higher orders. Estimating the separate terms by absolute values can lose the cancellations needed to determine their limit.
From finite matrices to one scalar integral
Section 3 gave an exact finite-matrix formula for the density. Appendix A proves the limiting step needed here. We use its conclusion first, and study the resulting function by finite-dimensional calculations. The appendix can be read after this section; its convergence proof is independent of the evaluation of that function.
Fix an odd positive integer p p p and a residue r r r modulo 4 p 4p 4 p with ( r , p ) = 1 (r,p)=1 ( r , p ) = 1 . Choose an odd lift q ^ ≡ r ( m o d p ) \widehat{q}\equiv r\pmod p q ≡ r ( mod p ) . The conclusions we use from Appendix A are the following. First, Proposition 5.1 below states, for every fixed m ≥ 0 m\ge0 m ≥ 0 ,
lim q → ∞ q ≡ r ( m o d 4 p ) q p p , q ( 2 m ) ( 0 ) ( 2 m ) ! = [ t 2 m ] G p ( t ) . (39) \lim_{\substack{q\to\infty\\ q\equiv r\pmod{4p}}}\frac{q p_{p,q}^{(2m)}(0)}{(2m)!}=[t^{2m}]G_p(t). \tag*{(39)} q → ∞ q ≡ r ( mod 4 p ) lim ( 2 m )! q p p , q ( 2 m ) ( 0 ) = [ t 2 m ] G p ( t ) . (39)
Second, the appendix identifies G p G_p G p with the angular average of two diagonal entries of the inverse square root defined below. Its scalar-shift cancellation removes the difference of two one-sided Fourier transforms, so only the real angular parameters c , h c,h c , h with h 2 + 4 c 2 = 1 h^2+4c^2=1 h 2 + 4 c 2 = 1 remain. This finite matrix formula defines G p G_p G p analytically near t = 0 t=0 t = 0 . At this stage G p G_p G p may still depend on the chosen residue; the notation suppresses that dependence.
The reason for making this reduction is that the matrices now have size 2 p 2p 2 p , independent of the increasing denominator q q q . One analytic function encodes all the fixed-numerator contact limits. Our next task is therefore to evaluate G p G_p G p : we turn its two selected matrix entries into a scalar determinant, then into a scalar integral. The subsequent sections evaluate that integral and show that the resulting Taylor coefficients do not depend on p p p or the residue. The later argument in § 8 passes from fixed numerators to arbitrary reduced rational parameters.
The main conversion in this section is
two matrix entries ⟶ determinant derivatives ⟶ a scalar integral . \text{two matrix entries}\ \longrightarrow\ \text{determinant derivatives}\ \longrightarrow\ \text{a scalar integral}. two matrix entries ⟶ determinant derivatives ⟶ a scalar integral .
The construction parallels Section 3. There the commuting pair was J t , T t J_t,T_t J t , T t ; here the corresponding pair is U ( s ) , L ( s ) U(s),L(s) U ( s ) , L ( s ) . In both cases, normalizing the second matrix gives an involution that separates two spectral groups, and a signed trace formula selects the required entries. Here the signed sum of eigenvalue derivatives is replaced by a difference of logarithmic determinant derivatives. The one-entry dependence of U U U on s s s will recover the two entries required in G p G_p G p .
Only a local separation into two groups is needed for this calculation. The analytic inverse square root gives an involution and two complementary projections, continuing the + 1 +1 + 1 and − 1 -1 − 1 eigenspaces at s = 0 s=0 s = 0 . We take the determinants of U U U on their images. To recover these determinants from a scalar integral, we then set s = i η s=i\eta s = i η , η > 0 \eta>0 η > 0 . A fixed Hermitian form, together with an explicit identity of known sign, identifies the roots inside and outside the unit circle. Jensen’s formula completes the conversion. Direct block symmetries give the reality needed in the final differentiation.
Theorem 5.3 states the value of the resulting integral; the following sections prove it. We use s s s for the matrix parameter and t t t for the normalized function G p ( t ) G_p(t) G p ( t ) , with s = t / p s=t/p s = t / p .
The limiting matrix and its inverse square root
Write p = 2 n + 1 p=2n+1 p = 2 n + 1 , put θ = π q ^ / p \theta=\pi\widehat{q}/p θ = π q / p , and number the coordinates of C p \mathbb{C}^p C p by 0 , … , p − 1 0,\ldots,p-1 0 , … , p − 1 . Let e r e_r e r be the r r r th coordinate vector, e r = ( − 1 ) r e_r=(-1)^r e r = ( − 1 ) r , and J r s = 1 r + s = p − 1 J_{rs}=1_{r+s=p-1} J r s = 1 r + s = p − 1 . Thus J J J reverses the coordinates; it is different from the tridiagonal matrix J J J in the preceding section. Take real c , h c,h c , h with h 2 + 4 c 2 = 1 h^2+4c^2=1 h 2 + 4 c 2 = 1 . To write the matrix compactly, put
κ ℓ = cot ( ℓ θ ) , a ℓ ± = ( − 1 ) ℓ + 1 2 ( κ ℓ ± i h ) , 1 ≤ ℓ < p , \kappa_\ell=\cot(\ell\theta),\qquad a_\ell^\pm=\frac{(-1)^{\ell+1}}{2}(\kappa_\ell\pm ih),\qquad1\le\ell<p, κ ℓ = cot ( ℓ θ ) , a ℓ ± = 2 ( − 1 ) ℓ + 1 ( κ ℓ ± ih ) , 1 ≤ ℓ < p ,
where i 2 = − 1 i^2=-1 i 2 = − 1 . The upper triangular matrix M = M ( c , h ) M=M(c,h) M = M ( c , h ) has the following explicit pattern:
M = ( c / 2 a 1 + − c a 2 + c ⋯ ( − 1 ) n c 0 − c / 2 a 2 − c a 3 − ⋯ a n + 1 − 0 0 c / 2 a 3 + − c ⋯ ( − 1 ) n + 1 c ⋮ ⋮ ⋮ ⋱ ⋱ ⋱ ⋮ 0 ⋯ 0 0 c / 2 a p − 2 + − c 0 ⋯ 0 0 0 − c / 2 a p − 1 − 0 ⋯ 0 0 0 0 c / 2 ) . M=\begin{pmatrix}
c/2 & a_1^+ & -c & a_2^+ & c & \cdots& (-1)^n c\\
0 & -c/2 & a_2^- & c & a_3^- & \cdots& a_{n+1}^-\\
0 & 0 & c/2 & a_3^+ & -c & \cdots& (-1)^{n+1}c\\
\vdots& \vdots& \vdots& \ddots& \ddots& \ddots& \vdots\\
0 & \cdots& 0 & 0 & c/2 & a_{p-2}^+ & -c\\
0 & \cdots& 0 & 0 & 0 & -c/2 & a_{p-1}^-\\
0 & \cdots& 0 & 0 & 0 & 0 & c/2
\end{pmatrix}. M = c /2 0 0 ⋮ 0 0 0 a 1 + − c /2 0 ⋮ ⋯ ⋯ ⋯ − c a 2 − c /2 ⋮ 0 0 0 a 2 + c a 3 + ⋱ 0 0 0 c a 3 − − c ⋱ c /2 0 0 ⋯ ⋯ ⋯ ⋱ a p − 2 + − c /2 0 ( − 1 ) n c a n + 1 − ( − 1 ) n + 1 c ⋮ − c a p − 1 − c /2 .
For precision, reading a row from its diagonal to the right, the successive entries in an even row r r r are
c / 2 , a r + 1 + , − c , a r + 2 + , c , a r + 3 + , − c , … , c/2,\ a_{r+1}^+,\ -c,\ a_{r+2}^+,\ c,\ a_{r+3}^+,\ -c,\ldots, c /2 , a r + 1 + , − c , a r + 2 + , c , a r + 3 + , − c , … ,
and in an odd row they are
− c / 2 , a r + 1 − , c , a r + 2 − , − c , a r + 3 − , c , … . -c/2,\ a_{r+1}^-,\ c,\ a_{r+2}^-,\ -c,\ a_{r+3}^-,\ c,\ldots. − c /2 , a r + 1 − , c , a r + 2 − , − c , a r + 3 − , c , … .
Each row stops at column p − 1 p-1 p − 1 ; all entries below its diagonal are zero. Set V = M ( c , − h ) V=M(c,-h) V = M ( c , − h ) , so V = M ‾ V=\overline{M} V = M for real c , h c,h c , h . For p = 1 p=1 p = 1 , the matrix is simply M = ( c / 2 ) M=(c/2) M = ( c /2 ) . For p = 3 p=3 p = 3 , with κ = cot θ \kappa=\cot\theta κ = cot θ , it is
M = ( c / 2 ( κ + i h ) / 2 − c 0 − c / 2 ( κ + i h ) / 2 0 0 c / 2 ) , V = M ‾ . (40) M=\begin{pmatrix}
c/2 & (\kappa+ih)/2 & -c\\
0 & -c/2 & (\kappa+ih)/2\\
0 & 0 & c/2
\end{pmatrix},\qquad V=\overline{M}. \tag*{(40)} M = c /2 0 0 ( κ + ih ) /2 − c /2 0 − c ( κ + ih ) /2 c /2 , V = M . (40)
For p = 5 p=5 p = 5 , the full matrix is
M = ( c / 2 a 1 + − c a 2 + c 0 − c / 2 a 2 − c a 2 + 0 0 c / 2 a 2 − − c 0 0 0 − c / 2 a 1 + 0 0 0 0 c / 2 ) . M=\begin{pmatrix}
c/2 & a_1^+ & -c & a_2^+ & c\\
0 & -c/2 & a_2^- & c & a_2^+\\
0 & 0 & c/2 & a_2^- & -c\\
0 & 0 & 0 & -c/2 & a_1^+\\
0 & 0 & 0 & 0 & c/2
\end{pmatrix}. M = c /2 0 0 0 0 a 1 + − c /2 0 0 0 − c a 2 − c /2 0 0 a 2 + c a 2 − − c /2 0 c a 2 + − c a 1 + c /2 .
Here the repeated entries follow from κ p − ℓ = − κ ℓ \kappa_{p-\ell}=-\kappa_\ell κ p − ℓ = − κ ℓ . The matrices M M M and V V V are also persymmetric , in the sense introduced in Section 3:
J M T J = M , J V T J = V . JM^{T}J=M,\qquad JV^{T}J=V. J M T J = M , J V T J = V .
For their cotangent entries this follows from cot ( ( p − ℓ ) θ ) = − cot ( ℓ θ ) \cot((p-\ell)\theta)=-\cot(\ell\theta) cot (( p − ℓ ) θ ) = − cot ( ℓ θ ) ; the remaining entries can be read directly from the displayed matrices. Consequently M J MJ M J , J M JM J M , V J VJ V J , J V JV J V are symmetric under ordinary transpose. These are complex symmetric matrices; no conjugation is included in these symmetry identities.
The limit in Appendix A involves an inverse square root of a matrix quadratic in the parameter. We organize that quadratic matrix as the square of a matrix linear in s s s :
L ( s ) = ( − i s M J J i s V ) , L ( s ) 2 = ( I − s 2 M 2 − i s ( M J − J V ) i s ( V J − J M ) I − s 2 V 2 ) . L(s)=\begin{pmatrix}
-isM & J\\
J & isV
\end{pmatrix},\qquad
L(s)^2=\begin{pmatrix}
I-s^2M^2 & -is(MJ-JV)\\
is(VJ-JM) & I-s^2V^2
\end{pmatrix}. L ( s ) = ( − i s M J J i s V ) , L ( s ) 2 = ( I − s 2 M 2 i s ( V J − J M ) − i s ( M J − J V ) I − s 2 V 2 ) .
Both off-diagonal combinations M J − J V MJ-JV M J − J V and V J − J M VJ-JM V J − J M are symmetric under ordinary transpose. The main proof uses the stronger underlying persymmetry: for R = diag ( J , J ) R=\operatorname{diag}(J,J) R = diag ( J , J ) , R L T R = L RL^TR=L R L T R = L , hence R Y T R = Y RY^TR=Y R Y T R = Y for the matrix function defined below. These identities supply the equality of entries used in Proposition 5.2 to extract a determinant derivative. They do not assert that L L L or Y Y Y is Hermitian.
Up to the diagonal change of basis explained next to (A.12), this is exactly the quadratic matrix of the appendix. The linear matrix L ( s ) L(s) L ( s ) will allow us to construct a commuting matrix with only one varying entry.
Separating the two continued spectral groups. At zero the linear matrix is the Hermitian involution
L ( 0 ) = ( 0 J J 0 ) , L ( 0 ) 2 = I 2 p . L(0)=\begin{pmatrix}0 & J \\ J & 0\end{pmatrix}, \qquad L(0)^2=I_{2p}. L ( 0 ) = ( 0 J J 0 ) , L ( 0 ) 2 = I 2 p .
Its + 1 +1 + 1 and − 1 -1 − 1 eigenspaces each have dimension p p p . Rather than choose and continue individual eigenvectors, we continue the projections onto these two spaces. The inverse square root below gives them directly and applies to complex s s s from the start.
The matrix Y ( s ) Y(s) Y ( s ) below has the same role as X ( t ) = ( T t 2 ) − 1 / 2 X(t)=(T_t^2)^{-1/2} X ( t ) = ( T t 2 ) − 1/2 in Proposition 3.2 : its selected entries generate the contact coefficients, now after the fixed-numerator limit. Since L ( 0 ) 2 = I 2 p L(0)^2=I_{2p} L ( 0 ) 2 = I 2 p , its analytic branch is fixed by Y ( 0 ) = I 2 p Y(0)=I_{2p} Y ( 0 ) = I 2 p . Define
Y ( s ) = ( L ( s ) 2 ) − 1 / 2 : = ∑ k ≥ 0 ( − 1 / 2 k ) ( L ( s ) 2 − I 2 p ) k , Y(s)=(L(s)^2)^{-1/2}:=\sum_{k\geq0}\binom{-1/2}{k}(L(s)^2-I_{2p})^k, Y ( s ) = ( L ( s ) 2 ) − 1/2 := k ≥ 0 ∑ ( k − 1/2 ) ( L ( s ) 2 − I 2 p ) k ,
S ( s ) = L ( s ) Y ( s ) . (41) S(s)=L(s)Y(s). \tag*{(41)} S ( s ) = L ( s ) Y ( s ) . (41)
The series converges for all sufficiently small ∣ s ∣ |s| ∣ s ∣ and fixes the branch by Y ( 0 ) = I 2 p Y(0)=I_{2p} Y ( 0 ) = I 2 p ; the powers are ordinary matrix powers. It satisfies Y ( s ) 2 L ( s ) 2 = I 2 p Y(s)^2L(s)^2=I_{2p} Y ( s ) 2 L ( s ) 2 = I 2 p and [ Y ( s ) , L ( s ) ] = 0 [Y(s),L(s)]=0 [ Y ( s ) , L ( s )] = 0 . In particular, S ( s ) 2 = I 2 p S(s)^2=I_{2p} S ( s ) 2 = I 2 p and S ( 0 ) = L ( 0 ) S(0)=L(0) S ( 0 ) = L ( 0 ) . Thus the matrices
P ± ( s ) = I 2 p ± S ( s ) 2 , E ± ( s ) = Ran P ± ( s ) \mathcal{P}_{\pm}(s)=\frac{I_{2p}\pm S(s)}{2}, \qquad E_{\pm}(s)=\operatorname{Ran}\mathcal{P}_{\pm}(s) P ± ( s ) = 2 I 2 p ± S ( s ) , E ± ( s ) = Ran P ± ( s )
are complementary projections and their images. Here Ran \operatorname{Ran} Ran means the image of the matrix. Indeed, P ± 2 = P ± \mathcal{P}_{\pm}^2=\mathcal{P}_{\pm} P ± 2 = P ± , P + P − = 0 \mathcal{P}_{+}\mathcal{P}_{-}=0 P + P − = 0 , and P + + P − = I 2 p \mathcal{P}_{+}+\mathcal{P}_{-}=I_{2p} P + + P − = I 2 p , directly from S 2 = I 2 p S^2=I_{2p} S 2 = I 2 p . A projection has only the eigenvalues 0 and 1, so its rank equals its trace. That trace is a continuous integer-valued function of s s s ; therefore both ranks remain p p p near zero. At zero their images are
E + ( 0 ) = { ( v , J v ) : v ∈ C p } , E − ( 0 ) = { ( v , − J v ) : v ∈ C p } . E_{+}(0)=\{(v,Jv):v\in\mathbb{C}^{p}\}, \qquad E_{-}(0)=\{(v,-Jv):v\in\mathbb{C}^{p}\}. E + ( 0 ) = {( v , J v ) : v ∈ C p } , E − ( 0 ) = {( v , − J v ) : v ∈ C p } .
The signs label these two continued spaces. The formulas require neither simple eigenvalues nor a choice of eigenvalue branches. The projections need not be orthogonal when s ≠ 0 s\ne0 s = 0 .
Appendix A selects one coordinate from each of the two frequency blocks. Thus the expression to be converted into a determinant derivative is the weighted average
I ( s ; c , h ) = c 2 { Y ( s ) 00 + Y ( s ) p p } . \mathcal{I}(s;c,h)=\frac{c}{2}\{Y(s)_{00}+Y(s)_{pp}\}. I ( s ; c , h ) = 2 c { Y ( s ) 00 + Y ( s ) pp } .
The index p p p is the first coordinate of the second block. Two block symmetries will be useful when we differentiate a real logarithmic modulus. For real c , h c,h c , h and complex s s s , the explicit formula for L L L gives
L ( − s ) = − ( I p 0 0 − I p ) L ( s ) ( I p 0 0 − I p ) , L(-s)=-\begin{pmatrix}I_p & 0 \\ 0 & -I_p\end{pmatrix}L(s)\begin{pmatrix}I_p & 0 \\ 0 & -I_p\end{pmatrix}, L ( − s ) = − ( I p 0 0 − I p ) L ( s ) ( I p 0 0 − I p ) ,
L ( s ‾ ) ‾ = ( 0 I p I p 0 ) L ( s ) ( 0 I p I p 0 ) . \overline{L(\overline{s})}=\begin{pmatrix}0 & I_p \\ I_p & 0\end{pmatrix}L(s)\begin{pmatrix}0 & I_p \\ I_p & 0\end{pmatrix}. L ( s ) = ( 0 I p I p 0 ) L ( s ) ( 0 I p I p 0 ) .
The first identity is immediate from the signs of the diagonal blocks; the second uses V = M ‾ V=\overline{M} V = M . Squaring these identities and applying the real-coefficient series (41) gives
Y ( − s ) 00 = Y ( s ) 00 , Y ( − s ) p p = Y ( s ) p p , Y(-s)_{00}=Y(s)_{00}, \qquad Y(-s)_{pp}=Y(s)_{pp}, Y ( − s ) 00 = Y ( s ) 00 , Y ( − s ) pp = Y ( s ) pp ,
Y ( s ‾ ) ‾ 00 = Y ( s ) p p , Y ( s ‾ ) ‾ p p = Y ( s ) 00 . \overline{Y(\overline{s})}_{00}=Y(s)_{pp}, \qquad\overline{Y(\overline{s})}_{pp}=Y(s)_{00}. Y ( s ) 00 = Y ( s ) pp , Y ( s ) pp = Y ( s ) 00 .
Consequently I ( − s ) = I ( s ) \mathcal{I}(-s)=\mathcal{I}(s) I ( − s ) = I ( s ) and I ( s ‾ ) ‾ = I ( s ) \overline{\mathcal{I}(\overline{s})}=\mathcal{I}(s) I ( s ) = I ( s ) : its Taylor series is even and has real coefficients. In particular, I ( i η ) \mathcal{I}(i\eta) I ( i η ) is real for small real η \eta η .
To define the angular average, put c ( ψ ) = 1 2 cos ψ c(\psi)=\frac{1}{2}\cos\psi c ( ψ ) = 2 1 cos ψ , h ( ψ ) = sin ψ h(\psi)=\sin\psi h ( ψ ) = sin ψ , and write I p ( s , ψ ) = I ( s ; c ( ψ ) , h ( ψ ) ) \mathcal{I}_p(s,\psi)=\mathcal{I}(s;c(\psi),h(\psi)) I p ( s , ψ ) = I ( s ; c ( ψ ) , h ( ψ )) . Then
G p ( t ) = 1 π ∫ − π / 2 π / 2 Y ( t / p ) 00 + Y ( t / p ) p p 2 d ψ = 1 π ∫ − π / 2 π / 2 I p ( t / p , ψ ) c ( ψ ) d ψ . (42) \begin{aligned}
G_p(t)&=\frac{1}{\pi}\int_{-\pi/2}^{\pi/2}\frac{Y(t/p)_{00}+Y(t/p)_{pp}}{2}\,d\psi\\
&=\frac{1}{\pi}\int_{-\pi/2}^{\pi/2}\frac{\mathcal{I}_p(t/p,\psi)}{c(\psi)}\,d\psi. \tag*{(42)}
\end{aligned} G p ( t ) = π 1 ∫ − π /2 π /2 2 Y ( t / p ) 00 + Y ( t / p ) pp d ψ = π 1 ∫ − π /2 π /2 c ( ψ ) I p ( t / p , ψ ) d ψ . (42)
The first line defines the values at the endpoints, where c = 0 c=0 c = 0 . For fixed p p p all entries of L L L are continuous on the closed angular interval. Thus the series (41) converges uniformly there for ∣ s ∣ |s| ∣ s ∣ sufficiently small, and (42) defines an analytic function near t = 0 t=0 t = 0 . This also explains why division by c c c in the second line creates no endpoint singularity.
For orientation, the case p = 1 p=1 p = 1 gives M = V = c / 2 M=V=c/2 M = V = c /2 and L ( s ) 2 = ( 1 − s 2 c 2 / 4 ) I 2 L(s)^2=(1-s^2c^2/4)I_2 L ( s ) 2 = ( 1 − s 2 c 2 /4 ) I 2 . Hence I ( s ) = c ( 1 − s 2 c 2 / 4 ) − 1 / 2 \mathcal{I}(s)=c(1-s^2c^2/4)^{-1/2} I ( s ) = c ( 1 − s 2 c 2 /4 ) − 1/2 . The same branch and the same normalization are used at every matrix size.
Proposition 5.1 (Reduction for a fixed odd numerator). Fix an odd positive integer p p p and a residue q 0 q_0 q 0 modulo 4 p 4p 4 p with ( p , q 0 ) = 1 (p,q_0)=1 ( p , q 0 ) = 1 . Define G p G_p G p by (42) , using any odd lift q ^ ≡ q 0 ( m o d p ) \widehat{q}\equiv q_0\pmod p q ≡ q 0 ( mod p ) in its cotangents. For every fixed m ≥ 0 m\geq0 m ≥ 0 ,
lim q → ∞ q ≡ q 0 ( m o d 4 p ) q P p , q ( 2 m ) ( 0 ) ( 2 m ) ! = [ t 2 m ] G p ( t ) . \lim_{\substack{q\to\infty\\q\equiv q_0\pmod{4p}}}\frac{q\mathcal{P}_{p,q}^{(2m)}(0)}{(2m)!}=[t^{2m}]G_p(t). q → ∞ q ≡ q 0 ( mod 4 p ) lim ( 2 m )! q P p , q ( 2 m ) ( 0 ) = [ t 2 m ] G p ( t ) . The reduction identifies the limit at each fixed Taylor order. Appendix A proves it by rewriting the selected diagonal coefficient as a cut pairing, subtracting the exact principal contribution, and passing to the limit in the regular rows.
A commuting matrix that extracts the required entries
We now turn the two entries in I \mathcal{I} I into a determinant derivative. For this we seek an invertible matrix U ( s ) U(s) U ( s ) commuting with L ( s ) L(s) L ( s ) and depending on s s s in only one entry. We write U s = d U ( s ) / d s U_s=dU(s)/ds U s = d U ( s ) / d s . A logarithmic determinant derivative is a trace involving U − 1 U s U^{-1}U_s U − 1 U s . The one-entry dependence makes this product have rank at most one, so the trace can select the required entries. Commutation ensures that U U U preserves the two spectral groups of L L L .
The matrix C e C_e C e will be a unitary change of coordinates matching M M M with − V -V − V , up to one endpoint term. Such a correction is necessary when c ≠ 0 c\ne0 c = 0 : Tr M = c / 2 \operatorname{Tr}M=c/2 Tr M = c /2 , whereas Tr ( − V ) = − c / 2 \operatorname{Tr}(-V)=-c/2 Tr ( − V ) = − c /2 . With E r = e r e r T E_r=e_re_r^T E r = e r e r T , we seek M C e + C e V = c E p − 1 MC_e+C_eV=cE_{p-1} M C e + C e V = c E p − 1 , so that all other entries cancel. The adjacent two-coordinate blocks below give this cancellation. For a real angle ϕ \phi ϕ , put
C ( ϕ ) = ( − cos ϕ − i h sin ϕ 2 c sin ϕ 2 c sin ϕ cos ϕ − i h sin ϕ ) . C(\phi)=\begin{pmatrix}
-\cos\phi-ih\sin\phi& 2c\sin\phi\\
2c\sin\phi& \cos\phi-ih\sin\phi
\end{pmatrix}. C ( ϕ ) = ( − cos ϕ − ih sin ϕ 2 c sin ϕ 2 c sin ϕ cos ϕ − ih sin ϕ ) .
The full p × p p\times p p × p block diagonal matrix is
C e = ( C ( θ ) 0 ⋯ 0 0 0 C ( 3 θ ) ⋱ ⋮ ⋮ ⋮ ⋱ ⋱ 0 0 0 ⋯ 0 C ( ( p − 2 ) θ ) 0 0 ⋯ 0 0 1 ) . C_e=
\begin{pmatrix}
C(\theta)&0&\cdots&0&0\\
0&C(3\theta)&\ddots&\vdots&\vdots\\
\vdots&\ddots&\ddots&0&0\\
0&\cdots&0&C((p-2)\theta)&0\\
0&\cdots&0&0&1
\end{pmatrix}. C e = C ( θ ) 0 ⋮ 0 0 0 C ( 3 θ ) ⋱ ⋯ ⋯ ⋯ ⋱ ⋱ 0 0 0 ⋮ 0 C (( p − 2 ) θ ) 0 0 ⋮ 0 0 1 .
The blocks act successively on coordinates ( 0 , 1 ) (0,1) ( 0 , 1 ) , ( 2 , 3 ) , … , ( p − 3 , p − 2 ) (2,3),\ldots,(p-3,p-2) ( 2 , 3 ) , … , ( p − 3 , p − 2 ) , and the final coordinate p − 1 p-1 p − 1 has diagonal entry 1 1 1 . For p = 1 p=1 p = 1 , there are no 2 × 2 2\times2 2 × 2 blocks and C e = ( 1 ) C_e=(1) C e = ( 1 ) . For p = 3 p=3 p = 3 , the full matrix is
C e = ( − cos θ − i h sin θ 2 c sin θ 0 2 c sin θ cos θ − i h sin θ 0 0 0 1 ) . C_e=
\begin{pmatrix}
-\cos\theta-ih\sin\theta&2c\sin\theta&0\\
2c\sin\theta&\cos\theta-ih\sin\theta&0\\
0&0&1
\end{pmatrix}. C e = − cos θ − ih sin θ 2 c sin θ 0 2 c sin θ cos θ − ih sin θ 0 0 0 1 .
Each block satisfies C ( ϕ ) ∗ C ( ϕ ) = I 2 C(\phi)^{*}C(\phi)=I_{2} C ( ϕ ) ∗ C ( ϕ ) = I 2 and has determinant − 1 -1 − 1 for real c , h c,h c , h . Hence C e C_{e} C e is unitary and det C e = ( − 1 ) n \det C_{e}=(-1)^{n} det C e = ( − 1 ) n . This unitary property will put the eigenvalues of U ( 0 ) U(0) U ( 0 ) on the unit circle; the cancellation checked next makes its deformation commute with L L L . Appendix § 10 verifies the cancellation directly from three local checks: within a pair, between two pairs, and at the final coordinate:
M C e + C e V = c E p − 1 . (43) MC_{e}+C_{e}V=cE_{p-1}. \tag*{(43)} M C e + C e V = c E p − 1 . (43)
Since C e C_{e} C e is unitary and fixes the final coordinate, this identity can also be written
C e − 1 M C e = − V + c E p − 1 . C_{e}^{-1}MC_{e}=-V+cE_{p-1}. C e − 1 M C e = − V + c E p − 1 .
Thus the unitary changes on adjacent pairs leave just one discrepancy, at the unpaired final coordinate. This is the endpoint term that determines the corner in U U U . Define C o = J C e J C_{o}=JC_{e}J C o = J C e J and N = e 0 e p − 1 T N=e_{0}e_{p-1}^{\mathsf{T}} N = e 0 e p − 1 T , and put
U ( s ) = ( 0 C e C o i s c N ) . U(s)=\begin{pmatrix} 0 & C_{e} \\ C_{o} & iscN \end{pmatrix}. U ( s ) = ( 0 C o C e i sc N ) .
At s = 0 s=0 s = 0 the reversal relation C o = J C e J C_{o}=JC_{e}J C o = J C e J already gives [ U ( 0 ) , L ( 0 ) ] = 0 [U(0),L(0)]=0 [ U ( 0 ) , L ( 0 )] = 0 . For nonzero s s s , the corner i s c N iscN i sc N is precisely what maintains commutation. Indeed, transpose and reversal of (43) give C o M + V C o = c E 0 C_{o}M+VC_{o}=cE_{0} C o M + V C o = c E 0 . Also [ N , V ] = 0 [N,V]=0 [ N , V ] = 0 , because the first column and last row of V V V contain only the same diagonal value c / 2 c/2 c /2 . The two off-diagonal blocks of [ U , L ] [U,L] [ U , L ] are
i s ( M C e + C e V − c E p − 1 ) , − i s ( C o M + V C o − c E 0 ) , is(MC_{e}+C_{e}V-cE_{p-1}),\qquad-is(C_{o}M+VC_{o}-cE_{0}), i s ( M C e + C e V − c E p − 1 ) , − i s ( C o M + V C o − c E 0 ) ,
and the diagonal blocks vanish by C o = J C e J C_{o}=JC_{e}J C o = J C e J and [ N , V ] = 0 [N,V]=0 [ N , V ] = 0 . Thus
[ U ( s ) , L ( s ) ] = 0. (44) [U(s),L(s)]=0. \tag*{(44)} [ U ( s ) , L ( s )] = 0. (44)
The singleton entries of C e C_{e} C e and C o C_{o} C o give C o − 1 N = N C e − 1 = N C_{o}^{-1}N=NC_{e}^{-1}=N C o − 1 N = N C e − 1 = N . Direct multiplication then gives
U ( s ) − 1 = ( − i s c N C o − 1 C e − 1 0 ) , det U ( s ) = − 1. (45) U(s)^{-1}=\begin{pmatrix} -iscN & C_{o}^{-1} \\ C_{e}^{-1} & 0 \end{pmatrix},\qquad\det U(s)=-1. \tag*{(45)} U ( s ) − 1 = ( − i sc N C e − 1 C o − 1 0 ) , det U ( s ) = − 1. (45)
The determinant is independent of the lower-right block: exchanging the two groups of p p p columns makes the matrix block triangular, so
det U ( s ) = ( − 1 ) p det C e det C o = − 1. \det U(s)=(-1)^{p}\det C_{e}\det C_{o}=-1. det U ( s ) = ( − 1 ) p det C e det C o = − 1.
At zero, both C e C_{e} C e and C o C_{o} C o are unitary, so U ( 0 ) U(0) U ( 0 ) is unitary in the ordinary inner product. We will use this only to locate the roots of its undeformed characteristic polynomial. Commutation will preserve the two analytic subspaces already defined; an explicit identity for the fixed form L ( 0 ) L(0) L ( 0 ) will later locate their roots after the imaginary shift. We record the block formulas needed for the determinant calculation:
U ( s ) = ( 0 C e C o i s c N ) , U ( s ) − 1 = ( − i s c N C o − 1 C e − 1 0 ) . (46) U(s)=\begin{pmatrix} 0 & C_{e} \\ C_{o} & iscN \end{pmatrix},\qquad U(s)^{-1}=\begin{pmatrix} -iscN & C_{o}^{-1} \\ C_{e}^{-1} & 0 \end{pmatrix}. \tag*{(46)} U ( s ) = ( 0 C o C e i sc N ) , U ( s ) − 1 = ( − i sc N C e − 1 C o − 1 0 ) . (46)
Turning the matrix entries into determinant derivatives
The full determinant of U U U is the constant − 1 -1 − 1 , whose logarithmic derivative is zero. To recover I I I , we therefore take the determinants on the two subspaces selected by S S S separately. Their logarithmic derivatives enter with opposite signs, just as the eigenvalue derivatives did in Proposition 3.2 .
Since U U U commutes with L L L , it commutes with every term in the series for Y Y Y , hence with S S S and P ± P_{\pm} P ± . Thus it preserves the complementary spaces E ± ( s ) = Ran P ± ( s ) E_{\pm}(s)=\operatorname{Ran}P_{\pm}(s) E ± ( s ) = Ran P ± ( s ) for every sufficiently small complex s s s .
Define two scalar functions by ordinary 2 p × 2 p 2p\times2p 2 p × 2 p determinants:
D ± ( s ) = det ( I − P ± ( s ) + U ( s ) P ± ( s ) ) . D_{\pm}(s)=\det\left(I-P_{\pm}(s)+U(s)P_{\pm}(s)\right). D ± ( s ) = det ( I − P ± ( s ) + U ( s ) P ± ( s ) ) .
The matrix inside acts as U U U on E ± E_{\pm} E ± and as the identity on the other subspace. Thus D ± D_{\pm} D ± is exactly the determinant of the restriction of U U U to E ± E_{\pm} E ± , without choosing a basis there. The displayed formula makes analyticity immediate. Its inverse is I − P ± + U − 1 P ± I-P_{\pm}+U^{-1}P_{\pm} I − P ± + U − 1 P ± , so D ± D_{\pm} D ± never vanishes near zero. Multiplying the two matrices inside the determinants gives U U U , hence D + D − = det U = − 1 D_{+}D_{-}=\det U=-1 D + D − = det U = − 1 .
The matrix F F F is the restriction of U ( 0 ) U(0) U ( 0 ) to the first of these subspaces, written in the coordinate v v v . Explicitly, put F = C e J F=C_eJ F = C e J ; then
U ( 0 ) ( v J v ) = ( F v J F v ) , U ( 0 ) ( v − J v ) = ( − F v J F v ) . U(0)\begin{pmatrix}v\\Jv\end{pmatrix}=\begin{pmatrix}Fv\\JFv\end{pmatrix},\qquad U(0)\begin{pmatrix}v\\-Jv\end{pmatrix}=\begin{pmatrix}-Fv\\JFv\end{pmatrix}. U ( 0 ) ( v J v ) = ( F v J F v ) , U ( 0 ) ( v − J v ) = ( − F v J F v ) .
Thus the two restrictions are F F F and − F -F − F . This explains the role of F F F : it records the starting matrix for the one-entry deformation U ( s ) U(s) U ( s ) . Since det J = ( − 1 ) n \det J=(-1)^n det J = ( − 1 ) n and det C e = ( − 1 ) n \det C_e=(-1)^n det C e = ( − 1 ) n , we have det F = 1 \det F=1 det F = 1 . As p p p is odd, D + ( 0 ) = 1 D_{+}(0)=1 D + ( 0 ) = 1 and D − ( 0 ) = − 1 D_{-}(0)=-1 D − ( 0 ) = − 1 . The labels + + + and − - − refer to the two subspaces of L L L .
Both C e C_e C e and J J J are unitary, so F = C e J F=C_eJ F = C e J is unitary. Moreover,
F 2 = C e C o , det ( z I 2 p − U ( 0 ) ) = det ( z I p − F ) det ( z I p + F ) = det ( z 2 I p − F 2 ) . F^2=C_eC_o,\qquad\det(zI_{2p}-U(0))=\det(zI_p-F)\det(zI_p+F)=\det(z^2I_p-F^2). F 2 = C e C o , det ( z I 2 p − U ( 0 )) = det ( z I p − F ) det ( z I p + F ) = det ( z 2 I p − F 2 ) .
This is why F 2 F^2 F 2 appears in the scalar polynomial below. The next proposition gives both parts of the conversion: the first identity extracts I \mathcal{I} I by a determinant derivative, and the second computes the characteristic polynomial of U ( s ) U(s) U ( s ) .
Proposition 5.2 (The two determinant identities). For sufficiently small s s s ,
I ( s ) = 1 i d d s log D + ( s ) D − ( s ) = 2 i d d s log D + ( s ) . (47) \mathcal{I}(s)=\frac{1}{i}\frac{d}{ds}\log\frac{D_{+}(s)}{D_{-}(s)}=\frac{2}{i}\frac{d}{ds}\log D_{+}(s). \tag*{(47)} I ( s ) = i 1 d s d log D − ( s ) D + ( s ) = i 2 d s d log D + ( s ) . (47)
Logarithmic derivatives mean D ± ′ / D ± D_{\pm}'/D_{\pm} D ± ′ / D ± , so no choice of logarithm affects this formula. In addition,
det ( z I 2 p − U ( s ) ) = det ( z 2 I p − F 2 ) − i p s c p z p . (48) \det(zI_{2p}-U(s))=\det(z^2I_p-F^2)-ipsc^pz^p. \tag*{(48)} det ( z I 2 p − U ( s )) = det ( z 2 I p − F 2 ) − i p s c p z p . (48)
Proof. 1. The single-entry derivative selects the required entries. The inverse already computed gives
U − 1 U s = i c ( 0 N 0 0 ) . (49) U^{-1}U_s=ic\begin{pmatrix}0&N\\0&0\end{pmatrix}. \tag*{(49)} U − 1 U s = i c ( 0 0 N 0 ) . (49)
We identify the entry of S S S selected by this matrix. With R = diag ( J , J ) \mathcal{R}=\operatorname{diag}(J,J) R = diag ( J , J ) , persymmetry gives R L T R = L \mathcal{R}L^{T}\mathcal{R}=L R L T R = L , hence R Y T R = Y \mathcal{R}Y^{T}\mathcal{R}=Y R Y T R = Y . In particular, Y 2 p − 1 , 2 p − 1 = Y p p Y_{2p-1,2p-1}=Y_{pp} Y 2 p − 1 , 2 p − 1 = Y pp . The last row of L L L has just the entries 1 1 1 and i s c / 2 isc/2 i sc /2 , in columns 0 0 0 and 2 p − 1 2p-1 2 p − 1 ; its first column has just the entries − i s c / 2 -isc/2 − i sc /2 and 1 1 1 , in rows 0 0 0 and 2 p − 1 2p-1 2 p − 1 . Computing the same entry of S = L Y = Y L S=LY=YL S = L Y = Y L in these two ways therefore gives
( L Y ) 2 p − 1 , 0 = Y 00 + i s c 2 Y 2 p − 1 , 0 , ( Y L ) 2 p − 1 , 0 = Y p p − i s c 2 Y 2 p − 1 , 0 . (LY)_{2p-1,0}=Y_{00}+\frac{isc}{2}Y_{2p-1,0},\qquad(YL)_{2p-1,0}=Y_{pp}-\frac{isc}{2}Y_{2p-1,0}. ( L Y ) 2 p − 1 , 0 = Y 00 + 2 i sc Y 2 p − 1 , 0 , ( Y L ) 2 p − 1 , 0 = Y pp − 2 i sc Y 2 p − 1 , 0 .
Equating these expressions, and then averaging them, yields
Y p p − Y 00 = i s c Y 2 p − 1 , 0 , Y_{pp}-Y_{00}=iscY_{2p-1,0}, Y pp − Y 00 = i sc Y 2 p − 1 , 0 ,
S 2 p − 1 , 0 = Y 00 + i s c 2 Y 2 p − 1 , 0 = Y p p − i s c 2 Y 2 p − 1 , 0 = Y 00 + Y p p 2 . (50) S_{2p-1,0}=Y_{00}+\frac{isc}{2}Y_{2p-1,0}
=Y_{pp}-\frac{isc}{2}Y_{2p-1,0}
=\frac{Y_{00}+Y_{pp}}{2}. \tag*{(50)} S 2 p − 1 , 0 = Y 00 + 2 i sc Y 2 p − 1 , 0 = Y pp − 2 i sc Y 2 p − 1 , 0 = 2 Y 00 + Y pp . (50)
It follows that Tr ( S U − 1 U s ) = i I \operatorname{Tr}(SU^{-1}U_s)=i\mathcal{I} Tr ( S U − 1 U s ) = i I .
2. Differentiate the full-space determinants. We must account for the dependence of the projection on s s s . For either sign, write P = P ± P=P_{\pm} P = P ± , Q = I − P Q=I-P Q = I − P , and W = Q + U P W=Q+UP W = Q + U P , so D ± = det W D_{\pm}=\det W D ± = det W . Since U U U commutes with P P P ,
W − 1 = Q + U − 1 P , W s = U s P + ( U − I ) P s . W^{-1}=Q+U^{-1}P,\qquad W_s=U_sP+(U-I)P_s. W − 1 = Q + U − 1 P , W s = U s P + ( U − I ) P s .
Differentiating P 2 = P P^2=P P 2 = P gives P s P + P P s = P s P_sP+PP_s=P_s P s P + P P s = P s . Multiplying on both sides by P P P , and then by Q Q Q , shows that
P P s P = 0 , Q P s Q = 0. PP_sP=0,\qquad QP_sQ=0. P P s P = 0 , Q P s Q = 0.
At a fixed s s s , choose a basis adapted to Ran P ⊕ Ran Q \operatorname{Ran} P \oplus\operatorname{Ran} Q Ran P ⊕ Ran Q , and write U P , U Q U_P,U_Q U P , U Q for the restrictions of U U U to these two subspaces. In this basis,
P = ( I p 0 0 0 ) , U = ( U P 0 0 U Q ) . P=\begin{pmatrix} I_p & 0 \\ 0 & 0 \end{pmatrix}, \qquad U=\begin{pmatrix} U_P & 0 \\ 0 & U_Q \end{pmatrix}. P = ( I p 0 0 0 ) , U = ( U P 0 0 U Q ) .
The two matrices whose product concerns us are explicitly
P s = ( 0 ( P s ) 12 ( P s ) 21 0 ) , P_s=\begin{pmatrix} 0 & (P_s)_{12} \\ (P_s)_{21} & 0 \end{pmatrix}, P s = ( 0 ( P s ) 21 ( P s ) 12 0 ) ,
W − 1 ( U − I ) = ( I p − U P − 1 0 0 U Q − I p ) . W^{-1}(U-I)=\begin{pmatrix} I_p-U_P^{-1} & 0 \\ 0 & U_Q-I_p \end{pmatrix}. W − 1 ( U − I ) = ( I p − U P − 1 0 0 U Q − I p ) .
Their product has zero diagonal blocks, and therefore
Tr ( W − 1 ( U − I ) P s ) = 0. \operatorname{Tr}\bigl(W^{-1}(U-I)P_s\bigr)=0. Tr ( W − 1 ( U − I ) P s ) = 0.
This basis is used only to display the blocks at the fixed parameter; we do not differentiate a moving basis. Orthogonality of the two subspaces is not needed.
The determinant differentiation formula is ( det W ) ′ / det W = Tr ( W − 1 W s ) (\det W)'/\det W=\operatorname{Tr}(W^{-1}W_s) ( det W ) ′ / det W = Tr ( W − 1 W s ) . Using the zero trace just proved, then cyclic invariance of trace and P W − 1 = P U − 1 PW^{-1}=PU^{-1} P W − 1 = P U − 1 , we obtain
D ± ′ D ± = Tr ( W − 1 W s ) = Tr ( W − 1 U s P ) = Tr ( P W − 1 U s ) = Tr ( P ± U − 1 U s ) . \begin{aligned}
\frac{D_\pm'}{D_\pm}=\operatorname{Tr}(W^{-1}W_s)=\operatorname{Tr}(W^{-1}U_sP) \\
&=\operatorname{Tr}(PW^{-1}U_s)=\operatorname{Tr}(P_\pm U^{-1}U_s).
\end{aligned} D ± D ± ′ = Tr ( W − 1 W s ) = Tr ( W − 1 U s P ) = Tr ( P W − 1 U s ) = Tr ( P ± U − 1 U s ) .
Subtract the two equations and use P + − P − = S P_+-P_- = S P + − P − = S :
d d s log D + D − = Tr ( S U − 1 U s ) = i I . \frac{d}{ds}\log\frac{D_+}{D_-}=\operatorname{Tr}(SU^{-1}U_s)=iI. d s d log D − D + = Tr ( S U − 1 U s ) = i I .
Since D + D − = − 1 D_+D_-=-1 D + D − = − 1 , their logarithmic derivatives sum to zero. This proves both equalities in (47) .
3. The characteristic polynomial has a single additional term. For z ≠ 0 z\ne0 z = 0 , use the upper-left block z I p zI_p z I p to eliminate the lower-left block − C o -C_o − C o in z I 2 p − U ( s ) zI_{2p}-U(s) z I 2 p − U ( s ) . Specifically, add z − 1 C o z^{-1}C_o z − 1 C o times the first block row to the second. The result is
( z I p − C e 0 z I p − i s c N − z − 1 C o C e ) . \begin{pmatrix}
zI_p & -C_e \\
0 & zI_p-iscN-z^{-1}C_oC_e
\end{pmatrix}. ( z I p 0 − C e z I p − i sc N − z − 1 C o C e ) .
Taking its determinant and absorbing the factor z p z^p z p gives
det ( z I 2 p − U ( s ) ) = z p det ( z I p − i s c N − z − 1 C o C e ) = det ( z 2 I p − C o C e − i s c z N ) . \det(zI_{2p}-U(s))=z^p\det(zI_p-iscN-z^{-1}C_oC_e)=\det(z^2I_p-C_oC_e-isc zN). det ( z I 2 p − U ( s )) = z p det ( z I p − i sc N − z − 1 C o C e ) = det ( z 2 I p − C o C e − i scz N ) .
The equality at z = 0 z=0 z = 0 follows by polynomial identity.
We use the rank-one determinant formula
det ( A + τ u v T ) = det A + τ v T adj ( A ) u . \det(A+\tau uv^T)=\det A+\tau v^T\operatorname{adj}(A)u. det ( A + τ u v T ) = det A + τ v T adj ( A ) u .
Here adj ( A ) \operatorname{adj}(A) adj ( A ) is the transpose of the cofactor matrix. The formula remains valid when A A A is singular. With N = e 0 e p − 1 T N=e_0e_{p-1}^T N = e 0 e p − 1 T , it yields
det ( z I 2 p − U ( s ) ) = det ( z 2 I p − C o C e ) − i s c z adj ( z 2 I p − C o C e ) p − 1 , 0 . \det(zI_{2p}-U(s))=\det(z^2I_p-C_oC_e)-isc z\operatorname{adj}(z^2I_p-C_oC_e)_{p-1,0}. det ( z I 2 p − U ( s )) = det ( z 2 I p − C o C e ) − i scz adj ( z 2 I p − C o C e ) p − 1 , 0 .
The required corner cofactor is
adj ( x I p − C o C e ) p − 1 , 0 = x n ∏ r = 0 p − 2 b r , b r = 2 c sin ( ( r + 1 ) θ ) . (51) \operatorname{adj}(xI_p-C_oC_e)_{p-1,0}
=x^n\prod_{r=0}^{p-2}b_r,\qquad b_r=2c\sin((r+1)\theta). \tag*{(51)} adj ( x I p − C o C e ) p − 1 , 0 = x n r = 0 ∏ p − 2 b r , b r = 2 c sin (( r + 1 ) θ ) . (51)
To verify it, factor x I p − C o C e = C o ( x C o − 1 − C e ) xI_p-C_oC_e=C_o(xC_o^{-1}-C_e) x I p − C o C e = C o ( x C o − 1 − C e ) . Since C o − 1 e 0 = e 0 C_o^{-1}e_0=e_0 C o − 1 e 0 = e 0 , the lower-left inverse corners agree. Multiplication by their determinants therefore gives
adj ( x I p − C o C e ) p − 1 , 0 = det C o adj ( x C o − 1 − C e ) p − 1 , 0 . \operatorname{adj}(xI_p-C_oC_e)_{p-1,0}
=\det C_o\,\operatorname{adj}(xC_o^{-1}-C_e)_{p-1,0}. adj ( x I p − C o C e ) p − 1 , 0 = det C o adj ( x C o − 1 − C e ) p − 1 , 0 .
This argument for nonsingular matrices extends to every x x x by polynomial identity. The matrix x C o − 1 − C e xC_o^{-1}-C_e x C o − 1 − C e is symmetric tridiagonal. Its successive off-diagonal entries are
− b 0 , x b 1 , − b 2 , x b 3 , … , x b p − 2 . -b_0,\ xb_1,\ -b_2,\ xb_3,\ldots,xb_{p-2}. − b 0 , x b 1 , − b 2 , x b 3 , … , x b p − 2 .
Reflection preserves the bond values because sin ( ( p − 1 − r ) θ ) = sin ( ( r + 1 ) θ ) \sin((p-1-r)\theta)=\sin((r+1)\theta) sin (( p − 1 − r ) θ ) = sin (( r + 1 ) θ ) for odd q ^ \widehat{q} q . The matrix C e C_e C e has blocks on the even bonds and C o − 1 C_o^{-1} C o − 1 on the odd bonds; the off-diagonal entry of the inverse of a symmetric two-by-two block of determinant − 1 -1 − 1 equals its original off-diagonal entry. The corner adjugate of a symmetric tridiagonal matrix is ( − 1 ) p − 1 (-1)^{p-1} ( − 1 ) p − 1 times the product of its off-diagonal entries: the defining minor is triangular with precisely these diagonal entries. Since p − 1 = 2 n p-1=2n p − 1 = 2 n , the product here is ( − 1 ) n x n ∏ b r (-1)^n x^n\prod b_r ( − 1 ) n x n ∏ b r . The factor det C o = ( − 1 ) n \det C_o=(-1)^n det C o = ( − 1 ) n cancels this sign, proving (51) .
Finally,
∏ j = 1 p − 1 2 sin ( π q ^ j / p ) = p , ∏ r = 0 p − 2 b r = p c p − 1 . \prod_{j=1}^{p-1}2\sin(\pi\widehat{q}j/p)=p,\qquad\prod_{r=0}^{p-2}b_r=pc^{p-1}. j = 1 ∏ p − 1 2 sin ( π q j / p ) = p , r = 0 ∏ p − 2 b r = p c p − 1 .
The absolute value of the first product is the usual product over the nontrivial p p p th roots of unity; its sign is positive because q ^ \widehat{q} q is odd and the terms for j , p − j j,p-j j , p − j are equal. Thus the additional term, with x = z 2 x=z^2 x = z 2 , is
− i s c z p − 1 z 2 n = − i p s c p z p . -iscz^{p-1}z^{2n}=-ipsc^pz^p. − i sc z p − 1 z 2 n = − i p s c p z p .
Also C o C e C_oC_e C o C e is similar to C e C o = F 2 C_eC_o=F^2 C e C o = F 2 , since conjugation by C e C_e C e interchanges the two products. This proves (48) . Empty products give the same calculation when p = 1 p=1 p = 1 . □
Which determinant factor corresponds to the inside roots?
Jensen’s formula distinguishes roots inside and outside the unit circle. We take s = i η s=i\eta s = i η , η > 0 \eta>0 η > 0 , and identify the factor D + D_+ D + with the inside roots. The fixed Hermitian form L ( 0 ) L(0) L ( 0 ) will identify which of the two continued subspaces contains them. Since the matrix function is analytic near zero, its values on this imaginary segment suffice to determine its Taylor series. For the rest of this subsection assume c > 0 c>0 c > 0 and write
D 0 ( z ) = det ( z 2 I p − F 2 ) , D η ( z ) = D 0 ( z ) + p η c p z p = det ( z I 2 p − U ( i η ) ) , η > 0. D_0(z)=\det(z^2I_p-F^2),\qquad D_\eta(z)=D_0(z)+p\eta c^p z^p=\det(zI_{2p}-U(i\eta)),\qquad\eta>0. D 0 ( z ) = det ( z 2 I p − F 2 ) , D η ( z ) = D 0 ( z ) + p η c p z p = det ( z I 2 p − U ( i η )) , η > 0.
The notation D η D_\eta D η denotes a characteristic polynomial in z z z ; D ± D_\pm D ± denote the two scalar restriction determinants. We will relate them using Jensen’s formula.
First, D η D_\eta D η has no root on the unit circle. The matrix U ( 0 ) U(0) U ( 0 ) is unitary and has determinant − 1 -1 − 1 , so its monic characteristic polynomial satisfies
D 0 ( z ) = − z 2 p D 0 ( 1 / z ‾ ) ‾ . D_0(z)=-z^{2p}\overline{D_0(1/\overline{z})}. D 0 ( z ) = − z 2 p D 0 ( 1/ z ) .
Thus e − i p k D 0 ( e i k ) e^{-ipk}D_0(e^{ik}) e − i p k D 0 ( e ik ) is purely imaginary, while p η c p p\eta c^p p η c p is positive. Their sum cannot vanish.
The absence of unit-circle roots does not yet identify which group is inside. For this we use the fixed Hermitian form x ∗ L ( 0 ) x x^*L(0)x x ∗ L ( 0 ) x . At zero it is positive on E + ( 0 ) E_+(0) E + ( 0 ) and negative on E − ( 0 ) E_-(0) E − ( 0 ) , because L ( 0 ) L(0) L ( 0 ) acts there as + I +I + I and − I -I − I , respectively. These signs persist on the continued subspaces for small complex s s s . Thus the sign of the form can label the two groups after the imaginary shift. For an explicit check, choose δ < 1 / 2 \delta<1/2 δ < 1/2 with ∥ P ± ( s ) − P ± ( 0 ) ∥ < δ \|\mathcal{P}_\pm(s)-\mathcal{P}_\pm(0)\|<\delta ∥ P ± ( s ) − P ± ( 0 ) ∥ < δ . For nonzero x ∈ E + ( s ) x\in E_+(s) x ∈ E + ( s ) one has ∥ ( I − P + ( 0 ) ) x ∥ ≤ δ ∥ x ∥ \|(I-\mathcal{P}_+(0))x\|\leq\delta\|x\| ∥ ( I − P + ( 0 )) x ∥ ≤ δ ∥ x ∥ , and hence
x ∗ L ( 0 ) x = ∥ x ∥ 2 − 2 ∥ ( I − P + ( 0 ) ) x ∥ 2 ≥ ( 1 − 2 δ 2 ) ∥ x ∥ 2 > 0. \begin{aligned}
x^*L(0)x=\|x\|^2-2\|(I-\mathcal{P}_+(0))x\|^2 \\
&\geq(1-2\delta^2)\|x\|^2>0.
\end{aligned} x ∗ L ( 0 ) x = ∥ x ∥ 2 − 2∥ ( I − P + ( 0 )) x ∥ 2 ≥ ( 1 − 2 δ 2 ) ∥ x ∥ 2 > 0.
The same calculation gives x ∗ L ( 0 ) x ≤ − ( 1 − 2 δ 2 ) ∥ x ∥ 2 < 0 x^*L(0)x\leq-(1-2\delta^2)\|x\|^2<0 x ∗ L ( 0 ) x ≤ − ( 1 − 2 δ 2 ) ∥ x ∥ 2 < 0 for nonzero x ∈ E − ( s ) x\in E_-(s) x ∈ E − ( s ) . The neighborhood can be chosen uniformly in the angular variable, by compactness and the uniformly convergent projection series. To connect these signs with the moduli of the roots, multiply the explicit blocks of U U U once more, now allowing complex s s s . This gives
U ( s ) ∗ L ( 0 ) U ( s ) − L ( 0 ) = − 2 c Im s ( 0 0 0 E p − 1 ) . U(s)^*L(0)U(s)-L(0)=-2c\,\operatorname{Im}s
\begin{pmatrix}
0 & 0\\
0 & E_{p-1}
\end{pmatrix}. U ( s ) ∗ L ( 0 ) U ( s ) − L ( 0 ) = − 2 c Im s ( 0 0 0 E p − 1 ) .
Indeed, the off-diagonal blocks equal J J J , while the last diagonal block is i c ( s − s ˉ ) E p − 1 ic(s-\bar{s})E_{p-1} i c ( s − s ˉ ) E p − 1 . Here C e ∗ J N = N ∗ J C e = E p − 1 C_e^*JN=N^*JC_e=E_{p-1} C e ∗ J N = N ∗ J C e = E p − 1 because the last entry of C e C_e C e is 1. Thus the departure from preservation of the fixed form occurs in only one coordinate, with a known sign. For an eigenvector U ( i η ) x = λ x U(i\eta)x=\lambda x U ( i η ) x = λ x with x ∈ E ± ( i η ) x\in E_{\pm}(i\eta) x ∈ E ± ( i η ) , the displayed identity gives
( ∣ λ ∣ 2 − 1 ) x ∗ L ( 0 ) x = − 2 c η ∣ x 2 p − 1 ∣ 2 ≤ 0. \left(|\lambda|^2-1\right)x^*L(0)x=-2c\eta|x_{2p-1}|^2\leq0. ( ∣ λ ∣ 2 − 1 ) x ∗ L ( 0 ) x = − 2 cη ∣ x 2 p − 1 ∣ 2 ≤ 0.
Thus the + + + group has ∣ λ ∣ ≤ 1 |\lambda|\leq1 ∣ λ ∣ ≤ 1 , and the − - − group has ∣ λ ∣ ≥ 1 |\lambda|\geq1 ∣ λ ∣ ≥ 1 . The absence of unit-circle roots makes both inequalities strict. Therefore D + ( i η ) D_+(i\eta) D + ( i η ) is the product of the p p p inside roots and D − ( i η ) D_-(i\eta) D − ( i η ) the product of the p p p outside roots, counted with multiplicity. No eigenvalue simplicity or choice of eigenvalue branches is required.
From the determinant factors to a scalar integral
We now recover the modulus of D + D_+ D + from the characteristic polynomial on the unit circle. The elementary polynomial form of Jensen’s formula does precisely this by distinguishing roots inside and outside the circle. For a monic polynomial D D D , put
m ( D ) = ∫ 0 2 π log ∣ D ( e i k ) ∣ d k 2 π . m(D)=\int_{0}^{2\pi}\log\left|D(e^{ik})\right|\frac{dk}{2\pi}. m ( D ) = ∫ 0 2 π log D ( e ik ) 2 π d k .
Jensen’s formula says that m ( D ) m(D) m ( D ) is the sum of log ∣ z j ∣ \log|z_j| log ∣ z j ∣ over its roots outside the unit disk. Applying this to the root assignment just proved, and using D + D − = − 1 D_+D_-=-1 D + D − = − 1 , gives
m ( D η ) = log ∣ D − ( i η ) ∣ = − log ∣ D + ( i η ) ∣ . (52) m(D_\eta)=\log|D_-(i\eta)|=-\log|D_+(i\eta)|. \tag*{(52)} m ( D η ) = log ∣ D − ( i η ) ∣ = − log ∣ D + ( i η ) ∣. (52)
This equality is asserted for sufficiently small positive η \eta η , where the restriction determinants are defined.
The block symmetries proved after the definition of I \mathcal{I} I show that I ( i η ) \mathcal{I}(i\eta) I ( i η ) is real. Therefore the logarithmic derivative of D + D_+ D + along this segment can be recovered from its modulus. Indeed, (47) gives
d d η log D + ( i η ) = D + ′ ( i η ) D + ( i η ) = − 1 2 I ( i η ) . \frac{d}{d\eta}\log D_+(i\eta)=\frac{D_+'(i\eta)}{D_+(i\eta)}=-\frac{1}{2}\mathcal{I}(i\eta). d η d log D + ( i η ) = D + ( i η ) D + ′ ( i η ) = − 2 1 I ( i η ) .
Taking real parts in this equality and differentiating (52) proves
I ( i η ) = 2 d d η m ( D η ) . (53) \mathcal{I}(i\eta)=2\frac{d}{d\eta}m(D_\eta). \tag*{(53)} I ( i η ) = 2 d η d m ( D η ) . (53)
It remains to put this identity into the angular average defining G p G_p G p . On the unit circle, D 0 D_0 D 0 has a fixed phase and a real remaining factor. Dividing by that phase and by c p c^p c p isolates a single real scalar function. For c = c ( ψ ) > 0 c=c(\psi)>0 c = c ( ψ ) > 0 and h = h ( ψ ) h=h(\psi) h = h ( ψ ) define
Δ p ( k , ψ ) = − i D 0 ( e i k ; ψ ) c ( ψ ) p e i p k ∈ R . \Delta_p(k,\psi)=-i\frac{D_0(e^{ik};\psi)}{c(\psi)^p e^{ipk}}\in\mathbb{R}. Δ p ( k , ψ ) = − i c ( ψ ) p e i p k D 0 ( e ik ; ψ ) ∈ R .
On the unit circle,
D η ( e i k ) = c p e i p k ( i Δ p ( k , ψ ) + p η ) . D_\eta(e^{ik})=c^p e^{ipk}\left(i\Delta_p(k,\psi)+p\eta\right). D η ( e ik ) = c p e i p k ( i Δ p ( k , ψ ) + p η ) .
Its logarithmic modulus is therefore p log c + 1 2 log ( Δ p 2 + p 2 η 2 ) p\log c+\frac{1}{2}\log\left(\Delta_p^2+p^2\eta^2\right) p log c + 2 1 log ( Δ p 2 + p 2 η 2 ) . For η > 0 \eta>0 η > 0 there are no zeros on the circle, so differentiation under the k k k integral is legitimate and gives
I p ( i η , ψ ) = 2 p 2 ∫ 0 2 π η Δ p ( k , ψ ) 2 + p 2 η 2 d k 2 π . \mathcal{I}_p(i\eta,\psi)=2p^2\int_{0}^{2\pi}\frac{\eta}{\Delta_p(k,\psi)^2+p^2\eta^2}\frac{dk}{2\pi}. I p ( i η , ψ ) = 2 p 2 ∫ 0 2 π Δ p ( k , ψ ) 2 + p 2 η 2 η 2 π d k .
The normalized variable in G p ( i y ) G_p(iy) G p ( i y ) is η = y / p \eta=y/p η = y / p . Substitution first gives
I p ( i y / p , ψ ) = 2 p ∫ 0 2 π y y 2 + Δ p ( k , ψ ) 2 d k 2 π , \mathcal{I}_p(iy/p,\psi)=2p\int_{0}^{2\pi}\frac{y}{y^2+\Delta_p(k,\psi)^2}\frac{dk}{2\pi}, I p ( i y / p , ψ ) = 2 p ∫ 0 2 π y 2 + Δ p ( k , ψ ) 2 y 2 π d k ,
and then (42) gives
G p ( i y ) = 2 p π ∫ − π / 2 π / 2 d ψ c ( ψ ) ∫ 0 2 π y y 2 + Δ p ( k , ψ ) 2 d k 2 π . (54) G_p(iy)=\frac{2p}{\pi}\int_{-\pi/2}^{\pi/2}\frac{d\psi}{c(\psi)}\int_{0}^{2\pi}\frac{y}{y^2+\Delta_p(k,\psi)^2}\frac{dk}{2\pi}. \tag*{(54)} G p ( i y ) = π 2 p ∫ − π /2 π /2 c ( ψ ) d ψ ∫ 0 2 π y 2 + Δ p ( k , ψ ) 2 y 2 π d k . (54)
This formula holds for sufficiently small positive y y y . The uniform analytic matrix definition supplies its endpoint values and proves its integrability. The displayed nonnegative kernel is useful away from y = 0 y=0 y = 0 ; it is not expanded term by term at zero, since Δ p \Delta_p Δ p has real zeros.
The scalar identity needed for the contact limits
The definition (42) shows directly that G p G_p G p is even and G p ( 0 ) = 1 G_p(0)=1 G p ( 0 ) = 1 . Its evaluation is the remaining task of the following sections.
Theorem 5.3 (Evaluation of the scalar integral). For every odd positive integer p p p and every odd q ^ \widehat{q} q coprime to p p p ,
G p ( t ) = 2 F 1 ( 1 2 , 1 2 ; 1 ; t 2 16 ) = ∑ m ≥ 0 ( 2 m m ) 2 256 m t 2 m . (55) G_p(t)=2F_1\left(\frac{1}{2},\frac{1}{2};1;\frac{t^2}{16}\right)=\sum_{m\geq0}\frac{\binom{2m}{m}^{2}}{256^{m}}t^{2m}. \tag*{(55)} G p ( t ) = 2 F 1 ( 2 1 , 2 1 ; 1 ; 16 t 2 ) = m ≥ 0 ∑ 25 6 m ( m 2 m ) 2 t 2 m . (55)
Combined with (39) , this identity evaluates every fixed-numerator contact limit. The arithmetic argument in § 8 will remove the restriction that the numerator is fixed.
A polynomial recurrence for the undeformed determinant
The matrix-to-integral reduction is complete. We now prepare two descriptions of the same polynomial D 0 D_0 D 0 for its evaluation in the following sections. The recurrence in this subsection gives its coefficients and a useful reversal identity. The tridiagonal determinant in the next subsection makes its dependence on the integration variables explicit.
The reason for returning to F 2 F^2 F 2 is its simple factorization F 2 = C e C o F^2=C_eC_o F 2 = C e C o : the factors contain alternating adjacent two-coordinate blocks. Their characteristic polynomial can therefore be built one coordinate at a time. The first description is the standard recurrence for the characteristic polynomial of a unitary matrix made from alternating two-coordinate blocks. Assume c > 0 c>0 c > 0 in this subsection. For 1 ≤ j ≤ p 1\leq j\leq p 1 ≤ j ≤ p put
α j − 1 = − cos ( j θ ) + i h sin ( j θ ) , ϱ j − 1 = 2 c sin ( j θ ) . (56) \alpha_{j-1}=-\cos(j\theta)+ih\sin(j\theta),\qquad\varrho_{j-1}=2c\sin(j\theta). \tag*{(56)} α j − 1 = − cos ( j θ ) + ih sin ( j θ ) , ϱ j − 1 = 2 c sin ( j θ ) . (56)
Then ∣ α j − 1 ∣ 2 + ϱ j − 1 2 = 1 |\alpha_{j-1}|^2+\varrho_{j-1}^2=1 ∣ α j − 1 ∣ 2 + ϱ j − 1 2 = 1 and ∣ α j − 1 ∣ < 1 |\alpha_{j-1}|<1 ∣ α j − 1 ∣ < 1 for j < p j<p j < p , while α p − 1 = 1 \alpha_{p-1}=1 α p − 1 = 1 and ϱ p − 1 = 0 \varrho_{p-1}=0 ϱ p − 1 = 0 . The matrix
Θ ( α , ϱ ) = ( α ‾ ϱ ϱ − α ) \Theta(\alpha,\varrho)=
\begin{pmatrix}
\overline{\alpha} & \varrho\\
\varrho& -\alpha
\end{pmatrix} Θ ( α , ϱ ) = ( α ϱ ϱ − α )
is exactly the two-coordinate block used above. Thus C e C_e C e contains the even blocks, and C o C_o C o contains the odd blocks with a singleton 1 1 1 in the first coordinate. For the latter assertion, reflect a block and use sin ( π q ^ − x ) = sin x \sin(\pi\widehat{q}-x)=\sin x sin ( π q − x ) = sin x and cos ( π q ^ − x ) = − cos x \cos(\pi\widehat{q}-x)=-\cos x cos ( π q − x ) = − cos x . It follows that F 2 = C e C o F^2=C_eC_o F 2 = C e C o .
The block product F 2 = C e C o F^2=C_eC_o F 2 = C e C o is a five-diagonal unitary matrix. Its characteristic polynomial can be computed by the recurrence below; see [21 ] . The coefficients (56) have possibly signed ϱ j \varrho_j ϱ j . A real diagonal conjugation, with successive sign ratios sign ϱ j \operatorname{sign}\varrho_j sign ϱ j , makes the interior ϱ j \varrho_j ϱ j positive without changing α j \alpha_j α j or the characteristic polynomial. We may therefore use the usual polynomial recurrence. Define Φ 0 = Φ 0 ∗ = 1 \Phi_0=\Phi_0^*=1 Φ 0 = Φ 0 ∗ = 1 and, for 0 ≤ j < p 0\leq j<p 0 ≤ j < p ,
Φ j + 1 ( x ) = x Φ j ( x ) − α j ‾ Φ j ∗ ( x ) , Φ j + 1 ∗ ( x ) = Φ j ∗ ( x ) − α j x Φ j ( x ) . (57) \begin{aligned}
\Phi_{j+1}(x)&=x\Phi_j(x)-\overline{\alpha_j}\Phi_j^*(x),\\
\Phi_{j+1}^*(x)&=\Phi_j^*(x)-\alpha_jx\Phi_j(x). \tag*{(57)}
\end{aligned} Φ j + 1 ( x ) Φ j + 1 ∗ ( x ) = x Φ j ( x ) − α j Φ j ∗ ( x ) , = Φ j ∗ ( x ) − α j x Φ j ( x ) . (57)
The star denotes polynomial reversal: Φ j ∗ ( x ) = x j Φ j ( 1 / x ‾ ) ‾ \Phi_j^*(x)=x^j\overline{\Phi_j(1/\overline{x})} Φ j ∗ ( x ) = x j Φ j ( 1/ x ) . Since the last coefficient is α p − 1 = 1 \alpha_{p-1}=1 α p − 1 = 1 , the last step of (57) gives the useful identity
Φ p ∗ = − Φ p . (58) \Phi_p^*=-\Phi_p. \tag*{(58)} Φ p ∗ = − Φ p . (58)
The characteristic-polynomial formula for this finite unitary matrix gives
D 0 ( z ) = det ( z 2 I p − F 2 ) = Φ p ( z 2 ) . (59) D_0(z)=\det(z^2I_p-F^2)=\Phi_p(z^2). \tag*{(59)} D 0 ( z ) = det ( z 2 I p − F 2 ) = Φ p ( z 2 ) . (59)
An explicit Laurent determinant
The second description writes Δ p \Delta_p Δ p as the determinant of an explicit tridiagonal matrix whose entries are Laurent polynomials. This exposes the boundary terms and special level sets used later. The positive variable w w w replaces the angular parameter so that the entries become rational functions, and a common denominator can be removed. Put a = e i θ a=e^{i\theta} a = e i θ , so a p = − 1 a^p=-1 a p = − 1 , and introduce w > 0 w>0 w > 0 by
c = w 1 + w 2 , h = w 2 − 1 1 + w 2 . c=\frac{w}{1+w^2},\qquad h=\frac{w^2-1}{1+w^2}. c = 1 + w 2 w , h = 1 + w 2 w 2 − 1 .
This parametrizes the same real interval c > 0 c>0 c > 0 , h 2 + 4 c 2 = 1 h^2+4c^2=1 h 2 + 4 c 2 = 1 . For z ≠ 0 z\ne0 z = 0 define
K ( w , z ) = z 2 C o − 1 − C e c z , Δ p ( w , z ) = − i ( − 1 ) n det K ( w , z ) . K(w,z)=\frac{z^2C_o^{-1}-C_e}{cz},\qquad\Delta_p(w,z)=-i(-1)^n\det K(w,z). K ( w , z ) = cz z 2 C o − 1 − C e , Δ p ( w , z ) = − i ( − 1 ) n det K ( w , z ) .
Indeed, multiplying z 2 C o − 1 − C e z^2C_o^{-1}-C_e z 2 C o − 1 − C e on the left by C o C_o C o shows that its determinant is ( − 1 ) n D 0 ( z ) (-1)^nD_0(z) ( − 1 ) n D 0 ( z ) . Hence this Δ p \Delta_p Δ p agrees with the scalar in (54) when z = e i θ z=e^{i\theta} z = e i θ .
The symbols A ( z ) = z + a z − 1 A(z)=z+az^{-1} A ( z ) = z + a z − 1 and B ( z ) = z + a − 1 z − 1 B(z)=z+a^{-1}z^{-1} B ( z ) = z + a − 1 z − 1 below are scalar functions of z z z . Let D D D be the diagonal matrix
D r r = { a r , r even , − a − r − 1 , r odd . D_{rr}=
\begin{cases}
a^r, & r\ \text{even},\\
-a^{-r-1}, & r\ \text{odd}.
\end{cases} D r r = { a r , − a − r − 1 , r even , r odd .
Direct substitution of the two-coordinate blocks gives
K ( w , z ) = w A ( z ) D + w − 1 B ( z ) D − 1 + T ( z ) , (60) K(w,z)=wA(z)D+w^{-1}B(z)D^{-1}+T(z), \tag*{(60)} K ( w , z ) = w A ( z ) D + w − 1 B ( z ) D − 1 + T ( z ) , (60)
where T ( z ) T(z) T ( z ) has zero diagonal and symmetric off-diagonal entries
T r , r + 1 = T r + 1 , r = { − 2 z − 1 sin ( ( r + 1 ) θ ) , r even , 2 z sin ( ( r + 1 ) θ ) , r odd . T_{r,r+1}=T_{r+1,r}=
\begin{cases}
-2z^{-1}\sin((r+1)\theta), & r\ \text{even},\\
2z\sin((r+1)\theta), & r\ \text{odd}.
\end{cases} T r , r + 1 = T r + 1 , r = { − 2 z − 1 sin (( r + 1 ) θ ) , 2 z sin (( r + 1 ) θ ) , r even , r odd .
In particular, with d r = w A ( z ) D r r + w − 1 B ( z ) D r r − 1 d_r=wA(z)D_{rr}+w^{-1}B(z)D^{-1}_{rr} d r = w A ( z ) D r r + w − 1 B ( z ) D r r − 1 and b r = T r , r + 1 b_r=T_{r,r+1} b r = T r , r + 1 , the full matrix is
K ( w , z ) = ( d 0 b 0 0 ⋯ 0 b 0 d 1 b 1 ⋱ ⋮ 0 b 1 d 2 ⋱ 0 ⋮ ⋱ ⋱ ⋱ b p − 2 0 ⋯ 0 b p − 2 d p − 1 ) . (61) K(w,z)=
\begin{pmatrix}
d_0 & b_0 & 0 & \cdots& 0\\
b_0 & d_1 & b_1 & \ddots& \vdots\\
0 & b_1 & d_2 & \ddots& 0\\
\vdots& \ddots& \ddots& \ddots& b_{p-2}\\
0 & \cdots& 0 & b_{p-2} & d_{p-1}
\end{pmatrix}. \tag*{(61)} K ( w , z ) = d 0 b 0 0 ⋮ 0 b 0 d 1 b 1 ⋱ ⋯ 0 b 1 d 2 ⋱ 0 ⋯ ⋱ ⋱ ⋱ b p − 2 0 ⋮ 0 b p − 2 d p − 1 . (61)
The convention also gives the 1 × 1 1\times1 1 × 1 matrix for p = 1 p=1 p = 1 . For example, when r r r is even its diagonal entry is
z ( w a r + w − 1 a − r ) + z − 1 ( w a r + 1 + w − 1 a − r − 1 ) , z(wa^r+w^{-1}a^{-r})+z^{-1}(wa^{r+1}+w^{-1}a^{-r-1}), z ( w a r + w − 1 a − r ) + z − 1 ( w a r + 1 + w − 1 a − r − 1 ) ,
as one checks directly from the block formula. The odd case uses D r r = − a − r − 1 D_{rr}=-a^{-r-1} D r r = − a − r − 1 . The singleton endpoints are included because a p = − 1 a^p=-1 a p = − 1 . This explicit Laurent matrix is the input for the boundary and critical-point analysis in the next section.
The compact family and its four singular levels
The scalar integral of the preceding section is built from the level curves f = λ f=\lambda f = λ . We now complete those curves by adding their ends, while carrying along the differential form that occurs in the integral. The output is Proposition 6.5: a connected genus-one family with singular levels 0 0 0 , 4 4 4 , − 4 -4 − 4 , ∞ \infty ∞ , together with the explicitly normalized residue form in (72) .
There are three calculations to verify. The face polynomials determine where the affine curves meet the boundary and how the pencil is resolved. The factorization test establishes connectedness. Finally, a complete intersection count proves that the nodes we find are the only finite critical points. Each calculation has a separate purpose; all three are needed before the period argument in § 7 .
Fix an odd integer p ≥ 3 p \ge3 p ≥ 3 , an odd integer q ^ \hat{q} q ^ coprime to p p p , and put θ = π q ^ / p \theta= \pi\hat{q}/p θ = π q ^ / p and a = e i θ a = e^{i\theta} a = e i θ . Thus a a a has exact order 2 p 2p 2 p . Matrix indices run from 0 to p − 1 p-1 p − 1 , and J J J is the matrix reversing their order. As in the finite-matrix construction, set
c = w 2 1 + w 2 , h = w 2 − 1 1 + w 2 , ϵ = ( − 1 ) ( p − 1 ) / 2 . c = \frac{w^2}{1+w^2}, \qquad h = \frac{w^2-1}{1+w^2}, \qquad\epsilon= (-1)^{(p-1)/2}. c = 1 + w 2 w 2 , h = 1 + w 2 w 2 − 1 , ϵ = ( − 1 ) ( p − 1 ) /2 .
These rational functions satisfy h 2 + 4 c 2 = 1 h^2+4c^2=1 h 2 + 4 c 2 = 1 and extend the real angular parametrization used in the matrix integral. The case p = 1 p=1 p = 1 will be evaluated directly at the end of the next section. The Laurent polynomial to be studied is
f ( w , z ) = Δ p ( w , z ) = − i ϵ det K ( w , z ) = − i det ( z 2 I − F 2 ) c 2 p , F = C ϵ J . (62) f(w,z)=\Delta_p(w,z)=-i\epsilon\det K(w,z)=-i\frac{\det(z^2I-F^2)}{c^{2p}}, \qquad F=C_\epsilon J. \tag*{(62)} f ( w , z ) = Δ p ( w , z ) = − i ϵ det K ( w , z ) = − i c 2 p det ( z 2 I − F 2 ) , F = C ϵ J . (62)
The exact matrix identity established in the preceding section is
K = w ( z + a z − 1 ) D + w − 1 ( z + a − 1 z − 1 ) D − 1 + T ( z ) , (63) K=w(z+az^{-1})D+w^{-1}(z+a^{-1}z^{-1})D^{-1}+T(z), \tag*{(63)} K = w ( z + a z − 1 ) D + w − 1 ( z + a − 1 z − 1 ) D − 1 + T ( z ) , (63)
The diagonal matrix D D D and the symmetric off-diagonal matrix T T T are those defined in (60) ; the complete tridiagonal shape of K K K is displayed in (61) . All exponents of both w w w and z z z in f f f are odd. Consequently f f f is invariant under ( w , z ) ↦ ( − w , − z ) (w,z)\mapsto(-w,-z) ( w , z ) ↦ ( − w , − z ) , and it is natural to work on the quotient torus
T = ( C ∗ ) 2 / { ( w , z ) ∼ ( − w , − z ) } , x = w z , y = w / z . (64) \mathbb{T}=(\mathbb{C}^{*})^2/\{(w,z)\sim(-w,-z)\}, \qquad x=wz,\quad y=w/z. \tag*{(64)} T = ( C ∗ ) 2 / {( w , z ) ∼ ( − w , − z )} , x = w z , y = w / z . (64)
Here C ∗ = C ∖ { 0 } \mathbb{C}^{*}=\mathbb{C}\setminus\{0\} C ∗ = C ∖ { 0 } . A Laurent polynomial is a finite sum of monomials with integer exponents. The monomials that are functions on this quotient have exponent lattice Λ = { ( k , l ) ∈ Z 2 : k + l is even } \Lambda=\{(k,l)\in\mathbb{Z}^2:k+l\text{ is even}\} Λ = {( k , l ) ∈ Z 2 : k + l is even } . The coordinates x , y x,y x , y identify T \mathbb{T} T with another copy of ( C ∗ ) 2 (\mathbb{C}^{*})^2 ( C ∗ ) 2 . We study the one-parameter family f = λ f=\lambda f = λ on this quotient; each level curve is called a fiber . The parity of Λ \Lambda Λ will be used in the irreducibility proof.
Four exact boundary polynomials
For a Laurent polynomial H = ∑ h k l w k z l H=\sum h_{kl}w^kz^l H = ∑ h k l w k z l , mark the point ( k , l ) (k,l) ( k , l ) whenever h k l ≠ 0 h_{kl}\ne0 h k l = 0 . Its Newton polygon is the smallest convex polygon containing these points, their convex hull. Negative coordinates record negative powers. The Newton polygon of f − λ f-\lambda f − λ is the square [ − p , p ] 2 [-p,p]^2 [ − p , p ] 2 in ( k , l ) (k,l) ( k , l ) coordinates, or the diamond with vertices ( p , 0 ) , ( 0 , p ) , ( − p , 0 ) , ( 0 , − p ) (p,0),(0,p),(-p,0),(0,-p) ( p , 0 ) , ( 0 , p ) , ( − p , 0 ) , ( 0 , − p ) in ( x , y ) (x,y) ( x , y ) exponents. The terms on an edge form its face polynomial. Up to a nonzero constant and a Laurent monomial, the four face polynomials are
( z 2 + a ) p , ( z 2 + a − 1 ) p , ( w 2 + 1 ) p , ( w 2 + 1 ) p . (65) (z^2+a)^p,\qquad(z^2+a^{-1})^p,\qquad(w^2+1)^p,\qquad(w^2+1)^p. \tag*{(65)} ( z 2 + a ) p , ( z 2 + a − 1 ) p , ( w 2 + 1 ) p , ( w 2 + 1 ) p . (65)
The first two follow by taking only the w w w or only the w − 1 w^{-1} w − 1 diagonal terms of (63) . The other two follow from the leading and constant coefficients of det ( z 2 I − F 2 ) \det(z^2I-F^2) det ( z 2 I − F 2 ) and c − p = ( w + w − 1 ) p c^{-p}=(w+w^{-1})^p c − p = ( w + w − 1 ) p .
We also need the following divisibilities, which control the terms immediately behind each face:
( w + w − 1 ) ∣ l ∣ ∣ [ z l ] f , l odd , (66) (w+w^{-1})^{|l|}\mid[z^l]f,\qquad l\text{ odd}, \tag*{(66)} ( w + w − 1 ) ∣ l ∣ ∣ [ z l ] f , l odd , (66)
( z + a z − 1 ) k ∣ [ w k ] f , k > 0 , (67) (z+az^{-1})^k\mid[w^k]f,\qquad k>0, \tag*{(67)} ( z + a z − 1 ) k ∣ [ w k ] f , k > 0 , (67)
( z + a − 1 z − 1 ) − k ∣ [ w k ] f , k < 0. (68) (z+a^{-1}z^{-1})^{-k}\mid[w^k]f,\qquad k<0. \tag*{(68)} ( z + a − 1 z − 1 ) − k ∣ [ w k ] f , k < 0. (68)
Here divisibility is in the appropriate one-variable Laurent ring. To prove the first assertion for l = p − 2 j > 0 l=p-2j>0 l = p − 2 j > 0 , write
det ( z 2 I − F 2 ) = ∑ j = 0 p A j ( c , h ) z 2 ( p − j ) . \det(z^2I-F^2)=\sum_{j=0}^{p}A_j(c,h)z^{2(p-j)}. det ( z 2 I − F 2 ) = j = 0 ∑ p A j ( c , h ) z 2 ( p − j ) .
Each A j A_j A j is a sum of j j j -row principal minors of F 2 F^2 F 2 , so its total degree in c , h c,h c , h is at most 2 j 2j 2 j . Substitution of c , h c,h c , h and division by c p c^p c p give the factor ( w + w − 1 ) p − 2 j (w+w^{-1})^{p-2j} ( w + w − 1 ) p − 2 j . The lower-half coefficients satisfy the same assertion by the polynomial reversal identity (58) . For the last two assertions, expand the tridiagonal determinant into matchings. A term with w w w exponent k > 0 k>0 k > 0 uses at least k k k positive-w w w diagonal factors, each containing z + a z − 1 z+az^{-1} z + a z − 1 ; the negative case is identical. This proves (66) –(68) term by term.
The zero fiber and the two other singular levels
We first exhibit four critical points on the covering torus, giving two nodes after the quotient. The coefficient of the simple factor at zero will also prove that these intersections are transverse.
Lemma 6.1 (The reducible zero fiber). There is a Laurent polynomial R R R such that
f = ( w + w − 1 ) R , R ( ± i , z ) = − i ϵ p ( z − z − 1 ) . (69) f=(w+w^{-1})R,\qquad R(\pm i,z)=-i\epsilon p(z-z^{-1}). \tag*{(69)} f = ( w + w − 1 ) R , R ( ± i , z ) = − i ϵ p ( z − z − 1 ) . (69) The zero fiber on T \mathbb{T} T therefore has two transverse nodes, represented by ( i , 1 ) (i,1) ( i , 1 ) and ( i , − 1 ) (i,-1) ( i , − 1 ) .
Proof. Write s r = sin ( r θ ) s_r=\sin(r\theta) s r = sin ( r θ ) , with s 0 = s p = 0 s_0=s_p=0 s 0 = s p = 0 . For δ = ± 1 \delta=\pm1 δ = ± 1 , put v r = δ r ( − 1 ) r ( r + 1 ) / 2 v_r=\delta^r(-1)^{r(r+1)/2} v r = δ r ( − 1 ) r ( r + 1 ) /2 and Q = diag ( v r ) Q=\operatorname{diag}(v_r) Q = diag ( v r ) . Direct substitution into (63) shows that Q K ( δ i , z ) Q QK(\delta i,z)Q Q K ( δ i , z ) Q is a negative weighted path Laplacian, with edge weights
e r = { 2 δ z − 1 s r + 1 , r even , 2 δ z s r + 1 , r odd . e_r=
\begin{cases}
2\delta z^{-1}s_{r+1}, & r\ \text{even},\\
2\delta z s_{r+1}, & r\ \text{odd}.
\end{cases} e r = { 2 δ z − 1 s r + 1 , 2 δ z s r + 1 , r even , r odd .
The product of these weights is
∏ r = 0 p − 2 e r = ∏ r = 1 p − 1 2 sin ( r θ ) = p . \prod_{r=0}^{p-2}e_r=\prod_{r=1}^{p-1}2\sin(r\theta)=p. r = 0 ∏ p − 2 e r = r = 1 ∏ p − 1 2 sin ( r θ ) = p .
For the last equality, pair r r r with p − r p-r p − r and use the elementary root-of-unity sine product. A path has a unique spanning tree; alternatively, deleting one row and column gives its product by successive elimination. Thus
adj K ( δ i , z ) = p v v T , det K ( δ i , z ) = 0. \operatorname{adj}K(\delta i,z)=pvv^{T},\qquad\det K(\delta i,z)=0. adj K ( δ i , z ) = p v v T , det K ( δ i , z ) = 0.
This proves divisibility by w + w − 1 w+w^{-1} w + w − 1 .
The entries of D D D are a permutation of 1 , a , … , a p − 1 1,a,\ldots,a^{p-1} 1 , a , … , a p − 1 , whence
Tr D = 2 1 − a , Tr D − 1 = − 2 a 1 − a . \operatorname{Tr}D=\frac{2}{1-a},\qquad\operatorname{Tr}D^{-1}=-\frac{2a}{1-a}. Tr D = 1 − a 2 , Tr D − 1 = − 1 − a 2 a .
As K w ( δ i , z ) = ( z + a z − 1 ) D + ( z + a − 1 z − 1 ) D − 1 K_w(\delta i,z)=(z+az^{-1})D+(z+a^{-1}z^{-1})D^{-1} K w ( δ i , z ) = ( z + a z − 1 ) D + ( z + a − 1 z − 1 ) D − 1 and v r 2 = 1 v_r^2=1 v r 2 = 1 , differentiation of the determinant gives
∂ w det K ∣ w = δ i = 2 p ( z − z − 1 ) . \left.\partial_w\det K\right|_{w=\delta i}=2p(z-z^{-1}). ∂ w det K ∣ w = δ i = 2 p ( z − z − 1 ) .
Since ∂ w ( w + w − 1 ) = 2 \partial_w(w+w^{-1})=2 ∂ w ( w + w − 1 ) = 2 at either endpoint, this is precisely (69) . The two roots z = ± 1 z=\pm1 z = ± 1 are simple in R ( δ i , z ) R(\delta i,z) R ( δ i , z ) , so the two components meet transversely. The simultaneous-sign quotient pairs these four points into the two asserted nodes. □
We next exhibit one critical point at each of two nonzero levels. The later intersection count will prove both their nondegeneracy and the absence of additional critical points:
f ( 1 , i ) = 4 ϵ , f ( 1 , − i ) = − 4 ϵ , d f ( 1 , i ) = d f ( 1 , − i ) = 0. (70) f(1,i)=4\epsilon,\qquad f(1,-i)=-4\epsilon,\qquad df(1,i)=df(1,-i)=0. \tag*{(70)} f ( 1 , i ) = 4 ϵ , f ( 1 , − i ) = − 4 ϵ , df ( 1 , i ) = df ( 1 , − i ) = 0. (70)
Here is a direct verification. At w = 1 w=1 w = 1 we have h = 0 h=0 h = 0 , c = 1 / 2 c=1/2 c = 1/2 . The polynomial recurrence (57) , evaluated at x = − 1 x=-1 x = − 1 , gives
Φ p ( − 1 ) = ∏ j = 1 p [ − 1 + ( − 1 ) j − 1 cos ( j θ ) ] = − 2 2 − p . \Phi_p(-1)=\prod_{j=1}^{p}\left[-1+(-1)^{j-1}\cos(j\theta)\right]=-2^{2-p}. Φ p ( − 1 ) = j = 1 ∏ p [ − 1 + ( − 1 ) j − 1 cos ( j θ ) ] = − 2 2 − p .
Indeed k = ( q ^ + p ) / 2 k=(\hat q+p)/2 k = ( q ^ + p ) /2 is coprime to p p p , and these factors are − 2 cos 2 ( π k j / p ) -2\cos^2(\pi k j/p) − 2 cos 2 ( π k j / p ) ; their product is the elementary odd-root cosine product. Inserting the value into (62) gives the two displayed levels. At h = 0 h=0 h = 0 the polynomial has real coefficients and satisfies (58) , so its normalized value on z = e i k z=e^{ik} z = e ik is a sum of odd sine modes. Its z z z derivative vanishes at z = ± i z=\pm i z = ± i . For real h h h , the terminal coefficient α p − 1 = 1 \alpha_{p-1}=1 α p − 1 = 1 gives Φ p ∗ = − Φ p \Phi_p^*=-\Phi_p Φ p ∗ = − Φ p . Since p p p is odd, Φ p ( − 1 ; h ) \Phi_p(-1;h) Φ p ( − 1 ; h ) is real. Conjugating the coefficients changes h h h to − h -h − h , so this value is even in h h h . The factor c − p c^{-p} c − p is even as well, and the w w w derivative vanishes at w = 1 w=1 w = 1 . This proves (70) .
The compactification and its two-form
We now resolve the pencil at its four boundary basepoints. The same coordinate substitutions will track the integration form. Keeping these calculations together is important: the later period equation must use the differential supplied by the matrix integral, with its normalization fixed.
Start with the product of the two projective lines with coordinates x , y x,y x , y , and blow up its four corners. The boundary becomes a cycle B 0 B_0 B 0 of eight projective lines. The original four lines have self-intersection − 2 -2 − 2 , and the four exceptional lines have self-intersection − 1 -1 − 1 . One checks the pole orders of a Laurent monomial by substituting x = u x=u x = u , y = u v y=uv y = uv , and the analogous coordinates at the other corners. The Newton diamond then says that the pole divisor of f f f is exactly p B 0 pB_0 p B 0 .
The four face polynomials in (65) give one basepoint on each of the four exceptional boundary lines. Here a basepoint is a point where the two functions defining the pencil both vanish. At such a point take local coordinates u , v u,v u , v , with boundary u = 0 u=0 u = 0 and face root v = 0 v=0 v = 0 . The pencil is represented by
F ( u , v ) , u p , ord ( 0 , 0 ) F ≥ p , F ( 0 , v ) = v p b ( v ) , b ( 0 ) ≠ 0. F(u,v),\quad u^p,\quad\operatorname{ord}_{(0,0)} F \ge p,\quad F(0,v)=v^p b(v),\quad b(0)\ne0. F ( u , v ) , u p , ord ( 0 , 0 ) F ≥ p , F ( 0 , v ) = v p b ( v ) , b ( 0 ) = 0.
For example, at w = ∞ w=\infty w = ∞ put u = w − 1 u=w^{-1} u = w − 1 and v = z − z 0 v=z-z_0 v = z − z 0 with z 0 2 = − a z_0^2=-a z 0 2 = − a , and use F = u p f F=u^p f F = u p f . The term [ w k ] f [w^k]f [ w k ] f contributes u p − k [ w k ] f u^{p-k}[w^k]f u p − k [ w k ] f . If k > 0 k>0 k > 0 , its coefficient vanishes to order at least k k k in v v v , by (67) ; if k ≤ 0 k\le0 k ≤ 0 , its u u u power already has order at least p p p . Thus every term has total order at least p p p . The other three faces work in the same way. The two roots in a covering coordinate represent one point on the quotient, and a neighborhood of either representative is a local coordinate neighborhood downstairs.
One blowup removes each basepoint. In the chart v = u s v=us v = u s , divide the two pencil generators by u p u^p u p ; the second becomes 1 1 1 . In the chart u = v t u=vt u = v t , divide by v p v^p v p ; the second becomes t p t^p t p , and the first is a unit at v = t = 0 v=t=0 v = t = 0 . No basepoint remains. Let X X X be the resulting compact smooth surface and let B B B be the strict boundary. We have a holomorphic map
f : X ⟶ P 1 , f − 1 ( ∞ ) = p B . (71) f:X\longrightarrow\mathbb{P}^1,\qquad f^{-1}(\infty)=pB. \tag*{(71)} f : X ⟶ P 1 , f − 1 ( ∞ ) = pB . (71)
The curve B B B is a cycle of eight projective lines, now all of self-intersection − 2 -2 − 2 . The four new exceptional curves are not components of B B B . On each, f f f is the degree-p p p polynomial F p ( 1 , s ) F_p(1,s) F p ( 1 , s ) , where F p F_p F p is the degree-p p p part of F F F .
We also keep track of a particular differential. Write d log w = d w / w d\log w=dw/w d log w = d w / w and put
Ω = d log w ∧ d log z = − 1 2 d log x ∧ d log y , η λ = Res f = λ Ω f − λ . (72) \Omega=d\log w\wedge d\log z=-\frac{1}{2}d\log x\wedge d\log y,\qquad\eta_\lambda=\operatorname{Res}_{f=\lambda}\frac{\Omega}{f-\lambda}. \tag*{(72)} Ω = d log w ∧ d log z = − 2 1 d log x ∧ d log y , η λ = Res f = λ f − λ Ω . (72)
Here the residue is a one-form on the level curve, not a number. At a smooth point of the fiber, d f ≠ 0 df\ne0 df = 0 , so the implicit function theorem supplies local coordinates t = f − λ , r t=f-\lambda,r t = f − λ , r , with the fiber given by t = 0 t=0 t = 0 . Write Ω = b ( t , r ) d t ∧ d r \Omega=b(t,r)\,dt\wedge dr Ω = b ( t , r ) d t ∧ d r . Then
Ω f − λ = d t t ∧ b ( t , r ) d r , η λ = b ( 0 , r ) d r . \frac{\Omega}{f-\lambda}=\frac{dt}{t}\wedge b(t,r)\,dr,\qquad\eta_\lambda=b(0,r)\,dr. f − λ Ω = t d t ∧ b ( t , r ) d r , η λ = b ( 0 , r ) d r .
Thus the residue removes the transverse factor d t / t dt/t d t / t and restricts the remaining coefficient to the curve. This is the usual Poincaré residue , with the ordinary one-variable normalization: integration in the positively oriented normal t t t -circle contributes 2 π i 2\pi i 2 π i times this coefficient.
Equivalently, extend η λ \eta_\lambda η λ locally to a one-form on the surface such that d f ∧ η λ = Ω df\wedge\eta_\lambda=\Omega df ∧ η λ = Ω . This wedge identity is understood for the extension; two choices differ by a multiple of d f df df , whose restriction to the level curve is zero. In the torus coordinates this gives the useful formulas
η λ = d z w z f w ∣ f = λ ( f w ≠ 0 ) , η λ = − d w w z f z ∣ f = λ ( f z ≠ 0 ) . \eta_\lambda=\left.\frac{dz}{wz f_w}\right|_{f=\lambda}\quad(f_w\ne0),\qquad\eta_\lambda=-\left.\frac{dw}{wz f_z}\right|_{f=\lambda}\quad(f_z\ne0). η λ = w z f w d z f = λ ( f w = 0 ) , η λ = − w z f z d w f = λ ( f z = 0 ) .
where f w = ∂ f / ∂ w f_w=\partial f/\partial w f w = ∂ f / ∂ w and f z = ∂ f / ∂ z f_z=\partial f/\partial z f z = ∂ f / ∂ z . They agree because f w d w + f z d z = 0 f_w\,dw+f_z\,dz=0 f w d w + f z d z = 0 along the curve. At a smooth point at least one derivative is nonzero, so these formulas fit together without introducing a pole where only one denominator vanishes. At the added points we use regular local coordinates t , r t,r t , r instead of w , z w,z w , z .
Lemma 6.2 (The two-form at the boundary). The form Ω \Omega Ω has a simple pole on every component of B B B and no other pole or zero. In particular it is regular and nowhere zero over the finite λ \lambda λ -plane.
Proof. On the product of projective lines the assertion is immediate from d log x ∧ d log y d\log x \wedge d\log y d log x ∧ d log y . At a corner, substitution x = u x=u x = u , y = u v y=uv y = uv gives d u / u ∧ d v / v du/u \wedge dv/v d u / u ∧ d v / v , so a corner blowup preserves the asserted logarithmic poles. At a smooth boundary point the form is a unit times d u / u ∧ d v du/u \wedge dv d u / u ∧ d v . In the chart v = u s v=us v = u s it becomes a unit times d u ∧ d s du \wedge ds d u ∧ d s , and in the chart u = v t u=vt u = v t it becomes a unit times d t / t ∧ d v dt/t \wedge dv d t / t ∧ d v . Thus a face blowup removes the exceptional pole and leaves only the pole on the strict boundary. □
An elementary factorization test
Lemma 6.3 (Connectedness of the finite fibers). Every finite torus fiber is reduced, meaning that its equation has no repeated component. For λ ≠ 0 \lambda\ne0 λ = 0 , f − λ f-\lambda f − λ is geometrically irreducible in the quotient Laurent ring. At zero the quotient-torus fiber has exactly two components, meeting at the two points of Lemma 6.1 .
Proof. The boundary coefficients sharply restrict any possible factor. We apply those restrictions to the Newton polygons of its factors. The Newton polygon of a product is the sum of its factors’ polygons, where the sum consists of all sums of pairs of points. Every summand of the square [ − p , p ] 2 [-p,p]^2 [ − p , p ] 2 is an axis-parallel rectangle or a segment. A rectangle in the exponent lattice Λ \Lambda Λ has horizontal and vertical widths 2 u 2u 2 u , 2 v 2v 2 v . Suppose u , v > 0 u,v>0 u , v > 0 , and call its upper-right, lower-right, lower-left, upper-left corner coefficients A , B , D , C A,B,D,C A , B , D , C . Its four faces divide the binomial powers in (65) ; their extreme coefficients give
B / A = a v , D / C = a − v , C / A = D / B = 1. B/A=a^v,\qquad D/C=a^{-v},\qquad C/A=D/B=1. B / A = a v , D / C = a − v , C / A = D / B = 1.
Consequently a 2 v = 1 a^{2v}=1 a 2 v = 1 , so p ∣ v p\mid v p ∣ v . The total vertical half-width is p p p ; at most one factor can have a two-dimensional polygon, counting repetitions.
A vertical segment factor would divide both z 2 + a z^2+a z 2 + a and z 2 + a − 1 z^2+a^{-1} z 2 + a − 1 , which have no common root because p ≥ 3 p\ge3 p ≥ 3 . A horizontal segment factor is a power of w 2 + 1 w^2+1 w 2 + 1 , up to an invertible Laurent monomial. Since f ( ± i , z ) = 0 f(\pm i,z)=0 f ( ± i , z ) = 0 , it is impossible for λ ≠ 0 \lambda\ne0 λ = 0 . For λ = 0 \lambda=0 λ = 0 it occurs exactly once, by (69) , and the remaining two-dimensional factor is irreducible. Notice that the factor on the quotient is w 2 + 1 w^2+1 w 2 + 1 : explicitly, f = ( w 2 + 1 ) ( R / w ) f=(w^2+1)(R/w) f = ( w 2 + 1 ) ( R / w ) and R / w R/w R / w belongs to the quotient Laurent ring. These arguments work over every algebraically closed extension of the coefficient field, so they prove geometric irreducibility as well as the reducedness assertion. □
Counting all critical points by intersections
We use only the following elementary facts about curve intersections. Curves of bidegrees ( r , s ) (r,s) ( r , s ) and ( r ′ , s ′ ) (r',s') ( r ′ , s ′ ) on a product of projective lines, with no common component, have total intersection r s ′ + r ′ s rs'+r's r s ′ + r ′ s . At an isolated intersection the multiplicity is at least the product of the two vanishing orders. Equality holds only if their lowest nonzero homogeneous terms have no common factor. The last statements follow by a point blowup: the intersection number is the product of the orders plus the sum of the intersections of the strict transforms on the exceptional line. Thus a common tangent direction makes the inequality strict. This is the same elementary intersection calculation used for blowups and rational scrolls; see [20 Sections 1.7, 1.10 and 2.9] .
Lemma 6.4 (The complete critical-point count). The compact family has exactly four finite critical points: two nodes on the zero fiber and one node on each of the fibers λ = 4 , − 4 \lambda=4,-4 λ = 4 , − 4 . It has no critical point on any of the four exceptional curves over a finite value.
Proof. 1. The derivative curves have no common component. The free degree-two torus cover is a local isomorphism, so the finite fibers remain reduced on the covering torus. The derivatives there have no common curve component. If they vanished along a curve, the restriction of f f f to its smooth part would be constant, say λ 0 \lambda_0 λ 0 . This would make that curve a repeated component of f − λ 0 f-\lambda_0 f − λ 0 in characteristic zero, contrary to Lemma 6.3.
Count the intersections at the boundary. For the count, temporarily return to the two-sheeted covering coordinates w w w , z z z . Clear the negative exponents in D w f = w ∂ w f D_wf = w\partial_w f D w f = w ∂ w f and D z f = z ∂ z f D_zf = z\partial_z f D z f = z ∂ z f , and homogenize separately in each variable. The resulting two curves have bidegrees ( 2 p , 2 p ) (2p,2p) ( 2 p , 2 p ) , with total intersection 8 p 2 8p^2 8 p 2 . Their corner coefficients are nonzero, so neither has a component in the boundary and they do not meet at a corner. On each of the four open boundary lines their common zeros are exactly the two face roots, giving eight boundary points.
At a face point write F = u p f F = u^p f F = u p f as above. Up to invertible analytic factors the local derivative ideal is
( F v , p F − u F u ) = ( F v , C ) , C = p F − u F u − v F v . (F_v,\ pF-uF_u)=(F_v,\ C),\qquad C=pF-uF_u-vF_v. ( F v , pF − u F u ) = ( F v , C ) , C = pF − u F u − v F v .
The order of F v F_v F v is at least p − 1 p-1 p − 1 . Since F F F has order p p p , Euler’s identity cancels its degree-p p p part in C C C , so C C C has order at least p + 1 p+1 p + 1 . Each boundary intersection therefore contributes at least p 2 − 1 p^2-1 p 2 − 1 . At most
8 p 2 − 8 ( p 2 − 1 ) = 8 8p^2-8(p^2-1)=8 8 p 2 − 8 ( p 2 − 1 ) = 8
intersections remain in the covering torus.
The known interior points exhaust the count. There are already eight distinct critical points there: ( w , z ) = ( ± i , ± 1 ) (w,z)=(\pm i,\pm1) ( w , z ) = ( ± i , ± 1 ) from (69) , and ( w , z ) = ( ± 1 , ± i ) (w,z)=(\pm1,\pm i) ( w , z ) = ( ± 1 , ± i ) from (70) and the separate sign changes. Every one contributes at least one. Hence all contributions are exactly the indicated minima. The eight torus intersections are simple, so their Hessians are invertible, and there are no other torus critical points. The free simultaneous-sign quotient pairs them into four points, with the asserted critical values.
No finite critical point appears on an added curve. Equality at a boundary point also proves that the initial homogeneous forms of F v F_v F v and C C C , of degrees p − 1 p-1 p − 1 and p + 1 p+1 p + 1 , are relatively prime. In the blowup chart,
f ( u , s ) = q 0 ( s ) + u q 1 ( s ) + O ( u 2 ) , q 0 ( s ) = F p ( 1 , s ) , q 1 ( s ) = F p + 1 ( 1 , s ) . f(u,s)=q_0(s)+u q_1(s)+O(u^2),\qquad q_0(s)=F_p(1,s),\quad q_1(s)=F_{p+1}(1,s). f ( u , s ) = q 0 ( s ) + u q 1 ( s ) + O ( u 2 ) , q 0 ( s ) = F p ( 1 , s ) , q 1 ( s ) = F p + 1 ( 1 , s ) .
Those initial forms, at u = 1 u=1 u = 1 , are q 0 ′ q_0' q 0 ′ and − q 1 -q_1 − q 1 . They have no common root. Consequently ( ∂ s f , ∂ u f ) = ( q 0 ′ , q 1 ) (\partial_s f,\partial_u f)=(q_0',q_1) ( ∂ s f , ∂ u f ) = ( q 0 ′ , q 1 ) never vanishes on the exceptional curve over a finite value. Its remaining point lies on B B B , over infinity. This proves the last assertion. □
Proposition 6.5 (The compact family and its singular fibers). Every smooth finite compact fiber is a connected curve of genus one. The zero fiber is two smooth projective lines meeting transversely twice. Each of the fibers at 4, −4 is an irreducible rational curve with one ordinary node. The fiber at infinity is the explicit multiple cycle p B pB pB of (71) . There is no other singular or multiple fiber.
Proof. A noncritical finite fiber is smooth and is connected by Lemma 6.3. The residue η λ \eta_\lambda η λ is holomorphic and nowhere zero on it: both assertions follow locally from the submersion coordinate f = λ f=\lambda f = λ and Lemma 6.2. On a compact connected Riemann surface the number of zeros of a holomorphic one-form is 2 g − 2 2g-2 2 g − 2 . Thus its genus is one.
Lemma 6.4 supplies all finite singularities and shows that their Hessians are invertible. We use the holomorphic Morse lemma : near a critical point with invertible Hessian, there are holomorphic coordinates X , Y X,Y X , Y in which f = λ 0 + X 2 + Y 2 f=\lambda_0+X^2+Y^2 f = λ 0 + X 2 + Y 2 . This is an analytic completion of squares. For a brief proof, first make a linear coordinate change so that f x x ( 0 ) ≠ 0 f_{xx}(0)\ne0 f xx ( 0 ) = 0 . The implicit function theorem solves f x = 0 f_x=0 f x = 0 as x = a ( y ) x=a(y) x = a ( y ) , and Taylor expansion gives f − f ( a ( y ) , y ) = ( x − a ( y ) ) 2 A ( x , y ) f-f(a(y),y)=(x-a(y))^2A(x,y) f − f ( a ( y ) , y ) = ( x − a ( y ) ) 2 A ( x , y ) with A ( 0 ) ≠ 0 A(0)\ne0 A ( 0 ) = 0 . A local holomorphic square root absorbs A A A . Invertibility of the Hessian makes the remaining one-variable quadratic term nonzero, so the same argument absorbs its higher terms. Putting u = X + i Y u=X+iY u = X + iY , v = X − i Y v=X-iY v = X − iY now writes the nearby levels as u v = λ − λ 0 uv=\lambda-\lambda_0 uv = λ − λ 0 in the same coordinate neighborhood. A nearby compact genus-one fiber is obtained by replacing each node by an annulus. At zero, the two components meet twice, so their graph is one cycle; neither component can have positive genus, because smoothing preserves that genus and adds the cycle’s one handle. Hence both components are projective lines. At 4 , − 4 4,-4 4 , − 4 , the irreducible fiber has one node; smoothing adds one handle, so its normalization is a projective line. A finite multiple component would make d f df df vanish along its smooth points, which the critical-point count excludes. The explicit resolution already describes the fiber at infinity.
The compact family is now fully specified for the period calculation: its finite nodal fibers, its multiple boundary fiber, and its residue form have all been checked. In particular, an unlisted finite singular fiber cannot enter the differential equation later.
One normalized period and its differential equation
The goal is to evaluate the single analytic function G p G_p G p , rather than calculate its coefficients separately. We will show that it is one normalized closed period, and that the singular fibers force its differential equation. The resulting uniqueness argument makes the answer independent of the fixed odd numerator p p p .
A period is the integral of a differential form over a closed curve. In a smooth algebraic family, parameter derivatives of the differential satisfy a linear relation modulo exact differentials. Integrating that relation around a closed curve removes the exact term. For genus one, the space of closed smooth complex one-forms modulo exact forms has dimension two, so periods satisfy an equation of order at most two; see [26 Section 1.2] and [11 Chapter 10] . Here we determine the equation itself, including its coefficients, from the following three inputs:
Proposition 7.2 identifies G p G_p G p with a closed period of the fixed residue form η λ \eta_\lambda η λ .
Lemma 7.3 gives an analytic period and a logarithmic companion at each finite nodal level.
The explicit calculation at infinity gives the order of the Wronskian there, despite the multiplicity of the boundary fiber.
The Wronskian count then excludes every additional scalar singularity. Finally, G p ( 0 ) = 1 G_p(0)=1 G p ( 0 ) = 1 selects the unique analytic solution.
Throughout we use the map f : X → P 1 f:X\to\mathbb{P}^{1} f : X → P 1 and boundary cycle B B B from § 6 . The defining identity d f ∧ η λ = Ω df\wedge\eta_\lambda=\Omega df ∧ η λ = Ω fixes both the residue sign and its scale. This normalization is retained in every local chart below; multiplying the form by a function of λ \lambda λ would change the period equation.
The upper Cauchy transform
We begin with a parameter in the upper half-plane, where the real integral has no pole. Its boundary value will then identify the closed period, including the multiplicative constant.
Let G p ( t ) G_p(t) G p ( t ) be the normalized matrix integral defined above. The finite-dimensional signed-determinant and Jensen argument gave, for y > 0 y>0 y > 0 sufficiently small,
G p ( i y ) = 2 p π 2 ∫ R × [ 0 , 2 π ] y y 2 + f ( e u , e i k ) 2 d u d k . (73) G_p(iy)=\frac{2p}{\pi^2}\int_{\mathbb{R}\times[0,2\pi]}\frac{y}{y^2+f(e^u,e^{ik})^2}\,du\,dk. \tag*{(73)} G p ( i y ) = π 2 2 p ∫ R × [ 0 , 2 π ] y 2 + f ( e u , e ik ) 2 y d u d k . (73)
The matrix definition also shows that G p G_p G p is holomorphic on a full disk about zero, with real even coefficients and G p ( 0 ) = 1 G_p(0)=1 G p ( 0 ) = 1 . These finite-matrix facts are the only analytic inputs from the earlier sections in the argument below.
For Im t > 0 \operatorname{Im}t>0 Im t > 0 define
H ( t ) = 2 p π 2 i ∫ R × [ 0 , 2 π ] d u d k f ( e u , e i k ) − t . (74) H(t)=\frac{2p}{\pi^2 i}\int_{\mathbb{R}\times[0,2\pi]}\frac{du\,dk}{f(e^u,e^{ik})-t}. \tag*{(74)} H ( t ) = π 2 i 2 p ∫ R × [ 0 , 2 π ] f ( e u , e ik ) − t d u d k . (74)
This formula uses the upper half-plane branch throughout.
Lemma 7.1. The integral (74) is absolutely convergent and holomorphic in the upper half-plane. On the upper half of a small disk about zero it equals G p G_p G p .
Proof. On the real cylinder w = e u w=e^u w = e u , z = e i k z=e^{ik} z = e ik , f f f is real and Ω = i d u ∧ d k \Omega=i\,du\wedge dk Ω = i d u ∧ d k . Thus no denominator in (74) vanishes for Im t > 0 \operatorname{Im}t>0 Im t > 0 . The compactification gives a uniform convergence estimate at both ends of the cylinder, including near every face root. At a smooth point of B B B there are complex coordinates s , v s,v s , v with
Ω = a d s ∧ d v s , f = b s − p , \Omega=a\frac{ds\wedge dv}{s},\qquad f=bs^{-p}, Ω = a s d s ∧ d v , f = b s − p ,
a , b a,b a , b nonvanishing holomorphic functions. Hence Ω / ( f − t ) \Omega/(f-t) Ω/ ( f − t ) is a regular multiple of s p − 1 d s ∧ d v s^{p-1}ds\wedge dv s p − 1 d s ∧ d v . At a crossing of two components it is a regular multiple of s p − 1 v p − 1 d s ∧ d v s^{p-1}v^{p-1}ds\wedge dv s p − 1 v p − 1 d s ∧ d v . The leading factors are units because the pencil has been completely resolved; no further basepoint is being suppressed. At the exceptional curves above the face roots, Lemma 6.2 gives a regular form without a pole. These estimates and all parameter derivatives are uniform for t t t in compact subsets of the upper half-plane. The closure of the real cylinder is contained in the compact real locus of X X X , so finitely many real charts prove absolute integrability and dominated holomorphic differentiation.
All z z z exponents of f f f are odd, and therefore f ( w , − z ) = − f ( w , z ) f(w,-z)=-f(w,z) f ( w , − z ) = − f ( w , z ) . Pairing k k k and k + π k+\pi k + π in (74) gives, at t = i y t=iy t = i y , precisely (73) . Thus H ( i y ) = G p ( i y ) H(iy)=G_p(iy) H ( i y ) = G p ( i y ) for a positive interval of small y y y . Both sides are holomorphic on the upper half-disk, so the identity theorem proves the assertion. □
Identifying the real integral with a closed period
The real level curves initially lie in a noncompact cylinder. We first check that the added points close every end, then orient the resulting circles by the residue form. Coarea and the boundary value of the Cauchy transform will give the exact factor relating their period to G p G_p G p .
The real structure on the compactified family is
σ ( w , z ) = ( w ‾ , 1 / z ‾ ) , f ∘ σ = f ‾ . (75) \sigma(w,z)=(\overline{w},1/\overline{z}),\qquad f\circ\sigma=\overline{f}. \tag*{(75)} σ ( w , z ) = ( w , 1/ z ) , f ∘ σ = f . (75)
The reality identity follows first from the unitary characteristic-polynomial identity on w > 0 w>0 w > 0 , ∣ z ∣ = 1 |z|=1 ∣ z ∣ = 1 , and then from algebraic continuation. The involution preserves the boundary fan and the four face basepoints, so extends to X X X .
We verify that the real cylinder represents the torus real locus exactly once. A quotient point fixed by σ \sigma σ has a lift satisfying either σ ( w , z ) = ( w , z ) \sigma(w,z)=(w,z) σ ( w , z ) = ( w , z ) or σ ( w , z ) = ( − w , − z ) \sigma(w,z)=(-w,-z) σ ( w , z ) = ( − w , − z ) . The second equality would require ∣ z ∣ 2 = − 1 |z|^2=-1 ∣ z ∣ 2 = − 1 and is impossible. In the first case w ∈ R ∗ w\in\mathbb{R}^* w ∈ R ∗ and ∣ z ∣ = 1 |z|=1 ∣ z ∣ = 1 ; the simultaneous-sign quotient has one and only one representative with w > 0 w>0 w > 0 .
At zero, the only singular points of the compact fiber are the two nodes in Lemma 6.1 . They are exchanged by σ \sigma σ : ( i , 1 ) (i,1) ( i , 1 ) goes to ( − i , 1 ) (-i,1) ( − i , 1 ) , which represents ( i , − 1 ) (i,-1) ( i , − 1 ) . In particular neither is real. It follows that for every real t t t in a sufficiently small interval about zero the compact real fiber is smooth. The points of this compact real fiber missing from the real torus are finitely many points on the four exceptional curves. Indeed the rest of X ∖ T X\setminus T X ∖ T lies in the infinity fiber. A real submersion chart at any of these finitely many points shows that the adjacent arcs of the real fiber lie in the torus. Restoring those points therefore closes all ends of the real level curves.
For real t t t , (72) and (75) imply σ ∗ η t ‾ = − η t \sigma^*\overline{\eta_t}=-\eta_t σ ∗ η t = − η t . Thus η t / i \eta_t/i η t / i is a real nowhere-vanishing one-form on the compact real fiber. It orients each of its circles. Let Γ t \Gamma_t Γ t be the sum of these oriented circles. It is an integral homology cycle on the compact fiber: the integration is over closed curves, including the finitely many added points.
Proposition 7.2. For sufficiently small nonzero real t t t , and hence as an analytic germ at zero,
G p ( t ) = 2 p π i ∫ Γ t η t . (76) G_p(t)=\frac{2p}{\pi i}\int_{\Gamma_t}\eta_t. \tag*{(76)} G p ( t ) = π i 2 p ∫ Γ t η t . (76) Along a regular real interval the homology class of Γ t \Gamma_t Γ t is transported continuously in the smooth fibers. The equality extends through zero using the analytic function G p G_p G p ; the complex fiber at zero itself is nodal, although its real circles avoid the nodes.
Proof. Let ν ( t ) \nu(t) ν ( t ) be the density of the pushforward of d u d k du\,dk d u d k under the real-valued function f f f . With the orientation just specified, coarea gives the following expression, written using the Dirac delta:
ν ( t ) : = ∫ δ ( f − t ) d u d k = 1 i ∫ Γ t η t . (77) \nu(t):=\int\delta(f-t)\,du\,dk=\frac{1}{i}\int_{\Gamma_t}\eta_t. \tag*{(77)} ν ( t ) := ∫ δ ( f − t ) d u d k = i 1 ∫ Γ t η t . (77)
The factor 1 / i 1/i 1/ i is fixed by Ω = i d u ∧ d k \Omega= i\,du \wedge dk Ω = i d u ∧ d k and d f ∧ η t = Ω df \wedge\eta_t = \Omega df ∧ η t = Ω . The finitely many added points cause no boundary term, since η t / i \eta_t/i η t / i extends there as the same nonzero one-form on the incoming and outgoing arcs.
The real part of (74) is
Re H ( t + i y ) = 2 p π 2 ∫ y ( f − t ) 2 + y 2 d u d k . \operatorname{Re} H(t+iy)=\frac{2p}{\pi^2}\int\frac{y}{(f-t)^2+y^2}\,du\,dk. Re H ( t + i y ) = π 2 2 p ∫ ( f − t ) 2 + y 2 y d u d k .
On a compact tubular neighborhood of the smooth compact real fiber, the ordinary Poisson-kernel limit and coarea yield ( 2 p / π ) ν ( t ) (2p/\pi)\nu(t) ( 2 p / π ) ν ( t ) . Outside that neighborhood the denominator is bounded away from zero in compact charts. Near B B B , the same calculation as in Lemma 7.1 gives a bound by a constant times y ∣ s ∣ 2 p − 1 y|s|^{2p-1} y ∣ s ∣ 2 p − 1 , with the analogous product at corners; that contribution tends to zero. Thus
Re H ( t + i 0 ) = 2 p π ν ( t ) . \operatorname{Re} H(t+i0)=\frac{2p}{\pi}\nu(t). Re H ( t + i 0 ) = π 2 p ν ( t ) .
By Lemma 7.1 , the upper germ is G p G_p G p , whose values on the real interval are real. Combining with (77) proves (76) . Continuous transport of the circles follows from the local real submersion charts and compactness; on either side of zero these are cycles in the smooth proper family. □
Nodal fibers: an analytic period and a logarithmic companion
The next lemma is the single local integration calculation used at all singular levels. At a finite node it gives a simple Wronskian pole. After a change of parameter, it will also determine the Wronskian’s order at infinity.
Lemma 7.3 (Periods in a nodal annulus). Suppose a family of compact genus-one curves degenerates to a cycle of N N N rational components. Here N = 1 N=1 N = 1 means one nodal rational curve, and N = 2 N=2 N = 2 means two projective lines meeting twice. Near each node the family has equation u v = s uv=s uv = s , up to a holomorphic change of coordinates and a nonzero analytic factor in s s s . Suppose ω s \omega_s ω s is holomorphic on each smooth compact fiber, depends holomorphically on s s s away from the nodes, and extends in a nodal chart in the form h ( u , v ) d u / u = − h ( u , v ) d v / v h(u,v)\,du/u=-h(u,v)\,dv/v h ( u , v ) d u / u = − h ( u , v ) d v / v along the fibers, with h ( 0 , 0 ) ≠ 0 h(0,0)\ne0 h ( 0 , 0 ) = 0 . Then a basis of closed periods has the form
A ( s ) , B ( s ) + κ A ( s ) log s , A , B ∈ C { s } , A ( 0 ) ≠ 0 , κ ≠ 0. (78) A(s),\qquad B(s)+\kappa A(s)\log s,\qquad A,B\in\mathbb{C}\{s\},\qquad A(0)\ne0,\qquad\kappa\ne0. \tag*{(78)} A ( s ) , B ( s ) + κ A ( s ) log s , A , B ∈ C { s } , A ( 0 ) = 0 , κ = 0. (78) The notation C { s } \mathbb{C}\{s\} C { s } means convergent power series. In an oriented integral homology basis, κ = ± N / ( 2 π i ) \kappa=\pm N/(2\pi i) κ = ± N / ( 2 π i ) .
Proof. Remove small disks about the two preimages of every node on the normalizations. The remaining pieces are spheres with two holes, joined by the annuli u v = s uv=s uv = s . A circle in any annulus represents one primitive homology class α \alpha α . Its period is analytic in s s s and has nonzero limit. More explicitly, expand h ( u , v ) = ∑ h i j u i v j h(u,v)=\sum h_{ij}u^iv^j h ( u , v ) = ∑ h ij u i v j in a fixed node chart. On v = s / u v=s/u v = s / u , the coefficient of d u / u du/u d u / u is r ( s ) = ∑ h j j s j r(s)=\sum h_{jj}s^j r ( s ) = ∑ h j j s j , so the circle integral is A ( s ) = 2 π i r ( s ) A(s)=2\pi i\,r(s) A ( s ) = 2 π i r ( s ) with r ( 0 ) ≠ 0 r(0)\ne0 r ( 0 ) = 0 . Choose a second cycle β \beta β passing once through every annulus. The constant Laurent coefficient contributes r ( s ) log s r(s)\log s r ( s ) log s on crossing an annulus. Every other Laurent term has an elementary antiderivative with no logarithm, and its endpoint evaluations are convergent power series in s s s . Integrals along the remaining compact arcs also depend holomorphically on s s s . The annular contributions have the same sign after consistent orientations: equivalently, as s s s turns once, each of the N N N annuli adds one copy of α \alpha α to β \beta β . Thus the total logarithmic coefficient is ± N A ( s ) / ( 2 π i ) \pm N A(s)/(2\pi i) ± N A ( s ) / ( 2 π i ) , which is nonzero. Replacing s s s in one chart by s s s times a unit adds only an analytic logarithm of that unit to B ( s ) B(s) B ( s ) . This proves (78) . □
At a finite critical level λ 0 \lambda_0 λ 0 , the Morse chart u v = λ − λ 0 uv=\lambda-\lambda_0 uv = λ − λ 0 and the form Ω = h ( u , v ) d u ∧ d v \Omega=h(u,v)\,du\wedge dv Ω = h ( u , v ) d u ∧ d v , with h ( 0 , 0 ) ≠ 0 h(0,0)\ne0 h ( 0 , 0 ) = 0 , give η λ = − h d u / u \eta_\lambda=-h\,du/u η λ = − h d u / u along the fibers. Hence Lemma 7.3 applies, with N = 2 N=2 N = 2 at zero and N = 1 N=1 N = 1 at 4 , − 4 4,-4 4 , − 4 .
The multiple fiber at infinity
At infinity we must track both the multiplicity p p p of the fiber and the scale of the residue form. A change of parameter first replaces the multiple boundary by a reduced nodal cycle. We then apply the same annulus lemma, after computing the power of the new parameter contributed by the form. Put τ = λ − 1 \tau=\lambda^{-1} τ = λ − 1 . Since the zero divisor of τ \tau τ is exactly p B pB pB , near a smooth point of B B B one has τ = u p \tau=u^p τ = u p times a unit, and near a crossing one has
τ = ( u v ) p times a unit . \tau=(uv)^p\text{ times a unit}. τ = ( uv ) p times a unit .
On a small coordinate disk the unit has an analytic p p p th root, which can be absorbed into u u u . Make the base change τ = s p \tau=s^p τ = s p and separate the local branches. At a smooth boundary point these branches are u = ζ s u=\zeta s u = ζ s ; at a node they are
u v = ζ s , ζ p = 1. (79) uv=\zeta s,\qquad\zeta^p=1. \tag*{(79)} uv = ζ s , ζ p = 1. (79)
These charts give a smooth surface over the s s s -disk, with a reduced nodal central fiber. This construction is the normalization of the fiber product; the local equations above describe it completely.
Its central fiber maps to B B B as an unbranched degree-p p p cover of nodal curves. In particular it is locally an isomorphism on each branch at every node. Each component of B B B is a sphere, which has no nontrivial connected unbranched cover, so each has p p p lifts. The central fiber is connected. To see this without a factorization theorem, suppose its compact components split into two disjoint closed sets. Disjoint neighborhoods of these sets contain all fibers sufficiently near zero, by properness. Both neighborhoods meet every sufficiently nearby fiber, by the local equations (79) ; those fibers would be disconnected. This contradicts the geometric irreducibility of the generic fiber in Lemma 6.3, which persists after the base change. The central fiber is therefore one cycle of 8 p 8p 8 p spheres.
The differential normalization can be read in the same charts. At a node write Ω = h d u / u ∧ d v / v \Omega=h\,du/u\wedge dv/v Ω = h d u / u ∧ d v / v with h h h a unit. On the branch u v = ζ s uv=\zeta s uv = ζ s this is
Ω = h d s s ∧ d v v . \Omega=h\frac{ds}{s}\wedge\frac{dv}{v}. Ω = h s d s ∧ v d v .
Since λ = s − p \lambda=s^{-p} λ = s − p , the defining identity d f ∧ η λ = Ω df\wedge\eta_\lambda=\Omega df ∧ η λ = Ω yields
η λ = − s p p h d v v . (80) \eta_\lambda=-\frac{s^p}{p}h\frac{dv}{v}. \tag*{(80)} η λ = − p s p h v d v . (80)
Consequently η λ / τ \eta_\lambda/\tau η λ / τ extends as a nonzero nodal differential on the reduced cycle: on the two branches of a node it has opposite simple residues, both nonzero. The same conclusion holds at smooth points of the central fiber. Applying Lemma 7.3 on this cover gives a period basis
y 1 = s p A ( s ) , y 2 = s p { B ( s ) + κ A ( s ) log s } , A ( 0 ) ≠ 0 , κ ≠ 0. (81) y_1=s^pA(s),\qquad y_2=s^p\{B(s)+\kappa A(s)\log s\},\qquad A(0)\ne0,\quad\kappa\ne0. \tag*{(81)} y 1 = s p A ( s ) , y 2 = s p { B ( s ) + κ A ( s ) log s } , A ( 0 ) = 0 , κ = 0. (81)
Constant factors such as − 1 / p -1/p − 1/ p have been included in A , B A,B A , B . We do not assume that A , B A,B A , B themselves are power series in τ \tau τ ; that descent is unnecessary for the next argument.
The Wronskian determines the differential equation
The local calculations now determine the Wronskian on the entire parameter sphere. We first prove that it has no zero at a smooth fiber. Only then do we divide by it to construct the differential equation. This order rules out apparent singularities as part of the proof, rather than assuming their absence.
Let
U = P 1 ∖ { 0 , 4 , − 4 , ∞ } . \mathcal{U}=\mathbb{P}^{1}\setminus\{0,4,-4,\infty\}. U = P 1 ∖ { 0 , 4 , − 4 , ∞ } .
Over a simply connected part of U \mathcal{U} U , transport an oriented homology basis γ 1 , γ 2 \gamma_1,\gamma_2 γ 1 , γ 2 of a compact fiber and set
y j ( λ ) = ∫ γ j ( λ ) η λ , W ( λ ) = y 1 y 2 ′ − y 1 ′ y 2 . y_j(\lambda)=\int_{\gamma_j(\lambda)}\eta_\lambda,\qquad W(\lambda)=y_1y_2'-y_1'y_2. y j ( λ ) = ∫ γ j ( λ ) η λ , W ( λ ) = y 1 y 2 ′ − y 1 ′ y 2 .
The prime here means differentiation in λ \lambda λ . Independence of the homology cycles alone does not prove W ≠ 0 W\ne0 W = 0 ; the following local calculations and zero-and-pole count establish that fact throughout U \mathcal{U} U .
Closed periods depend holomorphically on λ \lambda λ on U U U . A change of oriented integral homology basis has determinant one, because it preserves the intersection pairing. It therefore leaves W W W unchanged. Thus W W W is a single-valued holomorphic function on U U U .
At a finite singular level, (78) gives
W = κ A ( 0 ) 2 λ − λ 0 + O ( 1 ) . W=\frac{\kappa A(0)^2}{\lambda-\lambda_0}+O(1). W = λ − λ 0 κ A ( 0 ) 2 + O ( 1 ) .
In particular W W W has exactly a simple pole there. At infinity, differentiate (81) , using d / d λ = − s p + 1 ( d / d s ) / p \mathrm{d}/\mathrm{d}\lambda=-s^{p+1}(\mathrm{d}/\mathrm{d}s)/p d / d λ = − s p + 1 ( d / d s ) / p . One obtains
W = − κ p s 3 p A ( s ) 2 − 1 p s 3 p + 1 { A B ′ − A ′ B } = − κ A ( 0 ) 2 p τ 3 + o ( τ 3 ) . (82) W=-\frac{\kappa}{p}s^{3p}A(s)^2-\frac{1}{p}s^{3p+1}\{AB'-A'B\}=-\frac{\kappa A(0)^2}{p}\tau^3+o(\tau^3). \tag*{(82)} W = − p κ s 3 p A ( s ) 2 − p 1 s 3 p + 1 { A B ′ − A ′ B } = − p κ A ( 0 ) 2 τ 3 + o ( τ 3 ) . (82)
The primes on A , B A,B A , B in this display mean derivatives in s s s . Since W W W is already single-valued, its bound O ( ∣ τ ∣ 3 ) O(|\tau|^3) O ( ∣ τ ∣ 3 ) makes it holomorphic at τ = 0 \tau=0 τ = 0 , and the nonzero limit of W / τ 3 W/\tau^3 W / τ 3 proves that its zero has order exactly three. No statement about the separate descent of A A A and B B B is used.
It follows that W W W is meromorphic on the sphere, hence rational. It has three simple finite poles, at 0 0 0 , 4 4 4 , − 4 -4 − 4 , and a zero of order three at infinity. A rational function has the same total number of zeros and poles, counted with multiplicity. Hence there is no further zero, and
W ( λ ) = C λ ( λ 2 − 16 ) , C ≠ 0. (83) W(\lambda)=\frac{C}{\lambda(\lambda^2-16)},\qquad C\ne0. \tag*{(83)} W ( λ ) = λ ( λ 2 − 16 ) C , C = 0. (83)
This short divisor count is the crucial step: it rules out all apparent singularities of the scalar period equation.
The orders in this divisor count do not depend on p p p . At infinity, the factor s p s^p s p in the form becomes exactly λ − 1 \lambda^{-1} λ − 1 , while changing the number of annuli changes only a nonzero constant. This is where the numerator disappears from the differential equation.
To require both periods to satisfy the monic equation y ′ ′ + P y ′ + Q y = 0 y''+P y'+Qy=0 y ′′ + P y ′ + Q y = 0 (coefficient one on y ′ ′ y'' y ′′ ) is to solve
( y 1 ′ y 1 y 2 ′ y 2 ) ( P Q ) = − ( y 1 ′ ′ y 2 ′ ′ ) . \begin{pmatrix} y_1' & y_1 \\ y_2' & y_2 \end{pmatrix}
\begin{pmatrix} P \\ Q \end{pmatrix}
=-\begin{pmatrix} y_1'' \\ y_2'' \end{pmatrix}. ( y 1 ′ y 2 ′ y 1 y 2 ) ( P Q ) = − ( y 1 ′′ y 2 ′′ ) .
The determinant is − W -W − W . Wherever W ≠ 0 W\ne0 W = 0 , the unique solution is
P = − W ′ W , Q = y 1 ′ y 2 ′ ′ − y 1 ′ ′ y 2 ′ W . P=-\frac{W'}{W},\qquad Q=\frac{y_1'y_2''-y_1''y_2'}{W}. P = − W W ′ , Q = W y 1 ′ y 2 ′′ − y 1 ′′ y 2 ′ .
Thus P , Q P,Q P , Q are holomorphic wherever the periods are holomorphic and W W W is nonzero. A constant invertible change of the two periods gives the same equation, since it gives an equivalent linear system. Indeed, since W W W never vanishes on U U U , the two periods satisfy one monic second-order equation
y ′ ′ + P ( λ ) y ′ + Q ( λ ) y = 0. y''+P(\lambda)y'+Q(\lambda)y=0. y ′′ + P ( λ ) y ′ + Q ( λ ) y = 0.
Solving the two linear equations for P , Q P,Q P , Q shows that they are single-valued meromorphic functions, and
P = − W ′ W = 1 λ + 1 λ − 4 + 1 λ + 4 . (84) P=-\frac{W'}{W}=\frac{1}{\lambda}+\frac{1}{\lambda-4}+\frac{1}{\lambda+4}. \tag*{(84)} P = − W W ′ = λ 1 + λ − 4 1 + λ + 4 1 . (84)
They have no poles at ordinary points of U U U . At a finite singular level the analytic period A A A is nonzero at the central point, so Q = − ( A ′ ′ + P A ′ ) / A Q=-(A''+P A')/A Q = − ( A ′′ + P A ′ ) / A has at most a simple pole. At infinity take the first period in (81) . It has the form λ − 1 ( A ( 0 ) + O ( λ − 1 / p ) ) \lambda^{-1}(A(0)+O(\lambda^{-1/p})) λ − 1 ( A ( 0 ) + O ( λ − 1/ p )) on the cover. Together with (84) this gives
Q ( λ ) = λ − 2 + O ( λ − 2 − 1 / p ) . Q(\lambda)=\lambda^{-2}+O(\lambda^{-2-1/p}). Q ( λ ) = λ − 2 + O ( λ − 2 − 1/ p ) .
Because Q Q Q is single-valued, this is an ordinary meromorphic extension at infinity. In particular Q Q Q is rational, vanishes to order two there, and its leading coefficient is one.
Finally, Proposition 7.2 puts G p G_p G p in this period space. It is even and holomorphic at zero, with G p ( 0 ) = 1 G_p(0)=1 G p ( 0 ) = 1 . Hence G p ′ = O ( λ ) G_p'=O(\lambda) G p ′ = O ( λ ) , and the identity
Q = − G p ′ ′ + P G p ′ G p Q=-\frac{G_p''+PG_p'}{G_p} Q = − G p G p ′′ + P G p ′
shows that Q Q Q is bounded at zero: the possible pole 1 / λ 1/\lambda 1/ λ in P P P is canceled by G p ′ G_p' G p ′ . Its only possible finite poles are therefore simple poles at 4 , − 4 4,-4 4 , − 4 . Consequently ( λ 2 − 16 ) Q ( λ ) (\lambda^2-16)Q(\lambda) ( λ 2 − 16 ) Q ( λ ) has no finite pole and tends to 1 1 1 at infinity. It is identically 1 1 1 , so Q = ( λ 2 − 16 ) − 1 Q=(\lambda^2-16)^{-1} Q = ( λ 2 − 16 ) − 1 . We have proved
Theorem 7.4. Every closed period of the particular differential η λ \eta_\lambda η λ satisfies
y ′ ′ + ( 1 λ + 1 λ − 4 + 1 λ + 4 ) y ′ + y λ 2 − 16 = 0. (85) y''+\left(\frac{1}{\lambda}+\frac{1}{\lambda-4}+\frac{1}{\lambda+4}\right)y'+\frac{y}{\lambda^2-16}=0. \tag*{(85)} y ′′ + ( λ 1 + λ − 4 1 + λ + 4 1 ) y ′ + λ 2 − 16 y = 0. (85) There is no additional scalar singularity at a smooth fiber.
This determines the equation for the differential actually present in the matrix integral. In particular a parameter-dependent rescaling of a differential has not been introduced at any stage.
Evaluation and universality of the scalar integral
For p = 1 p=1 p = 1 , the matrices give M = V = c / 2 M=V=c/2 M = V = c /2 and L ( t ) 2 = ( 1 − t 2 c 2 / 4 ) I 2 L(t)^2=(1-t^2c^2/4)I_2 L ( t ) 2 = ( 1 − t 2 c 2 /4 ) I 2 , so
G 1 ( t ) = 1 π ∫ − π / 2 π / 2 d φ 1 − t 2 cos 2 φ / 16 = 2 π ∫ 0 π / 2 d φ 1 − ( t / 4 ) 2 sin 2 φ . G_1(t)=\frac{1}{\pi}\int_{-\pi/2}^{\pi/2}\frac{d\varphi}{\sqrt{1-t^2\cos^2\varphi/16}}=\frac{2}{\pi}\int_0^{\pi/2}\frac{d\varphi}{\sqrt{1-(t/4)^2\sin^2\varphi}}. G 1 ( t ) = π 1 ∫ − π /2 π /2 1 − t 2 cos 2 φ /16 d φ = π 2 ∫ 0 π /2 1 − ( t /4 ) 2 sin 2 φ d φ .
The last expression is the normalized complete elliptic integral of the first kind, with modulus t / 4 t/4 t /4 .
For general odd p p p , Proposition 7.2 puts G p G_p G p in the period space just computed. Multiplication of (85) by t ( 16 − t 2 ) t(16-t^2) t ( 16 − t 2 ) gives
t ( 16 − t 2 ) G p ′ ′ + ( 16 − 3 t 2 ) G p ′ − t G p = 0. t(16-t^2)G_p''+(16-3t^2)G_p'-tG_p=0. t ( 16 − t 2 ) G p ′′ + ( 16 − 3 t 2 ) G p ′ − t G p = 0.
Put μ = t 2 / 16 \mu=t^2/16 μ = t 2 /16 and write G p ( t ) = g p ( μ ) G_p(t)=g_p(\mu) G p ( t ) = g p ( μ ) , which is possible near zero because G p G_p G p is even and holomorphic. The chain rule gives G p ′ = t g p ′ / 8 G_p'=tg_p'/8 G p ′ = t g p ′ /8 and G p ′ ′ = g p ′ / 8 + t 2 g p ′ ′ / 64 G_p''=g_p'/8+t^2g_p''/64 G p ′′ = g p ′ /8 + t 2 g p ′′ /64 , where the derivatives of g p g_p g p are in μ \mu μ . Thus the equation becomes
μ ( 1 − μ ) g p ′ ′ + ( 1 − 2 μ ) g p ′ − 1 4 g p = 0. \mu(1-\mu)g_p''+(1-2\mu)g_p'-\frac{1}{4}g_p=0. μ ( 1 − μ ) g p ′′ + ( 1 − 2 μ ) g p ′ − 4 1 g p = 0.
This is the classical Picard–Fuchs equation for the Legendre family y 2 = x ( x − 1 ) ( x − μ ) y^2=x(x-1)(x-\mu) y 2 = x ( x − 1 ) ( x − μ ) ; see [26 Section 1.2] . The elementary integral G 1 G_1 G 1 is its normalized analytic period. Experts may recognize this as the period-theoretic content of a common-Jacobian description: the normalized differentials satisfy the same Picard–Fuchs equation.
To compare the solutions at zero directly, write G p ( t ) = ∑ j ≥ 0 a j t 2 j G_p(t)=\sum_{j\geq0}a_jt^{2j} G p ( t ) = ∑ j ≥ 0 a j t 2 j . Comparison of coefficients gives
64 ( j + 1 ) 2 a j + 1 = ( 2 j + 1 ) 2 a j , a 0 = 1. 64(j+1)^2a_{j+1}=(2j+1)^2a_j,\qquad a_0=1. 64 ( j + 1 ) 2 a j + 1 = ( 2 j + 1 ) 2 a j , a 0 = 1.
The same recurrence holds for G 1 G_1 G 1 , and it determines every coefficient from the constant term. Solving it gives
G p ( t ) = 2 F 1 ( 1 2 , 1 2 ; 1 ; t 2 16 ) = ∑ m ≥ 0 ( 2 m m ) 2 256 m t 2 m . (86) G_p(t)=\,{}_2F_1\left(\frac{1}{2},\frac{1}{2};1;\frac{t^2}{16}\right)=\sum_{m\geq0}\frac{\binom{2m}{m}^2}{256^m}t^{2m}. \tag*{(86)} G p ( t ) = 2 F 1 ( 2 1 , 2 1 ; 1 ; 16 t 2 ) = m ≥ 0 ∑ 25 6 m ( m 2 m ) 2 t 2 m . (86)
Consequently, as analytic germs,
G p ( t ) = G 1 ( t ) (87) G_p(t)=G_1(t) \tag*{(87)} G p ( t ) = G 1 ( t ) (87)
for every odd p p p .
From fixed numerators to all rational parameters
We have evaluated the scalar integral for each fixed odd p p p . Appendix A identifies its Taylor coefficients with the limits of q ρ p , q ( 2 m ) ( 0 ) q\rho_{p,q}^{(2m)}(0) q ρ p , q ( 2 m ) ( 0 ) as q → ∞ q \to\infty q → ∞ . This still leaves an arithmetic issue: Theorem 1.2 allows p p p to vary with q q q . The argument below resolves it without comparing spectra at nearby rational parameters. We instead rewrite each derivative as a convergent sum of matrix products whose dependence on the inverse of p p p modulo q q q is easy to control.
The new arithmetic parameter is α = p ‾ / q \alpha=\overline{p}/q α = p / q , where p p ‾ ≡ 1 ( m o d q ) p\overline{p}\equiv1 \pmod q p p ≡ 1 ( mod q ) . The argument has three steps: express a contact as a function of α \alpha α , control that function by sums over closed paths, and use the fixed-numerator limits to determine its one-sided values at a dense set of rational points. The same path estimate then gives uniformity as both p p p and q q q vary. In particular, this section does not require an error estimate uniform in p p p from the fixed-numerator calculation.
Corollary 8.1 (Evaluated fixed-numerator contacts). For each fixed odd positive integer p p p and every m ≥ 0 m \ge0 m ≥ 0 ,
lim q → ∞ ( p , q ) = 1 q ρ p , q ( 2 m ) ( 0 ) ( 2 m ) ! = ( 2 m m ) 2 256 m . (88) \lim_{\substack{q\to\infty\\(p,q)=1}}\frac{q\rho_{p,q}^{(2m)}(0)}{(2m)!}=\frac{\binom{2m}{m}^{2}}{256^{m}}. \tag*{(88)} q → ∞ ( p , q ) = 1 lim ( 2 m )! q ρ p , q ( 2 m ) ( 0 ) = 25 6 m ( m 2 m ) 2 . (88) Proof. Proposition 5.1 proves the reduction separately on each admissible residue class. Theorem 5.3 evaluates G p G_p G p and gives the same coefficient in every class. There are finitely many classes modulo 4 p 4p 4 p , which proves the limit without a residue restriction. The reduction proof in Appendix A uses no evaluation theorem; the evaluation enters only in this corollary. □
A permutation using the modular inverse
Write
γ m = ( 2 m ) ! ( 2 m m ) 2 256 m , m ≥ 1. \gamma_m=\frac{(2m)!\binom{2m}{m}^{2}}{256^m},\qquad m\ge1. γ m = 25 6 m ( 2 m )! ( m 2 m ) 2 , m ≥ 1.
Proposition 5.1 and Theorem 5.3 show that q ρ p , q ( 2 m ) ( 0 ) → γ m q\rho_{p,q}^{(2m)}(0)\to\gamma_m q ρ p , q ( 2 m ) ( 0 ) → γ m for each fixed odd p p p , through all admissible denominator classes. We now pass to moving numerators. For this arithmetic calculation only, expand the one-pencil formula using the auxiliary products
A = ε B R , C = ε R B , H = A + C , A C = B 2 . A=\varepsilon BR,\qquad C=\varepsilon RB,\qquad H=A+C,\qquad AC=B^2. A = εB R , C = εR B , H = A + C , A C = B 2 .
They are abbreviations formed from the matrices B , R B,R B , R of Section 3, rather than additional data. We first record the exact permutation of their entries, including both denominator parities. Let a = p ‾ a=\overline{p} a = p denote the inverse of p p p modulo q q q , with 1 ≤ a < q 1\le a<q 1 ≤ a < q . Write [ x ] q [x]_q [ x ] q for the unique representative of the integer x x x in { 0 , … , q − 1 } \{0,\ldots,q-1\} { 0 , … , q − 1 } . For r = 0 , … , q − 1 r=0,\ldots,q-1 r = 0 , … , q − 1 set
x r = [ a r ] q , ϵ r = 1 { 2 x r < q } , V r s = ϵ r − ϵ s 2 q sin ( π ( r − s ) / q ) ( r ≠ s ) , V r r = 0. x_r=[ar]_q,\qquad\epsilon_r=\mathbf{1}_{\{2x_r<q\}},\qquad V_{rs}=\frac{\epsilon_r-\epsilon_s}{2q\sin(\pi(r-s)/q)}\ (r\ne s),\qquad V_{rr}=0. x r = [ a r ] q , ϵ r = 1 { 2 x r < q } , V r s = 2 q sin ( π ( r − s ) / q ) ϵ r − ϵ s ( r = s ) , V r r = 0.
Define symmetric matrices
A ~ r s = V r s 1 { x r + x s < q } , H ~ r s = V r s ( 21 { x r + x s < q } − 1 ) . (89) \widetilde{A}_{rs}=V_{rs}\mathbf{1}_{\{x_r+x_s<q\}},\qquad\widetilde{H}_{rs}=V_{rs}\left(2\mathbf{1}_{\{x_r+x_s<q\}}-1\right). \tag*{(89)} A r s = V r s 1 { x r + x s < q } , H r s = V r s ( 2 1 { x r + x s < q } − 1 ) . (89)
Lemma 8.2. A word of length n n n in A , H A,H A , H means a product of n n n factors, each equal to A A A or H H H , in the specified order. For every such word w w w of even length,
⟨ e 1 , w ( A , H ) e 1 ⟩ = ⟨ e 0 , w ( A ~ , H ~ ) e 0 ⟩ . \langle e_1,w(A,H)e_1\rangle=\langle e_0,w(\widetilde{A},\widetilde{H})e_0\rangle. ⟨ e 1 , w ( A , H ) e 1 ⟩ = ⟨ e 0 , w ( A , H ) e 0 ⟩ . Thus every even Taylor coefficient of the finite formula (26) can be evaluated using A ~ , H ~ \widetilde{A},\widetilde{H} A , H and the first coordinate e 0 e_0 e 0 .
Proof. First let q = 2 M + 1 q=2M+1 q = 2 M + 1 . Label original matrix coordinate 2 i + 1 2i+1 2 i + 1 by x = i x=i x = i (0 ≤ i ≤ M 0\le i\le M 0 ≤ i ≤ M ), and original matrix coordinate 2 j 2j 2 j by x = M + j x=M+j x = M + j (1 ≤ j ≤ M 1\le j\le M 1 ≤ j ≤ M ). These labels cover 0 , … , q − 1 0,\ldots,q-1 0 , … , q − 1 , and the two index parities give ϵ = 1 , 0 \epsilon=1,0 ϵ = 1 , 0 . An odd-even A A A entry has i + j ≤ M i+j\le M i + j ≤ M , exactly the cutoff x o d d + x e v e n < q x_{\mathrm{odd}}+x_{\mathrm{even}}<q x odd + x even < q . The complementary H H H sector has the opposite sign. Permute the label x x x to r x = [ p ^ x ] q r_x=[\widehat{p}x]_q r x = [ p x ] q , writing p ^ x = r x + q k x \widehat{p}x=r_x+qk_x p x = r x + q k x . In an odd–even entry the original sine is
sin ( π p ^ ( i + M + 1 − j ) q ) = − sin ( π p ^ ( x o d d − x e v e n ) q ) . \sin\left(\frac{\pi\widehat{p}(i+M+1-j)}{q}\right)=-\sin\left(\frac{\pi\widehat{p}(x_{\mathrm{odd}}-x_{\mathrm{even}})}{q}\right). sin ( q π p ( i + M + 1 − j ) ) = − sin ( q π p ( x odd − x even ) ) .
Its alternating numerator gives the common sign ( − 1 ) M + 1 (-1)^{M+1} ( − 1 ) M + 1 and the two diagonal signs ( − 1 ) i + k x o d d (-1)^{i+k_{x_{\mathrm{odd}}}} ( − 1 ) i + k x odd , ( − 1 ) j + k x e v e n (-1)^{j+k_{x_{\mathrm{even}}}} ( − 1 ) j + k x even . Direct substitution gives the same signed conjugacy for H H H ; on the complementary sector its sign reverses exactly as in (89) . The distinguished coordinate 1 1 1 maps to 0 0 0 and has sign 1 1 1 . Conjugacy preserves first-coordinate entries and an even word removes the common sign.
Now let q = 2 M q=2M q = 2 M . The original odd-coordinate block contains the distinguished coordinate and is invariant under A , H A,H A , H . Up to simultaneous diagonal signs and one common sign, its matrices, for 0 ≤ i , j < M 0\le i,j<M 0 ≤ i , j < M , are
A i j e = 1 { i + j < M } 2 q cos ( π p ( i − j ) / q ) , H i j e = 21 { i + j < M } − 1 2 q cos ( π p ( i − j ) / q ) . (90) A^{\mathrm{e}}_{ij}=\frac{\mathbf{1}_{\{i+j<M\}}}{2q\cos(\pi p(i-j)/q)},\qquad H^{\mathrm{e}}_{ij}=\frac{2\mathbf{1}_{\{i+j<M\}}-1}{2q\cos(\pi p(i-j)/q)}. \tag*{(90)} A ij e = 2 q cos ( π p ( i − j ) / q ) 1 { i + j < M } , H ij e = 2 q cos ( π p ( i − j ) / q ) 2 1 { i + j < M } − 1 . (90)
Indeed, the numerator of an original A A A entry has sign ( − 1 ) M − i − j − 1 (-1)^{M-i-j-1} ( − 1 ) M − i − j − 1 , whereas its sine denominator equals ( − 1 ) ( p − 1 ) / 2 cos ( π p ( i − j ) / q ) (-1)^{(p-1)/2}\cos(\pi p(i-j)/q) ( − 1 ) ( p − 1 ) /2 cos ( π p ( i − j ) / q ) ; C = H − A C=H-A C = H − A has the opposite common sign on the complementary sector. The diagonal sign at i i i is ( − 1 ) i (-1)^i ( − 1 ) i .
In (89) , split the x x x labels into i i i and M + j M+j M + j . Since p p p is odd, p M ≡ M ( m o d q ) pM\equiv M\pmod q pM ≡ M ( mod q ) . Reindex by r = [ p x ] q r=[px]_q r = [ p x ] q and remove the diagonal signs from wrapping by q q q . The identity
sin ( π p ( i − j − M ) q ) = ( − 1 ) ( p + 1 ) / 2 cos ( π p ( i − j ) q ) \sin\left(\frac{\pi p(i-j-M)}{q}\right)=(-1)^{(p+1)/2}\cos\left(\frac{\pi p(i-j)}{q}\right) sin ( q π p ( i − j − M ) ) = ( − 1 ) ( p + 1 ) /2 cos ( q π p ( i − j ) )
shows that the resulting matrices, up to one common sign, are
( 0 A e A e 0 ) , ( 0 H e H e 0 ) . \begin{pmatrix}0&A^{\mathrm{e}}\\A^{\mathrm{e}}&0\end{pmatrix},\qquad
\begin{pmatrix}0&H^{\mathrm{e}}\\H^{\mathrm{e}}&0\end{pmatrix}. ( 0 A e A e 0 ) , ( 0 H e H e 0 ) .
Every even word has the same first-coordinate entry as the corresponding word in (90) . This proves the assertion for even q q q . □
For odd q q q , the lift p ^ \widehat{p} p in this calculation is p p p if p p p is odd and p + q p+q p + q otherwise; for even q q q , p p p is already odd. The exact finite generator uses this same lift.
An absolutely convergent sum over closed paths
The path notation is simply the usual formula for an entry of a matrix product. For example,
( K 1 ⋯ K n ) 00 = ∑ j 1 , … , j n − 1 ( K 1 ) 0 j 1 ( K 2 ) j 1 j 2 ⋯ ( K n ) j n − 1 0 . (K_1\cdots K_n)_{00}=\sum_{j_1,\ldots,j_{n-1}}(K_1)_{0j_1}(K_2)_{j_1j_2}\cdots(K_n)_{j_{n-1}0}. ( K 1 ⋯ K n ) 00 = j 1 , … , j n − 1 ∑ ( K 1 ) 0 j 1 ( K 2 ) j 1 j 2 ⋯ ( K n ) j n − 1 0 .
The initial and final indices agree, so each summand describes a closed path. For the infinite matrices below, the estimate following (91) proves that these sums converge absolutely. For 0 < α < 1 0<\alpha<1 0 < α < 1 , write { x } = x − ⌊ x ⌋ \{x\}=x-\lfloor x\rfloor { x } = x − ⌊ x ⌋ for the fractional part. Define indicator factors and kernels, indexed by j , k ∈ Z j,k\in\mathbb{Z} j , k ∈ Z , by
η j = 1 { j α } < 1 / 2 , χ j k = 1 { j α } + { k α } < 1 , K j k = η j − η k 2 π ( j − k ) ( j ≠ k ) , K j j = 0 , \eta_j=\mathbf{1}_{\{j\alpha\}<1/2},\qquad\chi_{jk}=\mathbf{1}_{\{j\alpha\}+\{k\alpha\}<1},\qquad K_{jk}=\frac{\eta_j-\eta_k}{2\pi(j-k)}\ (j\ne k),\qquad K_{jj}=0, η j = 1 { j α } < 1/2 , χ j k = 1 { j α } + { k α } < 1 , K j k = 2 π ( j − k ) η j − η k ( j = k ) , K j j = 0 ,
( A α ) j k = K j k χ j k , ( H α ) j k = K j k ( 2 χ j k − 1 ) , (A_\alpha)_{jk}=K_{jk}\chi_{jk},\qquad(H_\alpha)_{jk}=K_{jk}(2\chi_{jk}-1), ( A α ) j k = K j k χ j k , ( H α ) j k = K j k ( 2 χ j k − 1 ) ,
Φ m ( α ) = ( 2 m ) ! [ t 2 m ] [ ( I + t H α + t 2 A α ( H α − A α ) ) − 1 / 2 ] 00 . (91) \Phi_m(\alpha)=(2m)![t^{2m}]\left[(I+tH_\alpha+t^2A_\alpha(H_\alpha-A_\alpha))^{-1/2}\right]_{00}. \tag*{(91)} Φ m ( α ) = ( 2 m )! [ t 2 m ] [ ( I + t H α + t 2 A α ( H α − A α ) ) − 1/2 ] 00 . (91)
The products in the first two definitions are entrywise products, not matrix multiplication. The coefficient Φ m ( α ) \Phi_m(\alpha) Φ m ( α ) is a finite linear combination of sums over closed paths 0 = j 0 , j 1 , … , j n = 0 0=j_0,j_1,\ldots,j_n=0 0 = j 0 , j 1 , … , j n = 0 : each edge supplies one kernel entry. Every word contributing to [ t 2 m ] [t^{2m}] [ t 2 m ] has length exactly 2 m 2m 2 m : the linear term t H α tH_\alpha t H α supplies one factor, and the quadratic term t 2 A α ( H α − A α ) t^2A_\alpha(H_\alpha-A_\alpha) t 2 A α ( H α − A α ) supplies two. The next bound proves absolute convergence; no boundedness assertion for the infinite matrices is needed. For a closed word of length n ≥ 2 n\ge2 n ≥ 2 and a cutoff L ≥ 2 L\ge2 L ≥ 2 , deleting all edges longer than L L L changes its sum with initial and final coordinate zero by at most
C n L − 1 log − 2 L . (92) C_nL^{-1}\log^{-2}L. \tag*{(92)} C n L − 1 log − 2 L . (92)
Indeed, in a dyadic shell where the longest edge has length in [ R , 2 R ) [R,2R) [ R , 2 R ) , closure forces a second edge of length at least R / ( n − 1 ) R/(n-1) R / ( n − 1 ) . Choose their two positions; the first displacement has O ( R ) O(R) O ( R ) choices, the closing condition determines the second, and their factors cost O n ( R − 2 ) O_n(R^{-2}) O n ( R − 2 ) . The other n − 2 n-2 n − 2 displacements contribute harmonic sums. Summing the dyadic shells proves (92) , also for cyclic distances in the finite reindexed matrices. Short paths lift uniquely to integer closed paths when n L < q nL<q n L < q .
The exact rational Fourier-transform identity
Proposition 8.3 (Exact rational transfer). For every reduced p / q p/q p / q and every m ≥ 1 m\geq1 m ≥ 1 ,
q ρ p , q ( 2 m ) ( 0 ) = Φ m ( p ˉ / q ) , p p ˉ ≡ 1 ( m o d q ) . (93) q\rho_{p,q}^{(2m)}(0)=\Phi_m(\bar{p}/q),\qquad p\bar{p}\equiv1\pmod{q}. \tag*{(93)} q ρ p , q ( 2 m ) ( 0 ) = Φ m ( p ˉ / q ) , p p ˉ ≡ 1 ( mod q ) . (93) This is an equality at each finite denominator, with no limit or error term.
Proof. Put a = p ˉ a=\bar{p} a = p ˉ and α = a / q \alpha=a/q α = a / q . All the indicator factors in (91) are q q q -periodic in each index. Write an integer index as j = r + q ℓ j=r+q\ell j = r + q ℓ , where 0 ≤ r < q 0\leq r<q 0 ≤ r < q and ℓ ∈ Z \ell\in\mathbb{Z} ℓ ∈ Z . For finitely supported vectors define
( F f ) r ( ϑ ) = ∑ ℓ ∈ Z f r + q ℓ e − i ℓ ϑ . (\mathcal{F}f)_r(\vartheta)=\sum_{\ell\in\mathbb{Z}}f_{r+q\ell}e^{-i\ell\vartheta}. ( F f ) r ( ϑ ) = ℓ ∈ Z ∑ f r + q ℓ e − i ℓ ϑ .
Parseval’s identity extends this to a unitary map from ℓ 2 ( Z ) \ell^2(\mathbb{Z}) ℓ 2 ( Z ) to L 2 ( ( − π , π ) , C q ; d ϑ / ( 2 π ) ) L^2((-\pi,\pi),\mathbb{C}^q;d\vartheta/(2\pi)) L 2 (( − π , π ) , C q ; d ϑ / ( 2 π )) . The coordinate vector at index zero becomes the constant vector e 0 e_0 e 0 .
Here is the scalar Fourier identity used in every off-diagonal residue block. For δ ∉ Z \delta\notin\mathbb{Z} δ ∈ / Z ,
∑ n ∈ Z e − i n ϑ n + δ = π { cot ( π δ ) − i sgn ϑ } e i δ ϑ ( − π < ϑ < π ) . (94) \sum_{n\in\mathbb{Z}}\frac{e^{-in\vartheta}}{n+\delta}=\pi\{\cot(\pi\delta)-i\operatorname{sgn}\vartheta\}e^{i\delta\vartheta}\qquad(-\pi<\vartheta<\pi). \tag*{(94)} n ∈ Z ∑ n + δ e − in ϑ = π { cot ( π δ ) − i sgn ϑ } e i δ ϑ ( − π < ϑ < π ) . (94)
as an L 2 L^2 L 2 Fourier series, with the value at zero immaterial. To check it, integrate the right side against e i n ϑ / ( 2 π ) e^{in\vartheta}/(2\pi) e in ϑ / ( 2 π ) separately on ( − π , 0 ) (-\pi,0) ( − π , 0 ) and ( 0 , π ) (0,\pi) ( 0 , π ) . With b = n + δ b=n+\delta b = n + δ , the result is
( cot π δ + i ) ( 1 − e − i π b ) + ( cot π δ − i ) ( e i π b − 1 ) 2 i b = 1 b . \frac{(\cot\pi\delta+i)(1-e^{-i\pi b})+(\cot\pi\delta-i)(e^{i\pi b}-1)}{2ib}=\frac{1}{b}. 2 ib ( cot π δ + i ) ( 1 − e − iπ b ) + ( cot π δ − i ) ( e iπ b − 1 ) = b 1 .
In the last equality use cot x ± i = e ± i x / sin x \cot x\pm i=e^{\pm ix}/\sin x cot x ± i = e ± i x / sin x and e i n π = ( − 1 ) n e^{in\pi}=(-1)^n e inπ = ( − 1 ) n . This also proves that the convolution defined by the series is bounded, since its displayed multiplier has modulus π / ∣ sin π δ ∣ \pi/|\sin\pi\delta| π /∣ sin π δ ∣ .
For residue indices r ≠ s r\ne s r = s , set δ = ( r − s ) / q \delta=(r-s)/q δ = ( r − s ) / q . The kernel K K K of (91) therefore has multiplier
K ^ r s ( ϑ ) = η r − η s 2 q ( cot π ( r − s ) q − i sgn ϑ ) e i ( r − s ) ϑ / q . \widehat{K}_{rs}(\vartheta)=\frac{\eta_r-\eta_s}{2q}\left(\cot\frac{\pi(r-s)}{q}-i\operatorname{sgn}\vartheta\right)e^{i(r-s)\vartheta/q}. K r s ( ϑ ) = 2 q η r − η s ( cot q π ( r − s ) − i sgn ϑ ) e i ( r − s ) ϑ / q .
For r = s r=s r = s it is zero, including between distinct translates, because the two indicator values are equal. Define the diagonal unitary matrix
D ( ϑ ) r r = exp ( i r ( ϑ − π sgn ϑ ) q ) . D(\vartheta)_{rr}=\exp\left(\frac{ir(\vartheta-\pi\operatorname{sgn}\vartheta)}{q}\right). D ( ϑ ) r r = exp ( q i r ( ϑ − π sgn ϑ ) ) .
Since η r = ϵ r \eta_r=\epsilon_r η r = ϵ r and cot x − i sgn ϑ = e − i x sgn ϑ / sin x \cot x-i\operatorname{sgn}\vartheta=e^{-ix\operatorname{sgn}\vartheta}/\sin x cot x − i sgn ϑ = e − i x sgn ϑ / sin x , we have
K ^ = D V D − 1 , A ~ α = D A ~ D − 1 , H ~ α = D H ~ D − 1 \widehat{K}=DVD^{-1},\qquad\widetilde{A}_{\alpha}=D\widetilde{A}D^{-1},\qquad\widetilde{H}_{\alpha}=D\widetilde{H}D^{-1} K = D V D − 1 , A α = D A D − 1 , H α = D H D − 1
for almost every ϑ \vartheta ϑ .
The last two identities hold because the masks are periodic and are exactly the masks defining the finite matrices. In particular all operators used here are bounded for this fixed rational parameter. No bound uniform in q q q is asserted in this step.
For any word w w w of length 2 m 2m 2 m , its transformed product is thus D w ( A ~ , H ~ ) D − 1 Dw(\widetilde{A},\widetilde{H})D^{-1} D w ( A , H ) D − 1 . Because D 00 = 1 D_{00}=1 D 00 = 1 , Parseval gives
⟨ e 0 , w ( A α , H α ) e 0 ⟩ = 1 2 π ∫ − π π w ( A ~ , H ~ ) 00 d ϑ = w ( A ~ , H ~ ) 00 . \langle e_0,w(A_\alpha,H_\alpha)e_0\rangle=\frac{1}{2\pi}\int_{-\pi}^{\pi}w(\widetilde{A},\widetilde{H})_{00}\,d\vartheta=w(\widetilde{A},\widetilde{H})_{00}. ⟨ e 0 , w ( A α , H α ) e 0 ⟩ = 2 π 1 ∫ − π π w ( A , H ) 00 d ϑ = w ( A , H ) 00 .
The infinite matrix-product sums agree with these operator entries. Indeed, insert the projections onto [ − N , N ] [-N,N] [ − N , N ] between the factors: strong convergence of the projections and boundedness give convergence to the operator product. The absolute closed-path bound (92) gives the same limit for the sums of entries. This separates the conditionally convergent single-kernel Fourier series from the absolutely convergent closed products. Finally expand the fixed coefficient in (91) , apply Lemma 8.2 to each even word, and use Proposition 3.2. This proves (93) . □
One-sided limits and the rational point value
Truncation makes Φ m \Phi_m Φ m a uniform limit of finite step functions whose breakpoints are rational. It is continuous at every irrational and has finite left and right limits at every rational. At reduced b / d b/d b / d , its point value differs from either one-sided limit by
O m ( d − 1 log 2 m − 2 d ) . (95) O_m\left(d^{-1}\log^{2m-2}d\right). \tag*{(95)} O m ( d − 1 log 2 m − 2 d ) . (95)
Every contributing word has length 2 m 2m 2 m . For d d d sufficiently large, choose 2 ≤ L < d / ( 8 m ) 2 \le L < d/(8m) 2 ≤ L < d / ( 8 m ) comparable to d d d ; the finitely many smaller denominators are absorbed in the constant. A nonpermanent mask equality on a short path would require d ∣ j d \mid j d ∣ j , d ∣ 2 j d \mid2j d ∣ 2 j , or d ∣ j + k d \mid j+k d ∣ j + k . None can occur for such short paths except a literal permanent equality, whose mask value is the same in the one-sided limit. Apply (92) to the omitted paths to obtain (95) . The same argument proves stabilization from either side.
Completion of the contact theorem
Every reduced rational b / d b/d b / d with odd d d d can be approached using the fixed odd numerator d d d : choose q q q so d a = b q + 1 da=bq+1 d a = b q + 1 , giving a / q = b / d + 1 / ( d q ) a/q=b/d+1/(dq) a / q = b / d + 1/ ( d q ) . The fixed numerator theorem and the uniform path bound identify the right-hand limit with γ m \gamma_m γ m . The symmetry p ↦ q − p p\mapsto q-p p ↦ q − p supplies the left-hand limits: the identity P p , q = P q − p , q P_{p,q}=P_{q-p,q} P p , q = P q − p , q gives Φ m ( a / q ) = Φ m ( 1 − a / q ) \Phi_m(a/q)=\Phi_m(1-a/q) Φ m ( a / q ) = Φ m ( 1 − a / q ) at reduced rational points. Continuity at irrationals and the existence of one-sided limits extend this symmetry to the left and right limits. Reduced rationals of odd denominator are dense. At an irrational α \alpha α , choose such rationals approaching α \alpha α , and points sufficiently close to their one-sided limits. Continuity at α \alpha α forces Φ m ( α ) = γ m \Phi_m(\alpha)=\gamma_m Φ m ( α ) = γ m . All irrational values therefore equal γ m \gamma_m γ m . Approaching any rational from either side through irrationals then forces both of its one-sided limits to equal γ m \gamma_m γ m . Equations (93) –(95) prove the contact estimate in Theorem 1.2:
Proposition 8.4 (Uniform contact estimate). For every fixed m ≥ 1 m \ge1 m ≥ 1 ,
sup 1 ≤ p < q ( p , q ) = 1 q ∣ ρ p , q ( 2 m ) ( 0 ) − γ m ∣ = O m ( log 2 m − 2 q q ) . (96) \sup_{\substack{1\le p<q\\(p,q)=1}} q\left|\rho_{p,q}^{(2m)}(0)-\gamma_m\right|=O_m\left(\frac{\log^{2m-2}q}{q}\right). \tag*{(96)} 1 ≤ p < q ( p , q ) = 1 sup q ρ p , q ( 2 m ) ( 0 ) − γ m = O m ( q log 2 m − 2 q ) . (96)
The preceding proposition proves the derivative assertion of Theorem 1.2. Together with the local convergence and endpoint argument in Section 2, it completes the proof of Theorem 1.1.
This appendix proves Proposition 5.1: the Taylor coefficients of the finite coarea density converge to those of the explicit function G p G_p G p in § 5 . The evaluation of G p G_p G p belongs to the main text and is not used here.
Fix an odd positive integer p p p and a residue q 0 q_0 q 0 modulo 4 p 4p 4 p with ( p , q 0 ) = 1 (p,q_0)=1 ( p , q 0 ) = 1 . Let q → ∞ q\to\infty q → ∞ within this class, and fix an odd lift q ^ ≡ q 0 ( m o d p ) \hat q\equiv q_0\pmod p q ^ ≡ q 0 ( mod p ) . Constants may depend on p , q 0 p,q_0 p , q 0 and the fixed Taylor order, but not on q q q . The case p = 1 p=1 p = 1 , which has no singular inter-subinterval terms, is included separately at the end.
The finite quantity and the order of the proof
The quantity to be passed to the limit is one entry of a finite matrix coefficient. We begin with this finite quantity, before introducing the limiting Fourier matrices. Fix m ≥ 1 m\ge1 m ≥ 1 for now; the zeroth coefficient is the Last–Wilkinson identity and will also be recorded below.
Write R q R_q R q for coordinate reversal at denominator q q q . For the finite pencil of Proposition 3.2,
T q , t = ε R q + t B q , X 2 m , q = [ t 2 m ] ( T q , t 2 ) − 1 / 2 = F 2 m ( B q , R q ) , T_{q,t}=\varepsilon R_q+tB_q,\qquad X_{2m,q}=[t^{2m}](T_{q,t}^{2})^{-1/2}=F_{2m}(B_q,R_q), T q , t = ε R q + t B q , X 2 m , q = [ t 2 m ] ( T q , t 2 ) − 1/2 = F 2 m ( B q , R q ) ,
where the finite noncommutative polynomial is
F 2 m ( K , R ) = [ t 2 m ] ∑ k = 0 2 m ( − 1 / 2 k ) ( ε t ( R K + K R ) + t 2 K 2 ) k . F_{2m}(K,R)=[t^{2m}]\sum_{k=0}^{2m}\binom{-1/2}{k}\left(\varepsilon t(RK+KR)+t^2K^2\right)^k. F 2 m ( K , R ) = [ t 2 m ] k = 0 ∑ 2 m ( k − 1/2 ) ( εt ( R K + K R ) + t 2 K 2 ) k .
A word here is a product of the displayed factors in their specified order. Every contributing word contains exactly 2 m 2m 2 m occurrences of K K K ; reflection factors do not contribute powers of t t t . The number of linear terms is even, so their powers of ε \varepsilon ε cancel; F 2 m F_{2m} F 2 m is independent of this sign. No commutation of K K K with R R R is used. The identity from Section 3 reads [ t 2 m ] ( q ρ p , q ( t ) ) = ( X 2 m , q ) 11 [t^{2m}](q\rho_{p,q}(t))=(X_{2m,q})_{11} [ t 2 m ] ( q ρ p , q ( t )) = ( X 2 m , q ) 11 .
The entry ( X 2 m , q ) 11 (X_{2m,q})_{11} ( X 2 m , q ) 11 is not controlled by strong operator convergence. The useful extra fact is the position-independent cut identity of Proposition 3.2 . At a cut near the middle of the first subinterval, it pairs a row of X 2 m , q X_{2m,q} X 2 m , q with columns of B q B_q B q whose kernels are nonsingular. Thus we need convergence of these rows in L 2 L^2 L 2 , rather than convergence of an unregularized diagonal kernel.
The proof has the following order. Each step supplies exactly the input named in the next column.
caption:
:::
The scalar-shift identity is proved before the final cut calculation so that the last subsection can complete the proof without interruption.
The four levels of notation. The following table is a guide to the finite matrices, grid operators, and Fourier matrices. The latter two are defined in the next subsections; they are listed here to make their different roles visible.
caption:
:::
Grid operators are compared by their step-function isometries into H = L 2 ( R ) 2 p \mathcal{H}=L^2(\mathbb{R})^{2p} H = L 2 ( R ) 2 p . An identity on a grid space becomes its step-function projection P q P_q P q on H \mathcal{H} H , with P q → I P_q\to I P q → I strongly; it is not the ambient identity at finite q q q . In particular the scalar identity from (A.15) is an identity on the full grid block, or α m , p , q P q 00 \alpha_{m,p,q}P_q^{00} α m , p , q P q 00 after this embedding. Extend B q B_q B q by zero outside its finite interval blocks, but use the full-grid unitary reflection R q R_q R q in both pencils. It maps the selected finite sites onto one another, so this extension restricts to the original finite reversal. For m ≥ 1 m\ge1 m ≥ 1 , every word of F 2 m F_{2m} F 2 m contains B q B_q B q ; hence F 2 m ( B q , R q ) F_{2m}(B_q,R_q) F 2 m ( B q , R q ) agrees with the zero extension of the finite coefficient X 2 m , q X_{2m,q} X 2 m , q . At degree zero we use the full-grid identity, whose restriction to the finite sites is the original identity. Its value at the chosen cut is unchanged. This convention makes products and adjoints comparable without confusing a scalar entry with a block operator.
Rescaled grids: where the singular terms come from
A.2. Rescaled grids: where the singular terms come from. The q q q coordinates are divided into p p p subintervals, and each subinterval retains the two parities of the original matrix index. There are therefore 2 p 2p 2 p grid components. Their limiting Hilbert space is the direct sum of 2 p 2p 2 p copies of L 2 ( R ) L^2(\mathbb{R}) L 2 ( R ) . We keep the exact shifts of the grids: losing them would lose the cotangent terms in the limiting matrix.
Put ϵ r = ( − 1 ) r \epsilon_r=(-1)^r ϵ r = ( − 1 ) r for the subinterval orientations. Put G q = diag ( i j ) j = 1 q G_q=\operatorname{diag}(i^j)_{j=1}^q G q = diag ( i j ) j = 1 q and conjugate the denominator-q q q matrices by A ↦ G q A G q − 1 A\mapsto G_qAG_q^{-1} A ↦ G q A G q − 1 . In the operator calculations through the scalar-shift lemma, we suppress the conjugation marks: B q B_q B q , R q R_q R q , and all matrix functions formed from them are written in this basis. Take their full-grid extensions after conjugation. The reciprocal denominator-p p p matrix B B B defined later remains in its stated basis. The final cut calculation returns to the original finite matrices and marks conjugates by primes. Every allowed B q B_q B q entry here is therefore − i / [ 2 q sin ( π p ( i + j − 1 ) / ( 2 q ) ) ] -i/[2q\sin(\pi p(i+j-1)/(2q))] − i / [ 2 q sin ( π p ( i + j − 1 ) / ( 2 q ))] . Use x j = ( j − 1 / 2 ) / q x_j=(j-1/2)/q x j = ( j − 1/2 ) / q . In subinterval r r r write
x r ( z ) = r + 1 / 2 + ϵ r z p , − 1 2 < z < 1 2 , Δ = 2 p q . x_r(z)=\frac{r+1/2+\epsilon_r z}{p},\qquad-\frac{1}{2}<z<\frac{1}{2},\qquad\Delta=\frac{2p}{q}. x r ( z ) = p r + 1/2 + ϵ r z , − 2 1 < z < 2 1 , Δ = q 2 p .
No matrix index lies on a subinterval boundary: such an equality would give p ( 2 j − 1 ) = 2 q k p(2j-1)=2qk p ( 2 j − 1 ) = 2 q k , impossible because the left side is odd. The index parity ν ∈ { 0 , 1 } \nu\in\{0,1\} ν ∈ { 0 , 1 } (0 even, 1 odd) has, after reindexing its integer coordinate, the centered grid
Δ ( Z + β r , ν ) , β r , ν = ϵ r ( ν − 1 / 2 2 − q ( r + 1 / 2 ) 2 p ) . (97) \Delta(\mathbb{Z}+\beta_{r,\nu}),\qquad\beta_{r,\nu}=\epsilon_r\left(\frac{\nu-1/2}{2}-\frac{q(r+1/2)}{2p}\right). \tag*{(97)} Δ ( Z + β r , ν ) , β r , ν = ϵ r ( 2 ν − 1/2 − 2 p q ( r + 1/2 ) ) . (97)
These are signed real offsets; the parentheses are ordinary grouping. Only the grid Z + β r , ν \mathbb{Z}+\beta_{r,\nu} Z + β r , ν depends on the offset modulo integers. Along the fixed residue class these offsets are fixed modulo integers. For r < s r<s r < s , r + s r+s r + s odd, set ℓ = ( r + s + 1 ) / 2 \ell=(r+s+1)/2 ℓ = ( r + s + 1 ) /2 . The exact entry is
− i Δ 4 p ( − 1 ) ℓ ϵ r csc π ( z − w ) 2 , (98) -\frac{i\Delta}{4p}(-1)^\ell\epsilon_r\csc\frac{\pi(z-w)}{2}, \tag*{(98)} − 4 p i Δ ( − 1 ) ℓ ϵ r csc 2 π ( z − w ) , (98)
between opposite index parities. Its principal part is obtained by retaining the 1 / ( z − w ) 1/(z-w) 1/ ( z − w ) singularity and extending this term to the whole real line, rather than restricting it to the finite subinterval. Thus its full-line principal part is
− i Δ 2 π p ( − 1 ) ℓ ϵ r 1 z − w . (99) -\frac{i\Delta}{2\pi p}(-1)^\ell\epsilon_r\frac{1}{z-w}. \tag*{(99)} − 2 π p i Δ ( − 1 ) ℓ ϵ r z − w 1 . (99)
Call the operator with this full-line singular kernel S B , q S_{B,q} S B , q , the principal replacement of B q B_q B q . All its other subinterval blocks are zero. It retains the exact Cauchy singularities and their grid shifts. We compare the two pencils
T q , t = ϵ R q + t B q , T q , t p r = ϵ R q + t S B , q . T_{q,t}=\epsilon R_q+tB_q,\qquad T^{\mathrm{pr}}_{q,t}=\epsilon R_q+tS_{B,q}. T q , t = ϵ R q + t B q , T q , t pr = ϵ R q + t S B , q .
The reflection is identical in the two pencils, so their only difference is t ( B q − S B , q ) t(B_q-S_{B,q}) t ( B q − S B , q ) . This difference will have regular rows near the cut.
For output parity ν \nu ν and input parity 1 − ν 1-\nu 1 − ν , the difference between the grid offsets is
δ = β r , ν − β s , 1 − ν = − ϵ r q ℓ p . (100) \delta=\beta_{r,\nu}-\beta_{s,1-\nu}=-\epsilon_r\frac{q\ell}{p}. \tag*{(100)} δ = β r , ν − β s , 1 − ν = − ϵ r p q ℓ . (100)
Since 1 ≤ ℓ ≤ p − 1 1\leq\ell\leq p-1 1 ≤ ℓ ≤ p − 1 , its distance from the integers is at least 1 / p 1/p 1/ p . There are no diagonal-channel Cauchy poles because p p p is odd. For r < s r<s r < s of the same parity the exact kernel is
− i Δ 4 p ( − 1 ) ( r + s ) / 2 sec π ( z + w ) 2 . (101) -\frac{i\Delta}{4p}(-1)^{(r+s)/2}\sec\frac{\pi(z+w)}{2}. \tag*{(101)} − 4 p i Δ ( − 1 ) ( r + s ) /2 sec 2 π ( z + w ) . (101)
For r = s r=s r = s it is
− i Δ 4 p ϵ r sec π ( z + w ) 2 { 1 z < w , r even , 1 z > w , r odd . (102) -\frac{i\Delta}{4p}\epsilon_r\sec\frac{\pi(z+w)}{2}
\begin{cases}
1_{z<w}, & r\ \text{even},\\
1_{z>w}, & r\ \text{odd}.
\end{cases} \tag*{(102)} − 4 p i Δ ϵ r sec 2 π ( z + w ) { 1 z < w , 1 z > w , r even , r odd . (102)
All lower subinterval blocks vanish. These formulas follow simply by inserting p ( x r ( z ) + x s ( w ) ) / 2 = ( r + s + 1 ) / 2 + ( ϵ r z + ϵ s w ) / 2 p(x_r(z)+x_s(w))/2=(r+s+1)/2+(\epsilon_r z+\epsilon_s w)/2 p ( x r ( z ) + x s ( w )) /2 = ( r + s + 1 ) /2 + ( ϵ r z + ϵ s w ) /2 . In particular they retain every triangular index restriction.
For the full-line shifted Cauchy matrix D δ ( n , k ) = ( n − k + δ ) − 1 D_\delta(n,k)=(n-k+\delta)^{-1} D δ ( n , k ) = ( n − k + δ ) − 1 , sequence Fourier transformation gives
D δ ^ ( ϑ ) = π { cot ( π δ ) − i sgn ϑ } e i δ ϑ , − π < ϑ < π . (103) \widehat{D_\delta}(\vartheta)=\pi\{\cot(\pi\delta)-i\operatorname{sgn}\vartheta\}e^{i\delta\vartheta},\qquad-\pi<\vartheta<\pi. \tag*{(103)} D δ ( ϑ ) = π { cot ( π δ ) − i sgn ϑ } e i δ ϑ , − π < ϑ < π . (103)
One may verify this multiplier without a summation convention: integrating the displayed piecewise function on ( − π , 0 ) (-\pi,0) ( − π , 0 ) and ( 0 , π ) (0,\pi) ( 0 , π ) gives Fourier coefficient ( n + δ ) − 1 (n+\delta)^{-1} ( n + δ ) − 1 . Its norm is π / ∣ sin π δ ∣ ≤ C p \pi/|\sin\pi\delta|\le C_p π /∣ sin π δ ∣ ≤ C p . The diagonal gauges e − i β r , ν ϑ e^{-i\beta_{r,\nu}\vartheta} e − i β r , ν ϑ remove the exponential. The original coordinate reversal sends ( r , ν , z ) (r,\nu,z) ( r , ν , z ) to ( p − 1 − r , ν ′ , − z ) (p-1-r,\nu',-z) ( p − 1 − r , ν ′ , − z ) , with ν ′ ≡ q + 1 − ν ( m o d 2 ) \nu'\equiv q+1-\nu\pmod2 ν ′ ≡ q + 1 − ν ( mod 2 ) . Moreover
β r , ν + β p − 1 − r , ν ′ = ϵ r ( ( ν + ν ′ − 1 ) / 2 − q / 2 ) ∈ Z . (104) \beta_{r,\nu}+\beta_{p-1-r,\nu'}=\epsilon_r\left((\nu+\nu'-1)/2-q/2\right)\in\mathbb{Z}. \tag*{(104)} β r , ν + β p − 1 − r , ν ′ = ϵ r ( ( ν + ν ′ − 1 ) /2 − q /2 ) ∈ Z . (104)
Thus its integer reflection phase cancels exactly under these gauges. The remaining constant index-parity matrix anticommutes with σ x \sigma_x σ x and squares to I I I . The resulting two parity sectors will be calculated explicitly below; their even coefficients agree.
Uniform bounds and strong convergence
We need two properties: uniform operator bounds, and convergence on each fixed square-integrable input, for both the operators and their adjoints. These properties give inverse bounds and strong convergence for the two pencils near t = 0 t=0 t = 0 . These operator facts will be used in the resolvent comparison. The regular-row estimate later supplies the additional local information needed to evaluate the cut.
Embed each of the 2 p 2p 2 p full grid spaces isometrically into L 2 ( R ) L^2(\mathbb{R}) L 2 ( R ) by constant functions on grid intervals of length Δ \Delta Δ . Write I = ( − 1 / 2 , 1 / 2 ) I=(-1/2,1/2) I = ( − 1/2 , 1/2 ) . A finite subinterval block occupies a union of such intervals whose symmetric difference from I I I is contained in endpoint strips of width O ( Δ ) O(\Delta) O ( Δ ) . Extend its operator by zero outside that union. These strips preserve the isometry and disappear in the strong limits below.
Uniform operator bounds. We shall use the following row-and-column estimate for an integral kernel K ( x , y ) K(x,y) K ( x , y ) . If a positive weight b b b and a constant C 0 C_0 C 0 satisfy
∫ ∣ K ( x , y ) ∣ b ( y ) d y ≤ C 0 b ( x ) , ∫ ∣ K ( x , y ) ∣ b ( x ) d x ≤ C 0 b ( y ) , \int|K(x,y)|b(y)\,dy\le C_0b(x),\qquad\int|K(x,y)|b(x)\,dx\le C_0b(y), ∫ ∣ K ( x , y ) ∣ b ( y ) d y ≤ C 0 b ( x ) , ∫ ∣ K ( x , y ) ∣ b ( x ) d x ≤ C 0 b ( y ) ,
then the induced operator on L 2 L^2 L 2 has norm at most C 0 C_0 C 0 . This follows directly from the Cauchy–Schwarz inequality; replace integrals by sums for grid matrices. The constant weight b = 1 b=1 b = 1 suffices for a bounded kernel on a finite interval.
The true operators have uniform norm C p C_p C p directly from the displayed kernels. On opposite-parity subinterval blocks subtract 2 / ( π ( z − w ) ) 2/(\pi(z-w)) 2/ ( π ( z − w )) from the cosecant: its remainder is smooth on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] , hence has bounded norm by the row-and-column estimate. The principal part has norm bounded by (A.7). For secant blocks use
0 < sec π ( z + w ) 2 ≤ 1 1 − z − w + 1 1 + z + w . 0<\sec\frac{\pi(z+w)}{2}\le\frac{1}{1-z-w}+\frac{1}{1+z+w}. 0 < sec 2 π ( z + w ) ≤ 1 − z − w 1 + 1 + z + w 1 .
Every matrix index has endpoint distance at least
∣ p ( 2 j − 1 ) − 2 q k ∣ 2 q ≥ 1 2 q = Δ 4 p , \frac{|p(2j-1)-2qk|}{2q}\ge\frac{1}{2q}=\frac{\Delta}{4p}, 2 q ∣ p ( 2 j − 1 ) − 2 q k ∣ ≥ 2 q 1 = 4 p Δ ,
because p p p is odd. After reversing indices, each positive majorant is therefore a compression of 1 / ( n + m + γ ) 1/(n+m+\gamma) 1/ ( n + m + γ ) with γ ≥ 1 / ( 2 p ) \gamma\ge1/(2p) γ ≥ 1/ ( 2 p ) . These positive matrices have bounded norm uniformly in γ \gamma γ : compare with max ( 1 , 2 p ) / ( n + m + 1 ) \max(1,2p)/(n+m+1) max ( 1 , 2 p ) / ( n + m + 1 ) and use weights b n = ( n + 1 ) − 1 / 2 b_n=(n+1)^{-1/2} b n = ( n + 1 ) − 1/2 in the same estimate. Triangular masks preserve domination by the positive majorant. The finitely many blocks give the claimed bound for B q B_q B q . The full-line principal replacement S B , q S_{B,q} S B , q has the same type of bound, since it has finitely many blocks and (A.7) applies. The common full-grid reflection R q R_q R q is unitary. The true and principal pencils are therefore uniformly bounded near t = 0 t=0 t = 0 . Their inverse-square-root and inverse bounds will follow from the same Neumann estimate.
Strong convergence, including adjoints. On smooth inputs compactly supported inside I I I , shifted Cauchy quadrature converges to
PV ∫ f ( w ) z − w d w + π cot ( π δ ) f ( z ) . (105) \operatorname{PV}\int\frac{f(w)}{z-w}\,dw+\pi\cot(\pi\delta)f(z). \tag*{(105)} PV ∫ z − w f ( w ) d w + π cot ( π δ ) f ( z ) . (105)
Here the principal value is the symmetric exclusion of the pole:
PV ∫ f ( w ) z − w d w = lim ε ↓ 0 ∫ ∣ w − z ∣ > ε f ( w ) z − w d w . \operatorname{PV}\int\frac{f(w)}{z-w}\,dw=\lim_{\varepsilon\downarrow0}\int_{\lvert w-z\rvert>\varepsilon}\frac{f(w)}{z-w}\,dw. PV ∫ z − w f ( w ) d w = ε ↓ 0 lim ∫ ∣ w − z ∣ > ε z − w f ( w ) d w .
The principal value of an infinite grid sum is likewise its symmetric partial-sum limit. Indeed subtract f ( z ) f(z) f ( z ) near w = z w=z w = z ; for a common C β C^{\beta} C β bound the part ∣ w − z ∣ < ε \lvert w-z\rvert<\varepsilon ∣ w − z ∣ < ε is O p ( ε β ) O_p(\varepsilon^{\beta}) O p ( ε β ) , by dyadic annuli and the lower bound on the nearest shifted denominator. Away from this part ordinary Riemann sums converge. The constant part uses PV ∑ k ( k + δ ) − 1 = π cot π δ \operatorname{PV}\sum_k(k+\delta)^{-1}=\pi\cot\pi\delta PV ∑ k ( k + δ ) − 1 = π cot π δ . Smooth terms use ordinary quadrature. Output endpoint strips have vanishing L 2 L^2 L 2 mass because compact input support keeps their possible endpoint poles separated. For the principal full-line operators the omitted output tails are bounded by C / ( 1 + ∣ z ∣ ) C/(1+\lvert z\rvert) C / ( 1 + ∣ z ∣) . For true zero extensions add smooth inputs compactly supported outside I I I ; these are annihilated for sufficiently fine grids. Together with the interior inputs they span a dense subspace of L 2 ( R ) L^2(\mathbb{R}) L 2 ( R ) . For full-line principal operators the identical quadrature proof permits any compact input support on R \mathbb{R} R . Here strong convergence means convergence in L 2 L^2 L 2 on each fixed input. Thus both types converge strongly on their respective dense classes; the uniform bounds extend convergence to all L 2 L^2 L 2 inputs. Conjugate-transposing the kernels proves strong adjoint convergence by the same argument. The common full-grid reflections R q R_q R q converge strongly together with their adjoints to full-line reflection, with the channel permutation and constant phases specified by (A.8). Their finite compressions would instead converge to compressed reflection; those compressions are not used in the principal comparison.
The Fourier matrices and the limiting integral
We now compute the limiting operators. The change of variables below turns the secant and cosecant kernels into functions of a difference. This is the reason finite Fourier matrices occur.
Make the following unitary change of variables. To distinguish its real coordinate from the analytic parameter t t t , write the new coordinate as τ \tau τ and the input coordinate as σ \sigma σ . A function f ( y ) f(y) f ( y ) is sent to y ′ ( τ ) f ( y ( τ ) ) \sqrt{y'(\tau)}f(y(\tau)) y ′ ( τ ) f ( y ( τ )) , where
y ( τ ) = 1 2 + 2 π arctan ( tanh τ ) , z ( τ ) = y ( τ ) − 1 2 . y(\tau)=\frac{1}{2}+\frac{2}{\pi}\arctan(\tanh\tau),\qquad z(\tau)=y(\tau)-\frac{1}{2}. y ( τ ) = 2 1 + π 2 arctan ( tanh τ ) , z ( τ ) = y ( τ ) − 2 1 .
The exact identities
y ′ ( τ ) y ′ ( σ ) sec π ( z ( τ ) + z ( σ ) ) 2 = 2 π sech ( τ − σ ) , \sqrt{y'(\tau)y'(\sigma)}\sec\frac{\pi(z(\tau)+z(\sigma))}{2}=\frac{2}{\pi}\operatorname{sech}(\tau-\sigma), y ′ ( τ ) y ′ ( σ ) sec 2 π ( z ( τ ) + z ( σ )) = π 2 sech ( τ − σ ) ,
y ′ ( τ ) y ′ ( σ ) csc π ( z ( τ ) − z ( σ ) ) 2 = 2 π csch ( τ − σ ) (106) \sqrt{y'(\tau)y'(\sigma)}\csc\frac{\pi(z(\tau)-z(\sigma))}{2}=\frac{2}{\pi}\operatorname{csch}(\tau-\sigma) \tag*{(106)} y ′ ( τ ) y ′ ( σ ) csc 2 π ( z ( τ ) − z ( σ )) = π 2 csch ( τ − σ ) (106)
follow by substituting tan ( π z ( τ ) / 2 ) = tanh τ \tan(\pi z(\tau)/2)=\tanh\tau tan ( π z ( τ ) /2 ) = tanh τ .
The finite Fourier matrices. Here J J J is the p × p p\times p p × p reversal matrix, with coordinates numbered 0 , … , p − 1 0,\ldots,p-1 0 , … , p − 1 ; it is different from the tridiagonal matrix J J J of Section 3. The full and one-sided transformed kernels lead to the following quantities. For real Fourier frequency ξ \xi ξ , put
u ( ξ ) = ∫ − ∞ 0 sech τ 2 π e − i ξ τ d τ , v ( ξ ) = u ( − ξ ) , u(\xi)=\int_{-\infty}^{0}\frac{\operatorname{sech}\tau}{2\pi}e^{-i\xi\tau}\,d\tau,\qquad v(\xi)=u(-\xi), u ( ξ ) = ∫ − ∞ 0 2 π sech τ e − i ξ τ d τ , v ( ξ ) = u ( − ξ ) ,
c = u + v = 1 2 cosh ( π ξ / 2 ) , h = tanh ( π ξ / 2 ) , d = u − v , κ ℓ = cot ( π q ℓ / p ) . c=u+v=\frac{1}{2\cosh(\pi\xi/2)},\qquad h=\tanh(\pi\xi/2),\qquad d=u-v,\qquad\kappa_\ell=\cot(\pi q\ell/p). c = u + v = 2 cosh ( π ξ /2 ) 1 , h = tanh ( π ξ /2 ) , d = u − v , κ ℓ = cot ( π q ℓ / p ) .
The cotangents are constant on the chosen residue class, and cot ( π q ℓ / p ) = κ ℓ \cot(\pi q\ell/p)=\kappa_\ell cot ( π q ℓ / p ) = κ ℓ . All quantities u , v , c , h , d u,v,c,h,d u , v , c , h , d in the following matrices refer to this same frequency ξ \xi ξ . For these Fourier matrices, abbreviate
a ℓ ± = ( − 1 ) ℓ + 1 2 ( κ ℓ ± i h ) , 1 ≤ ℓ < p . a_{\ell}^{\pm}=\frac{(-1)^{\ell+1}}{2}(\kappa_{\ell}\pm ih),\qquad1\le\ell<p. a ℓ ± = 2 ( − 1 ) ℓ + 1 ( κ ℓ ± ih ) , 1 ≤ ℓ < p .
The full upper triangular matrix is
M ( ξ ) = ( u a 1 + − c a 2 + c ⋯ ( − 1 ) n c 0 − v a 2 − c a 3 − ⋯ a n + 1 − 0 0 u a 3 + − c ⋯ ( − 1 ) n + 1 c ⋮ ⋮ ⋮ ⋱ ⋱ ⋱ ⋮ 0 ⋯ 0 0 u a p − 2 + − c 0 ⋯ 0 0 0 − v a p − 1 − 0 ⋯ 0 0 0 0 u ) , p = 2 n + 1. (107) M(\xi)=\begin{pmatrix}
u & a_{1}^{+} & -c & a_{2}^{+} & c & \cdots& (-1)^{n}c\\
0 & -v & a_{2}^{-} & c & a_{3}^{-} & \cdots& a_{n+1}^{-}\\
0 & 0 & u & a_{3}^{+} & -c & \cdots& (-1)^{n+1}c\\
\vdots& \vdots& \vdots& \ddots& \ddots& \ddots& \vdots\\
0 & \cdots& 0 & 0 & u & a_{p-2}^{+} & -c\\
0 & \cdots& 0 & 0 & 0 & -v & a_{p-1}^{-}\\
0 & \cdots& 0 & 0 & 0 & 0 & u
\end{pmatrix},\qquad p=2n+1. \tag*{(107)} M ( ξ ) = u 0 0 ⋮ 0 0 0 a 1 + − v 0 ⋮ ⋯ ⋯ ⋯ − c a 2 − u ⋮ 0 0 0 a 2 + c a 3 + ⋱ 0 0 0 c a 3 − − c ⋱ u 0 0 ⋯ ⋯ ⋯ ⋱ a p − 2 + − v 0 ( − 1 ) n c a n + 1 − ( − 1 ) n + 1 c ⋮ − c a p − 1 − u , p = 2 n + 1. (107)
Its off-diagonal pattern is the same as the explicitly displayed matrix in § 5 ; its diagonal now alternates u , − v , u , … , − v , u u,-v,u,\ldots,-v,u u , − v , u , … , − v , u . For example, when p = 3 p=3 p = 3 it is
M ( ξ ) = ( u ( κ 1 + i h ) / 2 − c 0 − v ( κ 1 + i h ) / 2 0 0 u ) . M(\xi)=\begin{pmatrix}
u & (\kappa_{1}+ih)/2 & -c\\
0 & -v & (\kappa_{1}+ih)/2\\
0 & 0 & u
\end{pmatrix}. M ( ξ ) = u 0 0 ( κ 1 + ih ) /2 − v 0 − c ( κ 1 + ih ) /2 u .
Equivalently, with d = u − v d=u-v d = u − v and M 0 = M ( c , h ) M_{0}=M(c,h) M 0 = M ( c , h ) from § 5 ,
M = M 0 + d 2 I p , V = M ( c , − h ) − d 2 I p . M=M_{0}+\frac{d}{2}I_{p},\qquad V=M(c,-h)-\frac{d}{2}I_{p}. M = M 0 + 2 d I p , V = M ( c , − h ) − 2 d I p .
This makes the scalar difference between the Fourier and angular matrices explicit. For p = 1 p=1 p = 1 , take M = ( u ) M=(u) M = ( u ) and V = ( v ) V=(v) V = ( v ) ; no cotangent parameter occurs. Put V ( ξ ) = M ( − ξ ) V(\xi)=M(-\xi) V ( ξ ) = M ( − ξ ) , so V V V replaces u , v , h u,v,h u , v , h by v , u , − h v,u,-h v , u , − h . Define the 2 p × 2 p 2p\times2p 2 p × 2 p matrix polynomial and its coefficient blocks by
Q ( z ; ξ ) = I 2 p + z H + z 2 D , Q(z;\xi)=I_{2p}+z\mathcal{H}+z^{2}\mathcal{D}, Q ( z ; ξ ) = I 2 p + z H + z 2 D ,
H = ( 0 M J − J V V J − J M 0 ) , D = − diag ( M 2 , V 2 ) , \mathcal{H}=\begin{pmatrix}
0 & MJ-JV\\
VJ-JM & 0
\end{pmatrix},\qquad\mathcal{D}=-\operatorname{diag}(M^{2},V^{2}), H = ( 0 V J − J M M J − J V 0 ) , D = − diag ( M 2 , V 2 ) ,
D n ( ξ ) = upper-left p × p block of [ z n ] Q ( z ; ξ ) − 1 / 2 . (108) D_{n}(\xi)=\text{upper-left }p\times p\text{ block of }\left[z^{n}\right]Q(z;\xi)^{-1/2}. \tag*{(108)} D n ( ξ ) = upper-left p × p block of [ z n ] Q ( z ; ξ ) − 1/2 . (108)
The inverse square root is defined by its binomial series and equals I 2 p I_{2p} I 2 p at z = 0 z=0 z = 0 . The two p × p p\times p p × p blocks here pair frequencies ξ \xi ξ and − ξ -\xi − ξ . They should be distinguished from the original index-parity components of the discrete grids introduced below.
The apparent phase difference from § 5 is just a change of basis. For the same M , V M,V M , V , let
L d ( z ) = ( − i z M J J i z V ) , T = diag ( I p , i I p ) . L_{d}(z)=\begin{pmatrix}
-izM & J\\
J & izV
\end{pmatrix},\qquad T=\operatorname{diag}(I_{p},iI_{p}). L d ( z ) = ( − i z M J J i z V ) , T = diag ( I p , i I p ) .
Direct multiplication gives
L d ( z ) 2 = ( I − z 2 M 2 − i z ( M J − J V ) i z ( V J − J M ) I − z 2 V 2 ) , Q ( z ; ξ ) = T − 1 L d ( z ) 2 T . L_{d}(z)^{2}=\begin{pmatrix}
I-z^{2}M^{2} & -iz(MJ-JV)\\
iz(VJ-JM) & I-z^{2}V^{2}
\end{pmatrix},\qquad Q(z;\xi)=T^{-1}L_{d}(z)^{2}T. L d ( z ) 2 = ( I − z 2 M 2 i z ( V J − J M ) − i z ( M J − J V ) I − z 2 V 2 ) , Q ( z ; ξ ) = T − 1 L d ( z ) 2 T .
The same conjugation relates their inverse square roots, whose two diagonal blocks are therefore identical. When d = 0 d=0 d = 0 , this is exactly the matrix L L L used in § 5 .
For clarity, the angular function in that section is
G p ( t ) = 1 π ∫ − π / 2 π / 2 I ( t / p ; c , h ) c d ψ , c = 1 2 cos ψ , h = sin ψ , G_{p}(t)=\frac{1}{\pi}\int_{-\pi/2}^{\pi/2}\frac{\mathcal{I}(t/p;c,h)}{c}\,d\psi,\qquad c=\frac{1}{2}\cos\psi,\qquad h=\sin\psi, G p ( t ) = π 1 ∫ − π /2 π /2 c I ( t / p ; c , h ) d ψ , c = 2 1 cos ψ , h = sin ψ ,
I ( s ; c , h ) = c 2 { Y 0 ( s ) 00 + Y 0 ( s ) p p } , Y 0 ( s ) = ( L 0 ( s ) 2 ) − 1 / 2 . (109) \mathcal{I}(s;c,h)=\frac{c}{2}\{Y_{0}(s)_{00}+Y_{0}(s)_{pp}\},\qquad Y_{0}(s)=(L_{0}(s)^{2})^{-1/2}. \tag*{(109)} I ( s ; c , h ) = 2 c { Y 0 ( s ) 00 + Y 0 ( s ) pp } , Y 0 ( s ) = ( L 0 ( s ) 2 ) − 1/2 . (109)
Here L 0 L_{0} L 0 uses M 0 = M ( c , h ) M_{0}=M(c,h) M 0 = M ( c , h ) and V 0 = M ( c , − h ) V_{0}=M(c,-h) V 0 = M ( c , − h ) , with diagonal entries c / 2 , − c / 2 , … , c / 2 c/2,-c/2,\ldots,c/2 c /2 , − c /2 , … , c /2 . At c = 0 c=0 c = 0 , the ratio I / c \mathcal{I}/c I / c is defined by its displayed formula as the half-sum of two entries. Those finite matrices are continuous on the closed angular interval, so the binomial series is uniformly convergent there for small ∣ t ∣ |t| ∣ t ∣ . Thus G p G_{p} G p is analytic near zero.
Write σ x = ( 0 1 1 0 ) \sigma_x = \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix} σ x = ( 0 1 1 0 ) for the matrix exchanging the two index-parity components. The coordinate change z = z ( τ ) z = z(\tau) z = z ( τ ) is increasing, so it preserves each triangular order mask. Fourier transformation, with convention b ^ k ( ξ ) = ∫ R k ( τ ) e − i ξ τ d τ \widehat{b}_k(\xi) = \int R_k(\tau)e^{-i\xi\tau}\,d\tau b k ( ξ ) = ∫ R k ( τ ) e − i ξ τ d τ , now gives the Fourier multiplier B = − i σ x M ( D ) / p B = -i\sigma_x M(D)/p B = − i σ x M ( D ) / p , where σ x \sigma_x σ x acts on the two index-parity components as above, and M ( D ) M(D) M ( D ) denotes the operator with Fourier multiplier M ( ξ ) M(\xi) M ( ξ ) given in (A.11). Here the transform of the principal value kernel csch τ \operatorname{csch}\tau csch τ is − i π tanh ( π ξ / 2 ) -i\pi\tanh(\pi\xi/2) − iπ tanh ( π ξ /2 ) ; the delta correction is exactly the cotangent in (A.9). The same-parity secant kernels become the full or one-sided sech kernels in (A.11). The transformed reflection is reversal of the subinterval order and τ ↦ − τ \tau\mapsto-\tau τ ↦ − τ , with the constant parity matrix already described. The following calculation verifies its parity and frequency blocks explicitly.
Products with p × p p\times p p × p matrix-valued multipliers use the tensor-product convention; explicitly,
σ x M ± : = ( σ x ⊗ I p ) diag ( M ± , M ± ) = ( 0 M ± M ± 0 ) . \sigma_xM_{\pm}:=(\sigma_x\otimes I_p)\operatorname{diag}(M_{\pm},M_{\pm})=\begin{pmatrix}0&M_{\pm}\\M_{\pm}&0\end{pmatrix}. σ x M ± := ( σ x ⊗ I p ) diag ( M ± , M ± ) = ( 0 M ± M ± 0 ) .
Using (A.4) in (A.3) then gives the principal symbol − i σ x M ± / p -i\sigma_xM_{\pm}/p − i σ x M ± / p , where M ± M_{\pm} M ± is precisely (A.11) at u = v = c = 0 u=v=c=0 u = v = c = 0 , h = ± 1 h=\pm1 h = ± 1 .
Checking the parity and frequency blocks. The following calculation also fixes the factors of i i i in (A.12). Order the index-parity components as even, odd, put Z = diag ( 1 , − 1 ) Z=\operatorname{diag}(1,-1) Z = diag ( 1 , − 1 ) , and let σ x \sigma_x σ x exchange them. Conjugation by diag ( i j ) \operatorname{diag}(i^j) diag ( i j ) changes the original reversal to
R ′ = K q J R , K q = { i − q − 1 Z , q odd , i − q − 1 Z σ x , q even , ( R f ) ( ξ ) = f ( − ξ ) . R'=K_qJR,\qquad K_q=\begin{cases}i^{-q-1}Z,&q\text{ odd},\\i^{-q-1}Z\sigma_x,&q\text{ even},\end{cases}\qquad(Rf)(\xi)=f(-\xi). R ′ = K q J R , K q = { i − q − 1 Z , i − q − 1 Z σ x , q odd , q even , ( R f ) ( ξ ) = f ( − ξ ) .
Indeed its entry at ( j , q + 1 − j ) (j,q+1-j) ( j , q + 1 − j ) is i 2 j − q − 1 = ( − 1 ) j i − q − 1 i^{2j-q-1}=(-1)^ji^{-q-1} i 2 j − q − 1 = ( − 1 ) j i − q − 1 ; the parity is preserved for odd q q q and exchanged for even q q q . In both cases K q 2 = I K_q^2=I K q 2 = I and K q σ x = − σ x K q K_q\sigma_x=-\sigma_xK_q K q σ x = − σ x K q . Set Σ q = − i ε σ x K q \Sigma_q=-i\varepsilon\sigma_xK_q Σ q = − i ε σ x K q . Then Σ q 2 = I \Sigma_q^2=I Σ q 2 = I and its two eigenvalues are + 1 , − 1 +1,-1 + 1 , − 1 . The primed operators in this parity calculation are the same conjugated operators denoted B q B_q B q , R q R_q R q in the surrounding argument. From B ′ = − i σ x M ( D ) / p B'=-i\sigma_xM(D)/p B ′ = − i σ x M ( D ) / p one obtains, at frequency ξ \xi ξ ,
{ ε ( R q B q + B q R q ) } lim = 1 p Σ q { M ( ξ ) J − J M ( − ξ ) } R , ( B q 2 ) lim = − 1 p 2 M ( ξ ) 2 . \left\{\varepsilon(R_qB_q+B_qR_q)\right\}_{\lim}=\frac{1}{p}\Sigma_q\{M(\xi)J-JM(-\xi)\}R,\qquad(B_q^2)_{\lim}=-\frac{1}{p^2}M(\xi)^2. { ε ( R q B q + B q R q ) } l i m = p 1 Σ q { M ( ξ ) J − J M ( − ξ )} R , ( B q 2 ) l i m = − p 2 1 M ( ξ ) 2 .
On a Σ q \Sigma_q Σ q eigenspace with eigenvalue λ = ± 1 \lambda=\pm1 λ = ± 1 , pairing ξ \xi ξ and − ξ -\xi − ξ therefore gives exactly Q ( λ t / p ; ξ ) Q(\lambda t/p;\xi) Q ( λ t / p ; ξ ) . The linear coefficient exchanges the frequency blocks and the quadratic coefficient preserves them. Consequently every even coefficient is block diagonal in the frequency pair and is independent of λ \lambda λ . This proves both the use of (A.12) and the parity trace factor two used in the final cut calculation. The identical argument applies to the full-line principal symbols M + , M − M_+,M_- M + , M − .
The intermediate limit to be proved. The grid and Fourier calculations have now defined every term in the following form of Proposition 5.1:
lim q → ∞ [ q 2 m ( q p p , q ( t ) ) ] = 1 p 2 m ∫ R { ( D 2 m ( ξ ) M ( ξ ) ) 00 + ( D 2 m ( − ξ ) M ( − ξ ) ) 00 } d ξ . (110) \lim_{q\to\infty}\left[q^{2m}\left(qp_{p,q}(t)\right)\right]=\frac{1}{p^{2m}}\int_{\mathbb{R}}\left\{(D_{2m}(\xi)M(\xi))_{00}+(D_{2m}(-\xi)M(-\xi))_{00}\right\}\,d\xi. \tag*{(110)} q → ∞ lim [ q 2 m ( q p p , q ( t ) ) ] = p 2 m 1 ∫ R { ( D 2 m ( ξ ) M ( ξ ) ) 00 + ( D 2 m ( − ξ ) M ( − ξ ) ) 00 } d ξ . (110)
The two terms are paired before integration; absolute convergence refers to their sum. The scalar-shift lemma below then sets d = 0 d=0 d = 0 and turns this integral into (A.13). The case m = 0 m=0 m = 0 is immediate from q p p , q ( 0 ) = G p ( 0 ) = 1 qp_{p,q}(0)=G_p(0)=1 q p p , q ( 0 ) = G p ( 0 ) = 1 . It remains to prove the formula for m ≥ 1 m\ge1 m ≥ 1 .
The principal operator has a useful exact description at the reciprocal rational parameter q ^ / p \widehat q/p q / p . Its contribution is computed by Proposition 3.2 at the finite denominator p p p . This uses no contact-limit theorem and no induction assumption.
Let p ≥ 3 p\ge3 p ≥ 3 be odd, ( p , q ) = 1 (p,q)=1 ( p , q ) = 1 , and choose an odd integer q ^ ≡ q ( m o d p ) \widehat q\equiv q\pmod p q ≡ q ( mod p ) . Index the subintervals by r , s = 0 , … , p − 1 r,s=0,\ldots,p-1 r , s = 0 , … , p − 1 , set c r = ( − 1 ) r c_r=(-1)^r c r = ( − 1 ) r , and let J J J reverse the subintervals. Define the finite reciprocal-sine matrix for denominator p p p by
B r s = { ( − 1 ) ( s − r − 1 ) / 2 2 p sin ( π q ^ ( r + s + 1 ) / ( 2 p ) ) , r < s , r + s odd , 0 , otherwise . B_{rs}=\begin{cases}
\dfrac{(-1)^{(s-r-1)/2}}{2p\sin\!\left(\pi\widehat q(r+s+1)/(2p)\right)},&r<s,\quad r+s\text{ odd},\\
0,&\text{otherwise}.
\end{cases} B r s = ⎩ ⎨ ⎧ 2 p sin ( π q ( r + s + 1 ) / ( 2 p ) ) ( − 1 ) ( s − r − 1 ) /2 , 0 , r < s , r + s odd , otherwise .
The odd lift makes this the standard signed matrix for the rational parameter q ^ / p \widehat{q}/p q / p . For this odd denominator, the signed sine coefficients occur in equal pairs h p − j = h j h_{p-j}=h_j h p − j = h j , so their product is positive and ε = 1 \varepsilon=1 ε = 1 . Thus the linear coefficient in the finite generator of Section 3 is B J + J B BJ+JB B J + J B .
Let
( M + ) r s = { ( − 1 ) ℓ + 1 2 [ cot π q q ^ p + i ε r ] , r < s , r + s odd , ℓ = ( r + s + 1 ) / 2 , 0 , otherwise , M − = M + ‾ . (M_+)_{rs}=
\begin{cases}
\dfrac{(-1)^{\ell+1}}{2}\left[\cot\dfrac{\pi q\widehat{q}}{p}+i\varepsilon_r\right], & r<s,\ r+s\ \text{odd},\ \ell=(r+s+1)/2,\\
0, & \text{otherwise},
\end{cases}
\qquad M_-=\overline{M_+}. ( M + ) r s = ⎩ ⎨ ⎧ 2 ( − 1 ) ℓ + 1 [ cot p π q q + i ε r ] , 0 , r < s , r + s odd , ℓ = ( r + s + 1 ) /2 , otherwise , M − = M + .
These are the unnormalized full-line principal symbols at positive and negative Fourier frequencies; the normalized symbols are M ± / p M_\pm/p M ± / p .
Put
D r = exp ( i ε r π q ^ ( r + 1 / 2 ) 2 p ) , T r = { 1 , r even , i , r odd , D_r=\exp\left(\frac{i\varepsilon_r\pi\widehat{q}(r+1/2)}{2p}\right),
\qquad
T_r=
\begin{cases}
1, & r\ \text{even},\\
i, & r\ \text{odd},
\end{cases} D r = exp ( 2 p i ε r π q ( r + 1/2 ) ) , T r = { 1 , i , r even , r odd ,
G r e c = diag ( D r T r ) , κ 0 = e − i π q ^ / 2 ∈ { i , − i } . G_{\mathrm{rec}}=\operatorname{diag}(D_rT_r),\qquad
\kappa_0=e^{-i\pi\widehat{q}/2}\in\{i,-i\}. G rec = diag ( D r T r ) , κ 0 = e − iπ q /2 ∈ { i , − i } .
Lemma A.1 (The exact principal subtraction). For every m ≥ 0 m\geq0 m ≥ 0 , the block of the principal coefficient from subinterval zero to itself is scalar:
( F 2 m ( S B , q , R q ) ) 00 = α m , p , q I , α m , p , q = p q ^ , p ( 2 m ) ( 0 ) ( 2 m ) ! , (111) (F_{2m}(S_{B,q},R_q))_{00}=\alpha_{m,p,q}I,\qquad
\alpha_{m,p,q}=\frac{p_{\widehat{q},p}^{(2m)}(0)}{(2m)!}, \tag*{(111)} ( F 2 m ( S B , q , R q ) ) 00 = α m , p , q I , α m , p , q = ( 2 m )! p q , p ( 2 m ) ( 0 ) , (111) Here the subscript 00 means the block from subinterval zero to itself, retaining both index-parity components; I I I is the identity on that full-line block. The number α m , p , q \alpha_{m,p,q} α m , p , q is a known finite contact at denominator p p p , not the limit being proved at denominator q q q .
Proof. We verify a finite change of basis and then apply the finite rooted generator at denominator p p p . All matrices in this calculation have size p p p or 2 p 2p 2 p ; no limit is taken.
The reciprocal change of basis. The exact identities are
M + p = i G r e c B G r e c − 1 , M − p = − i G ‾ r e c B G ‾ r e c − 1 , G r e c − 1 J G ‾ r e c = κ 0 J (112) \frac{M_+}{p}=iG_{\mathrm{rec}}BG_{\mathrm{rec}}^{-1},\qquad
\frac{M_-}{p}=-i\overline{G}_{\mathrm{rec}}B\overline{G}_{\mathrm{rec}}^{-1},\qquad
G_{\mathrm{rec}}^{-1}J\overline{G}_{\mathrm{rec}}=\kappa_0J \tag*{(112)} p M + = i G rec B G rec − 1 , p M − = − i G rec B G rec − 1 , G rec − 1 J G rec = κ 0 J (112)
To verify the first identity, take an allowed entry and put θ = π q ( r + s + 1 ) / ( 2 p ) \theta=\pi q(r+s+1)/(2p) θ = π q ( r + s + 1 ) / ( 2 p ) . Since r , s r,s r , s have opposite parity,
( G r e c ) r r ( G r e c ) s s = − i ε r e i ε r θ , ( − 1 ) ℓ + 1 ( − 1 ) ( s − r − 1 ) / 2 = ε r , cot θ + i ε r = e i ε r θ sin θ . \frac{(G_{\mathrm{rec}})_{rr}}{(G_{\mathrm{rec}})_{ss}}
=-i\varepsilon_r e^{i\varepsilon_r\theta},\qquad
\frac{(-1)^{\ell+1}}{(-1)^{(s-r-1)/2}}=\varepsilon_r,\qquad
\cot\theta+i\varepsilon_r=\frac{e^{i\varepsilon_r\theta}}{\sin\theta}. ( G rec ) ss ( G rec ) r r = − i ε r e i ε r θ , ( − 1 ) ( s − r − 1 ) /2 ( − 1 ) ℓ + 1 = ε r , cot θ + i ε r = sin θ e i ε r θ .
These give ( M + / p ) r s = i ( ( G r e c ) r r / ( G r e c ) s s ) B r s (M_+/p)_{rs}=i((G_{\mathrm{rec}})_{rr}/(G_{\mathrm{rec}})_{ss})B_{rs} ( M + / p ) r s = i (( G rec ) r r / ( G rec ) ss ) B r s . Replacing q q q by q ^ \widehat{q} q does not change its cotangent. Complex conjugation gives the second identity. For the third, p − 1 − r p-1-r p − 1 − r has the same parity as r r r , so
( G ‾ r e c ) p − 1 − r , p − 1 − r ( G r e c ) r r = ε r e − i ε r π q ^ / 2 = e − i π q ^ / 2 . \frac{(\overline{G}_{\mathrm{rec}})_{p-1-r,p-1-r}}{(G_{\mathrm{rec}})_{rr}}
=\varepsilon_r e^{-i\varepsilon_r\pi\widehat{q}/2}
=e^{-i\pi\widehat{q}/2}. ( G rec ) r r ( G rec ) p − 1 − r , p − 1 − r = ε r e − i ε r π q /2 = e − iπ q /2 .
Identification with the finite generator. Consider the normalized principal matrix polynomial
Q ( z ) = I + z H + z 2 D , H = 1 p ( 0 M + J − J M − M − J − J M + 0 ) , D = − 1 p 2 diag ( M + 2 , M − 2 ) . Q(z)=I+zH+z^2D,\qquad
H=\frac{1}{p}
\begin{pmatrix}
0 & M_+J-JM_-\\
M_-J-JM_+ & 0
\end{pmatrix},
\qquad
D=-\frac{1}{p^2}\operatorname{diag}(M_+^2,M_-^2). Q ( z ) = I + z H + z 2 D , H = p 1 ( 0 M − J − J M + M + J − J M − 0 ) , D = − p 2 1 diag ( M + 2 , M − 2 ) .
With W = diag ( G r e c , G ‾ r e c ) W=\operatorname{diag}(G_{\mathrm{rec}},\overline{G}_{\mathrm{rec}}) W = diag ( G rec , G rec ) and H f = B J + J B H_f=BJ+JB H f = B J + J B , (A.16) gives
W − 1 H W = ( 0 s H f s H f 0 ) , W − 1 D W = diag ( B 2 , B 2 ) , s = i κ 0 ∈ { 1 , − 1 } . (113) W^{-1}HW=
\begin{pmatrix}
0 & sH_f\\
sH_f & 0
\end{pmatrix},
\qquad
W^{-1}DW=\operatorname{diag}(B^2,B^2),\qquad
s=i\kappa_0\in\{1,-1\}. \tag*{(113)} W − 1 H W = ( 0 s H f s H f 0 ) , W − 1 D W = diag ( B 2 , B 2 ) , s = i κ 0 ∈ { 1 , − 1 } . (113)
Indeed the upper off-diagonal block is i κ 0 ( B J + J B ) i\kappa_0(BJ+JB) i κ 0 ( B J + J B ) and the lower is − i κ ‾ 0 ( B J + J B ) -i\overline{\kappa}_0(BJ+JB) − i κ 0 ( B J + J B ) , which are equal. The remaining block calculation can be written explicitly. Put
H b l k = 1 2 ( I p I p I p − I p ) , H b l k − 1 = H b l k . H_{\mathrm{blk}}=\frac{1}{\sqrt{2}}
\begin{pmatrix}
I_p & I_p\\
I_p & -I_p
\end{pmatrix},
\qquad
H_{\mathrm{blk}}^{-1}=H_{\mathrm{blk}}. H blk = 2 1 ( I p I p I p − I p ) , H blk − 1 = H blk .
Then
H b l k − 1 W − 1 Q ( z ) W H b l k = diag ( I + s z H f + z 2 B 2 , I − s z H f + z 2 B 2 ) . H_{\mathrm{blk}}^{-1}W^{-1}Q(z)WH_{\mathrm{blk}}
=\operatorname{diag}\left(I+szH_f+z^2B^2,\ I-szH_f+z^2B^2\right). H blk − 1 W − 1 Q ( z ) W H blk = diag ( I + sz H f + z 2 B 2 , I − sz H f + z 2 B 2 ) .
The two blocks differ only by z ↦ − z z \mapsto-z z ↦ − z , so their even inverse-square-root coefficients are identical. Their first diagonal entry is therefore also the first diagonal entry before the block change of basis. The finite rooted generator of Proposition 3.2 , applied at denominator p p p , proves
[ z 2 m ] ( Q ( z ) − 1 / 2 ) 00 = p p q , p ( 2 m ) ( 0 ) ( 2 m ) ! . (114) [z^{2m}]\left(\mathcal{Q}(z)^{-1/2}\right)_{00}=\frac{p\,p_{q,p}^{(2m)}(0)}{(2m)!}. \tag*{(114)} [ z 2 m ] ( Q ( z ) − 1/2 ) 00 = ( 2 m )! p p q , p ( 2 m ) ( 0 ) . (114)
The block for the first subinterval is scalar in the two index parity sectors, and its off-diagonal frequency coefficient is zero in even degree. Swapping positive and negative Fourier frequencies leaves (A.18) unchanged. The original fractional-translation phase gauges do not alter a scalar same-channel diagonal identity. Hence (A.18) supplies the exact principal regularization constant at every Fourier frequency for every fixed odd p p p .
Only the first-subinterval block is used at the cut. This argument does not claim that the other subinterval blocks are scalar. This gives (A.15) on the full shifted grids. □
The exact algebra of the full-grid principal operators
The row argument will place a principal resolvent to the left of the smooth remainder B q − S B , q B_q-SB_{,q} B q − S B , q . To control its effect on rows, we need more than its operator norm: we need the exact kernel types of this left factor. The following finite Fourier algebra shows that inversion creates only Cauchy kernels, local identifications, and their reflected versions.
Fix the odd numerator and a denominator class as above. Number the N = 2 p N=2p N = 2 p grid channels by a a a , and write their coordinates as z a , n = Δ ( n + β a ) z_{a,n}=\Delta(n+\beta_a) z a , n = Δ ( n + β a ) , n ∈ Z n\in\mathbb{Z} n ∈ Z . Let
( F f ) a ( ϑ ) = ∑ n ∈ Z f a ( n ) e − i n ϑ , G ( ϑ ) = diag ( e − i β a ϑ ) , D = diag ( e 2 π i β a ) . (\mathcal{F}f)_a(\vartheta)=\sum_{n\in\mathbb{Z}}f_a(n)e^{-in\vartheta},\qquad G(\vartheta)=\operatorname{diag}(e^{-i\beta_a\vartheta}),\qquad D=\operatorname{diag}(e^{2\pi i\beta_a}). ( F f ) a ( ϑ ) = n ∈ Z ∑ f a ( n ) e − in ϑ , G ( ϑ ) = diag ( e − i β a ϑ ) , D = diag ( e 2 π i β a ) .
All channel offsets may be chosen with representatives fixed in the denominator class: replacing q q q by q + 4 p q+4p q + 4 p changes β r , ν \beta_{r,\nu} β r , ν by the integer − ϵ r ( 2 r + 1 ) -\epsilon_r(2r+1) − ϵ r ( 2 r + 1 ) . The constant reversal phases are fixed as well, since q q q is fixed modulo four. The sign ϵ \epsilon ϵ is one for odd q q q , and ( − 1 ) ( p − 1 ) / 2 (-1)^{(p-1)/2} ( − 1 ) ( p − 1 ) /2 for even q q q ; this follows by pairing h j h_j h j with h q − j h_{q-j} h q − j and, for even q q q , retaining the central factor h q / 2 = 2 sin ( π p / 2 ) h_{q/2}=2\sin(\pi p/2) h q /2 = 2 sin ( π p /2 ) . Thus all finite matrices below are independent of the increasing denominator within the chosen class.
The gauge G G G need not be periodic; we use it only on − π < ϑ < π -\pi<\vartheta<\pi − π < ϑ < π . Its endpoint mismatch will be retained explicitly.
For the inverse calculations, let Γ + \Gamma_+ Γ + and Γ − \Gamma_- Γ − be the counterclockwise circles of radius 1 / 4 1/4 1/4 centered at + 1 +1 + 1 and − 1 -1 − 1 , and put Γ = Γ + ∪ Γ − \Gamma=\Gamma_+\cup\Gamma_- Γ = Γ + ∪ Γ − . Write ζ \zeta ζ for their complex coordinate and set
g ( ζ ) = { 1 / ζ , ζ near Γ + , − 1 / ζ , ζ near Γ − . g(\zeta)=
\begin{cases}
1/\zeta, & \zeta\text{ near }\Gamma_+,\\
-1/\zeta, & \zeta\text{ near }\Gamma_-.
\end{cases} g ( ζ ) = { 1/ ζ , − 1/ ζ , ζ near Γ + , ζ near Γ − .
These contours separate the two spectral groups of the common reflection ϵ R q \epsilon R_q ϵ R q . Small norm perturbations leave the groups inside the circles. The same scalar function g g g gives the inverse square root of the square of either pencil: near + 1 +1 + 1 the chosen square root of ζ 2 \zeta^2 ζ 2 is ζ \zeta ζ , whereas near − 1 -1 − 1 it is − ζ -\zeta − ζ .
Lemma A.2 (Principal resolvents have the same kernel types). Let S B , q S_{B,q} S B , q be the exact full-grid Cauchy principal part of B q B_q B q , and let R q R_q R q be the exact full-grid reversal, including its channel permutation and constant phases. Set
T q , t p r = ϵ R q + t S B , q . T^{\mathrm{pr}}_{q,t}=\epsilon R_q+tS_{B,q}. T q , t pr = ϵ R q + t S B , q . For sufficiently small complex t t t , uniformly in the denominator in the chosen class, each channel block of
( ζ I − T q , t p r ) − 1 , ζ ∈ Γ , (\zeta I-T^{\mathrm{pr}}_{q,t})^{-1},\qquad\zeta\in\Gamma, ( ζ I − T q , t pr ) − 1 , ζ ∈ Γ , is a sum of the following four types:
shifted Cauchy convolution, with entries κ / ( n − k + β a − β b ) \kappa/(n-k+\beta_a-\beta_b) κ / ( n − k + β a − β b ) ;
aligned local identification, when β a − β b \beta_a-\beta_b β a − β b is an integer;
reflected shifted Cauchy convolution, with entries κ / ( n + k + β a + β b ) \kappa/(n+k+\beta_a+\beta_b) κ / ( n + k + β a + β b ) ;
(4) aligned local reflection, when β a + β b \beta_a+\beta_b β a + β b is an integer.
The Cauchy types apply only to noninteger shifts. There is no singular integer-shift Cauchy kernel and no extra kernel ( − 1 ) n − k / ( n − k + δ ) (-1)^{n-k}/(n-k+\delta) ( − 1 ) n − k / ( n − k + δ ) or its reflected version. All scalar coefficients in these representations are bounded by C p C_p C p , uniformly for ζ ∈ Γ \zeta\in\Gamma ζ ∈ Γ and on a sufficiently small closed t t t -disk.
Proof. There are three steps: retain the endpoint twists of the gauge, show that these twists survive inversion of a finite matrix, and invert each block explicitly.
The endpoint and reflection twists. Under the gauge G G G , the full-grid principal part has constant matrix multipliers B + , B − B_+,B_- B + , B − on the two half-intervals ϑ > 0 \vartheta>0 ϑ > 0 , ϑ < 0 \vartheta<0 ϑ < 0 . For an original Cauchy block put δ = β a − β b ∉ Z \delta=\beta_a-\beta_b\notin\mathbb{Z} δ = β a − β b ∈ / Z . Its multiplier before gauging is a scalar multiple of
π { cot ( π δ ) − i sgn ϑ } e i δ ϑ . \pi\{\cot(\pi\delta)-i\operatorname{sgn}\vartheta\}e^{i\delta\vartheta}. π { cot ( π δ ) − i sgn ϑ } e i δ ϑ .
The values at π \pi π and − π -\pi − π agree. Since cot ( π δ ) + i = e 2 π i δ ( cot ( π δ ) − i ) \cot(\pi\delta)+i=e^{2\pi i\delta}(\cot(\pi\delta)-i) cot ( π δ ) + i = e 2 π i δ ( cot ( π δ ) − i ) , the gauged constants satisfy
B − = D B + D − 1 . (115) B_-=DB_+D^{-1}. \tag*{(115)} B − = D B + D − 1 . (115)
Zero blocks obey the same identity.
Write a ↦ σ ( a ) a\mapsto\sigma(a) a ↦ σ ( a ) for the reversal of channels. The exact grid reversal sends f σ ( a ) ( k ) f_{\sigma(a)}(k) f σ ( a ) ( k ) to the output with k = − n − m a k=-n-m_a k = − n − m a , where m a = β a + β σ ( a ) ∈ Z m_a=\beta_a+\beta_{\sigma(a)}\in\mathbb{Z} m a = β a + β σ ( a ) ∈ Z . Thus its Fourier action has the form
( R q f ^ ) a ( ϑ ) = ω a e i m a ϑ f ^ σ ( a ) ( − ϑ ) . (\widehat{R_qf})_a(\vartheta)=\omega_a e^{im_a\vartheta}\widehat{f}_{\sigma(a)}(-\vartheta). ( R q f ) a ( ϑ ) = ω a e i m a ϑ f σ ( a ) ( − ϑ ) .
The identity m a = β a + β σ ( a ) m_a=\beta_a+\beta_{\sigma(a)} m a = β a + β σ ( a ) shows that conjugating by G G G removes this exponential exactly. The gauged reversal is R c h J R_{\mathrm{ch}}J R ch J , where ( J g ) ( ϑ ) = g ( − ϑ ) (Jg)(\vartheta)=g(-\vartheta) ( J g ) ( ϑ ) = g ( − ϑ ) and ( R c h ) a , σ ( a ) = ω a (R_{\mathrm{ch}})_{a,\sigma(a)}=\omega_a ( R ch ) a , σ ( a ) = ω a . In particular
R c h 2 = I , D R c h D = R c h . (116) R_{\mathrm{ch}}^2=I,\qquad DR_{\mathrm{ch}}D=R_{\mathrm{ch}}. \tag*{(116)} R ch 2 = I , D R ch D = R ch . (116)
The second equality follows entrywise from e 2 π i m a = 1 e^{2\pi i m_a}=1 e 2 π i m a = 1 .
A finite matrix algebra closed under inversion. More generally consider gauged operators C g + E J g C_g+\mathcal{E}Jg C g + E J g , with constants C ± , E ± C_\pm,\mathcal{E}_\pm C ± , E ± on the positive and negative half-intervals, satisfying
C − = D C + D − 1 , E − = D E + D . (117) C_-=DC_+D^{-1},\qquad\mathcal{E}_-=D\mathcal{E}_+D. \tag*{(117)} C − = D C + D − 1 , E − = D E + D . (117)
Pair the vectors g ( ϑ ) , g ( − ϑ ) g(\vartheta),g(-\vartheta) g ( ϑ ) , g ( − ϑ ) for 0 < ϑ < π 0<\vartheta<\pi 0 < ϑ < π . The operator becomes the single finite matrix
M = ( C + E + E − C − ) . \mathcal{M}=
\begin{pmatrix}
C_+ & \mathcal{E}_+\\
\mathcal{E}_- & C_-
\end{pmatrix}. M = ( C + E − E + C − ) .
The two conditions in (A.21) are exactly
[ M , Λ ] = 0 , Λ = ( 0 D − 1 D 0 ) , Λ 2 = I . [\mathcal{M},\Lambda]=0,\qquad
\Lambda=
\begin{pmatrix}
0 & D^{-1}\\
D & 0
\end{pmatrix},
\qquad
\Lambda^2=I. [ M , Λ ] = 0 , Λ = ( 0 D D − 1 0 ) , Λ 2 = I .
This proves closure under sums, products and inverses whenever the finite matrix is invertible. The principal pencil belongs to this algebra, with
C ± = t B ± , E ± = ε R c h , T ( t ) = ( t B + ε R c h ε R c h t B − ) . C_\pm=tB_\pm,\qquad\mathcal{E}_\pm=\varepsilon R_{\mathrm{ch}},\qquad
\mathcal{T}(t)=
\begin{pmatrix}
tB_+ & \varepsilon R_{\mathrm{ch}}\\
\varepsilon R_{\mathrm{ch}} & tB_-
\end{pmatrix}. C ± = t B ± , E ± = ε R ch , T ( t ) = ( t B + ε R ch ε R ch t B − ) .
Equations (A.19) and (A.20) verify the two conditions. Consequently every inverse ( ζ I − T ( t ) ) − 1 (\zeta I-\mathcal{T}(t))^{-1} ( ζ I − T ( t ) ) − 1 retains the twists whenever it exists.
Inverse Fourier transformation, including integer shifts. We treat a noninteger shift and an integer shift separately. This is where local identifications enter; they must not be replaced by singular integer-shift Cauchy kernels. For an ordinary block set δ = β a − β b \delta=\beta_a-\beta_b δ = β a − β b and let c + , c − c_+,c_- c + , c − be the corresponding gauged entries. Removing the gauge gives e i δ ϑ c ± e^{i\delta\vartheta}c_\pm e i δ ϑ c ± , with c − = e 2 π i δ c + c_-=e^{2\pi i\delta}c_+ c − = e 2 π i δ c + . If δ ∉ Z \delta\notin\mathbb{Z} δ ∈ / Z , put
κ = c + sin ( π δ ) e i π δ π . \kappa=\frac{c_+\sin(\pi\delta)e^{i\pi\delta}}{\pi}. κ = π c + sin ( π δ ) e iπ δ .
Then the multiplier is precisely κ π ( cot π δ − i sgn ϑ ) e i δ ϑ \kappa\pi(\cot\pi\delta-i\operatorname{sgn}\vartheta)e^{i\delta\vartheta} κπ ( cot π δ − i sgn ϑ ) e i δ ϑ , whose Fourier coefficient at index j j j is κ / ( j + δ ) \kappa/(j+\delta) κ / ( j + δ ) . If δ = m ∈ Z \delta=m\in\mathbb{Z} δ = m ∈ Z , the twist instead forces c − = c + c_-=c_+ c − = c + . The multiplier is c + e i m ϑ c_+e^{im\vartheta} c + e im ϑ and its coefficient is c + 1 j = − m c_{+1j}=-m c + 1 j = − m . The resulting operator is f b ( n ) ↦ c + f b ( n + m ) f_b(n) \mapsto c_{+}f_b(n+m) f b ( n ) ↦ c + f b ( n + m ) , which identifies the same coordinate z a , n = z b , n + m z_{a,n}=z_{b,n+m} z a , n = z b , n + m on the two grids.
For a reflected block the calculation is identical with δ = β a + β b \delta=\beta_a+\beta_b δ = β a + β b , because removing the gauge gives e i ( β a + β b ) ϑ ( E ± ) a b f ^ b ( − ϑ ) e^{i(\beta_a+\beta_b)\vartheta}(\mathcal{E}_{\pm})_{ab}\widehat{f}_b(-\vartheta) e i ( β a + β b ) ϑ ( E ± ) ab f b ( − ϑ ) . The resulting kernel depends on n + k n+k n + k rather than n − k n-k n − k . For an integer sum m m m it sends f b f_b f b to f b ( − n − m ) f_b(-n-m) f b ( − n − m ) , the coordinate reflection w = − z w=-z w = − z . These calculations exhaust the four asserted types. In particular the endpoint compatibility has eliminated the possible extra oscillatory Fourier coefficients.
Uniform bounds for the resulting coefficients. Finally Z ( 0 ) \mathcal{Z}(0) Z ( 0 ) is a self-adjoint involution: it is the finite matrix of the unitary reflection ε R q \varepsilon R_q ε R q after unitary Fourier changes of basis. Its resolvent on Γ \Gamma Γ has norm at most 4. Choose the t t t -disk so that ∥ Z ( t ) − Z ( 0 ) ∥ ≤ 1 / 8 \|\mathcal{Z}(t)-\mathcal{Z}(0)\|\leq1/8 ∥ Z ( t ) − Z ( 0 ) ∥ ≤ 1/8 . The Neumann series then gives
∥ ( ζ I − Z ( t ) ) − 1 ∥ ≤ 8 , ζ ∈ Γ . \|(\zeta I-\mathcal{Z}(t))^{-1}\|\leq8,\qquad\zeta\in\Gamma. ∥ ( ζ I − Z ( t ) ) − 1 ∥ ≤ 8 , ζ ∈ Γ.
Every c + c_{+} c + is bounded by this matrix norm, and ∣ κ ∣ ≤ ∣ c + ∣ / π |\kappa|\leq|c_{+}|/\pi ∣ κ ∣ ≤ ∣ c + ∣/ π . The noninteger fractional shifts range through a fixed finite set in the chosen denominator class; integer parts merely renumber the grid indices. This proves all uniform coefficient estimates and completes the proof. □
After the exact subtraction, the rows converge
Strong operator convergence does not evaluate a changing matrix diagonal. We instead regard a row as a function of the input coordinate. If a matrix K q K_q K q acts between grids of spacing Δ \Delta Δ , its normalized row at output index j j j is the step function
r q , j ( w ) = ( K q ) j k Δ on input cell k , ∥ r q , j ∥ L 2 2 = 1 Δ ∑ k ∣ ( K q ) j k ∣ 2 . r_{q,j}(w)=\frac{(K_q)_{jk}}{\Delta}\quad\text{on input cell }k,\qquad\|r_{q,j}\|_{L^2}^{2}=\frac{1}{\Delta}\sum_k |(K_q)_{jk}|^2. r q , j ( w ) = Δ ( K q ) j k on input cell k , ∥ r q , j ∥ L 2 2 = Δ 1 k ∑ ∣ ( K q ) j k ∣ 2 .
For several grid components, the norm sums over all components. Interpolate in the output coordinate between adjacent rows of each component. We call these rows regular near zero if they converge locally uniformly as L 2 L^2 L 2 -valued functions and satisfy a common Hölder estimate
∥ r q ( x ) − r q ( x ′ ) ∥ L 2 ≤ C ∣ x − x ′ ∣ β , β > 0 , \|r_q(x)-r_q(x')\|_{L^2}\leq C|x-x'|^\beta,\qquad\beta>0, ∥ r q ( x ) − r q ( x ′ ) ∥ L 2 ≤ C ∣ x − x ′ ∣ β , β > 0 ,
on compact output neighborhoods. The estimate concerns the interpolated rows; the input step functions retain their isometric grid embedding.
First check the single remainder. On every compact output neighborhood W W W of 0 inside I I I , the single remainder E q = B q − S B , q E_q=B_q-SB_{,q} E q = B q − S B , q has regular rows. For (A.2), subtracting (A.3) removes its only pole and leaves a smooth kernel. For (A.5)–(A.6), the secant and its output derivative are uniformly bounded on W × I W\times I W × I . Moving the triangular jump changes a row on an interval of length ∣ x − x ′ ∣ |x-x'| ∣ x − x ′ ∣ , giving row difference C ∣ x − x ′ ∣ 1 / 2 C|x-x'|^{1/2} C ∣ x − x ′ ∣ 1/2 . Outside I I I , the subtracted full-line Cauchy row and its output derivative are square integrable, uniformly on W W W . The rows therefore have common C β C^\beta C β bounds for any 0 < β < 1 / 2 0<\beta<1/2 0 < β < 1/2 , and ordinary quadrature proves their convergence. This holds on all 2 p 2p 2 p channels at once.
The following argument passes the entire inverse square root to the limit near t = 0 t=0 t = 0 . It uses one resolvent identity in place of a separate expansion at each Taylor order. Throughout it, R q R_q R q is the same full-grid reflection in the true and principal pencils. All operator identities are first taken on the grid spaces. After their isometric embedding, a grid identity becomes the step-function projection P q P_q P q , not the ambient identity.
Lemma A.3 (Analytic regular rows). There is a number τ > 0 \tau>0 τ > 0 , depending only on the fixed numerator and residue class, such that
T q , t = ε R q + t B q , T q , t p r = ε R q + t S B , q , T_{q,t}=\varepsilon R_q+tB_q,\qquad T^{\mathrm{pr}}_{q,t}=\varepsilon R_q+tS B_{,q}, T q , t = ε R q + t B q , T q , t pr = ε R q + tS B , q , X q ( t ) = ( T q 2 ) − 1 / 2 , X q p r ( t ) = ( ( T q p r ) 2 ) − 1 / 2 X_q(t)=(T_q^2)^{-1/2},\qquad X_q^{\mathrm{pr}}(t)=((T^{\mathrm{pr}}_q)^2)^{-1/2} X q ( t ) = ( T q 2 ) − 1/2 , X q pr ( t ) = (( T q pr ) 2 ) − 1/2 are defined by their branches equal to the grid identity at zero. The difference X q ( t ) − X q p r ( t ) X_q(t)-X_q^{\mathrm{pr}}(t) X q ( t ) − X q pr ( t ) has regular rows near zero, locally uniformly jointly in the output coordinate and ∣ t ∣ < τ |t|<\tau ∣ t ∣ < τ . Its limiting row functions are holomorphic in t t t . In particular, every fixed Taylor coefficient of this difference has regular rows on the same output neighborhood.
Proof. We first obtain uniform inverse bounds, then isolate the single regular remainder and estimate its rows.
Uniform resolvent bounds and a compact contour formula. The preceding estimates give ∥ B q ∥ , ∥ S B , q ∥ ≤ C p \lVert B_q\rVert,\lVert S_{B,q}\rVert\le C_p ∥ B q ∥ , ∥ S B , q ∥ ≤ C p . The common reflection Q q = ε R q Q_q=\varepsilon R_q Q q = ε R q is a self-adjoint involution. Choose τ \tau τ with τ C p < 1 / 8 \tau C_p<1/8 τ C p < 1/8 . The two resolvents
R q ( ζ , t ) = ( ζ I − T q , t ) − 1 , R q p r ( ζ , t ) = ( ζ I − T q , t p r ) − 1 R_q(\zeta,t)=(\zeta I-T_{q,t})^{-1},\qquad R_q^{\mathrm{pr}}(\zeta,t)=(\zeta I-T_{q,t}^{\mathrm{pr}})^{-1} R q ( ζ , t ) = ( ζ I − T q , t ) − 1 , R q pr ( ζ , t ) = ( ζ I − T q , t pr ) − 1
then exist for ζ ∈ Γ \zeta\in\Gamma ζ ∈ Γ , ∣ t ∣ ≤ τ |t|\le\tau ∣ t ∣ ≤ τ , and have norm at most 8. Indeed, ( ζ I − Q q ) − 1 (\zeta I-Q_q)^{-1} ( ζ I − Q q ) − 1 has norm at most 4, and the perturbation in either pencil has norm less than 1 / 8 1/8 1/8 .
The same Neumann series prove strong convergence with the adjoints, uniformly for ζ ∈ Γ \zeta\in\Gamma ζ ∈ Γ and on the closed t t t -disk. To keep track of the grid identity, put R q , 0 ( ζ ) = ( ζ I − Q q ) − 1 R_{q,0}(\zeta)=(\zeta I-Q_q)^{-1} R q , 0 ( ζ ) = ( ζ I − Q q ) − 1 and use hats for the isometric embeddings into the common L 2 L^2 L 2 space. Then
R ^ q , 0 ( ζ ) = ζ P q + ε R ^ q ζ 2 − 1 , \widehat{R}_{q,0}(\zeta)=\frac{\zeta P_q+\varepsilon\widehat{R}_q}{\zeta^2-1}, R q , 0 ( ζ ) = ζ 2 − 1 ζ P q + ε R q ,
and
R ^ q ( ζ , t ) = R ^ q , 0 + ∑ k ≥ 1 t k ( R ^ q , 0 B ^ q ) k R ^ q , 0 . \widehat{R}_q(\zeta,t)=\widehat{R}_{q,0}+\sum_{k\ge1}t^k(\widehat{R}_{q,0}\widehat{B}_q)^k\widehat{R}_{q,0}. R q ( ζ , t ) = R q , 0 + k ≥ 1 ∑ t k ( R q , 0 B q ) k R q , 0 .
Here P q → I P_q\to I P q → I strongly; the same formula holds for the principal pencil. Each fixed term converges strongly with its adjoint, and the geometric tails are uniform. This proves the asserted convergences without replacing a grid identity by the ambient identity.
The inverse square roots are given by
X q ( t ) = 1 2 π i ∫ Γ g ( ζ ) R q ( ζ , t ) d ζ , X q p r ( t ) = 1 2 π i ∫ Γ g ( ζ ) R q p r ( ζ , t ) d ζ . X_q(t)=\frac{1}{2\pi i}\int_\Gamma g(\zeta)R_q(\zeta,t)\,d\zeta,\qquad X_q^{\mathrm{pr}}(t)=\frac{1}{2\pi i}\int_\Gamma g(\zeta)R_q^{\mathrm{pr}}(\zeta,t)\,d\zeta. X q ( t ) = 2 π i 1 ∫ Γ g ( ζ ) R q ( ζ , t ) d ζ , X q pr ( t ) = 2 π i 1 ∫ Γ g ( ζ ) R q pr ( ζ , t ) d ζ .
For completeness, these are exactly the original binomial branches. On Γ \Gamma Γ , ∣ ζ 2 − 1 ∣ ≤ 9 / 16 |\zeta^2-1|\le9/16 ∣ ζ 2 − 1∣ ≤ 9/16 , and
g ( ζ ) = ∑ k ≥ 0 ( − 1 / 2 k ) ( ζ 2 − 1 ) k g(\zeta)=\sum_{k\ge0}\binom{-1/2}{k}(\zeta^2-1)^k g ( ζ ) = k ≥ 0 ∑ ( k − 1/2 ) ( ζ 2 − 1 ) k
uniformly on both circles. For a polynomial a a a , the elementary resolvent Cauchy formula is
1 2 π i ∫ Γ a ( ζ ) ( ζ I − T ) − 1 d ζ = a ( T ) . \frac{1}{2\pi i}\int_\Gamma a(\zeta)(\zeta I-T)^{-1}\,d\zeta=a(T). 2 π i 1 ∫ Γ a ( ζ ) ( ζ I − T ) − 1 d ζ = a ( T ) .
One proves it by deforming the two circles to a large circle and using the large-ζ \zeta ζ Neumann series; no spectrum lies between the contours because the perturbation of Q q Q_q Q q has norm below 1 / 8 1/8 1/8 . Termwise integration of the uniformly convergent scalar series therefore gives ∑ k ( − 1 / 2 k ) ( T 2 − I ) k \sum_k\binom{-1/2}{k}(T^2-I)^k ∑ k ( k − 1/2 ) ( T 2 − I ) k . This operator series converges, since ∥ T 2 − I ∥ ≤ 2 / 8 + 1 / 64 = 17 / 64 < 1 \lVert T^2-I\rVert\le2/8+1/64=17/64<1 ∥ T 2 − I ∥ ≤ 2/8 + 1/64 = 17/64 < 1 for both pencils.
One resolvent identity isolates the regular remainder. Since T q , t − T q , t p r = t E q T_{q,t}-T_{q,t}^{\mathrm{pr}}=tE_q T q , t − T q , t pr = t E q , the inverse identity gives
R q − R q p r = t R q p r E q R q . R_q-R_q^{\mathrm{pr}}=tR_q^{\mathrm{pr}}E_qR_q. R q − R q pr = t R q pr E q R q .
Thus
X q ( t ) − X q p r ( t ) = t 2 π i ∫ Γ g ( ζ ) R q p r ( ζ , t ) E q R q ( ζ , t ) d ζ . (118) X_q(t)-X_q^{\mathrm{pr}}(t)=\frac{t}{2\pi i}\int_\Gamma g(\zeta)R_q^{\mathrm{pr}}(\zeta,t)E_qR_q(\zeta,t)\,d\zeta. \tag*{(118)} X q ( t ) − X q pr ( t ) = 2 π i t ∫ Γ g ( ζ ) R q pr ( ζ , t ) E q R q ( ζ , t ) d ζ . (118)
The order of the factors is useful: E q E_q E q has regular rows, the bounded right factor preserves their L 2 L^2 L 2 control, and the left principal resolvent has exactly the kernel types classified in Lemma A.2. Thus only one regular remainder occurs, rather than a separate product estimate for every Taylor order. Lemma A.2 expresses the left factor R q p r R_q^{\mathrm{pr}} R q pr as a finite sum of aligned local operators, shifted Cauchy blocks, and their compositions with reflection. Their coefficients are bounded by C p C_p C p uniformly for ζ ∈ Γ \zeta\in\Gamma ζ ∈ Γ , ∣ t ∣ ≤ τ |t|\le\tau ∣ t ∣ ≤ τ , and all denominators in the class. The right factor R q R_q R q is bounded and converges strongly with its adjoint, uniformly on the same compact parameter set. After integer relabeling of the grids, the finite gauged matrices and the sign ε \varepsilon ε are fixed in the denominator class. The scalar coefficients of the left-factor representation therefore depend on ζ \zeta ζ , t t t but not on q q q ; their grid realizations converge by the same quadrature as the original principal kernels.
The integrand has convergent regular rows. We give the local estimate once, uniformly in the parameters. There are two operations to check.
Right multiplication. Right multiplication of a regular row by a bounded operator preserves its Hölder bound. Strong adjoint convergence gives locally uniform convergence of these rows: the continuous limit row field has compact range on a compact output interval, and uniformly bounded strongly convergent operators converge uniformly on a compact set by a finite-net argument. The same argument is uniform for ζ ∈ Γ \zeta\in\Gamma ζ ∈ Γ and ∣ t ∣ ≤ τ |t| \le\tau ∣ t ∣ ≤ τ . Consequently K q = E q R q K_q = E_qR_q K q = E q R q has regular rows, with norms and Hölder constants bounded by C ρ C_\rho C ρ on this compact parameter set. It also converges strongly with its adjoint, by the preceding operator convergence of E q E_q E q and R q R_q R q .
The classified left factors. A local left factor just selects an aligned row, and reflection permutes channels and replaces x x x by − x -x − x . For a reflected Cauchy block, first reflect the intermediate coordinate; the same argument below then applies on a symmetric neighborhood. For a left Cauchy block H q H_q H q , use K q = E q R q K_q = E_qR_q K q = E q R q and choose a smooth cutoff χ \chi χ supported in W W W , equal to one on a smaller symmetric neighborhood W 0 W_0 W 0 . Split
H q K q = H q χ K q + H q ( 1 − χ ) K q . H_qK_q = H_q\chi K_q + H_q(1-\chi)K_q. H q K q = H q χ K q + H q ( 1 − χ ) K q .
Near the Cauchy pole. In the first term, apply the Cauchy kernel to the Hilbert-valued row field of χ K q \chi K_q χ K q . Subtract its value at the output point near the pole. The contribution of ∣ x − w ∣ < ϵ |x-w| < \epsilon ∣ x − w ∣ < ϵ is at most C ϵ β C\epsilon^\beta C ϵ β times the row bound, uniformly in the mesh: use dyadic annuli and the fixed positive separation of the nonintegral grid shifts from the integers. For the subtracted constant, scalar quadrature ((A.9)) gives the cotangent term and the smooth principal-value transform of a cutoff. Away from the pole, ordinary Hilbert-valued quadrature converges uniformly on compact output and parameter sets. Sending the mesh to zero and then ϵ \epsilon ϵ to zero proves locally uniform convergence.
For the output modulus, write h = ∣ x − x ′ ∣ ≥ Δ h = |x-x'| \ge\Delta h = ∣ x − x ′ ∣ ≥ Δ and split at radius 2 h 2h 2 h . The near part is bounded by C h β ( 1 + ∣ log h ∣ ) Ch^\beta(1+|\log h|) C h β ( 1 + ∣ log h ∣ ) ; on the far part the kernel difference is bounded by C h / ∣ x − w ∣ 2 Ch/|x-w|^2 C h /∣ x − w ∣ 2 , and
C h ∫ 2 h 1 r β − 2 d r = O β ( h β ) . Ch\int_{2h}^{1} r^{\beta-2}\,dr = O_\beta(h^\beta). C h ∫ 2 h 1 r β − 2 d r = O β ( h β ) .
Thus every exponent 0 < γ < β 0 < \gamma< \beta 0 < γ < β gives a common C γ C^\gamma C γ bound. For h < Δ h < \Delta h < Δ , linear interpolation gives the same estimate from the bound at adjacent grid sites.
Away from the Cauchy pole. For the exit term, the input is separated from W 0 W_0 W 0 by some d > 0 d > 0 d > 0 . Its Cauchy row g q ( x ) g_q(x) g q ( x ) and output derivative satisfy
∑ ∣ x − w k ∣ ≥ d Δ ∣ x − w k ∣ 2 ≤ C ρ d , ∑ ∣ x − w k ∣ ≥ d Δ ∣ x − w k ∣ 4 ≤ C ρ d 3 . (119) \sum_{|x-w_k|\ge d}\frac{\Delta}{|x-w_k|^2}\le\frac{C_\rho}{d},\qquad
\sum_{|x-w_k|\ge d}\frac{\Delta}{|x-w_k|^4}\le\frac{C_\rho}{d^3}. \tag*{(119)} ∣ x − w k ∣ ≥ d ∑ ∣ x − w k ∣ 2 Δ ≤ d C ρ , ∣ x − w k ∣ ≥ d ∑ ∣ x − w k ∣ 4 Δ ≤ d 3 C ρ . (119)
Consequently g q g_q g q is a convergent uniformly Lipschitz L 2 L^2 L 2 -valued row field. Its product with K q K_q K q is handled by strong adjoint convergence of K q K_q K q , uniformly on the compact range of g q g_q g q . No regularity of the rows of K q K_q K q outside W W W is required. This proves the same conclusion for the exit term.
The left coefficients, the right product, and g g g are uniformly bounded on Γ × { ∣ t ∣ ≤ τ } \Gamma\times\{|t| \le\tau\} Γ × { ∣ t ∣ ≤ τ } . The integrand in part 2 therefore has regular-row norms and C γ C^\gamma C γ constants at most C ρ C_\rho C ρ on this compact parameter set. There is only one Cauchy application, so the neighborhood W 0 W_0 W 0 and exponent γ \gamma γ do not depend on a Taylor order.
Integrate and extract coefficients. The contour Γ \Gamma Γ has finite length, and the rows converge locally uniformly jointly in the output coordinate, ζ ∈ Γ \zeta\in\Gamma ζ ∈ Γ , and ∣ t ∣ ≤ τ |t| \le\tau ∣ t ∣ ≤ τ . Integration over this fixed compact contour therefore preserves locally uniform convergence and the common Hölder bound. The difference in part 2 has regular rows throughout the same t t t -disk.
For each fixed grid, row extraction followed by its step-function embedding is a bounded complex-linear map into L 2 L^2 L 2 , with norm of order Δ − 1 / 2 \Delta^{-1/2} Δ − 1/2 . The norm-convergent inverse series and their integral therefore give L 2 L^2 L 2 -valued holomorphic rows. The locally uniform limit satisfies the same Cauchy integral formula in t t t , hence is holomorphic. In particular, for any 0 < ρ < τ 0<\rho<\tau 0 < ρ < τ , its n n n th Taylor coefficient is obtained from
1 2 π i ∫ ∣ t ∣ = ρ X q ( t ) − X q p r ( t ) t n + 1 d t \frac{1}{2\pi i}\int_{\lvert t\rvert=\rho}\frac{X_q(t)-X_q^{\mathrm{pr}}(t)}{t^{n+1}}\,\mathrm{d}t 2 π i 1 ∫ ∣ t ∣ = ρ t n + 1 X q ( t ) − X q pr ( t ) d t
Locally uniform row convergence and the common C γ C^\gamma C γ bound pass through this integral. The neighborhood and exponent are the same for all orders; the coefficient bounds may grow with the order through the factor ρ − n \rho^{-n} ρ − n . This proves the coefficient assertion and the lemma. □
Conclusion for the block used at the cut. The inverse-square-root series converges in a common small disk, uniformly in q q q within the chosen residue class. Hence the exact coefficient identities (A.15) also give the analytic even-block identity
( X q p r ( t ) + X q p r ( − t ) 2 ) 00 = α p , q ( t ) I , α p , q ( t ) = ∑ m ≥ 0 α m , p , q t 2 m . \left(\frac{X_q^{\mathrm{pr}}(t)+X_q^{\mathrm{pr}}(-t)}{2}\right)^{00}=\alpha_{p,q}(t)I,\qquad\alpha_{p,q}(t)=\sum_{m\geq0}\alpha_{m,p,q}t^{2m}. ( 2 X q pr ( t ) + X q pr ( − t ) ) 00 = α p , q ( t ) I , α p , q ( t ) = m ≥ 0 ∑ α m , p , q t 2 m .
Here 00 00 00 denotes the whole block from subinterval zero to itself, retaining both index parities. On the fixed residue class the coefficients α m , p , q \alpha_{m,p,q} α m , p , q are the same finite reciprocal contacts. Taking the even part in Lemma A.3 and restricting its input to subinterval zero proves regular rows for the true even block minus α p , q ( t ) I \alpha_{p,q}(t)I α p , q ( t ) I .
In the coefficient notation used below, write X 2 m , q 00 X_{2m,q}^{00} X 2 m , q 00 for the finite coefficient block with both indices in subinterval zero, including both index parities. Its regularized rows
X 2 m , q 00 − α m , p , q I X_{2m,q}^{00}-\alpha_{m,p,q}I X 2 m , q 00 − α m , p , q I
converge in L 2 L^2 L 2 locally uniformly at the subinterval center. The identity here is a grid identity, or its projection after embedding. The true full-grid pencil restricts to the conjugated finite pencil on its selected finite sites; on the complementary sites B q = 0 B_q=0 B q = 0 and its inverse square root is the identity. Thus its rows sampled at the cut are exactly the conjugated finite rows needed below. No scalar identity is claimed for other input-subinterval blocks of the principal operator. The regular global difference is restricted only after it has been estimated.
The Fourier calculation produces two one-sided transforms u , v u,v u , v . Their sum c = u + v c=u+v c = u + v has a simple formula, whereas their difference d = u − v d=u-v d = u − v does not. We show that the paired expression needed for the limit is independent of d d d . This is an exact finite-matrix identity, valid before integration and at every Taylor coefficient.
Lemma A.4 (Scalar-shift cancellation). Let J J J reverse p p p coordinates. Let M 0 , V 0 M_0,V_0 M 0 , V 0 be upper triangular matrices with first diagonal entry c / 2 c/2 c /2 , satisfying
J M 0 T J = M 0 , J V 0 T J = V 0 . JM_0^{T}J=M_0,\qquad JV_0^{T}J=V_0. J M 0 T J = M 0 , J V 0 T J = V 0 . For a scalar d d d , put
M = M 0 + d 2 I , V = V 0 − d 2 I , u = c + d 2 , v = c − d 2 , M=M_0+\frac{d}{2}I,\qquad V=V_0-\frac{d}{2}I,\qquad u=\frac{c+d}{2},\qquad v=\frac{c-d}{2}, M = M 0 + 2 d I , V = V 0 − 2 d I , u = 2 c + d , v = 2 c − d , and define
Q ( t ) = ( I − t 2 M 2 t ( M J − J V ) t ( V J − J M ) I − t 2 V 2 ) , X ( t ) = Q ( t ) − 1 / 2 , X ( 0 ) = I . Q(t)=\begin{pmatrix}I-t^2M^2&t(MJ-JV)\\ t(VJ-JM)&I-t^2V^2\end{pmatrix},\qquad X(t)=Q(t)^{-1/2},\qquad X(0)=I. Q ( t ) = ( I − t 2 M 2 t ( V J − J M ) t ( M J − J V ) I − t 2 V 2 ) , X ( t ) = Q ( t ) − 1/2 , X ( 0 ) = I . Then
I ( t ) = u X 0 , 0 ( t ) + v X p , p ( t ) (120) \mathcal{I}(t)=uX_{0,0}(t)+vX_{p,p}(t) \tag*{(120)} I ( t ) = u X 0 , 0 ( t ) + v X p , p ( t ) (120) is independent of d d d near t = 0 t=0 t = 0 . Thus u , v u,v u , v can both be replaced by c / 2 c/2 c /2 in this paired expression before extracting any coefficient.
Proof. The reason is that d d d becomes a scalar shift of a simpler matrix. Introduce
L d ( t ) = ( − i t M J J i t V ) = L 0 ( t ) − i t d 2 I . (121) L_d(t)=\begin{pmatrix} -itM & J \\ J & itV \end{pmatrix}=L_0(t)-\frac{itd}{2}I. \tag*{(121)} L d ( t ) = ( − i tM J J i t V ) = L 0 ( t ) − 2 i t d I . (121)
Here L 0 L_0 L 0 means the same construction with d = 0 d=0 d = 0 . The relation with Q Q Q uses only a diagonal change of basis:
L d ( t ) 2 = ( I − t 2 M 2 − i t ( M J − J V ) i t ( V J − J M ) I − t 2 V 2 ) , T = diag ( I p , i I p ) , L_d(t)^2=\begin{pmatrix} I-t^2M^2 & -it(MJ-JV) \\ it(VJ-JM) & I-t^2V^2 \end{pmatrix}, \qquad T=\operatorname{diag}(I_p,iI_p), L d ( t ) 2 = ( I − t 2 M 2 i t ( V J − J M ) − i t ( M J − J V ) I − t 2 V 2 ) , T = diag ( I p , i I p ) ,
Q ( t ) = T − 1 L d ( t ) 2 T , X ( t ) = T − 1 Y d ( t ) T , Y d ( t ) = ( L d ( t ) 2 ) − 1 / 2 . Q(t)=T^{-1}L_d(t)^2T,\qquad X(t)=T^{-1}Y_d(t)T,\qquad Y_d(t)=\left(L_d(t)^2\right)^{-1/2}. Q ( t ) = T − 1 L d ( t ) 2 T , X ( t ) = T − 1 Y d ( t ) T , Y d ( t ) = ( L d ( t ) 2 ) − 1/2 .
The inverse square roots are fixed by their value I I I at zero; conjugating their convergent binomial series proves the last equality. In particular, X X X and Y d Y_d Y d have the same two diagonal blocks. It therefore suffices to prove independence of u ( Y d ) 0 , 0 + v ( Y d ) p , p u(Y_d)_{0,0}+v(Y_d)_{p,p} u ( Y d ) 0 , 0 + v ( Y d ) p , p .
At t = 0 t=0 t = 0 , L d ( 0 ) = ( 0 J J 0 ) L_d(0)=\begin{pmatrix} 0 & J \\ J & 0 \end{pmatrix} L d ( 0 ) = ( 0 J J 0 ) has its spectrum at + 1 , − 1 +1,-1 + 1 , − 1 . For fixed d d d , shrink ∣ t ∣ |t| ∣ t ∣ so that the spectra of both L 0 ( t ) L_0(t) L 0 ( t ) and L d ( t ) L_d(t) L d ( t ) lie strictly inside the two circles Γ + , Γ − \Gamma_+,\Gamma_- Γ + , Γ − defined above. Their projectors are
P ± , d ( t ) = 1 2 π i ∫ Γ ± ( ζ I − L d ( t ) ) − 1 d ζ . P_{\pm,d}(t)=\frac{1}{2\pi i}\int_{\Gamma_\pm}(\zeta I-L_d(t))^{-1}\,d\zeta. P ± , d ( t ) = 2 π i 1 ∫ Γ ± ( ζ I − L d ( t ) ) − 1 d ζ .
Since L d = L 0 − i t d I / 2 L_d=L_0-itdI/2 L d = L 0 − i t d I /2 , the change of variable ω = ζ + i t d / 2 \omega=\zeta+itd/2 ω = ζ + i t d /2 replaces this integral by the resolvent integral for L 0 L_0 L 0 on the translated circle Γ ± + i t d / 2 \Gamma_\pm+itd/2 Γ ± + i t d /2 . For sufficiently small ∣ t ∣ |t| ∣ t ∣ , the translated circle deforms back to Γ ± \Gamma_\pm Γ ± without crossing the spectrum of L 0 L_0 L 0 . Thus P ± , d = P ± , 0 P_{\pm,d}=P_{\pm,0} P ± , d = P ± , 0 . The analytic sign is the difference of these projectors, so
S d ( t ) : = Y d ( t ) L d ( t ) = P + , d ( t ) − P − , d ( t ) = S 0 ( t ) . (122) S_d(t):=Y_d(t)L_d(t)=P_{+,d}(t)-P_{-,d}(t)=S_0(t). \tag*{(122)} S d ( t ) := Y d ( t ) L d ( t ) = P + , d ( t ) − P − , d ( t ) = S 0 ( t ) . (122)
This argument requires neither simple eigenvalues nor diagonalizability.
We recover the desired weighted entries from S d S_d S d . For K = diag ( J , J ) K=\operatorname{diag}(J,J) K = diag ( J , J ) , the reversal identities imply L d T = K L d K L_d^T=KL_dK L d T = K L d K . The same identity holds for Y d Y_d Y d , term by term in its binomial series. Comparing entries gives
( Y d ) 0 , 2 p − 1 = ( Y d ) p , p − 1 . (123) (Y_d)_{0,2p-1}=(Y_d)_{p,p-1}. \tag*{(123)} ( Y d ) 0 , 2 p − 1 = ( Y d ) p , p − 1 . (123)
The first columns of the upper triangular matrices M , V M,V M , V contain only their first diagonal entries, namely u , v u,v u , v . The ( 0 , 0 ) (0,0) ( 0 , 0 ) and ( p , p ) (p,p) ( p , p ) entries of S d = Y d L d S_d=Y_dL_d S d = Y d L d are therefore
( S d ) 0 , 0 = − i t u ( Y d ) 0 , 0 + ( Y d ) 0 , 2 p − 1 , (S_d)_{0,0}=-itu(Y_d)_{0,0}+(Y_d)_{0,2p-1}, ( S d ) 0 , 0 = − i t u ( Y d ) 0 , 0 + ( Y d ) 0 , 2 p − 1 ,
( S d ) p , p = i t v ( Y d ) p , p + ( Y d ) p , p − 1 . (S_d)_{p,p}=itv(Y_d)_{p,p}+(Y_d)_{p,p-1}. ( S d ) p , p = i t v ( Y d ) p , p + ( Y d ) p , p − 1 .
Subtracting cancels the last terms by (A.27). Using (A.26), we obtain
u ( Y d ) 0 , 0 + v ( Y d ) p , p = i t [ ( S 0 ) 0 , 0 − ( S 0 ) p , p ] . (124) u(Y_d)_{0,0}+v(Y_d)_{p,p}=\frac{i}{t}\left[(S_0)_{0,0}-(S_0)_{p,p}\right]. \tag*{(124)} u ( Y d ) 0 , 0 + v ( Y d ) p , p = t i [ ( S 0 ) 0 , 0 − ( S 0 ) p , p ] . (124)
The quotient has a removable singularity at zero, as its equality with the analytic left side shows. Its right side is independent of d d d , which proves the lemma. The conclusion holds at every coefficient; each coefficient is a polynomial in d d d . □
Application to the limiting matrices. Take M 0 = M ( c , h ) M_0=M(c,h) M 0 = M ( c , h ) and V 0 = M ( c , − h ) V_0=M(c,-h) V 0 = M ( c , − h ) from the explicit matrix pattern in § 5 , and put d = u − v d=u-v d = u − v . The matrices in (A.11) satisfy
M = M 0 + d 2 I p , V = V 0 − d 2 I p . M=M_0+\frac{d}{2}I_p,\qquad V=V_0-\frac{d}{2}I_p. M = M 0 + 2 d I p , V = V 0 − 2 d I p .
Their off-diagonal entries are independent of d d d . The relation κ p − ℓ = − κ ℓ \kappa_{p-\ell}=-\kappa_\ell κ p − ℓ = − κ ℓ gives a p − ℓ ∓ = a ℓ ± a^\mp_{p-\ell}=a^\pm_\ell a p − ℓ ∓ = a ℓ ± in the displayed pattern. It is exactly the equality of opposite-parity entries reflected across the anti-diagonal. The alternating c c c entries and the diagonal have the same symmetry because p − 1 p-1 p − 1 is even. Thus J M 0 T J = M 0 JM_0^TJ=M_0 J M 0 T J = M 0 and J V 0 T J = V 0 JV_0^TJ=V_0 J V 0 T J = V 0 , and the lemma applies at each frequency. It removes u − v u-v u − v pointwise from the paired expression; only the explicit sum c = u + v c=u+v c = u + v remains. This algebraic cancellation does not require h 2 + 4 c 2 = 1 h^2+4c^2=1 h 2 + 4 c 2 = 1 or a higher cotangent identity.
The cut limit, with all normalization factors
We can now take the limit in the original finite quantity. In this subsection B q B_q B q and X 2 m , q X_{2m,q} X 2 m , q again denote the original finite matrices. For any such matrix A A A , write A ′ = G q A G q − 1 A' = G_q A G_q^{-1} A ′ = G q A G q − 1 ; all row and column limits below concern these conjugated forms. The regular-row lemma justifies the row–column pairing; the Fourier calculation evaluates that same pairing. Keeping these two roles separate is useful: a Fourier formula alone would not justify sampling an unregularized diagonal.
Fix m ≥ 1 m \ge1 m ≥ 1 and let q q q tend to infinity in the chosen denominator class. Choose cut indices j q / q → 1 / ( 2 p ) j_q/q \to1/(2p) j q / q → 1/ ( 2 p ) , in the middle of subinterval zero. The exact position-independent matrix expression is
( X 2 m , q ) 11 = q h j q { ( X 2 m , q B q ) j q , j q + 1 − ( X 2 m , q B q ) j q + 1 , j q } , h j = 2 sin ( π p j / q ) . (125) (X_{2m,q})_{11}=qh_{j_q}\{(X_{2m,q}B_q)_{j_q,j_q+1}-(X_{2m,q}B_q)_{j_q+1,j_q}\},\qquad h_j=2\sin(\pi pj/q). \tag*{(125)} ( X 2 m , q ) 11 = q h j q {( X 2 m , q B q ) j q , j q + 1 − ( X 2 m , q B q ) j q + 1 , j q } , h j = 2 sin ( π p j / q ) . (125)
Either B q ′ B'_q B q ′ column at this cut is supported to its left, entirely in subinterval zero. On that support its sine argument lies near [ π / 4 , π / 2 ] [\pi/4,\pi/2] [ π /4 , π /2 ] ; hence it has uniformly bounded normalized L 2 L^2 L 2 columns and converges to the one-sided continuum column. The squared L 2 L^2 L 2 norm of the difference between columns at cuts z , z ′ z,z' z , z ′ is O ( ∣ z − z ′ ∣ ) O(|z-z'|) O ( ∣ z − z ′ ∣ ) : the smooth kernel has bounded derivatives, and the moving jump affects an interval of that length. Cauchy–Schwarz and the regularized block for the first subinterval rows prove convergence of (A.29) using the regularized primed rows ( X 2 m , q ′ ) 00 − α m , p , q I (X'_{2m,q})_{00}-\alpha_{m,p,q}I ( X 2 m , q ′ ) 00 − α m , p , q I : the paired columns have no support in any other input subinterval. The subtracted identity contributes exactly α m , p , q \alpha_{m,p,q} α m , p , q , because q h j ( B q ) j , j + 1 = 1 qh_j(B_q)_{j,j+1}=1 q h j ( B q ) j , j + 1 = 1 and ( B q ) j + 1 , j = 0 (B_q)_{j+1,j}=0 ( B q ) j + 1 , j = 0 .
The exact subtraction and the continuous pairing
For the remainder of this calculation write α = α m , p , q \alpha=\alpha_{m,p,q} α = α m , p , q ; it is constant along the chosen denominator class. Put C m , q = ( X 2 m , q − α I ) B q C_{m,q}=(X_{2m,q}-\alpha I)B_q C m , q = ( X 2 m , q − α I ) B q . Then C m , q ′ = ( X 2 m , q ′ − α I ) B q ′ C'_{m,q}=(X'_{2m,q}-\alpha I)B'_q C m , q ′ = ( X 2 m , q ′ − α I ) B q ′ . The exact subtraction and the two conjugating phases give
( X 2 m , q ) 11 − α = i q h j q { ( C m , q ′ ) j q , j q + 1 + ( C m , q ′ ) j q + 1 , j q } . (X_{2m,q})_{11}-\alpha=iqh_{j_q}\{(C'_{m,q})_{j_q,j_q+1}+(C'_{m,q})_{j_q+1,j_q}\}. ( X 2 m , q ) 11 − α = i q h j q {( C m , q ′ ) j q , j q + 1 + ( C m , q ′ ) j q + 1 , j q } .
Let K m l i n ( z , w ) \mathcal{K}^{\mathrm{lin}}_m(z,w) K m lin ( z , w ) denote the limiting 2 × 2 2\times2 2 × 2 kernel of this product in subinterval zero, with the two indices recording original index parity. We use only the block from subinterval zero to itself, where the scalar principal subtraction has been proved. The normalization is matrix entry divided by Δ \Delta Δ . This kernel is continuous near ( 0 , 0 ) (0,0) ( 0 , 0 ) : the regularized rows converge in L 2 L^2 L 2 , the cut columns converge in L 2 L^2 L 2 , and their pairings are continuous in both coordinates. Only this regularized product is evaluated on the diagonal.
The integrable Fourier multiplier of the same pairing
Under the unitary half-density change of variables, the same product has kernel
K m c m ( τ , σ ) = y ′ ( τ ) y ′ ( σ ) K m l i n ( z ( τ ) , z ( σ ) ) , K m c m ( 0 , 0 ) = 2 π K m l i n ( 0 , 0 ) . \mathcal{K}^{\mathrm{cm}}_m(\tau,\sigma)=\sqrt{y'(\tau)y'(\sigma)}\,\mathcal{K}^{\mathrm{lin}}_m(z(\tau),z(\sigma)),\qquad\mathcal{K}^{\mathrm{cm}}_m(0,0)=\frac{2}{\pi}\mathcal{K}^{\mathrm{lin}}_m(0,0). K m cm ( τ , σ ) = y ′ ( τ ) y ′ ( σ ) K m lin ( z ( τ ) , z ( σ )) , K m cm ( 0 , 0 ) = π 2 K m lin ( 0 , 0 ) .
Here z ( τ ) = y ( τ ) − 1 / 2 z(\tau)=y(\tau)-1/2 z ( τ ) = y ( τ ) − 1/2 , so the chosen cut is z = 0 z=0 z = 0 , or τ = 0 \tau=0 τ = 0 . After this coordinate change the limiting product is a convolution operator. A hat denotes its Fourier multiplier, with inverse Fourier factor ( 2 π ) − 1 (2\pi)^{-1} ( 2 π ) − 1 . The parity and frequency calculation above gives, for the sampled block, the two multiplier factors
{ p − 2 m ( D 2 m ( ξ ) ) 00 − α } I 2 , − i p u ( ξ ) σ x . \{p^{-2m}(D_{2m}(\xi))_{00}-\alpha\}I_2,\qquad-\frac{i}{p}u(\xi)\sigma_x. { p − 2 m ( D 2 m ( ξ ) ) 00 − α } I 2 , − p i u ( ξ ) σ x .
Consequently the Fourier multiplier of the product kernel is
K ^ m c m ( ξ ) = − i p 2 m + 1 σ x b m ( ξ ) , b m ( ξ ) = ( D 2 m ( ξ ) M ( ξ ) ) 00 − p 2 m α u ( ξ ) . \widehat{\mathcal{K}}^{\mathrm{cm}}_m(\xi)=-\frac{i}{p^{2m+1}}\sigma_x b_m(\xi),\qquad b_m(\xi)=(D_{2m}(\xi)M(\xi))_{00}-p^{2m}\alpha u(\xi). K m cm ( ξ ) = − p 2 m + 1 i σ x b m ( ξ ) , b m ( ξ ) = ( D 2 m ( ξ ) M ( ξ ) ) 00 − p 2 m α u ( ξ ) .
The equality uses the upper triangular form of M M M : ( D 2 m M ) 00 = ( D 2 m ) 00 u (D_{2m}M)_{00}=(D_{2m})_{00}u ( D 2 m M ) 00 = ( D 2 m ) 00 u . Moreover b m b_m b m is integrable. Integration by parts in the half-sech integral gives u = O ( ( 1 + ∣ ξ ∣ ) − 1 ) u=O((1+|\xi|)^{-1}) u = O (( 1 + ∣ ξ ∣ ) − 1 ) . The exact principal contact (A.15) gives ( D 2 m ) 00 − p 2 m α = O ( ( 1 + ∣ ξ ∣ ) − 1 ) (D_{2m})_{00}-p^{2m}\alpha=O((1+|\xi|)^{-1}) ( D 2 m ) 00 − p 2 m α = O (( 1 + ∣ ξ ∣ ) − 1 ) at both ends; this also follows by inserting the principal limiting values in the fixed coefficient polynomial of (A.12). Thus b m = O ( ( 1 + ∣ ξ ∣ ) − 2 ) b_m=O((1+|\xi|)^{-2}) b m = O (( 1 + ∣ ξ ∣ ) − 2 ) . Fourier inversion therefore gives a continuous kernel for the same limiting operator. It agrees almost everywhere with the kernel defined by row–column pairing. Continuity of both kernels near the cut makes them agree at the cut as well.
The discrete, coordinate, parity, and Fourier factors
Write tr 2 \operatorname{tr}_2 tr 2 for the trace over the two original index parities; it is not a trace over the paired frequencies. Using q Δ = 2 p q\Delta= 2p q Δ = 2 p , h j q → 2 h_{j_q} \to2 h j q → 2 , y ′ ( 0 ) = 2 / π y'(0) = 2/\pi y ′ ( 0 ) = 2/ π , and the Fourier convention already fixed, the complete calculation is
lim q → ∞ { ( X 2 m , q ) 11 − α } = lim q → ∞ i ( q Δ ) h j q { ( C m , q ′ ) j q , j q + 1 Δ + ( C m , q ′ ) j q + 1 , j q Δ } = ( 2 p ) ( 2 i ) tr 2 { σ x K m l i n ( 0 , 0 ) } = ( 2 p ) ( 2 i ) 2 / π tr 2 { σ x K m m c ( 0 , 0 ) } = ( 2 p ) ( 2 i ) 2 / π 1 2 π ∫ R tr 2 { σ x K ^ m ( ξ ) } d ξ = ( 2 p ) ( 2 i ) 2 / π − i p 2 m + 1 tr 2 I 2 2 π ∫ R b m ( ξ ) d ξ = 2 p 2 m ∫ R b m ( ξ ) d ξ . \lim_{q\to\infty}\{(X_{2m,q})_{11}-\alpha\}
= \lim_{q\to\infty} i(q\Delta)h_{j_q}\left\{\frac{(C'_{m,q})_{j_q,j_q+1}}{\Delta}+\frac{(C'_{m,q})_{j_q+1,j_q}}{\Delta}\right\}
= (2p)(2i)\operatorname{tr}_2\{\sigma_x K_m^{\mathrm{lin}}(0,0)\}
= \frac{(2p)(2i)}{2/\pi}\operatorname{tr}_2\{\sigma_x K_m^{\mathrm{mc}}(0,0)\}
= \frac{(2p)(2i)}{2/\pi}\frac{1}{2\pi}\int_{\mathbb{R}}\operatorname{tr}_2\{\sigma_x\widehat{K}_m(\xi)\}\,d\xi
= \frac{(2p)(2i)}{2/\pi}\frac{-i}{p^{2m+1}}\frac{\operatorname{tr}_2 I_2}{2\pi}\int_{\mathbb{R}}b_m(\xi)\,d\xi
= \frac{2}{p^{2m}}\int_{\mathbb{R}}b_m(\xi)\,d\xi. q → ∞ lim {( X 2 m , q ) 11 − α } = q → ∞ lim i ( q Δ ) h j q { Δ ( C m , q ′ ) j q , j q + 1 + Δ ( C m , q ′ ) j q + 1 , j q } = ( 2 p ) ( 2 i ) tr 2 { σ x K m lin ( 0 , 0 )} = 2/ π ( 2 p ) ( 2 i ) tr 2 { σ x K m mc ( 0 , 0 )} = 2/ π ( 2 p ) ( 2 i ) 2 π 1 ∫ R tr 2 { σ x K m ( ξ )} d ξ = 2/ π ( 2 p ) ( 2 i ) p 2 m + 1 − i 2 π tr 2 I 2 ∫ R b m ( ξ ) d ξ = p 2 m 2 ∫ R b m ( ξ ) d ξ .
The second line combines the grid spacing and the limiting bond coefficient; the third is the half-density Jacobian; and the last two lines exhibit the multiplier, parity trace tr 2 I 2 = 2 \operatorname{tr}_2 I_2=2 tr 2 I 2 = 2 , and Fourier factor ( 2 π ) − 1 (2\pi)^{-1} ( 2 π ) − 1 together. Adding back α \alpha α gives
α m , p , q + 2 p 2 m ∫ R { ( D 2 m M ) 00 − p 2 m α m , p , q u } d ξ . (126) \alpha_{m,p,q}+\frac{2}{p^{2m}}\int_{\mathbb{R}}\{(D_{2m}M)_{00}-p^{2m}\alpha_{m,p,q}u\}\,d\xi. \tag*{(126)} α m , p , q + p 2 m 2 ∫ R {( D 2 m M ) 00 − p 2 m α m , p , q u } d ξ . (126)
Write f m ( ξ ) = ( D 2 m ( ξ ) M ( ξ ) ) 00 f_m(\xi)=(D_{2m}(\xi)M(\xi))_{00} f m ( ξ ) = ( D 2 m ( ξ ) M ( ξ ) ) 00 and symmetrize the absolutely convergent regularized integral. Since u ( ξ ) + u ( − ξ ) = c ( ξ ) u(\xi)+u(-\xi)=c(\xi) u ( ξ ) + u ( − ξ ) = c ( ξ ) and ∫ R c ( ξ ) d ξ = 1 \int_{\mathbb{R}}c(\xi)\,d\xi=1 ∫ R c ( ξ ) d ξ = 1 , its value is
α m , p , q + 1 p 2 m ∫ R { f m ( ξ ) + f m ( − ξ ) − p 2 m α m , p , q c ( ξ ) } d ξ = 1 p 2 m ∫ R { f m ( ξ ) + f m ( − ξ ) } d ξ . \alpha_{m,p,q}+\frac{1}{p^{2m}}\int_{\mathbb{R}}\{f_m(\xi)+f_m(-\xi)-p^{2m}\alpha_{m,p,q}c(\xi)\}\,d\xi
=\frac{1}{p^{2m}}\int_{\mathbb{R}}\{f_m(\xi)+f_m(-\xi)\}\,d\xi. α m , p , q + p 2 m 1 ∫ R { f m ( ξ ) + f m ( − ξ ) − p 2 m α m , p , q c ( ξ )} d ξ = p 2 m 1 ∫ R { f m ( ξ ) + f m ( − ξ )} d ξ .
This shows explicitly how the exact principal contact disappears, and proves (A.14), including m = 0 m=0 m = 0 .
The scalar-shift cancellation and the angular integral
Since V ( ξ ) = M ( − ξ ) V(\xi)=M(-\xi) V ( ξ ) = M ( − ξ ) , replacing ξ \xi ξ by − ξ -\xi − ξ exchanges the two frequency blocks. Lemma A.4 therefore gives
f m ( ξ ) + f m ( − ξ ) = [ s 2 m ] I ( s ; c ( ξ ) , h ( ξ ) ) . f_m(\xi)+f_m(-\xi)=[s^{2m}]\mathcal{I}(s;c(\xi),h(\xi)). f m ( ξ ) + f m ( − ξ ) = [ s 2 m ] I ( s ; c ( ξ ) , h ( ξ )) .
The factor p − 2 m p^{-2m} p − 2 m is absorbed by replacing s s s by t / p t/p t / p . Finally set h ( ξ ) = tanh ( π ξ / 2 ) = sin ψ h(\xi)=\tanh(\pi\xi/2)=\sin\psi h ( ξ ) = tanh ( π ξ /2 ) = sin ψ . Then c ( ξ ) = 1 2 cos ψ c(\xi)=\frac{1}{2}\cos\psi c ( ξ ) = 2 1 cos ψ and d ξ = d ψ / ( π c ( ξ ) ) d\xi=d\psi/(\pi c(\xi)) d ξ = d ψ / ( π c ( ξ )) , so the last steps are
lim q → ∞ ( X 2 m , q ) 11 = 1 p 2 m ∫ R { f m ( ξ ) + f m ( − ξ ) } d ξ = ∫ R [ t 2 m ] I ( t / p ; c ( ξ ) , h ( ξ ) ) d ξ = 1 π ∫ − π / 2 π / 2 [ t 2 m ] I ( t / p ; 1 2 cos ψ , sin ψ ) 1 2 cos ψ d ψ = [ t 2 m ] G p ( t ) . \lim_{q\to\infty}(X_{2m,q})_{11}
=\frac{1}{p^{2m}}\int_{\mathbb{R}}\{f_m(\xi)+f_m(-\xi)\}\,d\xi
=\int_{\mathbb{R}}[t^{2m}]\mathcal{I}(t/p;c(\xi),h(\xi))\,d\xi
=\frac{1}{\pi}\int_{-\pi/2}^{\pi/2}[t^{2m}]\frac{\mathcal{I}(t/p;\frac{1}{2}\cos\psi,\sin\psi)}{\frac{1}{2}\cos\psi}\,d\psi
=[t^{2m}]G_p(t). q → ∞ lim ( X 2 m , q ) 11 = p 2 m 1 ∫ R { f m ( ξ ) + f m ( − ξ )} d ξ = ∫ R [ t 2 m ] I ( t / p ; c ( ξ ) , h ( ξ )) d ξ = π 1 ∫ − π /2 π /2 [ t 2 m ] 2 1 cos ψ I ( t / p ; 2 1 cos ψ , sin ψ ) d ψ = [ t 2 m ] G p ( t ) .
At cos ψ = 0 \cos\psi=0 cos ψ = 0 , the quotient is the half-sum of the two selected entries of Y 0 Y_0 Y 0 , as in (A.13). Those entries are analytic in a common disk in t t t , uniformly on the closed angular interval. Their Cauchy coefficient formula therefore justifies taking the coefficient through the integral, including its endpoints.
This proves Proposition 5.1 on the chosen residue class, without evaluating G p G_p G p .
The nonsingular case p = 1 p=1 p = 1 and a normalization check. For p = 1 p=1 p = 1 there are no opposite-parity subinterval pairs and hence no Cauchy principal blocks. Use the zero principal operator, whose coefficient constant is α m , 1 = 1 m = 0 \alpha_{m,1}=1_{m=0} α m , 1 = 1 m = 0 . All remaining bounds, row arguments, and the matrix calculation apply without change. This proves the case p = 1 p=1 p = 1 as well; no denominator-one reciprocal contact is needed. As a check on the normalization, for p = 1 p=1 p = 1 the scalar matrix gives [ s 2 ] J ( s ; c , h ) = c 3 / 8 [s^2]\mathcal{J}(s;c,h)=c^3/8 [ s 2 ] J ( s ; c , h ) = c 3 /8 . Since ∫ R c ( ξ ) 3 d ξ = 1 / 8 \int_{\mathbb{R}}c(\xi)^3\,\mathrm{d}\xi=1/8 ∫ R c ( ξ ) 3 d ξ = 1/8 , the displayed Fourier formula yields
lim q → ∞ q ρ 1 , q ′ ′ ( 0 ) 2 ! = 1 64 , lim q → ∞ q ρ 1 , q ′ ′ ( 0 ) = 1 32 . \lim_{q\to\infty}\frac{q\rho_{1,q}^{\prime\prime}(0)}{2!}=\frac{1}{64},\qquad\lim_{q\to\infty}q\rho_{1,q}^{\prime\prime}(0)=\frac{1}{32}. q → ∞ lim 2 ! q ρ 1 , q ′′ ( 0 ) = 64 1 , q → ∞ lim q ρ 1 , q ′′ ( 0 ) = 32 1 .
in agreement with Proposition 4.1 . □
Appendix B. Verification of the block identity
We prove
M C e + C e V = c E p − 1 . MC_e+C_eV=cE_{p-1}. M C e + C e V = c E p − 1 .
The useful observation is that C e C_e C e is block diagonal. Each block of the product can therefore be checked independently; there is no sum over intermediate coordinate pairs.
Proof. For p = 1 p=1 p = 1 , the identity is ( c / 2 ) ⋅ 1 + 1 ⋅ ( c / 2 ) = c (c/2)\cdot1+1\cdot(c/2)=c ( c /2 ) ⋅ 1 + 1 ⋅ ( c /2 ) = c . Now let p = 2 n + 1 ≥ 3 p=2n+1\geq3 p = 2 n + 1 ≥ 3 , with θ = q ^ π / p \theta=\widehat{q}\pi/p θ = q π / p , where q ^ \widehat{q} q is odd and coprime to p p p . We take real c , h c,h c , h with h 2 + 4 c 2 = 1 h^2+4c^2=1 h 2 + 4 c 2 = 1 , so V = M ‾ V=\overline{M} V = M ; the bar denotes entrywise complex conjugation.
Group the coordinates as ( 0 , 1 ) , ( 2 , 3 ) , … , ( 2 n − 2 , 2 n − 1 ) (0,1),(2,3),\ldots,(2n-2,2n-1) ( 0 , 1 ) , ( 2 , 3 ) , … , ( 2 n − 2 , 2 n − 1 ) , followed by the singleton 2 n 2n 2 n . Put ϕ j = ( 2 j + 1 ) θ \phi_j=(2j+1)\theta ϕ j = ( 2 j + 1 ) θ , 0 ≤ j < n 0\leq j<n 0 ≤ j < n . The paired blocks of C e C_e C e are the matrices C ( ϕ j ) C(\phi_j) C ( ϕ j ) displayed in § 5 ; its last entry is 1 1 1 . There are three cases.
Within one coordinate pair. The diagonal block of M M M is
D ( ϕ ) = ( c / 2 ( cot ϕ + i h ) / 2 0 − c / 2 ) . D(\phi)=\begin{pmatrix} c/2 & (\cot\phi+ih)/2 \\ 0 & -c/2 \end{pmatrix}. D ( ϕ ) = ( c /2 0 ( cot ϕ + ih ) /2 − c /2 ) .
Multiplication gives D ( ϕ ) C ( ϕ ) + C ( ϕ ) D ( ϕ ) ‾ = 0 D(\phi)C(\phi)+C(\phi)\overline{D(\phi)}=0 D ( ϕ ) C ( ϕ ) + C ( ϕ ) D ( ϕ ) = 0 . For example, its first diagonal entry is
c ( − cos ϕ − i h sin ϕ ) + c sin ϕ ( cot ϕ + i h ) = 0. c(-\cos\phi-ih\sin\phi)+c\sin\phi(\cot\phi+ih)=0. c ( − cos ϕ − ih sin ϕ ) + c sin ϕ ( cot ϕ + ih ) = 0.
The other diagonal entry cancels with the opposite signs; the upper entry is i h ( cos ϕ − cot ϕ sin ϕ ) = 0 ih(\cos\phi-\cot\phi\sin\phi)=0 ih ( cos ϕ − cot ϕ sin ϕ ) = 0 , and the lower entry is zero.
Between two different coordinate pairs. For j < k j<k j < k , put ϕ = ϕ j \phi=\phi_j ϕ = ϕ j , ψ = ϕ k \psi=\phi_k ψ = ϕ k , and x = ( ϕ + ψ ) / 2 = ( j + k + 1 ) θ x=(\phi+\psi)/2=(j+k+1)\theta x = ( ϕ + ψ ) /2 = ( j + k + 1 ) θ . Reading the entries of M M M gives the block ( − 1 ) j + k B ( x ) (-1)^{j+k}B(x) ( − 1 ) j + k B ( x ) , where
B ( x ) = ( c ( cot x + i h ) / 2 ( cot x − i h ) / 2 − c ) . B(x)=\begin{pmatrix} c & (\cot x+ih)/2 \\ (\cot x-ih)/2 & -c \end{pmatrix}. B ( x ) = ( c ( cot x − ih ) /2 ( cot x + ih ) /2 − c ) .
The corresponding block of M C e + C e V MC_e+C_eV M C e + C e V , apart from this common sign, is B ( x ) C ( ψ ) + C ( ϕ ) B ( x ) ‾ B(x)C(\psi)+C(\phi)\overline{B(x)} B ( x ) C ( ψ ) + C ( ϕ ) B ( x ) . Define
X = cot x ( sin ϕ + sin ψ ) − ( cos ϕ + cos ψ ) , X=\cot x(\sin\phi+\sin\psi)-(\cos\phi+\cos\psi), X = cot x ( sin ϕ + sin ψ ) − ( cos ϕ + cos ψ ) ,
Y = cot x ( cos ψ − cos ϕ ) + ( sin ψ − sin ϕ ) . Y=\cot x(\cos\psi-\cos\phi)+(\sin\psi-\sin\phi). Y = cot x ( cos ψ − cos ϕ ) + ( sin ψ − sin ϕ ) .
Using h 2 + 4 c 2 = 1 h^2+4c^2=1 h 2 + 4 c 2 = 1 , direct multiplication yields
B ( x ) C ( ψ ) + C ( ϕ ) B ( x ) ‾ = ( c X ( Y − i h X ) / 2 ( − Y − i h X ) / 2 c X ) . B(x)C(\psi)+C(\phi)\overline{B(x)}
=
\begin{pmatrix}
cX & (Y-ihX)/2 \\
(-Y-ihX)/2 & cX
\end{pmatrix}. B ( x ) C ( ψ ) + C ( ϕ ) B ( x ) = ( c X ( − Y − ih X ) /2 ( Y − ih X ) /2 c X ) .
Both X X X and Y Y Y vanish. Indeed, with d = ( ψ − ϕ ) / 2 d=(\psi-\phi)/2 d = ( ψ − ϕ ) /2 ,
sin ϕ + sin ψ = 2 sin x cos d , cos ϕ + cos ψ = 2 cos x cos d , \sin\phi+\sin\psi=2\sin x\cos d,\qquad
\cos\phi+\cos\psi=2\cos x\cos d, sin ϕ + sin ψ = 2 sin x cos d , cos ϕ + cos ψ = 2 cos x cos d ,
cos ψ − cos ϕ = − 2 sin x sin d , sin ψ − sin ϕ = 2 cos x sin d . \cos\psi-\cos\phi=-2\sin x\sin d,\qquad
\sin\psi-\sin\phi=2\cos x\sin d. cos ψ − cos ϕ = − 2 sin x sin d , sin ψ − sin ϕ = 2 cos x sin d .
These identities cancel all four entries at once. Blocks below the paired diagonal vanish because M , V M,V M , V are block upper triangular.
Between a pair and the final coordinate. The last column of M M M in pair j j j , apart from ( − 1 ) j + n (-1)^{j+n} ( − 1 ) j + n , is
m = ( c ( κ − i h ) / 2 ) , κ = cot ( ( j + n + 1 ) θ ) . m=\begin{pmatrix}c\\(\kappa-ih)/2\end{pmatrix},\qquad\kappa=\cot((j+n+1)\theta). m = ( c ( κ − ih ) /2 ) , κ = cot (( j + n + 1 ) θ ) .
Since ( j + n + 1 ) θ = ( ϕ j + p θ ) / 2 (j+n+1)\theta=(\phi_j+p\theta)/2 ( j + n + 1 ) θ = ( ϕ j + pθ ) /2 and p θ p\theta pθ is an odd multiple of π \pi π , one has κ = − tan ( ϕ j / 2 ) \kappa=-\tan(\phi_j/2) κ = − tan ( ϕ j /2 ) . Using h 2 + 4 c 2 = 1 h^2+4c^2=1 h 2 + 4 c 2 = 1 , the required column is
m + C ( ϕ j ) m ‾ = ( c ( 1 − cos ϕ j + κ sin ϕ j ) { κ ( 1 + cos ϕ j ) + sin ϕ j + i h ( cos ϕ j − 1 − κ sin ϕ j ) } / 2 ) = 0. m+C(\phi_j)\overline{m}=\begin{pmatrix}c(1-\cos\phi_j+\kappa\sin\phi_j)\\\{\kappa(1+\cos\phi_j)+\sin\phi_j+ih(\cos\phi_j-1-\kappa\sin\phi_j)\}/2\end{pmatrix}=0. m + C ( ϕ j ) m = ( c ( 1 − cos ϕ j + κ sin ϕ j ) { κ ( 1 + cos ϕ j ) + sin ϕ j + ih ( cos ϕ j − 1 − κ sin ϕ j )} /2 ) = 0.
The equality follows from the half-angle identities κ sin ϕ j = cos ϕ j − 1 \kappa\sin\phi_j=\cos\phi_j-1 κ sin ϕ j = cos ϕ j − 1 and κ ( 1 + cos ϕ j ) = − sin ϕ j \kappa(1+\cos\phi_j)=-\sin\phi_j κ ( 1 + cos ϕ j ) = − sin ϕ j .
The final diagonal entry is ( c / 2 ) ⋅ 1 + 1 ⋅ ( c / 2 ) = c (c/2)\cdot1+1\cdot(c/2)=c ( c /2 ) ⋅ 1 + 1 ⋅ ( c /2 ) = c . All cotangents above are finite by coprimality; no primality of p p p is needed. We have proved
M C e + C e V = diag ( 0 2 , … , 0 2 , c ) = c E p − 1 , MCe+C_eV=\operatorname{diag}(0_2,\ldots,0_2,c)=cE_{p-1}, M C e + C e V = diag ( 0 2 , … , 0 2 , c ) = c E p − 1 ,
including c = 0 c=0 c = 0 directly. □
Appendix C. Stieltjes positivity for the lower-order bounds
The bandwidth implication in Section 2 uses only the positive-real estimate of Lemma 2.1 . The optional Section 4 uses more information: the Taylor coefficients are moments of a probability measure, and the square has the same positivity property. We give the complete proof here. It is independent of the matrix limit and the period calculation.
The two positive representations
C.1. The two positive representations. We use the following Stieltjes convention: a function F F F is a Stieltjes probability transform on [ 0 , 1 ] [0,1] [ 0 , 1 ] if
F ( z ) = ∫ [ 0 , 1 ] d μ ( x ) 1 − x z , μ ≥ 0 , μ ( [ 0 , 1 ] ) = 1. (127) F(z)=\int_{[0,1]}\frac{d\mu(x)}{1-xz},\qquad\mu\geq0,\qquad\mu([0,1])=1. \tag*{(127)} F ( z ) = ∫ [ 0 , 1 ] 1 − x z d μ ( x ) , μ ≥ 0 , μ ([ 0 , 1 ]) = 1. (127)
The function is holomorphic on C ∖ [ 1 , ∞ ) \mathbb{C}\setminus[1,\infty) C ∖ [ 1 , ∞ ) , and its Taylor coefficients are the moments ∫ 0 1 x m d μ ( x ) \int_0^1 x^m\,d\mu(x) ∫ 0 1 x m d μ ( x ) . These lie between zero and one because μ \mu μ is a probability measure. We will prove this representation for two functions at once.
Theorem C.1 (The density and its square). Let P P P be a monic polynomial with simple real zeros, and suppose that it has no critical value in ( − A , A ) (-A,A) ( − A , A ) , where A > 0 A>0 A > 0 . For ∣ t ∣ < A |t|<A ∣ t ∣ < A , define
ρ ( t ) = ∑ P ( E ) = t 1 ∣ P ′ ( E ) ∣ , C = 1 ρ ( 0 ) , F ( z ) = C 2 ( ρ ( A z ) + ρ ( − A z ) ) . \rho(t)=\sum_{P(E)=t}\frac{1}{|P'(E)|},\qquad C=\frac{1}{\rho(0)},\qquad\mathcal{F}(z)=\frac{C}{2}\bigl(\rho(A\sqrt{z})+\rho(-A\sqrt{z})\bigr). ρ ( t ) = P ( E ) = t ∑ ∣ P ′ ( E ) ∣ 1 , C = ρ ( 0 ) 1 , F ( z ) = 2 C ( ρ ( A z ) + ρ ( − A z ) ) . The sum defining ρ \rho ρ is over the real roots, and its analytic continuation is specified in the proof. Both F \mathcal{F} F and F 2 \mathcal{F}^2 F 2 have the representation (C.1), generally with different measures.
Proof. We explain the signs of the inverse branches first, then obtain the positive measures by Cauchy’s formula.
Inverse branches and their signs. Write P ( E ) = ∏ j = 1 d ( E − λ j ) P(E)=\prod_{j=1}^{d}(E-\lambda_j) P ( E ) = ∏ j = 1 d ( E − λ j ) , with λ 1 < ⋯ < λ d \lambda_1<\cdots<\lambda_d λ 1 < ⋯ < λ d , and put σ j = sgn P ′ ( λ j ) \sigma_j=\operatorname{sgn}P'(\lambda_j) σ j = sgn P ′ ( λ j ) . Rolle’s theorem places one critical point in each interval ( λ j , λ j + 1 ) (\lambda_j,\lambda_{j+1}) ( λ j , λ j + 1 ) . These are all d − 1 d-1 d − 1 critical points and are simple. In particular all critical values are real.
Let E j ( t ) E_j(t) E j ( t ) be the inverse branch with E j ( 0 ) = λ j E_j(0)=\lambda_j E j ( 0 ) = λ j . No critical value lies in ( − A , A ) (-A,A) ( − A , A ) , so all these branches remain real and simple there, and P ′ ( E j ( t ) ) P'(E_j(t)) P ′ ( E j ( t )) keeps the sign σ j \sigma_j σ j . Consequently
ρ ( t ) = ∑ j = 1 d σ j E j ′ ( t ) = ∑ j = 1 d σ j P ′ ( E j ( t ) ) (128) \rho(t)=\sum_{j=1}^{d}\sigma_jE_j'(t)=\sum_{j=1}^{d}\frac{\sigma_j}{P'(E_j(t))} \tag*{(128)} ρ ( t ) = j = 1 ∑ d σ j E j ′ ( t ) = j = 1 ∑ d P ′ ( E j ( t )) σ j (128)
is analytic near that interval and agrees with the stated positive inverse-slope sum. We use (C.2) for analytic continuation; absolute values are used only on the real interval.
Each branch continues throughout the upper parameter half-plane, which is simply connected and contains no critical value. For small ε > 0 \varepsilon> 0 ε > 0 ,
E j ( i ε ) = λ j + i ε P ′ ( λ j ) + O ( ε 2 ) . E_j(i\varepsilon)=\lambda_j+\frac{i\varepsilon}{P'(\lambda_j)}+O(\varepsilon^2). E j ( i ε ) = λ j + P ′ ( λ j ) i ε + O ( ε 2 ) .
Moreover, an inverse image of a nonreal parameter cannot be real. Thus sgn Im E j ( t ) = σ j \operatorname{sgn}\operatorname{Im} E_j(t)=\sigma_j sgn Im E j ( t ) = σ j throughout the upper half-plane. Conjugation gives the continuation in the lower half-plane. Together these continuations define ρ \rho ρ on
C ∖ ( ( − ∞ , − A ] ∪ [ A , ∞ ) ) . \mathbb{C}\setminus\left((-\infty,-A]\cup[A,\infty)\right). C ∖ ( ( − ∞ , − A ] ∪ [ A , ∞ ) ) .
Boundary values on the real axis. Take a real s ≠ 0 s\ne0 s = 0 which is not a critical value. A real root r r r of P ( E ) = s P(E)=s P ( E ) = s satisfies
E ( s + i ε ) = r + i ε P ′ ( r ) + O ( ε 2 ) . E(s+i\varepsilon)=r+\frac{i\varepsilon}{P'(r)}+O(\varepsilon^2). E ( s + i ε ) = r + P ′ ( r ) i ε + O ( ε 2 ) .
The preceding half-plane rule shows that its coefficient in (C.2) has the sign of P ′ ( r ) P'(r) P ′ ( r ) . Its contribution to ρ ( s + i 0 ) \rho(s+i0) ρ ( s + i 0 ) is therefore 1 / ∣ P ′ ( r ) ∣ 1/\lvert P'(r)\rvert 1/ ∣ P ′ ( r )∣ .
The nonreal roots occur in conjugate pairs. Let w w w be the member in the upper E E E -half-plane. The two branches have coefficients + 1 +1 + 1 and − 1 -1 − 1 , so their combined contribution is 2 i Im ( 1 / P ′ ( w ) ) 2i\operatorname{Im}(1/P'(w)) 2 i Im ( 1/ P ′ ( w )) . To determine its sign, set
L ( w ) = P ′ ( w ) P ( w ) = ∑ j = 1 d 1 w − λ j . L(w)=\frac{P'(w)}{P(w)}=\sum_{j=1}^{d}\frac{1}{w-\lambda_j}. L ( w ) = P ( w ) P ′ ( w ) = j = 1 ∑ d w − λ j 1 .
Since P ( w ) = s P(w)=s P ( w ) = s ,
Im 1 P ′ ( w ) = Im ( w ) s ∣ L ( w ) ∣ 2 ∑ j = 1 d 1 ∣ w − λ j ∣ 2 . (129) \operatorname{Im}\frac{1}{P'(w)}
=\frac{\operatorname{Im}(w)}{s\lvert L(w)\rvert^2}\sum_{j=1}^{d}\frac{1}{\lvert w-\lambda_j\rvert^2}. \tag*{(129)} Im P ′ ( w ) 1 = s ∣ L ( w ) ∣ 2 Im ( w ) j = 1 ∑ d ∣ w − λ j ∣ 2 1 . (129)
This has the sign of s s s . Adding the real roots and conjugate pairs proves
Re ρ ( s + i 0 ) ≥ 0 , s Im ρ ( s + i 0 ) ≥ 0. (130) \operatorname{Re}\rho(s+i0)\ge0,\qquad s\operatorname{Im}\rho(s+i0)\ge0. \tag*{(130)} Re ρ ( s + i 0 ) ≥ 0 , s Im ρ ( s + i 0 ) ≥ 0. (130)
Evenness makes F \mathcal{F} F single valued and holomorphic on C ∖ [ 1 , ∞ ) \mathbb{C}\setminus[1,\infty) C ∖ [ 1 , ∞ ) , including at z = 0 z=0 z = 0 . For u > 1 u>1 u > 1 , A u + i 0 A\sqrt{u}+i0 A u + i 0 approaches the positive parameter axis from above, whereas − A u + i 0 -A\sqrt{u}+i0 − A u + i 0 approaches the negative axis from below. By conjugation and (C.4), both terms have nonnegative real and imaginary parts. Hence, away from critical values,
Re F ( u + i 0 ) ≥ 0 , Im F ( u + i 0 ) ≥ 0 , Im F ( u + i 0 ) 2 ≥ 0. (131) \operatorname{Re}\mathcal{F}(u+i0)\ge0,\qquad\operatorname{Im}\mathcal{F}(u+i0)\ge0,\qquad\operatorname{Im}\mathcal{F}(u+i0)^2\ge0. \tag*{(131)} Re F ( u + i 0 ) ≥ 0 , Im F ( u + i 0 ) ≥ 0 , Im F ( u + i 0 ) 2 ≥ 0. (131)
The last inequality is the reason that the square also has a positive representation: its imaginary part is twice the product of the first two nonnegative quantities.
Critical values and positive atoms. Between two consecutive zeros the polynomial has one sign. Its critical point is a local maximum if the critical value is positive, and a local minimum if it is negative. As the level approaches that value from the side containing zero, the two real inverse branches therefore contribute a positive multiple of the inverse square root of the distance to the critical value. This also follows directly by inverting
P ( E ) = s c + 1 2 P ′ ′ ( E c ) ( E − E c ) 2 + O ( ( E − E c ) 3 ) . P(E)=s_c+\frac{1}{2}P''(E_c)(E-E_c)^2+O((E-E_c)^3). P ( E ) = s c + 2 1 P ′′ ( E c ) ( E − E c ) 2 + O (( E − E c ) 3 ) .
Thus at a transformed critical value r = s c 2 / A 2 ≥ 1 r=s_c^2/A^2\ge1 r = s c 2 / A 2 ≥ 1 ,
F ( z ) = γ r ( 1 − z / r ) − 1 / 2 + O ( 1 ) , γ r > 0. \mathcal{F}(z)=\gamma_r(1-z/r)^{-1/2}+O(1),\qquad\gamma_r>0. F ( z ) = γ r ( 1 − z / r ) − 1/2 + O ( 1 ) , γ r > 0.
If several critical values give the same r r r , their positive leading coefficients are added. It follows that
F ( z ) 2 = γ r 2 1 − z / r + O ( ∣ z − r ∣ − 1 / 2 ) . \mathcal{F}(z)^2=\frac{\gamma_r^2}{1-z/r}+O(\lvert z-r\rvert^{-1/2}). F ( z ) 2 = 1 − z / r γ r 2 + O (∣ z − r ∣ − 1/2 ) .
The pole has a positive coefficient. The remaining singularity is integrable along the cut. The function F \mathcal{F} F itself has only integrable inverse-square-root singularities.
Cauchy’s formula. If d ≥ 2 d \ge2 d ≥ 2 , inverse roots at infinity have size O ( ∣ t ∣ 1 / d ) O(|t|^{1/d}) O ( ∣ t ∣ 1/ d ) and inverse derivatives have size O ( ∣ t ∣ − ( d − 1 ) / d ) O(|t|^{-(d-1)/d}) O ( ∣ t ∣ − ( d − 1 ) / d ) . Consequently
F ( z ) = O ( ∣ z ∣ − ( d − 1 ) / ( 2 d ) ) , F ( z ) 2 = O ( ∣ z ∣ − ( d − 1 ) / d ) . \mathcal{F}(z) = O(|z|^{-(d-1)/(2d)}), \qquad\mathcal{F}(z)^2 = O(|z|^{-(d-1)/d}). F ( z ) = O ( ∣ z ∣ − ( d − 1 ) / ( 2 d ) ) , F ( z ) 2 = O ( ∣ z ∣ − ( d − 1 ) / d ) .
For either H = F H=\mathcal{F} H = F or H = F 2 H=\mathcal{F}^2 H = F 2 , subtract its finitely many pole terms and apply Cauchy’s formula around the cut [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) . The large circle contributes zero by this decay. Small circles around the remaining inverse-square-root singularities contribute zero because their integrals are O ( ε 1 / 2 ) O(\varepsilon^{1/2}) O ( ε 1/2 ) . The lower boundary value is the conjugate of the upper one. We obtain
H ( z ) = ∑ r α r 1 − z / r + 1 π ∫ 1 ∞ Im H ( u + i 0 ) u − z d u , α r ≥ 0. (132) H(z)=\sum_r\frac{\alpha_r}{1-z/r}+\frac{1}{\pi}\int_1^\infty\frac{\operatorname{Im}H(u+i0)}{u-z}\,du,\qquad\alpha_r\ge0. \tag*{(132)} H ( z ) = r ∑ 1 − z / r α r + π 1 ∫ 1 ∞ u − z Im H ( u + i 0 ) d u , α r ≥ 0. (132)
For H = F H=\mathcal{F} H = F the sum is empty; for H = F 2 H=\mathcal{F}^2 H = F 2 its coefficients are α r = γ r 2 \alpha_r=\gamma_r^2 α r = γ r 2 . By (C.5) the integral has nonnegative density. Under x = 1 / u x=1/u x = 1/ u , formula (C.6) becomes (C.1), with the positive measure (where δ a \delta_a δ a denotes unit point mass at a a a )
d μ H ( x ) = ∑ r α r d δ 1 / r ( x ) + 1 ( 0 , 1 ) ( x ) Im H ( 1 / x + i 0 ) π x d x . d\mu_H(x)=\sum_r\alpha_r\,d\delta_{1/r}(x)+\mathbf{1}_{(0,1)}(x)\frac{\operatorname{Im}H(1/x+i0)}{\pi x}\,dx. d μ H ( x ) = r ∑ α r d δ 1/ r ( x ) + 1 ( 0 , 1 ) ( x ) π x Im H ( 1/ x + i 0 ) d x .
Its mass is H ( 0 ) = 1 H(0)=1 H ( 0 ) = 1 , so it is a probability measure. Finally, if d = 1 d=1 d = 1 , then F = 1 \mathcal{F}=1 F = 1 and both measures are δ 0 \delta_0 δ 0 .
Moments and an alternative local-convergence argument
Apply Theorem C.1 with P = P p , q P=P_{p,q} P = P p , q , A = 4 A=4 A = 4 , and C = q C=q C = q . The function F p , q F_{p,q} F p , q from (13) and its square have Stieltjes probability representations. In particular,
a m ( p , q ) = ∫ 0 1 x m d μ p , q ( x ) , 0 ≤ a m ( p , q ) ≤ 1. (133) a_m(p,q)=\int_0^1 x^m\,d\mu_{p,q}(x),\qquad0\le a_m(p,q)\le1. \tag*{(133)} a m ( p , q ) = ∫ 0 1 x m d μ p , q ( x ) , 0 ≤ a m ( p , q ) ≤ 1. (133)
The square representation also gives F p , q ( u 2 ) ≤ ( 1 − u 2 ) − 1 F_{p,q}(u^2)\le(1-u^2)^{-1} F p , q ( u 2 ) ≤ ( 1 − u 2 ) − 1 . Its positive square root recovers (17) . Thus the two representations give an alternative proof of the local convergence used in Section 2, which we record for comparison.
Write
β m = ( 2 m m ) 2 16 m , F ∗ ( z ) = ∑ m ≥ 0 β m z m = G ( 4 z ) . \beta_m=\frac{\binom{2m}{m}^2}{16^m},\qquad F_*(z)=\sum_{m\ge0}\beta_m z^m=\mathcal{G}(4\sqrt{z}). β m = 1 6 m ( m 2 m ) 2 , F ∗ ( z ) = m ≥ 0 ∑ β m z m = G ( 4 z ) .
The factor 16 m 16^m 1 6 m here comes from t = 4 z t=4\sqrt{z} t = 4 z : these β m \beta_m β m are the coefficients in z z z , rather than the Taylor coefficients in t t t . The contact limits say exactly that each a m ( p , q ) → β m a_m(p,q)\to\beta_m a m ( p , q ) → β m uniformly in the numerator; for m = 0 m=0 m = 0 this is the exact identity a 0 = β 0 = 1 a_0=\beta_0=1 a 0 = β 0 = 1 . On ∣ z ∣ ≤ r < 1 |z|\le r<1 ∣ z ∣ ≤ r < 1 , the coefficient bound (C.7) bounds every series tail by ∑ m > M r m \sum_{m>M}r^m ∑ m > M r m . Convergence of the finitely many coefficients up to M M M , followed by M → ∞ M\to\infty M → ∞ , proves uniform convergence to F ∗ F_* F ∗ on that disk. This proves the asserted local profile.
AI disclosure
As explained in the introduction, the main idea of connecting Thouless’s bandwidth conjecture with the matrix-tree theorem and square-lattice spanning-tree entropy, the conjectured limiting coarea density, and the idea of generalizing the Last–Wilkinson identity to higher contacts are due to the author. The author used OpenAI’s ChatGPT and Codex to assist with developing proof strategies, deriving and checking intermediate calculations, searching the literature, and revising the exposition. The author takes full responsibility for all mathematical arguments and the final content of this paper.
Acknowledgements
The author thanks his former colleague and friend, the theoretical physicist Wei Zhu, for explaining the physical background of Thouless’s bandwidth conjecture and for many helpful discussions. He also thanks Prof. Xiaoping Yuan sincerely. Without their constant encouragement, this work would not have been completed.
References
School of Mathematics, Sichuan University, Chengdu, 610065, China
Email address: zhaoyq@scu.edu.cn