On $(p,q)$-binomial coefficient ratios for complex parameters

Extracted source text · arXiv 2607.08163

\documentclass[12pt,reqno]{amsart}
\usepackage[british]{babel}


\usepackage{graphicx}
\usepackage{amsmath, amssymb, mathtools,cite} 
\usepackage{geometry, url, flafter}
\usepackage{microtype}
\geometry{margin=1in}


\newcommand{\pqbinom}[3]{\binom{#1}{#2}_{\! #3}}
\newcommand{\nthreequarter}{n^{3/4}}
\newcommand{\Bhat}{\widehat{B}}
\newcommand{\ud}{\,\mathrm{d}}
\DeclareMathOperator{\Real}{Re}
\DeclareMathOperator{\Imag}{Im}
\DeclareMathOperator{\sgn}{sgn}
\DeclareMathOperator{\Li}{Li}
\DeclareMathOperator{\Arg}{Arg}
\DeclareMathOperator{\csch}{csch}
\DeclareMathOperator{\dist}{dist}

\newtheorem{theorem}{Theorem}[section]
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{example}[theorem]{Example}

\newtheorem*{theoremA}{Theorem A}
\newtheorem*{theoremB}{Theorem B}

\theoremstyle{remark}
\newtheorem*{remark}{Remark}
\numberwithin{equation}{section}

\title{On $(p,q)$-binomial coefficient ratios for complex parameters}

\begin{document}
\author{P. {\AA}hag} \address{Department
  of Mathematics and Mathematical Statistics, Ume{\aa} University,
  SE-901 87 Ume{\aa}, Sweden}
 \email{per.ahag@umu.se}

\author{R. Czy{\.z}}\address{Faculty of Mathematics and Computer
  Science, Jagiellonian University, \L ojasiewicza 6, 30-348 Krak\'ow,
  Poland}\email{rafal.czyz@im.uj.edu.pl}

\author{P. H. Lundow}
\address{Department of Mathematics and Mathematical Statistics,
  Ume{\aa} University, SE-901 87 Ume{\aa}, Sweden}
\email{per-hakan.lundow@umu.se}

\begin{abstract}
  We prove local asymptotics for near-central complex $(p,q)$-binomial
  coefficient moduli ratios allowing an imaginary parameter
  perturbation of order $n^{-3/4}$ at a $\sqrt{n}$ length scale from
  the centre. Moreover, we obtain ratio asymptotics for a smaller imaginary
perturbation of order $n^{-5/4}$ at the length scale $n^{3/4}$. These results were obtained by reducing the
  two-parameter coefficients to just one parameter, giving a
  branch-free logarithmic representation of the second-order ratio
  and, hence, uniform complex curvature asymptotes for near-central
  ratios. 
\end{abstract}

\subjclass[2020]{Primary 05A30; Secondary 33D45, 41A60, 05A16}
\keywords{$(p,q)$-binomial coefficients; Gaussian polynomials;
Rogers--Szeg\H{o} polynomials; complex asymptotics;
near-central coefficient ratios; discrete curvature; quartic scaling}

\maketitle



\section{Introduction}\label{sec:intro}

For integers $0\le k\le n$ and complex parameters $p,q$ with
$p\ne0$ and
\[
p^j\ne q^j\qquad (1\le j\le n),
\]
the $(p,q)$-binomial coefficients are defined by
\begin{equation}\label{eq:intro-pq-def}
\pqbinom{n}{k}{p,q}
:=
\prod_{j=1}^{k}\frac{p^{\,n-k+j}-q^{\,n-k+j}}{p^j-q^j}.
\end{equation}
The empty product is one.  The non-vanishing condition is used only to
make \eqref{eq:intro-pq-def} literal.  It is automatic for the
asymptotic choices considered below.  For $p=1$, one obtains the usual
$q$-binomial coefficients, or Gaussian polynomials, after the standard
polynomial cancellation \cite{GasperRahman2004,DLMF17,StanleyEC1}.  If
$q$ is a prime power, then $\pqbinom{n}{k}{q}$ counts the
$k$-dimensional subspaces of $\mathbb F_q^n$ \cite{StanleyEC1}.
Equivalently, the coefficient of $q^\ell$ in $\pqbinom{n}{k}{q}$ counts
the $k$-element subsets of $\{1,\ldots,n\}$ with element sum
$\binom{k+1}{2}+\ell$.  The unimodality and strict unimodality of
these coefficient sequences have been studied from several different
viewpoints \cite{Dhand2014,OHara1990,PakPanova2013}.  The
$(p,q)$-binomial coefficient is the homogeneous symmetric form of the
same polynomial.

The polynomial
\[
P_n(w;p,q):=\sum_{k=0}^{n}\pqbinom{n}{k}{p,q}w^k
\]
is the $(p,q)$-Rogers--Szeg\H{o} polynomial. Its one-parameter form is
\begin{equation}\label{eq:intro-RS-polynomial}
P_n(w;q):=P_n(w;1,q)=\sum_{k=0}^{n}\pqbinom{n}{k}{q}w^k.
\end{equation}

Szeg\H{o}'s 1926 note~\cite{Szego1926} is a classical source for the
theta-function interpretation. He studies an orthogonalisation problem on the
unit circle with an elliptic theta-function weight of the form
\begin{equation}\label{eq:intro-szego-theta-weight}
f_\eta(\theta)=1+2\sum_{m=1}^{\infty}\eta^{m^2}\cos(m\theta),
\qquad 0<\eta<1,
\end{equation}
and shows that the resulting orthogonal polynomials are simply related to
finite polynomials whose coefficients are $q$-binomial coefficients. He also
connects these finite polynomials with earlier trigonometric identities of
Rogers \cite[pp.~243--244]{Szego1926}. The term
``Rogers--Szeg\H{o} polynomial'' is used here in this sense: it refers to the
connection among $q$-binomial coefficients, theta functions, and
orthogonality on the unit circle. Atakishiyev and Nagiyev later studied an
orthogonality relation on the full real line and related the resulting
Rogers--Szeg\H{o} functions to Stieltjes--Wigert functions by the ordinary
Fourier transform \cite{AtakishiyevNagiyev1994}.

Lubinsky and Saff~\cite{LubinskySaff1987} used Rogers--Szeg\H{o} polynomials
to analyse Pad\'e approximants of partial theta functions and the zero
distribution of these polynomials on the unit circle. Their work is global in
the variable $z$ and assumes fixed $|q|=1$. Here $q/p$ depends on $n$,
approaches $1$, and the estimates concern coefficients near the centre.

Corcino~\cite{Corcino2008} studied algebraic properties of
$(p,q)$-binomial coefficients, including recurrences, generating functions,
inverse relations, and orthogonality relations. Jagannathan and Sridhar gave
the corresponding $(p,q)$-Rogers--Szeg\H{o} polynomial and its oscillator
interpretation \cite{JagannathanSridhar2010}.

Positive $(p,q)$-binomial distributions were later used by Lundow and
Rosengren~\cite{LundowRosengren2010,LundowRosengren2013}, after determining
asymptotic properties of the coefficients, to approximate Ising-model
magnetisation distributions. In the mean-field case this distribution is
obtained exactly in the limit $q/p\to1$, with a simple relation between the
temperature and $p,q$. Moreover, $(p,q)$-binomial distributions approximate
the magnetisation distribution of high-dimensional systems ($d\ge5$) very
accurately, with the correct limiting shapes, although the relation between
temperature and the $(p,q)$-parameters is less clear.

The positive-parameter unimodal or bimodal structure was proved by Su and
Wang \cite{SuWang2012}. The real transition function relating $p,q$ to the
locations of the distribution maxima was studied by the present authors in
\cite{AhagCzyzLundow2024}, and complex branches of the associated
generalised Lambert $W$-type function were studied in
\cite{AhagCzyzLundow2025}. These papers lead directly to the present complex
coefficient problem.

The present paper studies local complex asymptotics of the coefficients. The
Ising-model and transition-function references indicate the origin of the
problem but are not used in the proofs. The argument first separates the two
parameters, then represents the second-order ratio by an absolutely convergent
logarithmic series that does not require choosing a logarithm branch, estimates
the complex curvature, and finally sums the resulting second-order differences. The
introduction states only the two principal coefficient-ratio results.

Set
\[
r:=\frac qp.
\]
Then, by Lemma~\ref{lem:pq-factor}, for every $0\le k\le n$,
\[
\pqbinom{n}{k}{p,q}=p^{\,k(n-k)}\pqbinom{n}{k}{r}.
\]
Thus every local question splits into the explicit quadratic prefactor
$p^{k(n-k)}$ and the one-parameter coefficients
\begin{equation}\label{eq:intro-r-binomial}
\pqbinom{n}{k}{r}
:=\prod_{j=1}^{k}\frac{1-r^{\,n-k+j}}{1-r^j}.
\end{equation}
For $0\le k\le n-1$ put
\begin{equation}\label{eq:intro-rho}
\rho_k(r):=
\frac{\pqbinom{n}{k+1}{r}}{\pqbinom{n}{k}{r}}
=\frac{1-r^{n-k}}{1-r^{k+1}},
\end{equation}
and for $1\le k\le n-1$ put
\begin{equation}\label{eq:intro-R}
R(n,k;r):=
\frac{\pqbinom{n}{k-1}{r}\,\pqbinom{n}{k+1}{r}}
     {\pqbinom{n}{k}{r}^{2}}
=\frac{\rho_k(r)}{\rho_{k-1}(r)}
=\frac{(1-r^{n-k})(1-r^k)}{(1-r^{n-k+1})(1-r^{k+1})}.
\end{equation}
For the full $(p,q)$-coefficients we use the corresponding notation
\begin{equation}\label{eq:intro-full-ratios}
\rho_k(p,q):=\frac{\pqbinom{n}{k+1}{p,q}}{\pqbinom{n}{k}{p,q}},
\qquad
R(n,k;p,q):=\frac{\rho_k(p,q)}{\rho_{k-1}(p,q)}.
\end{equation}
The reduction gives
\begin{equation}\label{eq:intro-ratio-reduction}
\rho_k(p,q)=p^{\,n-2k-1}\rho_k(r),
\qquad
R(n,k;p,q)=p^{-2}R(n,k;r).
\end{equation}
Since complex coefficients have no natural order, we work with the modulus and
its discrete second-order difference.  Define
\begin{equation}\label{eq:intro-g}
g(k):=\log\left|p^{\,k(n-k)}\pqbinom{n}{k}{r}\right|
=\log\left|\pqbinom{n}{k}{p,q}\right|,
\qquad 0\le k\le n.
\end{equation}
Then, for $1\le k\le n-1$,
\begin{equation}\label{eq:intro-second-diff}
\Delta^2g(k):=g(k+1)-2g(k)+g(k-1)=\log|R(n,k;r)|-2\log|p|.
\end{equation}
If $n$ is even, $\kappa:=n/2$, and $0\le\ell\le\kappa$, then
\begin{equation}\label{eq:intro-middle-product}
\frac{\pqbinom{n}{\kappa+\ell}{r}}{\pqbinom{n}{\kappa}{r}}
=
\prod_{j=0}^{\ell-1}\rho_{\kappa+j}(r),
\end{equation}
with negative $\ell$ handled by the coefficient symmetry.

For the main complex estimates we use $n$-dependent parameters of the form
\begin{equation}\label{eq:intro-parameters}
p=\exp\!\left(-\frac{u}{n}-i\,\frac{\beta}{n^{3/4}}\right),
\qquad
q=\exp\!\left(-\frac{z}{n}-i\,\frac{\alpha}{n^{3/4}}\right),
\end{equation}
with fixed real parameters $u,z,\alpha,\beta$ satisfying $z-u>0$.  Hence
\begin{equation}\label{eq:intro-r-exponent}
r=\frac qp=\exp(-t),
\qquad
t=\frac{v}{n}+i\,\frac{\delta}{n^{3/4}},
\qquad
v:=z-u>0,
\qquad
\delta:=\alpha-\beta.
\end{equation}

\begin{definition}\label{def:admissible-set}
Fix constants
\[
0<v_{\min}\le v_{\max}<\infty,
\qquad
0\le \delta_{\max}<\infty.
\]
The admissible set for the exponent parameters $(v,\delta)$ is
\[
\mathcal D:=\{(v,\delta)\in\mathbb R^2:
v_{\min}\le v\le v_{\max},\ |\delta|\le \delta_{\max}\}.
\]
\end{definition}
All constants implicit in $O_{\mathcal D,L}(\cdot)$ are independent of $n$ and of
the chosen $(v,\delta)\in\mathcal D$.  For $z>0$ write
\begin{equation}\label{eq:intro-AB}
A(z):=z\left(\coth\frac z4-1\right),
\qquad
B(z):=\frac{z^3\sinh(z/2)}{8\sinh^4(z/4)}.
\end{equation}
Then $B(z)>0$.

The exponent $3/4$ has two roles.  On the positive real axis it is the
normalisation at which the quadratic term can disappear and a quartic term
remains.  Indeed, at $2u=A(z)$ the leading real expansion is quartic for
$\ell=O(n^{3/4})$.  In the complex problem it is also the normalisation for which the
argument of $e^{\ell t}$ has order one: if
$\Imag t$ is of order $n^{-3/4}$ and $\ell\sim x n^{3/4}$, then
$\ell\Imag t$ is of order one.

More generally, write abstractly
\[
t=\frac{v_0}{n^b}+i\frac{\delta_0}{n^a},
\qquad v_0>0,
\]
and take a coefficient range $\ell=xn^\sigma$.  The imaginary part first
appears through
\[
\ell\Imag t=\delta_0 x n^{\sigma-a}.
\]
Thus, for the quartic normalisation $\sigma=3/4$, an imaginary part of order $n^{-a}$ has a
finite non-zero limit only for $a=3/4$; it disappears from the leading term
when $a>3/4$; and for $a<3/4$ it has no fixed pointwise limit without
additional arithmetic or subsequence restrictions.  For $b=1$, $n t$ has a
non-zero finite limit and the functions $A$ and $B$ above govern the central
expansion.  If $b>1$, the radial part tends more rapidly to the ordinary
binomial case; if $b<1$, $n t$ diverges and the fixed-$A,B$ expansion does not
apply.  Theorem A treats $b=1$, $a=3/4$ for deviations of order $\sqrt n$.
Theorem B treats the imaginary part of order $n^{-5/4}$, namely
$b=1$, $a=5/4$, at the quartic normalisation.

\begin{theoremA}
Assume that $n$ is even and write $\kappa:=n/2$. Let $p,q,r,t$ be
given by \eqref{eq:intro-parameters}--\eqref{eq:intro-r-exponent} for fixed real
$u,z,\alpha,\beta$ with $v=z-u>0$. Put
\[
\tau:=v+i\delta n^{1/4},
\qquad
w:=\frac{\tau}{2},
\]
and define
\[
F(w):=\frac{1}{e^w-1},
\qquad
G(w):=\frac{e^w}{(e^w-1)^2},
\qquad
H(w):=\frac{e^w(e^w+1)}{(e^w-1)^3}.
\]
For every fixed $L>0$, uniformly for all integers $\ell$ with
$|\ell|\le L\sqrt n$, and with $x:=\ell/\sqrt n$,
\begin{align}
\log\left|
\frac{\pqbinom{n}{\kappa+\ell}{p,q}}
     {\pqbinom{n}{\kappa}{p,q}}
\right|
&=
x^2\left(u-C_n(v,\delta)\right)
+\frac{x^2}{2n}\Real\left(\tau^2G(w)\right)
-\frac{x^4}{12n}\Real\left(\tau^3H(w)\right)
\nonumber\\
&\quad+O_{v,\delta,L}(\mathcal E_n),
\label{eq:main-sqrt-pq}\\[1mm]
\log\left|
\frac{\pqbinom{n}{\kappa+\ell}{r}}{\pqbinom{n}{\kappa}{r}}
\right|
&=
-x^2 C_n(v,\delta)
+\frac{x^2}{2n}\Real\left(\tau^2G(w)\right)
-\frac{x^4}{12n}\Real\left(\tau^3H(w)\right)
\nonumber\\
&\quad+O_{v,\delta,L}(\mathcal E_n),
\label{eq:main-sqrt-r}
\end{align}
where
\[
\mathcal E_n:=n^3|t|^5+n^2|t|^4+n|t|^3
\]
and
\begin{equation}\label{eq:main-Cn}
C_n(v,\delta)
:=\Real\left(\tau F(w)\right)
=\Real\left(
\frac{v+i\delta n^{1/4}}
{e^{v/2+i\delta n^{1/4}/2}-1}
\right).
\end{equation}
If $\delta=0$, then $C_n(v,0)=A(v)/2$. If $\delta\ne0$ and
$\vartheta:=\delta n^{1/4}/2$, then
\begin{equation}\label{eq:main-Cn-explicit}
C_n(v,\delta)=
\frac{v(e^{v/2}\cos\vartheta-1)
+\delta n^{1/4}e^{v/2}\sin\vartheta}
{e^{v}-2e^{v/2}\cos\vartheta+1}.
\end{equation}
\end{theoremA}

\begin{remark}
The constant $L$ in Theorem A is fixed as $n\to\infty$. Thus the theorem is
uniform for bounded $x=\ell/\sqrt n$ and gives a local expansion about the
central coefficient. It is not intended to approximate the entire coefficient
array or to locate maxima that leave this range as $n$ increases. The numerical
illustrations below show both the accuracy near the centre and the deterioration
farther away from it.
\end{remark}

\begin{theoremB}
Let $z>0$, let $\alpha,\gamma\in\mathbb R$, and assume that $n$ is even. Put
$\kappa:=n/2$ and
\[
t:=\frac zn+i\frac{\alpha}{n^{5/4}},
\qquad
r:=e^{-t},
\qquad
u:=\frac{A(z)}2+\frac{\gamma}{\sqrt n},
\]
\[
p:=e^{-u/n},
\qquad
q:=p \,r.
\]
For every fixed $L>0$, uniformly for all integer sequences $\ell_n$ such that
$x_n:=\ell_n/n^{3/4}$ satisfies $|x_n|\le L$,
\begin{equation}\label{eq:intro-small-imaginary-perturbation}
\left|
\frac{\pqbinom{n}{\kappa+\ell_n}{p,q}}
     {\pqbinom{n}{\kappa}{p,q}}
\right|
=
\exp\!\left(
\left(\gamma+\frac{\alpha^2}{4}A''(z)\right)x_n^2
-\frac{B(z)}{12}x_n^4
+O_{z,\alpha,\gamma,L}(n^{-1/2})
\right).
\end{equation}
The same formula holds with $\kappa-\ell_n$ in place of
$\kappa+\ell_n$. In particular, if $x_n\to x$, then
\[
\left|
\frac{\pqbinom{n}{\kappa\pm\ell_n}{p,q}}
     {\pqbinom{n}{\kappa}{p,q}}
\right|
\longrightarrow
\exp\!\left(
\left(\gamma+\frac{\alpha^2}{4}A''(z)\right)x^2
-\frac{B(z)}{12}x^4
\right).
\]
For $\gamma=0$, the imaginary perturbation contributes the quadratic term
$\alpha^2A''(z)x^2/4$ at the real critical value $2u=A(z)$.
\end{theoremB}

\begin{figure}[tbp]
  \begin{minipage}{0.45\textwidth}
    \centering
    \includegraphics[width=0.9\textwidth]{fig1.pdf}
  \end{minipage}%
  \begin{minipage}{0.45\textwidth}
    \centering
    \includegraphics[width=0.9\textwidth]{fig2.pdf}
  \end{minipage}
  \caption{Numerical illustration of Theorem A with $u=1$, $z=2$,
    $v=1$, $\delta=3/4$, and $x=\ell/\sqrt n$. In each panel, the
    blue curve is the modulus of the centred $(p,q)$-binomial coefficient
    ratio and the red curve is the approximation in
    \eqref{eq:main-sqrt-pq}. Left: $n=450$, with approximation
    $\exp(-0.0269x^2-0.00292x^4)$. Right: $n=496$, with approximation
    $\exp(-0.000590x^2-0.00264x^4)$, which is nearly quartic.}
  \label{fig:theoremA-central}
\end{figure}

\begin{figure}[tbp]
  \begin{minipage}{0.45\textwidth}
    \centering
    \includegraphics[width=0.9\textwidth]{fig3.pdf}
  \end{minipage}%
  \begin{minipage}{0.45\textwidth}
    \centering
    \includegraphics[width=0.9\textwidth]{fig4.pdf}
  \end{minipage}
  \caption{The same comparison as in Figure~\ref{fig:theoremA-central},
    farther into the bimodal regime. Left: $n=750$, with approximation
    $\exp(0.129x^2-0.00171x^4)$. Right: $n=800$, with approximation
    $\exp(0.152x^2-0.00160x^4)$. The increasing discrepancy away from the
    centre is consistent with the local range stated in Theorem A.}
  \label{fig:theoremA-bimodal}
\end{figure}

\begin{figure}[tbp]
  \begin{minipage}{0.45\textwidth}
    \centering
    \includegraphics[width=0.9\textwidth]{fig5.pdf}
  \end{minipage}%
  \begin{minipage}{0.45\textwidth}
    \centering
    \includegraphics[width=0.9\textwidth]{fig6.pdf}
  \end{minipage}
  \caption{Numerical illustration of Theorem B with
    $x=\ell/n^{3/4}$. The blue, green, and orange curves are the moduli of
    the centred $(p,q)$-binomial coefficient ratios for $n=256$, $512$,
    and $1024$, respectively; the red curve is the limiting function.
    Left: $z=\alpha=1$ and
    $\gamma=-\alpha^2A''(z)/4\approx-0.0406$, giving the limit
    $\exp(-1.33x^4)$. Right: $z=\alpha=1$ and $\gamma=2$, giving the
    limit $\exp(2.04x^2-1.33x^4)$.}
  \label{fig:theoremB-limit}
\end{figure}

The paper is organised as follows.
Section~\ref{sec:one-parameter} proves the one-parameter local formulae,
including the absolutely convergent logarithmic series for the second-order ratio.
Section~\ref{sec:reduction} transfers these formulae to the full
$(p,q)$-coefficients and establishes self-reciprocity.
Section~\ref{sec:complex-central} contains the complex estimates near the
centre, including the result used in Theorem A.
Section~\ref{sec:small-imaginary-perturbation} treats the imaginary
perturbation of order $n^{-5/4}$ and proves Theorem B.
Section~\ref{sec:conclusion} gives further problems.

\section{The one-parameter local formulae}\label{sec:one-parameter}

This section proves the one-parameter formulae used below.
Lemma~\ref{lem:R-hyperbolic} gives the hyperbolic
form of the second-order ratio. Proposition~\ref{prop:Lambda} gives the logarithmic
series for the second-order ratio; its exponential is exactly $R(n,k;q)$, while
its real part is $\log|R(n,k;q)|$. Proposition~\ref{prop:complex-local} and
Corollary~\ref{cor:real-Lambda} give the complex and real local estimates
near the centre.

We use $q$ as the one-parameter variable in this section. For an integer
$n\ge0$ and $|q|<1$ define
\[
\pqbinom{n}{k}{q}
:=
\prod_{j=1}^{k}\frac{1-q^{n-k+j}}{1-q^j}
\qquad (0\le k\le n),
\]
where the empty product is one.
Since $|q|<1$, all factors $1-q^j$ with $j\ge1$ are non-zero. For
$n\ge2$ and $0\le k\le n-1$ the adjacent ratio is
\begin{equation}\label{eq:one-param-rho}
\frac{\pqbinom{n}{k+1}{q}}{\pqbinom{n}{k}{q}}
=\frac{1-q^{n-k}}{1-q^{k+1}}.
\end{equation}
Consequently, for $1\le k\le n-1$ the curvature ratio is
\begin{equation}\label{eq:R-basic}
R(n,k;q)
:=
\frac{\pqbinom{n}{k-1}{q}\,\pqbinom{n}{k+1}{q}}
     {\pqbinom{n}{k}{q}^{2}}
=
\frac{(1-q^{n-k})(1-q^k)}{(1-q^{n-k+1})(1-q^{k+1})}.
\end{equation}

The same ratio will be used in a hyperbolic form when $q=e^{-t}$. This
form is especially convenient on the real axis, where Taylor expansion of
$\log\sinh$ gives the constants $A$ and $B$.

\begin{lemma}\label{lem:R-hyperbolic}
Let $t\in\mathbb C$ with $\Real t>0$, let $q=e^{-t}$, and let
$1\le k\le n-1$. Then
\begin{equation}\label{eq:R-sinh}
R(n,k;e^{-t})
=
e^t\,
\frac{\sinh(\tfrac{k t}{2})\sinh(\tfrac{(n-k)t}{2})}
{\sinh(\tfrac{(k+1)t}{2})\sinh(\tfrac{(n-k+1)t}{2})}.
\end{equation}
\end{lemma}

\begin{proof}
Use $1-e^{-x}=2e^{-x/2}\sinh(x/2)$ in \eqref{eq:R-basic}. The
factors $2$ cancel. The exponential part is
\[
\frac{e^{-k t/2}e^{-(n-k)t/2}}
{e^{-(k+1)t/2}e^{-(n-k+1)t/2}}=e^t,
\]
and the remaining factors give \eqref{eq:R-sinh}.
\end{proof}

For complex $q$ it is safer not to choose a logarithm of the quotient
\eqref{eq:R-basic}. The following series is an ordinary absolutely
convergent series. It defines a logarithm of $R(n,k;q)$ only through its
exponential identity.

\begin{proposition}\label{prop:Lambda}
Let $t\in\mathbb C$ with $\Real t>0$, let $q=e^{-t}$, and let
$1\le k\le n-1$. Define
\begin{equation}\label{eq:Lambda-def}
\Lambda(n,k;q)
:=
-\sum_{m=1}^{\infty}\frac{1-q^m}{m}\left(q^{mk}+q^{m(n-k)}\right).
\end{equation}
Then the series converges absolutely and
\begin{equation}\label{eq:Lambda-exp}
e^{\Lambda(n,k;q)}=R(n,k;q).
\end{equation}
Moreover,
\begin{equation}\label{eq:Lambda-real}
\Real\Lambda(n,k;q)=\log|R(n,k;q)|.
\end{equation}
If
\[
\mu:=\min\{k,n-k\}\,\Real t,
\]
then $\mu>0$ and, for every integer $M\ge0$,
\begin{equation}\label{eq:Lambda-remainder}
\left|
\sum_{m=M+1}^{\infty}\frac{1-q^m}{m}\left(q^{mk}+q^{m(n-k)}\right)
\right|
\le
\frac{4e^{-(M+1)\mu}}{(M+1)(1-e^{-\mu})}.
\end{equation}
\end{proposition}

\begin{proof}
Since $|q|<1$ and $1\le k\le n-1$, one has
\[
|1-q^m|\le 1+|q|^m\le2,
\qquad
|q|^{mk}+|q|^{m(n-k)}\le2e^{-m\mu}.
\]
Hence
\[
\left|
\sum_{m=M+1}^{\infty}\frac{1-q^m}{m}\left(q^{mk}+q^{m(n-k)}\right)
\right|
\le
4\sum_{m=M+1}^{\infty}\frac{e^{-m\mu}}{m}.
\]
Since
\[
\sum_{m=M+1}^{\infty}\frac{e^{-m\mu}}{m}
\le
\frac{e^{-(M+1)\mu}}{M+1}\sum_{j=0}^{\infty}e^{-j\mu}
=
\frac{e^{-(M+1)\mu}}{(M+1)(1-e^{-\mu})},
\]
this proves \eqref{eq:Lambda-remainder}. In particular, the series
\eqref{eq:Lambda-def} converges absolutely.

For $j\ge1$, the principal logarithm satisfies
\[
\log(1-q^j)=-\sum_{m=1}^{\infty}\frac{q^{mj}}{m},
\]
because $|q^j|<1$ and hence $1-q^j$ lies in the right half-plane. Thus
\begin{align*}
\Lambda(n,k;q)
&=
\log(1-q^k)-\log(1-q^{k+1})
+\log(1-q^{n-k})-\log(1-q^{n-k+1}).
\end{align*}
Exponentiating this identity gives
\[
e^{\Lambda(n,k;q)}
=
\frac{(1-q^k)(1-q^{n-k})}{(1-q^{k+1})(1-q^{n-k+1})}
=
R(n,k;q),
\]
which proves \eqref{eq:Lambda-exp}. Taking absolute values in
\eqref{eq:Lambda-exp} gives
\[
e^{\Real\Lambda(n,k;q)}=|R(n,k;q)|,
\]
and therefore \eqref{eq:Lambda-real}.
\end{proof}

We now return to the quotient parameter $r$ and let it depend on $n$ as in
the introduction. In the central range both $k\Real t$ and
$(n-k)\Real t$ stay bounded away from zero, while $t$ itself is small.
This permits the expansion of $1-e^{-m t}$ inside \eqref{eq:Lambda-def}.

\begin{proposition}\label{prop:complex-local}
Fix an admissible set $\mathcal D$ and a constant $L>0$. For
\[
r=\exp(-t),
\qquad
t=\frac{v}{n}+i\,\frac{\delta}{n^{3/4}},
\qquad
(v,\delta)\in \mathcal D,
\]
and $1\le k\le n-1$, define
\begin{align}
S_1(n,k;r)
&:=
\frac{1}{e^{k t}-1}
+
\frac{1}{e^{(n-k)t}-1},\label{eq:S1-def}\\[1mm]
S_2(n,k;r)
&:=
\frac{e^{k t}}{(e^{k t}-1)^2}
+
\frac{e^{(n-k)t}}{(e^{(n-k)t}-1)^2}.\label{eq:S2-def}
\end{align}
Then, for all sufficiently large $n$ and all integers $k$ with
\[
|k-n/2|\le L n^{3/4},
\]
one has $1\le k\le n-1$ and
\begin{align}
\Lambda(n,k;r)
&=
-t S_1(n,k;r)
+O_{\mathcal D,L}(n^{-3/2}),\label{eq:complex-first-order}\\[1mm]
\Lambda(n,k;r)
&=
-t S_1(n,k;r)
+\frac{t^2}{2}\,S_2(n,k;r)
+O_{\mathcal D,L}(|t|^3).\label{eq:complex-second-order}
\end{align}
\end{proposition}

\begin{proof}
For all sufficiently large $n$, the condition $|k-n/2|\le L n^{3/4}$
implies
\[
k\ge \frac n4,
\qquad
n-k\ge \frac n4.
\]
Set
\[
a:=e^{-k t},
\qquad
b:=e^{-(n-k)t},
\qquad
\sigma:=e^{-v_{\min}/4}.
\]
Then $0<\sigma<1$ and
\[
|a|=e^{-k\Real t}\le \sigma,
\qquad
|b|=e^{-(n-k)\Real t}\le \sigma.
\]

For the first expansion, use the identity
\[
1-e^{-x}=x-x^2\int_0^1(1-s)e^{-sx}\ud s.
\]
If $\Real x\ge0$, this gives
\[
|1-e^{-x}-x|\le \frac{|x|^2}{2}.
\]
With $x=m t$ we may write
\[
1-e^{-m t}=m t+\varepsilon_m,
\qquad
|\varepsilon_m|\le \frac{m^2|t|^2}{2}.
\]
Substitution in \eqref{eq:Lambda-def} gives
\[
\Lambda(n,k;r)
=
-t\sum_{m=1}^{\infty}(a^m+b^m)
-
\sum_{m=1}^{\infty}\frac{\varepsilon_m}{m}(a^m+b^m).
\]
The remainder is bounded by
\[
\left|
\sum_{m=1}^{\infty}\frac{\varepsilon_m}{m}(a^m+b^m)
\right|
\le
\frac{|t|^2}{2}\sum_{m=1}^{\infty}m(|a|^m+|b|^m)
\le
|t|^2\sum_{m=1}^{\infty}m\sigma^m
=
\frac{\sigma}{(1-\sigma)^2}|t|^2.
\]
Moreover,
\[
\sum_{m=1}^{\infty}(a^m+b^m)
=
\frac{a}{1-a}+\frac{b}{1-b}
=
\frac{1}{e^{k t}-1}+\frac{1}{e^{(n-k)t}-1}
=
S_1(n,k;r).
\]
Since $|t|=O_{\mathcal D}(n^{-3/4})$, this proves
\eqref{eq:complex-first-order}.

For the second expansion, use
\[
1-e^{-x}
=
x-\frac{x^2}{2}
+x^3\int_0^1\frac{(1-s)^2}{2}e^{-sx}\ud s.
\]
For $\Real x\ge0$,
\[
\left|1-e^{-x}-x+\frac{x^2}{2}\right|
\le
\frac{|x|^3}{6}.
\]
Thus
\[
1-e^{-m t}
=
m t-\frac{m^2t^2}{2}+\eta_m,
\qquad
|\eta_m|\le \frac{m^3|t|^3}{6}.
\]
Substitution in \eqref{eq:Lambda-def} gives
\[
\Lambda(n,k;r)
=
-t\sum_{m=1}^{\infty}(a^m+b^m)
+\frac{t^2}{2}\sum_{m=1}^{\infty}m(a^m+b^m)
-
\sum_{m=1}^{\infty}\frac{\eta_m}{m}(a^m+b^m).
\]
The last sum is
\[
O_{\mathcal D,L}\!\left(
|t|^3\sum_{m=1}^{\infty}m^2\sigma^m
\right)
=O_{\mathcal D,L}(|t|^3).
\]
Furthermore,
\[
\sum_{m=1}^{\infty}m(a^m+b^m)
=
\frac{a}{(1-a)^2}+\frac{b}{(1-b)^2}
=
\frac{e^{k t}}{(e^{k t}-1)^2}
+
\frac{e^{(n-k)t}}{(e^{(n-k)t}-1)^2}
=
S_2(n,k;r).
\]
This proves \eqref{eq:complex-second-order}.
\end{proof}

On the real axis the hyperbolic formula gives a sharper expansion. This is
where the two functions $A$ and $B$ enter the paper.

\begin{corollary}\label{cor:real-Lambda}
Fix constants $0<z_0\le z_1<\infty$ and $L>0$. Let
$z\in[z_0,z_1]$ and let $q=e^{-z/n}$. Assume that $n$ is even and write
$\kappa:=n/2$. Then, uniformly for $z\in[z_0,z_1]$, all sufficiently
large even $n$, and all integers $\ell$ with $|\ell|\le L n^{3/4}$,
\begin{equation}\label{eq:real-logR}
\log R\!\left(n,\kappa+\ell;e^{-z/n}\right)
=
-\frac{A(z)}{n}
-\frac{B(z)}{n^3}\,\ell^2
+O_{z_0,z_1,L}(n^{-2}),
\end{equation}
where $A$ and $B$ are the functions defined in \eqref{eq:intro-AB}. Consequently,
\begin{equation}\label{eq:real-logR-centred}
\log R\!\left(n,\kappa+\ell;e^{-z/n}\right)-\log R\!\left(n,\kappa;e^{-z/n}\right)
=
-\frac{B(z)}{n^3}\,\ell^2
+O_{z_0,z_1,L}(n^{-2}).
\end{equation}
\end{corollary}

\begin{proof}
Set
\[
a:=\frac z4,
\qquad
h:=\frac{z}{2n},
\qquad
\Delta:=\frac{z\ell}{2n}.
\]
Since $z\in[z_0,z_1]$ and $|\ell|\le L n^{3/4}$, one has
$\Delta=O_{z_0,z_1,L}(n^{-1/4})$ and $h=O_{z_1}(n^{-1})$. For all
sufficiently large $n$, the points $a\pm\Delta$ and
$a+h\pm\Delta$ lie in a compact subinterval of $(0,\infty)$ that
only depends on $z_0,z_1$, and $L$.

By Lemma~\ref{lem:R-hyperbolic},
\[
\log R\!\left(n,\kappa+\ell;e^{-z/n}\right)
=
\frac zn
+\psi(a+\Delta)+\psi(a-\Delta)
-\psi(a+h+\Delta)-\psi(a+h-\Delta),
\]
where $\psi(y):=\log\sinh y$. Taylor's formula in the variable $h$
gives, uniformly for the above values of $y$,
\[
\psi(y)-\psi(y+h)
=
-h\psi'(y)+O_{z_0,z_1,L}(n^{-2}).
\]
Therefore
\[
\log R
=
\frac zn
-h\left(\psi'(a+\Delta)+\psi'(a-\Delta)\right)
+O_{z_0,z_1,L}(n^{-2}).
\]
Taylor's formula in the variable $\Delta$ gives
\[
\psi'(a+\Delta)+\psi'(a-\Delta)
=
2\psi'(a)+\psi^{(3)}(a)\Delta^2+O_{z_0,z_1,L}(\Delta^4).
\]
Since $h\Delta^4=O_{z_0,z_1,L}(n^{-2})$, substitution yields
\[
\log R
=
\frac zn-2h\psi'(a)-h\psi^{(3)}(a)\Delta^2
+O_{z_0,z_1,L}(n^{-2}).
\]
Now $\psi'(a)=\coth a$, and hence
\[
\frac zn-2h\psi'(a)
=
-\frac{z}{n}\left(\coth\frac z4-1\right)
=
-\frac{A(z)}{n}.
\]
Moreover,
\[
\psi^{(3)}(a)=2\coth(a)\csch^2(a)=\frac{\sinh(2a)}{\sinh^4(a)},
\]
and therefore
\[
-h\psi^{(3)}(a)\Delta^2
=
-\frac{z}{2n}\,\psi^{(3)}\!\left(\frac z4\right)
\frac{z^2\ell^2}{4n^2}
=
-\frac{B(z)}{n^3}\,\ell^2.
\]
This proves \eqref{eq:real-logR}. The same estimate with $\ell=0$ is
valid uniformly; subtracting it from \eqref{eq:real-logR} gives
\eqref{eq:real-logR-centred}.
\end{proof}



\section{Reduction from $(p,q)$ to one parameter}\label{sec:reduction}

This section transfers the one-parameter formulae to the full
$(p,q)$-coefficients. Lemma~\ref{lem:pq-factor} separates the explicit factor
$p^{k(n-k)}$ from the one-parameter coefficient. Corollary~\ref{cor:selfreciprocal}
gives the self-reciprocal generating polynomial, and
Corollary~\ref{cor:local-ratios} gives the adjacent-ratio and second-order ratio
reductions used in Sections~\ref{sec:complex-central} and
\ref{sec:small-imaginary-perturbation}.

The first lemma is elementary, but it fixes the hypotheses needed later and
separates the explicit $p$-factor from the one-parameter coefficient.

\begin{lemma}\label{lem:pq-factor}
Let $p,q\in\mathbb C$ with $p\ne0$ and $p^j\ne q^j$ for $1\le j\le n$.
Set $r:=q/p$. Then $r^j\ne1$ for $1\le j\le n$, and for every
$0\le k\le n$,
\begin{equation}\label{eq:pq-factor}
\pqbinom{n}{k}{p,q}
=
p^{\,k(n-k)}\prod_{j=1}^{k}\frac{1-r^{n-k+j}}{1-r^j}
=
p^{\,k(n-k)}\pqbinom{n}{k}{r}.
\end{equation}
\end{lemma}

\begin{proof}
For each factor in \eqref{eq:intro-pq-def},
\[
\frac{p^{\,n-k+j}-q^{\,n-k+j}}{p^j-q^j}
=
\frac{p^{\,n-k+j}(1-r^{n-k+j})}{p^j(1-r^j)}
=
p^{\,n-k}\frac{1-r^{n-k+j}}{1-r^j}.
\]
The denominators are non-zero because $r^j\ne1$ for $1\le j\le n$.
Multiplication over $j=1,\ldots,k$ gives the factor $p^{k(n-k)}$ and proves
\eqref{eq:pq-factor}. For $k=0$ both products are empty and the identity is
$1=1$.
\end{proof}

The same factorisation also gives the coefficient symmetry. We state it in
two forms, first as an ordinary polynomial identity and then in centred
Laurent form.

\begin{corollary}\label{cor:selfreciprocal}
Assume the hypotheses of Lemma~\ref{lem:pq-factor}. Let
\[
P_n(w):=\sum_{k=0}^{n}\pqbinom{n}{k}{p,q}\,w^k.
\]
Then
\begin{equation}\label{eq:self-recip-polynomial}
w^nP_n(w^{-1})=P_n(w).
\end{equation}
In particular, the zeros of $P_n$ occur in reciprocal pairs, with multiplicity. If $n$ is even
and
\begin{equation}\label{eq:Z-centred}
\mathcal Z_n(w):=\sum_{k=0}^{n}\pqbinom{n}{k}{p,q}\,w^{k-n/2},
\end{equation}
then
\begin{equation}\label{eq:self-recip}
\mathcal Z_n(w^{-1})=\mathcal Z_n(w).
\end{equation}
For odd $n$, the same centred identity is an identity for the corresponding
Laurent polynomial in $w^{1/2}$.
\end{corollary}

\begin{proof}
By Lemma~\ref{lem:pq-factor},
\[
\pqbinom{n}{k}{p,q}=p^{\,k(n-k)}\pqbinom{n}{k}{r}.
\]
Since $r^j\ne1$ for $1\le j\le n$,
\[
\pqbinom{n}{k}{r}
=
\frac{(r;r)_n}{(r;r)_k(r;r)_{n-k}},
\qquad
(r;r)_m:=\prod_{j=1}^{m}(1-r^j),
\]
with the convention $(r;r)_0=1$. Hence
$\pqbinom{n}{k}{r}=\pqbinom{n}{n-k}{r}$.  The factor
$p^{\,k(n-k)}$ is also invariant under $k\mapsto n-k$. Therefore
\[
\pqbinom{n}{k}{p,q}=\pqbinom{n}{n-k}{p,q}\qquad (0\le k\le n).
\]
This coefficient symmetry is exactly \eqref{eq:self-recip-polynomial}.
Since the constant and leading coefficients of $P_n$ are both one, zero is
not a zero of $P_n$; consequently every zero is accompanied by its reciprocal, with the same multiplicity.
If $n$ is even, multiplying \eqref{eq:self-recip-polynomial} by $w^{-n/2}$
gives \eqref{eq:self-recip}. The odd case is the same formal identity after
writing the centre with powers of $w^{1/2}$.
\end{proof}

The next corollary gives the way in which the adjacent and second-order ratios change under
the same reduction. Here the explicit quadratic factor in $p$ enters the
discrete second-order difference of the modulus.


\begin{corollary}\label{cor:local-ratios}
Assume the hypotheses of Lemma~\ref{lem:pq-factor} and set $r=q/p$. Then
\begin{align}
\rho_k(p,q)&=p^{\,n-2k-1}\rho_k(r),\qquad 0\le k\le n-1,\label{eq:rho-red}\\
R(n,k;p,q)&=p^{-2}R(n,k;r),\qquad 1\le k\le n-1.\label{eq:R-red}
\end{align}
Consequently, with $g$ as in \eqref{eq:intro-g}, for $1\le k\le n-1$,
\begin{equation}\label{eq:g-second-diff}
\Delta^2 g(k)
:=
g(k+1)-2g(k)+g(k-1)
=
\log|R(n,k;r)|-2\log|p|.
\end{equation}
\end{corollary}

\begin{proof}
From Lemma~\ref{lem:pq-factor},
\[
\frac{\pqbinom{n}{k+1}{p,q}}{\pqbinom{n}{k}{p,q}}
=
p^{(k+1)(n-k-1)-k(n-k)}
\frac{\pqbinom{n}{k+1}{r}}{\pqbinom{n}{k}{r}}
=
p^{\,n-2k-1}\rho_k(r),
\]
which proves \eqref{eq:rho-red}. Dividing the formula for $k$ by the same
formula with $k-1$ gives
\[
\frac{\rho_k(p,q)}{\rho_{k-1}(p,q)}
=
p^{-2}\frac{\rho_k(r)}{\rho_{k-1}(r)},
\]
and hence \eqref{eq:R-red}. Finally,
\[
\Delta^2 g(k)
=
\log\left|\frac{\pqbinom{n}{k-1}{p,q}\pqbinom{n}{k+1}{p,q}}
{\pqbinom{n}{k}{p,q}^{\,2}}\right|
=
\log|R(n,k;p,q)|,
\]
and \eqref{eq:g-second-diff} follows from \eqref{eq:R-red}.
\end{proof}


\section{Complex local estimates near the centre}\label{sec:complex-central}

This section treats the complex case near $k=n/2$. We use the
$n$-dependent parameters from \eqref{eq:intro-parameters}--\eqref{eq:intro-r-exponent}.
Thus $r=e^{-t}$ with $\Real t=v/n>0$. In estimates involving $g$
or the full coefficients we assume $|p|=e^{-u/n}$. The argument of $p$ does
not enter these absolute-value estimates.

Since the coefficients are complex, no order estimate for a positive sequence
is used. Theorem~\ref{thm:complex-curvature} estimates $\log |R(n,k;r)|$
and $\Delta^2 g(k)$; its proof uses Proposition~\ref{prop:complex-local},
\eqref{eq:Lambda-real}, and Corollary~\ref{cor:local-ratios}. Corollary~\ref{cor:complex-center}
gives the central value. Lemma~\ref{lem:complex-middle-exact} and
Theorem~\ref{thm:complex-middle} give centred ratios. Theorem~\ref{thm:complex-sqrt}
gives the estimate for $|\ell|=O(\sqrt n)$ and identifies the coefficient of
$x^2$ for the modulus of the full coefficient.

\begin{theorem}\label{thm:complex-curvature}
Fix an admissible set $\mathcal D$, a constant $L>0$, and a real number $u$.
For each $(v,\delta)\in \mathcal D$, set
\[
r=e^{-t},\qquad
 t=\frac{v}{n}+i\frac{\delta}{n^{3/4}},
\]
and let $p$ satisfy $|p|=e^{-u/n}$ in the definition of $g$ in
\eqref{eq:intro-g}. Then, for all sufficiently large $n$, all
$(v,\delta)\in \mathcal D$, and all integers $k$ with
\[
|k-n/2|\le L n^{3/4},
\]
one has $1\le k\le n-1$ and
\begin{align}
\log\left|R(n,k;r)\right|
&=
-\Real\left(t S_1(n,k;r)\right)
+O_{\mathcal D,L}(n^{-3/2}),\label{eq:complex-curvature-R}\\[1mm]
\Delta^2 g(k)
&=
\frac{2u}{n}
-\Real\left(t S_1(n,k;r)\right)
+O_{\mathcal D,L}(n^{-3/2}).\label{eq:complex-curvature-g}
\end{align}
In particular,
\[
\log\left|R(n,k;r)\right|=O_{\mathcal D,L}(n^{-3/4}),
\qquad
\Delta^2 g(k)=O_{\mathcal D,L,u}(n^{-3/4}).
\]
\end{theorem}

\begin{proof}
For all sufficiently large $n$, the central condition implies
$k\ge n/4$ and $n-k\ge n/4$, and hence $1\le k\le n-1$. Proposition~\ref{prop:complex-local} gives
\[
\Lambda(n,k;r)=-t\,S_1(n,k;r)+O_{\mathcal D,L}(n^{-3/2}).
\]
Taking real parts and using Proposition~\ref{prop:Lambda},
\[
\Real \Lambda(n,k;r)=\log\left|R(n,k;r)\right|,
\]
proves \eqref{eq:complex-curvature-R}. Corollary~\ref{cor:local-ratios} gives
\[
\Delta^2 g(k)=\log\left|R(n,k;r)\right|-2\log|p|.
\]
Since $|p|=e^{-u/n}$, one has $-2\log|p|=2u/n$, and
\eqref{eq:complex-curvature-g} follows.

It remains only to prove the stated bound. On the central range,
\[
\Real(k t)=\frac{k v}{n}\ge \frac{v_{\min}}4,
\qquad
\Real((n-k)t)=\frac{(n-k)v}{n}\ge \frac{v_{\min}}4.
\]
Consequently,
\[
|e^{k t}-1|\ge e^{v_{\min}/4}-1,
\qquad
|e^{(n-k)t}-1|\ge e^{v_{\min}/4}-1,
\]
and therefore $S_1(n,k;r)=O_{\mathcal D,L}(1)$. Since
$|t|=O_{\mathcal D}(n^{-3/4})$ and $u$ is fixed, the two
$O(n^{-3/4})$ bounds follow.
\end{proof}

At the exact centre the two terms in $S_1$ are equal. This gives the next
corollary. The real refinement uses the sharper real-axis expansion from
Corollary~\ref{cor:real-Lambda}.

\begin{corollary}\label{cor:complex-center}
Fix an admissible set $\mathcal D$ and a real number $u$. For each
$(v,\delta)\in \mathcal D$, set
\[
r=e^{-t},\qquad
 t=\frac{v}{n}+i\frac{\delta}{n^{3/4}},
\]
and let $p$ satisfy $|p|=e^{-u/n}$ in the definition of $g$ in
\eqref{eq:intro-g}. Assume that $n$ is even and write $\kappa:=n/2$.
Then, uniformly for all $(v,\delta)\in \mathcal D$ and all sufficiently large
even $n$,
\begin{equation}\label{eq:complex-center}
\Delta^2 g(\kappa)
=
\frac{2u}{n}
-
2\Real\left(\frac{t}{e^{\kappa t}-1}\right)
+O_{\mathcal D}(n^{-3/2}).
\end{equation}
If $\delta=0$, then, uniformly for all $(v,0)\in \mathcal D$,
\begin{equation}\label{eq:complex-center-real}
\Delta^2 g(\kappa)
=
\frac{2u-A(v)}{n}
+O_{\mathcal D}(n^{-2}).
\end{equation}
\end{corollary}

\begin{proof}
Set $k=\kappa$ in Theorem~\ref{thm:complex-curvature}. Then
\[
S_1(n,\kappa;r)=\frac{2}{e^{\kappa t}-1},
\]
which gives \eqref{eq:complex-center}. If $\delta=0$, then Corollary~\ref{cor:real-Lambda}, with $\ell=0$, gives
\[
\log R(n,\kappa;e^{-v/n})=-\frac{A(v)}{n}+O_{\mathcal D}(n^{-2}).
\]
Together with Corollary~\ref{cor:local-ratios} and $-2\log|p|=2u/n$, this gives
\eqref{eq:complex-center-real}.
\end{proof}

Curvature gives second-order differences. For centred ratios it is more direct to
multiply adjacent ratios and then expand the resulting product. The next
lemma gives an exact logarithmic form of that product.

\begin{lemma}\label{lem:complex-middle-exact}
Assume that $n$ is even and write $\kappa:=n/2$. Let
$r=e^{-t}$ with $\Real t>0$. Let $0\le \ell\le \kappa$ and set
\begin{equation}\label{eq:Epm-def}
E_-(n,\ell;r):=e^{-(\kappa-\ell+1)t},
\qquad
E_0(n;r):=e^{-(\kappa+1)t},
\qquad
E_+(n,\ell;r):=e^{-(\kappa+\ell+1)t}.
\end{equation}
Then
\begin{equation}\label{eq:complex-middle-exact}
\log\left|\frac{\pqbinom{n}{\kappa+\ell}{r}}{\pqbinom{n}{\kappa}{r}}\right|
=
-\Real\sum_{m=1}^{\infty}
\frac{E_-(n,\ell;r)^m-2E_0(n;r)^m+E_+(n,\ell;r)^m}
{m\,(1-e^{-m t})}.
\end{equation}
The series converges absolutely.
\end{lemma}

\begin{proof}
Since $\Real t>0$, one has $|r|<1$. The case $\ell=0$ gives zero on both
sides, so assume first that $1\le \ell\le\kappa$. From \eqref{eq:intro-rho},
\[
\rho_k(r)=\frac{1-r^{n-k}}{1-r^{k+1}},
\]
and therefore
\[
\frac{\pqbinom{n}{\kappa+\ell}{r}}{\pqbinom{n}{\kappa}{r}}
=
\prod_{s=0}^{\ell-1}\rho_{\kappa+s}(r)
=
\prod_{j=1}^{\ell}\frac{1-r^{\kappa-\ell+j}}{1-r^{\kappa+j}}.
\]
For $|w|<1$,
\[
\log|1-w|=-\Real\sum_{m=1}^{\infty}\frac{w^m}{m}.
\]
Applying this identity to each factor gives
\[
\log\left|\frac{\pqbinom{n}{\kappa+\ell}{r}}{\pqbinom{n}{\kappa}{r}}\right|
=
-\Real\sum_{m=1}^{\infty}\frac{1}{m}
\sum_{j=1}^{\ell}\left(r^{m(\kappa-\ell+j)}-r^{m(\kappa+j)}\right).
\]
The inner sum is
\[
\sum_{j=1}^{\ell}r^{m(\kappa-\ell+j)}
-
\sum_{j=1}^{\ell}r^{m(\kappa+j)}
=
\frac{r^{m(\kappa-\ell+1)}(1-r^{m\ell})^2}{1-r^m}
=
\frac{E_-^m-2E_0^m+E_+^m}{1-e^{-m t}},
\]
which proves \eqref{eq:complex-middle-exact}.

For absolute convergence, set
\[
\sigma:=|r|^{\kappa-\ell+1}<1.
\]
Then
\[
|E_-(n,\ell;r)|,\ |E_0(n;r)|,\ |E_+(n,\ell;r)|\le \sigma,
\qquad
|1-r^m|\ge 1-|r|^m\ge 1-|r|.
\]
Therefore
\[
\left|\frac{E_-^m-2E_0^m+E_+^m}{m(1-e^{-m t})}\right|
\le
\frac{4\sigma^m}{m(1-|r|)},
\]
and $\sum_{m\ge1}\sigma^m/m$ converges. This proves absolute convergence.
\end{proof}

We next keep one further term in the expansion of
$(1-e^{-m t})^{-1}$. This adds the series $\Phi_0$ to the
$\Phi_1$- and $\Phi_2$-terms.

\begin{theorem}\label{thm:complex-middle}
Fix an admissible set $\mathcal D$, a constant $L>0$, and a real number $u$.
Assume that $n$ is even and write $\kappa:=n/2$. Let
$r=e^{-t}$ with $t$ as in \eqref{eq:intro-r-exponent} and
$(v,\delta)\in \mathcal D$, and let $p$ satisfy $|p|=e^{-u/n}$ in the definition
of $g$ in \eqref{eq:intro-g}. Then, uniformly for all $(v,\delta)\in \mathcal D$,
all sufficiently large even $n$, and every integer $\ell$ with
$|\ell|\le L n^{3/4}$,
\begin{align}
\log\left|\frac{\pqbinom{n}{\kappa+\ell}{r}}{\pqbinom{n}{\kappa}{r}}\right|
&=
-\Real\left(
\frac{\Phi_2(n,|\ell|;r)}{t}
+\frac{1}{2}\,\Phi_1(n,|\ell|;r)
+\frac{t}{12}\,\Phi_0(n,|\ell|;r)
\right)
\nonumber\\
&\quad
+O_{\mathcal D,L}(n^{-9/4}),\label{eq:complex-middle-main}\\[1mm]
g(\kappa+\ell)-g(\kappa)
&=
\frac{u}{n}\,\ell^2
-\Real\left(
\frac{\Phi_2(n,|\ell|;r)}{t}
+\frac{1}{2}\,\Phi_1(n,|\ell|;r)
+\frac{t}{12}\,\Phi_0(n,|\ell|;r)
\right)
\nonumber\\
&\quad
+O_{\mathcal D,L}(n^{-9/4}),\label{eq:complex-middle-g}
\end{align}
where
\begin{align}
\Phi_0(n,\ell;r)
&:=
\sum_{m=1}^{\infty}
\left(E_-(n,\ell;r)^m-2E_0(n;r)^m+E_+(n,\ell;r)^m\right),\label{eq:Phi0-def}\\[1mm]
\Phi_1(n,\ell;r)
&:=
\sum_{m=1}^{\infty}\frac{E_-(n,\ell;r)^m-2E_0(n;r)^m+E_+(n,\ell;r)^m}{m},\label{eq:Phi1-def}\\[1mm]
\Phi_2(n,\ell;r)
&:=
\sum_{m=1}^{\infty}\frac{E_-(n,\ell;r)^m-2E_0(n;r)^m+E_+(n,\ell;r)^m}{m^2}.\label{eq:Phi2-def}
\end{align}
The three series converge absolutely, and
\begin{align}
\Phi_0(n,\ell;r)
&=
\frac{E_-(n,\ell;r)}{1-E_-(n,\ell;r)}
-2\,\frac{E_0(n;r)}{1-E_0(n;r)}
+\frac{E_+(n,\ell;r)}{1-E_+(n,\ell;r)},\label{eq:Phi0-rational}\\[1mm]
\Phi_1(n,\ell;r)
&=
-\log\left(1-E_-(n,\ell;r)\right)
+2\log\left(1-E_0(n;r)\right)
-\log\left(1-E_+(n,\ell;r)\right),\label{eq:Phi1-log}\\[1mm]
\Phi_2(n,\ell;r)
&=
\Li_2\left(E_-(n,\ell;r)\right)
-2\Li_2\left(E_0(n;r)\right)
+\Li_2\left(E_+(n,\ell;r)\right).\label{eq:Phi2-dilog}
\end{align}
Here the logarithms in \eqref{eq:Phi1-log} are principal logarithms.
\end{theorem}

\begin{proof}
The one-parameter coefficients satisfy
$\pqbinom{n}{\kappa+\ell}{r}=\pqbinom{n}{\kappa-\ell}{r}$, and hence
$g(\kappa+\ell)=g(\kappa-\ell)$. It is therefore enough to consider
$0\le \ell\le L n^{3/4}$. For all sufficiently large $n$ this range is
contained in $0\le\ell\le\kappa$.
Let
\[
\sigma:=e^{-v_{\min}/4}<1.
\]
For all sufficiently large $n$ and all such $\ell$ one has
$\kappa-\ell+1\ge n/4$, hence
\[
|E_-(n,\ell;r)|,\ |E_0(n;r)|,\ |E_+(n,\ell;r)|\le \sigma.
\]
In particular, the series defining $\Phi_0$, $\Phi_1$, and $\Phi_2$ converge
absolutely.

Choose
\[
M:=\lfloor n^{1/8}\rfloor.
\]
Since $|t|=O_{\mathcal D}(n^{-3/4})$, one has
$M|t|=O_{\mathcal D}(n^{-5/8})$. The function
\[
x\longmapsto
\frac{1}{1-e^{-x}}-\frac1x-\frac12-\frac{x}{12}
\]
is analytic at $x=0$. Therefore, uniformly for $1\le m\le M$,
\begin{equation}\label{eq:inv-den-expansion}
\frac{1}{1-e^{-m t}}
=
\frac{1}{m t}+\frac12+\frac{m t}{12}
+O_{\mathcal D}(m^3|t|^3).
\end{equation}

Set
\[
c_m:=E_-(n,\ell;r)^m-2E_0(n;r)^m+E_+(n,\ell;r)^m.
\]
By Lemma~\ref{lem:complex-middle-exact},
\[
\log\left|\frac{\pqbinom{n}{\kappa+\ell}{r}}{\pqbinom{n}{\kappa}{r}}\right|
=
-\Real\sum_{m=1}^{\infty}\frac{c_m}{m(1-e^{-m t})}.
\]
For $m\le M$, insert \eqref{eq:inv-den-expansion}. Since $|c_m|\le4\sigma^m$,
\[
\sum_{m=1}^{M}
\left|\frac{c_m}{m}\,O_{\mathcal D}(m^3|t|^3)\right|
\le
4|t|^3\sum_{m=1}^{\infty}m^2\sigma^m
=
O_{\mathcal D,L}(n^{-9/4}).
\]
Hence
\[
\sum_{m=1}^{M}\frac{c_m}{m(1-e^{-m t})}
=
\frac{1}{t}\sum_{m=1}^{M}\frac{c_m}{m^2}
+
\frac12\sum_{m=1}^{M}\frac{c_m}{m}
+
\frac{t}{12}\sum_{m=1}^{M}c_m
+O_{\mathcal D,L}(n^{-9/4}).
\]

It remains to estimate the remaining part. For $m\ge1$,
\[
|1-e^{-m t}|
\ge 1-e^{-m v/n}
\ge c_{\mathcal D}\min\left(\frac{m}{n},1\right),
\]
where $c_{\mathcal D}>0$ depends only on $v_{\min}$. Therefore
\[
\sum_{m=M+1}^{\infty}
\left|\frac{c_m}{m(1-e^{-m t})}\right|
\le
C_{\mathcal D}\sum_{m=M+1}^{\infty}
\frac{\sigma^m}{m\min(m/n,1)}
=
O_{\mathcal D}\!\left(e^{-c_{\mathcal D} n^{1/8}}\right).
\]
Also,
\[
\sum_{m=M+1}^{\infty}|c_m|
+
\sum_{m=M+1}^{\infty}\left|\frac{c_m}{m}\right|
\le
C_{\mathcal D}\sum_{m=M+1}^{\infty}\sigma^m
=
O_{\mathcal D}\!\left(e^{-c_{\mathcal D} n^{1/8}}\right),
\]
and, since $|t|\ge v_{\min}/n$,
\[
\frac{1}{|t|}
\sum_{m=M+1}^{\infty}\left|\frac{c_m}{m^2}\right|
\le
\frac{4}{|t|}\sum_{m=M+1}^{\infty}\frac{\sigma^m}{m^2}
=
O_{\mathcal D}\!\left(e^{-c_{\mathcal D} n^{1/8}}\right).
\]
The exponential bounds absorb the polynomial factors in $n$. Combining the
estimates proves \eqref{eq:complex-middle-main}. The formula
\eqref{eq:complex-middle-g} follows from
\[
g(\kappa+\ell)-g(\kappa)
=
\log\left|\frac{\pqbinom{n}{\kappa+\ell}{r}}{\pqbinom{n}{\kappa}{r}}\right|
+\frac{u}{n}\,\ell^2,
\]
which is the contribution of the factor $p^{k(n-k)}$.
Finally, \eqref{eq:Phi0-rational}, \eqref{eq:Phi1-log}, and
\eqref{eq:Phi2-dilog} follow from the absolutely convergent power series for
$w/(1-w)$, $-\log(1-w)$, and $\Li_2(w)$ on $|w|<1$.
\end{proof}


The following symmetric Green identity converts second-order differences to
centred ratios. It is used in Theorem~\ref{thm:complex-sqrt} and again in
Section~\ref{sec:small-imaginary-perturbation}.

\begin{lemma}\label{lem:green}
Let $f:\mathbb Z\to\mathbb C$ and define
\[
(\Delta^2 f)(k):=f(k+1)-2f(k)+f(k-1).
\]
Fix an integer $\kappa$.
Then for every integer $m\ge1$,
\begin{equation}\label{eq:green-symmetric}
f(\kappa+m)+f(\kappa-m)-2f(\kappa)
=
\sum_{j=-(m-1)}^{m-1}(m-|j|)\,\Delta^2 f(\kappa+j).
\end{equation}
\end{lemma}

\begin{proof}
Set $\Delta f(k):=f(k+1)-f(k)$.
Then
\[
f(\kappa+m)-f(\kappa)=\sum_{u=0}^{m-1}\Delta f(\kappa+u),
\qquad
f(\kappa-m)-f(\kappa)=-\sum_{u=1}^{m}\Delta f(\kappa-u).
\]
Adding these two identities gives
\[
f(\kappa+m)+f(\kappa-m)-2f(\kappa)
=
\sum_{u=0}^{m-1}\left(\Delta f(\kappa+u)-\Delta f(\kappa-u-1)\right).
\]
For each $u\ge0$,
\[
\Delta f(\kappa+u)-\Delta f(\kappa-u-1)
=
\sum_{j=-u}^{u}\Delta^2 f(\kappa+j)
\]
by telescoping.
Summing over $u=0,\dots,m-1$ and exchanging the order of summation, the coefficient of $\Delta^2 f(\kappa+j)$ is the number of integers $u\in\{0,\dots,m-1\}$ with $|j|\le u$, namely $m-|j|$.
This proves \eqref{eq:green-symmetric}.
\end{proof}

We also give a complex estimate for deviations of order $\sqrt n$. It follows
from the second expansion in Proposition~\ref{prop:complex-local} and the
same Green identity.

\begin{theorem}\label{thm:complex-sqrt}
Let $z>0$, $\alpha\in\mathbb R$, and
\[
r=\exp(-t),
\qquad
t:=\frac zn+i\frac{\alpha}{n^{3/4}}.
\]
Assume that $n$ is even and write $\kappa:=n/2$. Put
\[
\tau:=z+i\alpha n^{1/4},
\qquad
w:=\frac{\tau}{2},
\]
and define
\[
F(w):=\frac{1}{e^w-1},
\qquad
G(w):=\frac{e^w}{(e^w-1)^2},
\qquad
H(w):=\frac{e^w(e^w+1)}{(e^w-1)^3}.
\]
For each fixed $L>0$, uniformly for all integers $\ell$ with
$|\ell|\le L\sqrt n$,
\begin{align}
\log\left|\frac{\pqbinom{n}{\kappa+\ell}{r}}{\pqbinom{n}{\kappa}{r}}\right|
&=
-\ell^2\Real\left(tF(w)\right)
+\frac{\ell^2}{2}\Real\left(t^2G(w)\right)
-\frac{\ell^2(\ell^2-1)}{12}\Real\left(t^3H(w)\right)
\nonumber\\
&\quad
+O_{z,\alpha,L}\!\left(n^3|t|^5+n^2|t|^4+n|t|^3\right).
\label{eq:complex-sqrt}
\end{align}
Equivalently, writing $x=\ell/\sqrt n$, one has
\begin{align}
\log\left|\frac{\pqbinom{n}{\kappa+\ell}{r}}{\pqbinom{n}{\kappa}{r}}\right|
&=
-x^2 C_n(z,\alpha)
+\frac{x^2}{2n}\Real\left(\tau^2G(w)\right)
-\frac{x^4}{12n}\Real\left(\tau^3H(w)\right)
\nonumber\\
&\quad
+O_{z,\alpha,L}\!\left(n^3|t|^5+n^2|t|^4+n|t|^3\right),
\label{eq:complex-sqrt-x}
\end{align}
where
\begin{equation}\label{eq:Cn-def}
C_n(z,\alpha):=
\Real\left(\tau F(w)\right)
=
\Real\left(
\frac{z+i\alpha n^{1/4}}{e^{z/2+i\alpha n^{1/4}/2}-1}
\right).
\end{equation}
\end{theorem}

\begin{proof}
The quantity on the left of \eqref{eq:complex-sqrt} is even in $\ell$, and
the right-hand side has the same symmetry. It is therefore enough to consider
$0\le \ell\le L\sqrt n$. The case $\ell=0$ is immediate.
Set
\[
f_n(k):=\log \left|\pqbinom{n}{k}{r}\right|.
\]
Then
\[
\Delta^2 f_n(k)=\log|R(n,k;r)|.
\]
For $|j|\le L\sqrt n$ and all sufficiently large $n$,
\[
\frac z4\le \Real(w\pm jt)\le \frac{3z}{4}.
\]
On this vertical strip the required derivatives of $F$ and $G$ are bounded,
with constants depending only on $z$, $\alpha$, and $L$. Since $F''=H$,
Taylor's formula gives
\[
F(w+jt)+F(w-jt)
=
2F(w)+j^2t^2H(w)+O_{z,\alpha,L}(j^4|t|^4)
\]
and
\[
G(w+jt)+G(w-jt)
=
2G(w)+O_{z,\alpha,L}(j^2|t|^2).
\]
The second expansion in Proposition~\ref{prop:complex-local}, applied to
the fixed pair $(z,\alpha)$, gives
\begin{align*}
\Delta^2 f_n(\kappa+j)
&=
-\Real\left[
t\left(F(w+jt)+F(w-jt)\right)
\right]\\
&\quad+
\frac12\Real\left[
t^2\left(G(w+jt)+G(w-jt)\right)
\right]
+O_{z,\alpha,L}(|t|^3).
\end{align*}
Hence
\begin{align*}
\Delta^2 f_n(\kappa+j)
&=
-2\Real\left(tF(w)\right)
+\Real\left(t^2G(w)\right)
-j^2\Real\left(t^3H(w)\right)\\
&\quad
+O_{z,\alpha,L}\!\left(j^4|t|^5+j^2|t|^4+|t|^3\right)
\end{align*}
uniformly for $|j|\le L\sqrt n$.

Lemma~\ref{lem:green}, applied to $f_n$, gives
\[
2\left(f_n(\kappa+\ell)-f_n(\kappa)\right)
=
\sum_{j=-(\ell-1)}^{\ell-1}(\ell-|j|)\Delta^2 f_n(\kappa+j).
\]
Using
\[
\sum_{j=-(\ell-1)}^{\ell-1}(\ell-|j|)=\ell^2,
\qquad
\sum_{j=-(\ell-1)}^{\ell-1}(\ell-|j|)j^2
=
\frac{\ell^2(\ell^2-1)}6,
\]
and
\[
\sum_{j=-(\ell-1)}^{\ell-1}(\ell-|j|)j^4=O(\ell^6),
\]
we obtain
\begin{align*}
f_n(\kappa+\ell)-f_n(\kappa)
&=
-\ell^2\Real\left(tF(w)\right)
+\frac{\ell^2}{2}\Real\left(t^2G(w)\right)\\
&\quad
-\frac{\ell^2(\ell^2-1)}{12}\Real\left(t^3H(w)\right)\\
&\quad
+O_{z,\alpha,L}\!\left(\ell^6|t|^5+\ell^4|t|^4+\ell^2|t|^3\right).
\end{align*}
Since $\ell=O_L(\sqrt n)$, this proves \eqref{eq:complex-sqrt}. Replacing
$t$ by $\tau/n$ and $\ell$ by $x\sqrt n$ gives
\eqref{eq:complex-sqrt-x}; the difference between
$\ell^2(\ell^2-1)t^3/12$ and $x^4\tau^3/(12n)$ is absorbed by the
error term.
\end{proof}

\begin{remark}
The preceding theorem gives a direct comparison with the real critical value
for the full $(p,q)$-coefficient. Assume that $|p|=e^{-u/n}$ and let
$g(k)=\log\left|\pqbinom{n}{k}{p,q}\right|$. If $|x|\le L$ with
$x=\ell/\sqrt n$ and fixed $L$, then \eqref{eq:complex-sqrt-x} and the
factor $p^{k(n-k)}$ give
\begin{align}
g(\kappa+\ell)-g(\kappa)
&=
x^2\left(u-C_n(z,\alpha)\right)
+\frac{x^2}{2n}\Real\left(\tau^2G(w)\right)
-\frac{x^4}{12n}\Real\left(\tau^3H(w)\right)
\nonumber\\
&\quad
+O_{z,\alpha,L}\!\left(n^3|t|^5+n^2|t|^4+n|t|^3\right).
\label{eq:complex-sqrt-full-modulus}
\end{align}
For $\alpha=0$,
\[
C_n(z,0)=\frac{z}{e^{z/2}-1}=\frac{A(z)}2.
\]
Thus the choice $u=A(z)/2$ makes the first term in
\eqref{eq:complex-sqrt-full-modulus} vanish when $r$ is real. If
$\alpha\ne0$, write $b=\alpha n^{1/4}/2$. Then
\begin{equation}\label{eq:Cn-explicit}
C_n(z,\alpha)=
\frac{z(e^{z/2}\cos b-1)+\alpha n^{1/4}e^{z/2}\sin b}
{e^z-2e^{z/2}\cos b+1}.
\end{equation}
Hence the same choice $u=A(z)/2$ gives the coefficient
$A(z)/2-C_n(z,\alpha)$, which depends on $n$ when $\alpha\ne0$.
\end{remark}


\section{An imaginary perturbation of order $n^{-5/4}$ at the real critical value}\label{sec:small-imaginary-perturbation}

The main complex theorem treats $\Imag t$ of order $n^{-3/4}$ in $r=q/p$.
For comparison with the positive real critical value, we also treat an
imaginary perturbation of order $n^{-5/4}$.  For $|\ell|=O(n^{3/4})$ this perturbation changes
the quadratic coefficient but leaves the quartic coefficient unchanged.

\begin{proposition}\label{prop:small-imaginary-perturbation}
Let $z>0$, let $\alpha,\gamma\in\mathbb R$, and assume that $n$ is even. Put
$\kappa:=n/2$ and
\[
t:=\frac zn+i\frac{\alpha}{n^{5/4}},
\qquad
r:=e^{-t},
\qquad
u:=\frac{A(z)}2+\frac{\gamma}{\sqrt n},
\]
\[
p:=e^{-u/n},
\qquad
q:=p r.
\]
For every fixed $L>0$, uniformly for all integer sequences $\ell_n$ such that
$x_n:=\ell_n/n^{3/4}$ satisfies $|x_n|\le L$,
\begin{equation}\label{eq:small-imaginary-perturbation}
\left|
\frac{\pqbinom{n}{\kappa+\ell_n}{p,q}}
{\pqbinom{n}{\kappa}{p,q}}
\right|
=
\exp\!\left(
\left(\gamma+\frac{\alpha^2}{4}A''(z)\right)x_n^2
-\frac{B(z)}{12}x_n^4
+O_{z,\alpha,\gamma,L}(n^{-1/2})\right).
\end{equation}
The same formula holds with $\kappa-\ell_n$ in place of $\kappa+\ell_n$.
In particular, if $x_n\to x$, then
\[
\left|
\frac{\pqbinom{n}{\kappa\pm \ell_n}{p,q}}
{\pqbinom{n}{\kappa}{p,q}}
\right|
\longrightarrow
\exp\!\left(
\left(\gamma+\frac{\alpha^2}{4}A''(z)\right)x^2
-\frac{B(z)}{12}x^4
\right).
\]
For $\gamma=0$, the imaginary perturbation contributes the quadratic term
$\alpha^2A''(z)x^2/4$ at the real critical value $2u=A(z)$.
\end{proposition}

\begin{proof}
Let
\[
\tau:=nt=z+i\alpha n^{-1/4}.
\]
For $\tau$ in a neighbourhood of $z$ define
\[
\mathcal A(\tau):=\tau\left(\coth\frac{\tau}{4}-1\right),
\qquad
\mathcal B(\tau):=\frac{\tau^3\sinh(\tau/2)}{8\sinh^4(\tau/4)}.
\]
We first prove the local curvature estimate
\begin{equation}\label{eq:small-complex-logR}
\log\left|R(n,\kappa+j;e^{-t})\right|
=
-\frac{\Real\mathcal A(\tau)}{n}
-\frac{\Real\mathcal B(\tau)}{n^3}j^2
+O_{z,\alpha,L}(n^{-2})
\end{equation}
uniformly for $|j|\le L n^{3/4}$.
For all sufficiently large $n$, the points used below lie in a compact subset
of the domain on which one branch of $\log\sinh$ is analytic. Put
\[
a:=\frac{\tau}{4},
\qquad
h:=\frac{\tau}{2n},
\qquad
\Delta:=\frac{\tau j}{2n},
\]
and write $\psi(y):=\log\sinh y$ on this compact set. Lemma~\ref{lem:R-hyperbolic}
gives
\[
\log\left|R(n,\kappa+j;e^{-t})\right|
=
\Real\left(
\frac{\tau}{n}
+\psi(a+\Delta)+\psi(a-\Delta)
-\psi(a+h+\Delta)-\psi(a+h-\Delta)
\right).
\]
Taylor's formula in $h$ and then in $\Delta$ gives, uniformly in the
stated range of $j$,
\[
\log|R|
=
\Real\left(
\frac{\tau}{n}
-h\left(\psi'(a+\Delta)+\psi'(a-\Delta)\right)
\right)
+O_{z,\alpha,L}(n^{-2})
\]
and
\[
\psi'(a+\Delta)+\psi'(a-\Delta)
=
2\psi'(a)+\psi^{(3)}(a)\Delta^2+O_{z,\alpha,L}(\Delta^4).
\]
Since $h\Delta^4=O_{z,\alpha,L}(n^{-2})$, and since
$\psi'(a)=\coth a$ and
\[
\psi^{(3)}(a)=\frac{\sinh(2a)}{\sinh^4 a},
\]
this proves \eqref{eq:small-complex-logR}.

Let
\[
g_n(k):=\log\left|\pqbinom{n}{k}{p,q}\right|.
\]
By Corollary~\ref{cor:local-ratios},
\[
\Delta^2g_n(\kappa+j)
=
\log\left|R(n,\kappa+j;e^{-t})\right|+\frac{2u}{n}.
\]
Together with \eqref{eq:small-complex-logR}, this gives
\[
\Delta^2g_n(\kappa+j)
=
\frac{2u-\Real\mathcal A(\tau)}{n}
-\frac{\Real\mathcal B(\tau)}{n^3}j^2
+O_{z,\alpha,L}(n^{-2}).
\]
Applying Lemma~\ref{lem:green} gives, uniformly for $|\ell|\le L n^{3/4}$,
\begin{align*}
g_n(\kappa+\ell)-g_n(\kappa)
&=
\frac{2u-\Real\mathcal A(\tau)}{2n}\ell^2
-\frac{\Real\mathcal B(\tau)}{12n^3}\ell^2(\ell^2-1)\\
&\quad
+O_{z,\alpha,L}\left(\frac{\ell^2}{n^2}\right).
\end{align*}
Now $\mathcal A$ and $\mathcal B$ are analytic near $z$ and real on the real
axis. Hence
\[
\Real\mathcal A(z+i\alpha n^{-1/4})
=
A(z)-\frac{\alpha^2}{2\sqrt n}A''(z)+O_{z,\alpha}(n^{-1}),
\]
and
\[
\Real\mathcal B(z+i\alpha n^{-1/4})=B(z)+O_{z,\alpha}(n^{-1/2}).
\]
Since $u=A(z)/2+\gamma n^{-1/2}$ and $\ell=x_n n^{3/4}$, we obtain
\[
g_n(\kappa+\ell)-g_n(\kappa)
=
\left(\gamma+\frac{\alpha^2}{4}A''(z)\right)x_n^2
-\frac{B(z)}{12}x_n^4
+O_{z,\alpha,\gamma,L}(n^{-1/2}).
\]
This proves \eqref{eq:small-imaginary-perturbation}. The formula with
$\kappa-\ell_n$ follows from the coefficient symmetry.
\end{proof}


\subsection*{Proof of Theorems A and B}

\begin{proof}
The one-parameter formula \eqref{eq:main-sqrt-r} is
Theorem~\ref{thm:complex-sqrt}, with $z=v$ and $\alpha=\delta$.
By Lemma~\ref{lem:pq-factor},
\[
\log\left|
\frac{\pqbinom{n}{\kappa+\ell}{p,q}}
     {\pqbinom{n}{\kappa}{p,q}}
\right|
=
\log\left|
\frac{\pqbinom{n}{\kappa+\ell}{r}}{\pqbinom{n}{\kappa}{r}}
\right|
+\frac{u}{n}\ell^2.
\]
Since $x=\ell/\sqrt n$, this identity and
\eqref{eq:main-sqrt-r} give \eqref{eq:main-sqrt-pq}. The identity
$C_n(v,0)=A(v)/2$ follows from
\[
\frac{v}{e^{v/2}-1}
=
\frac{v}{2}\left(\coth\frac{v}{4}-1\right).
\]
Finally, \eqref{eq:main-Cn-explicit} follows by multiplying the numerator
and denominator in \eqref{eq:main-Cn} by
\[
e^{v/2-i\delta n^{1/4}/2}-1
\]
and taking real parts. This proves Theorem A.

Theorem B is Proposition~\ref{prop:small-imaginary-perturbation}; equation
\eqref{eq:intro-small-imaginary-perturbation} is the same as
\eqref{eq:small-imaginary-perturbation}.
\end{proof}


\section{Conclusion and future work}\label{sec:conclusion}

We have proved local complex estimates for near-central
$(p,q)$-binomial coefficients. The exact factorisation separates the explicit
quadratic factor from the one-parameter coefficient array. The second-order ratio
has an absolutely convergent logarithmic representation that does not require
choosing a logarithm branch, and the curvature estimates give centred complex
ratios. The imaginary perturbation of order $n^{-5/4}$ shows that a term which
is absent from $\ell\Imag t$ at leading order can nevertheless change the
quadratic coefficient at the real critical value.

One continuation is to obtain summation estimates for generating polynomials
without assuming positivity. In the positive real case, such arguments rely on
a uniform tail estimate. For complex coefficients, one also has to control
cancellation away from the centre.

A separate problem is to locate maxima of the coefficient moduli. The exact
adjacent ratio is
\[
\frac{\left|\pqbinom{n}{k+1}{p,q}\right|}
     {\left|\pqbinom{n}{k}{p,q}\right|}
=
|p|^{\,n-2k-1}
\frac{|1-r^{n-k}|}{|1-r^{k+1}|},
\qquad r=\frac qp.
\]
A maximum occurs where this ratio crosses $1$. Deriving an asymptotic formula
for that crossing when the maxima move by order $n^{3/4}$ requires a separate
analysis; it does not follow from the bounded-$x$ expansion in Theorem A.

A further problem is the full two-exponent theory for
\[
t=\frac{v_0}{n^b}+i\frac{\delta_0}{n^a}.
\]
The exact centred identities remain valid when $\Real t>0$, but the limiting
functions and the appropriate coefficient normalisations depend on $a$ and
$b$. A complete classification would have to include $b=1$, where $n t$ has
a finite non-zero limit; $b>1$, where the radial part is closer to the ordinary
binomial case; $b<1$, where $n t$ diverges; and the oscillatory behaviour of
the arguments when $a<3/4$ under the quartic normalisation.

\section*{Acknowledgments}
The first-named author was partially funded by the Swedish Research
Council under grant agreement no.~2025-05053 and by the Swedish Energy Agency
under project no.~P2025-04323. The second-named author was partially
funded by the National Science Centre, Poland, under the Weave-UNISONO
programme, grant no.~UMO-2025/07/Y/ST1/00146.

\begin{thebibliography}{99}
\bibitem{AhagCzyzLundow2024}
P. {\AA}hag, R. Czy{\.z}, and P. H. Lundow,
``On a generalised Lambert $W$ branch transition function arising from
$p,q$-binomial coefficients,''
\emph{Appl. Math. Comput.} \textbf{462} (2024), 128347;
\url{https://doi.org/10.1016/j.amc.2023.128347}.

\bibitem{AhagCzyzLundow2025}
P. {\AA}hag, R. Czy{\.z}, and P. H. Lundow,
``Complex branches of a generalized Lambert $W$ function arising from
$p,q$-binomial coefficients,''
\emph{Ann. Polon. Math.} \textbf{134} (2025), no.~1, 1--18;
\url{https://doi.org/10.4064/ap240830-1-7}.


\bibitem{AtakishiyevNagiyev1994}
N. M. Atakishiyev and Sh. M. Nagiyev,
``On the Rogers--Szeg\H{o} polynomials,''
\emph{J. Phys. A: Math. Gen.} \textbf{27} (1994), no.~17, L611--L615;
\url{https://doi.org/10.1088/0305-4470/27/17/003}.

\bibitem{Corcino2008}
R. B. Corcino,
``On $p,q$-binomial coefficients,''
\emph{Integers} \textbf{8} (2008), A29, 16 pp.

\bibitem{Dhand2014}
V. Dhand,
``A combinatorial proof of strict unimodality for $q$-binomial coefficients,''
\emph{Discrete Math.} \textbf{335} (2014), 20--24;
\url{https://doi.org/10.1016/j.disc.2014.07.001}.

\bibitem{GasperRahman2004}
G. Gasper and M. Rahman,
\emph{Basic Hypergeometric Series},
2nd ed., Encyclopedia of Mathematics and its Applications, vol.~96,
Cambridge University Press, Cambridge, 2004;
\url{https://doi.org/10.1017/CBO9780511526251}.

\bibitem{JagannathanSridhar2010}
R. Jagannathan and R. Sridhar,
``$(p,q)$-Rogers--Szeg\H{o} polynomial and the $(p,q)$-oscillator,''
in \emph{The Legacy of Alladi Ramakrishnan in the Mathematical Sciences},
K. Alladi, J. R. Klauder, and C. R. Rao (eds.),
Springer, New York, 2010, 491--501;
\url{https://doi.org/10.1007/978-1-4419-6263-8_29}.

\bibitem{LubinskySaff1987}
D. S. Lubinsky and E. B. Saff,
``Convergence of Pad\'e approximants of partial theta functions and the
Rogers--Szeg\H{o} polynomials,''
\emph{Constr. Approx.} \textbf{3} (1987), 331--361;
\url{https://doi.org/10.1007/BF01890574}.

\bibitem{LundowRosengren2010}
P. H. Lundow and A. Rosengren,
``On the $p,q$-binomial distribution and the Ising model,''
\emph{Philos. Mag.} \textbf{90} (2010), no.~24, 3313--3353;
\url{https://doi.org/10.1080/14786435.2010.484406}.

\bibitem{LundowRosengren2013}
P. H. Lundow and A. Rosengren,
``The $p,q$-binomial distribution applied to the 5d Ising model,''
\emph{Philos. Mag.} \textbf{93} (2013), no.~14, 1755--1770;
\url{https://doi.org/10.1080/14786435.2012.750770}.

\bibitem{DLMF17}
NIST Digital Library of Mathematical Functions,
\emph{\S17.2(ii) $q$-Binomial Coefficients},
National Institute of Standards and Technology,
\url{https://dlmf.nist.gov/17.2#ii}.

\bibitem{OHara1990}
K. M. O'Hara,
``Unimodality of Gaussian coefficients: a constructive proof,''
\emph{J. Combin. Theory Ser. A} \textbf{53} (1990), no.~1, 29--52;
\url{https://doi.org/10.1016/0097-3165(90)90018-R}.

\bibitem{PakPanova2013}
I. Pak and G. Panova,
``Strict unimodality of $q$-binomial coefficients,''
\emph{C. R. Math. Acad. Sci. Paris} \textbf{351} (2013), no.~11--12, 415--418;
\url{https://doi.org/10.1016/j.crma.2013.06.008}.

\bibitem{StanleyEC1}
R. P. Stanley,
\emph{Enumerative Combinatorics. Vol.~1},
2nd ed., Cambridge Studies in Advanced Mathematics, vol.~49,
Cambridge University Press, Cambridge, 2012;
\url{https://doi.org/10.1017/CBO9781139058520}.

\bibitem{SuWang2012}
X.-T. Su and Y. Wang,
``Proof of a conjecture of Lundow and Rosengren on the bimodality of
$p,q$-binomial coefficients,''
\emph{J. Math. Anal. Appl.} \textbf{391} (2012), no.~2, 653--656;
\url{https://doi.org/10.1016/j.jmaa.2012.02.049}.

\bibitem{Szego1926}
G. Szeg\H{o},
``Ein Beitrag zur Theorie der Thetafunktionen,''
\emph{Sitzungsberichte der Preussischen Akademie der Wissenschaften,
Physikalisch-Mathematische Klasse} (1926), 242--252.

\end{thebibliography}

\end{document}