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

\subsection{The question}

The torus partition function of a compact unitary two-dimensional
conformal field theory is invariant under the modular group and expands
in characters of the Virasoro algebra with nonnegative integer
multiplicities. The modular bootstrap studies the constraints that
these two properties place on the spectrum; the most basic is an upper
bound on the dimension $\Delta_{\rm gap}$ of the lightest nonvacuum
primary \cite{Hellerman,FriedanKeller,CLY,HMR}. Most such bounds use
only positivity of the multiplicities, and therefore hold for a larger
class of ``partition functions'' with real nonnegative multiplicities.
Integrality is harder to exploit \cite{KaidiPerlmutter}, but it can be
decisive.

A striking instance was found recently by Benjamin, Langhoff, Li and
Thaler \cite{BLLT}. For central charge $c$ slightly above one, and for
parity-symmetric spectra without conserved currents beyond the Virasoro
algebra, the numerical bound from positivity alone is
$\Delta_{\rm gap}\leq c/6+1/3$ \cite{CLY}, close to $\frac12$. Requiring in addition that the number of
primaries in certain windows be an integer, they obtained numerical
bounds that are smaller by an order of magnitude at
$c-1=3\times10^{-3}$, and suggested that the gap might vanish as
$c\to1^+$. No compact unitary theory with $c$ slightly above one is
known, and whether one exists is open. At $c=1$, by contrast, the
compact boson has a positive gap at every radius. The purpose of this
paper is to prove that, in the class studied in \cite{BLLT}, a positive
gap cannot persist as $c\to1^+$.

\subsection{The result}

Write $c=1+12a$. We consider functions
\[
 Z(\tau)=|\chi_{\vac}(\tau)|^2+\sum_j\chi_{h_j}(\tau)\,
 \overline{\chi_{\bar h_j}(\tau)},
 \qquad\chi_h=\frac{q^{h-a/2}}{\eta},\quad
 \chi_{\vac}=\frac{q^{-a/2}(1-q)}{\eta},
\]
on the upper half-plane, where each pair of weights $(h_j,\bar h_j)$ is
repeated according to its multiplicity. We call $Z$ \emph{admissible of
gap $g$} if it is invariant under $\tau\mapsto-1/\tau$ and
$\tau\mapsto\tau+1$, the sum converges for all $\tau$, the spins
$h_j-\bar h_j$ are integers, the spectrum is invariant under
$(h,\bar h)\mapsto(\bar h,h)$, and
\[
 h_j>0,\qquad\bar h_j>0,\qquad\Delta_j=h_j+\bar h_j\geq g
\]
for every $j$; see Definition~\ref{def:admissible}. The first two
inequalities say that the vacuum is the only state at the unitarity
bound $\Delta=|s|$. No lower bound on the twist $\Delta_j-|s_j|$ is
assumed, so currents are allowed to emerge in a limit.

\begin{maintheorem}\label{thm:main}
For every $g>0$ there exists $\epsilon(g)>0$ such that no admissible
partition function of gap $g$ has central charge in
$(1,1+\epsilon(g))$. Equivalently, if $Z_n$ are admissible at central
charges $c_n\downarrow1$, then their gaps tend to zero.
\end{maintheorem}

The statement concerns partition functions and makes no reference to
correlation functions. It does not assert that admissible partition
functions with $c$ close to one exist. The threshold $\epsilon(g)$ is
obtained by contradiction from a compactness argument, and the proof
gives no numerical value for it.

\subsection{Outline of the proof}

Suppose that admissible partition functions of gap $g$ exist along a
sequence $a\downarrow0$. The argument has five steps.

\smallskip\noindent\emph{Limits} (Sections~\ref{sec:setup}
and~\ref{sec:classification}). Modular invariance and positivity bound
the number of states below any energy uniformly in $a$, so a
subsequence converges to a modular-invariant function at $c=1$ which is
still a nonnegative integer combination of the characters
$q^h\bar q^{\bar h}/|\eta|^2$ plus the vacuum term
$|1-q|^2/|\eta|^2$. This is a stronger positivity property than
positivity in the irreducible characters at $c=1$. The momenta of such
a function form a self-dual crystalline measure in the plane, and the
theorem of Eberhardt and Levine \cite{EL} reduces it to a finite
combination of compact-boson partition functions $Z_r$. Examining the
coefficients of a few monomials, we find that the limit is
\[
 \tfrac12(Z_R+Z_t),\qquad 1\leq R,t\leq(2g)^{-1/2},
\]
with at least one of $2R$, $2t$ an integer
(Theorem~\ref{thm:classification}). The half-integrality comes from
the scalar of dimension two that the vacuum character subtracts.

\smallskip\noindent\emph{An obstruction at zero momentum}
(Section~\ref{sec:rigidity}). In the variables $h=P^2$,
$\bar h=Q^2$ a character is a Gaussian in the momentum
$\sqrt{h-a/2}$, and modular invariance becomes invariance under a
Fourier transform. At $a=0$ the momenta of $\frac12(Z_R+Z_t)$ lie on
two circle lattices. At $a>0$ the vacuum sits at
$P^2=Q^2=-a/2$; expanded about the origin, this displacement is the
distribution $-\frac a4(\partial_P^2+\partial_Q^2)\delta_0$. We show
that a distribution supported on a finite union of circle lattices and
invariant under the Fourier transform and under the spin phase is a
polynomial in the boost generator applied to the lattice sums
(Theorems~\ref{thm:circle-union} and~\ref{thm:circle-jets}); such a
distribution has no derivative terms at the origin. A linear-algebra
argument, which is possible because the rigidity statement holds
without any growth condition, converts this into a finite family of
Gaussian test functions that detect an origin derivative
(Theorem~\ref{thm:origin-certificate}). The consequence is
Theorem~\ref{thm:moving-localization}: it is impossible that every
state with $\Delta\leq u_*\log(1/a)$, for some $u_*>1/(2\pi)$, lies
within $a^p$ in dimension, with the same spin, of a point of a fixed
finite union of circle lattices together with one circle lattice of
variable radius. The logarithmic window is what makes the tail of the
spectrum negligible against $a$.

\smallskip\noindent\emph{Localization by positive kernels}
(Section~\ref{sec:kernels}). To produce such a localization we use
functions $F\geq0$ that are Gaussians times trigonometric polynomials
vanishing on the two limiting lattices and at the vacuum; a functional
of this type for the self-dual boson, built from a triangular function
and its Fourier transform, appears in \cite[Section~3.2]{ACHT}. The
Fourier
transform of $F$ is a finite sum of narrow Gaussian peaks, and the
exact identity $\mathcal L_a(F)=\mathcal L_a(\mathbb SF)$ between the
spectral sums bounds $\sum_jF(X_j)$ by $O(a\log^2(1/a))$ once the
negative peaks are matched by states whose positions are known. For a
limit $\frac12(Z_R+Z_t)$ in which $t$ is not a half-integer, the
trigonometric polynomial can be chosen, depending on the arithmetic of
$t$, so that this is the case, and one obtains the forbidden
localization. Hence both radii of a limit are half-integers
(Theorem~\ref{thm:half-integer-limits}).

\smallskip\noindent\emph{The scale of the departure}
(Sections~\ref{sec:kernels} and~\ref{sec:collision}). The finitely many
remaining limits lie on a common rational grid. Four scalar states at
one fixed energy now play the role of the anchors, and their
displacement $\delta_a$ controls the distance of the whole spectrum
from the grid. By the obstruction, $\delta_a$ cannot be bounded by a
power of $a$ (Theorem~\ref{thm:four-scalar-slow}). A quantitative form
of the rigidity theorem then shows that, on each fixed finite set of
states, the spectrum agrees to a high power of $\delta_a$ with that of
$\frac12(Z_R+Z_{t_a})$ for a half-integer $R$ and a radius $t_a$ with
$|t_a-T|$ comparable to $\delta_a$, where $T$ is the half-integer limit
of $t_a$ (Theorem~\ref{thm:comparison} and
\eqref{eq:col-theta-delta}).

\smallskip\noindent\emph{Resolving the collision}
(Section~\ref{sec:collision}). Since $t_a$ is not a half-integer, the
kernels of the third step apply at the radius $t_a$, but their peaks
and zeros now collide as $t_a\to T$. Using Gaussians narrow enough to
resolve the collisions costs only powers of $|t_a-T|^{-1}$, which are
smaller than any negative power of $a$ by the fourth step. The result
is again a localization within a power of $a$ of fixed lattices and the
moving lattice $\Lambda_{t_a}$, contradicting
Theorem~\ref{thm:moving-localization}.

\smallskip
Section~\ref{sec:effective} discusses what is needed to make
$\epsilon(g)$ effective. All test functions used in the proof are
finite or compact superpositions of Gaussians, for which the modular
identity is exact; the argument does not use a finite number of
derivatives of the modular crossing equation at $\tau=i$, and no formal
power series in $c-1$ is summed.

\subsection{Remarks}

The hypotheses are those of a torus partition function, and the limits
that have to be excluded reflect this. The average of two compact-boson
partition functions at different radii is not the partition function of
a conformal field theory, but it satisfies every condition that
survives the limit, and we know of no argument that discards it using
the torus data alone. These averages are responsible for the moving
radius in the second step and for most of the work in the last two.

Parity symmetry is included in the definition because it is part of
the setting of \cite{BLLT}. No step of the proof uses it: the
classification of Section~\ref{sec:classification} does not assume it,
and states of opposite spin are treated separately throughout
(Remark~\ref{rem:col-parity}). The assumption that no nonvacuum state has
$h=0$ or $\bar h=0$ is essential to the method: it makes the long
multiplicities at $c>1$ nonnegative term by term, and this positivity
is inherited by the limit.

\subsection*{Notation}
Section~\ref{sec:setup} fixes the conventions. Throughout,
$a=(c-1)/12$, $L=\log(1/a)$, $X=(P,Q)$ are momenta and
$(u,v)=(P+Q,P-Q)$ light-cone coordinates, $\Lambda_r$ is the circle
lattice of radius $r$, and a radius is called a half-integer if twice
it is an integer.