Irrational CFT, Fusion Kernels, and Crossing

When Virasoro intermediate representations form a continuum, changing the OPE channel is an integral transform rather than multiplication by a finite fusing matrix. The fusion kernel is the change-of-basis coefficient between normalized chiral conformal blocks. It is kinematical data fixed by the chiral algebra and external representations; it is not an OPE coefficient and does not by itself solve the full, nonchiral crossing equation.

Evidence cutoff. This account reflects sources consulted through 2026-08-09. Existence, invertibility, unitarity, and contour statements are asserted only for the domains and normalizations specified below; analytic continuations can require additional residue terms.

Required background. Liouville theory and the Virasoro bootstrap supplies the continuous Virasoro spectrum and DOZZ data. Chiral blocks, sewing, and modularity supplies the finite-dimensional rational analogue and the Moore–Seiberg consistency conditions.

Helpful background. Bounded, compact, and integral operators provides the operator language for kernels, measures, adjoints, and domains.

The channel transform

Fix four external Virasoro representations and normalize the $s$-channel block by

$$\mathcal F_{P_s}^{(s)}(z) =z^{h(P_s)-h_1-h_2}\bigl(1+O(z)\bigr).$$

Likewise normalize the $t$-channel block near $z=1$. For a continuum $P\geq0$, write

$$\mathcal F_{P_s}^{(s)}(z) =\int_0^\infty d\nu_t(P_t)\, \mathsf F_{P_sP_t} \begin{bmatrix} P_3&P_2\\[-2pt] P_4&P_1 \end{bmatrix} \mathcal F_{P_t}^{(t)}(1-z).$$

Here $d\nu_t(P_t)=\mu_t(P_t)dP_t$ is the declared Plancherel measure and $\mathsf F$ is the fusion kernel. External labels and the central charge are suppressed except in the bracket. In other conventions the measure is absorbed into the kernel; comparing two tabulated kernels without undoing this choice generally gives a false normalization discrepancy.

There is also a block-normalization freedom. If

$$\mathcal F_P^{(s)}\mapsto a_s(P)\mathcal F_P^{(s)}, \qquad \mathcal F_P^{(t)}\mapsto a_t(P)\mathcal F_P^{(t)},$$

then

$$\mathsf F_{P_sP_t} \mapsto \frac{a_s(P_s)}{a_t(P_t)}\mathsf F_{P_sP_t},$$

with a corresponding change if $a_t$ is absorbed into $d\nu_t$. Only a transform stated together with its block normalization and measure is reproducible.

Plancherel unitarity and its domain

Let $\delta_\nu(P,P')$ denote the delta distribution adapted to $d\nu$:

$$\int d\nu(P')\,\delta_\nu(P,P')f(P')=f(P).$$

If the fusion transform is unitary on the specified principal-series space, its kernel obeys

$$\int_0^\infty d\nu_t(P_t)\, \mathsf F_{P_sP_t}\mathsf F_{P_s'P_t}^{*} =\delta_{\nu_s}(P_s,P_s').$$

The inverse relation integrates over $P_s$. This statement has three essential hypotheses:

1. the external and intermediate momenta lie in the principal-series domain for which the transform is defined;
2. the integration contour has not crossed singularities of the kernel or blocks;
3. the complex conjugation and measures are those of the chosen Hilbert-space normalization.

It should not be extrapolated unchanged to degenerate external fields, nonnormalizable momenta, or an analytically continued central charge. In the 1999 construction, Ponsot and Teschner explicitly began from conjectured invertible fusion transformations, derived functional constraints, and identified the solution with Racah coefficients of a noncompact quantum group; the note also states which results were being announced rather than proved there. Those assumptions and the principal-series spectrum are recorded in Ponsot and Teschner 1999, §§1–3, pp. 1–7.

The degenerate limit is a finite matrix

If one external representation contains a level-two null vector, the conformal blocks solve a hypergeometric equation and only two intermediate representations are allowed. The integral transform degenerates to a $2\times2$ connection matrix.

One row is the exact Gauss connection formula

$$\begin{aligned} {}_2F_1(a,b;c;z) ={}&A\,{}_2F_1 \bigl(a,b;a+b-c+1;1-z\bigr)\\ &+B(1-z)^{c-a-b} {}_2F_1 \bigl(c-a,c-b;c-a-b+1;1-z\bigr), \end{aligned}$$

where

$$A=\frac{\Gamma(c)\Gamma(c-a-b)} {\Gamma(c-a)\Gamma(c-b)}, \qquad B=\frac{\Gamma(c)\Gamma(a+b-c)} {\Gamma(a)\Gamma(b)}.$$

The second independent $s$-channel solution supplies the other row. In a BPZ four-point function, $a,b,c$ are linear combinations of $b^{\pm1}$ and the external Liouville momenta. This finite connection problem is not a different mechanism: it is the residue or degenerate limit of the generic fusion transform. It provides a sharp failure test because the continuum kernel must reproduce these gamma-function coefficients when a momentum approaches a degenerate value. The Virasoro degenerate fusing matrix and its role in the shift equations are derived in Ribault 2018, §§2.3.2–2.4.3, pp. 45–63.

At resonant values such as $c-a-b\in\mathbb Z$, the displayed connection formula is singular term by term and its limiting solutions can contain logarithms. One must take the limit of the complete connection problem rather than interpreting an individual divergent gamma factor as a physical OPE coefficient.

Analytic continuation and discrete residues

The generic transform is initially defined with a contour separating two families of poles. Schematically,

$$\mathcal F_s =\int_{\mathcal C}d\nu(P_t)\, \mathsf F_{st}\mathcal F_t.$$

As an external momentum is varied, a pole $P_*(\alpha_i)$ may cross $\mathcal C$. Keeping the analytically continued function while restoring the original contour gives

$$\mathcal F_s =\int_{\mathcal C}d\nu(P_t)\, \mathsf F_{st}\mathcal F_t \mathbin{\pm}2\pi i \sum_{P_*} \operatorname*{Res}_{P_t=P_*} \bigl[d\nu(P_t)\mathsf F_{st}\mathcal F_t\bigr].$$

The sign is set by the contour orientation and direction of crossing. These residues can be the discrete terms expected from degenerate fusion rules or from nonnormalizable external states. Ponsot and Teschner give an explicit pole-separating contour for their special-function representation in Ponsot and Teschner 1999, Appendix, pp. 14–15.

Three checks are indispensable:

• list both pole families and state which side of $\mathcal C$ each occupies;
• continue external parameters along a specified path, since different paths can wind singular loci;
• compare the residues with the allowed discrete intermediate weights and their block normalizations.

Writing only “analytically continue the kernel” suppresses the information that decides the answer.

From a chiral transform to full crossing

Consider a diagonal four-point function

$$G(z,\bar z) =\int d\nu_s(P_s)\, \rho_s(P_s) \mathcal F_{P_s}^{(s)}(z) \overline{\mathcal F_{P_s}^{(s)}(z)}.$$

Applying the fusion transform to both chiral factors yields a double integral in $P_t$ and $P_t'$. To recover a diagonal $t$-channel expression, the dynamical density $\rho_s$ must combine with the two kernels so that the off-diagonal terms cancel and the diagonal coefficient becomes the correct $\rho_t$. In Liouville theory, this step uses the DOZZ structure constants together with the orthogonality properties of the fusion kernel.

Therefore:

• the fusion kernel changes a basis of chiral blocks;
• the three-point constants determine which blocks occur and with what dynamical weight;
• the two-point normalization determines the inverse metric inserted in the OPE;
• the contour and measure define the direct integral;
• crossing is the equality of the complete nonchiral decompositions.

A kernel satisfying pentagon identities is powerful consistency data, but it does not select a unique spectrum or set of OPE coefficients by itself.

Failure tests

Normalization test. Compare the $z\to0$ and $z\to1$ leading powers before comparing kernel values. A hidden rescaling of blocks changes the kernel by a ratio of normalization functions.

Degenerate test. Approach a level-two degenerate momentum and recover the two hypergeometric connection coefficients, including residues and signs.

Inverse test. Compose the forward and inverse transforms on a smooth compactly supported function of $P$. The result should reproduce the function in the $d\nu$ norm, not with a bare Lebesgue delta unless $d\nu=dP$.

Pole test. Move one external momentum across a known singular locus. A bare contour integral that changes discontinuously while the residue-corrected expression is continuous has exposed a missing discrete term.

Crossing test. Do not test the kernel in isolation. Transform a complete four-point integrand with its two-point inverse and three-point constants, then compare the resulting $t$-channel density.

The distributional care required when kernels act on delta-normalized correlators is developed on Distributional correlators, zero modes, and normalization.

References

• Ponsot, Bénédicte, and Jörg Teschner. “Liouville Bootstrap via Harmonic Analysis on a Noncompact Quantum Group.” Preprint, 1999. arXiv:hep-th/9911110.
• Ribault, Sylvain. “Conformal Field Theory on the Plane.” SciPost Physics Lecture Notes 1 (2018). DOI.