Large-N CFT Data and Vector Models

In a vector CFT, large $N$ suppresses fluctuations of normalized singlet composites and organizes their correlators in powers of $1/N$. At leading order this produces generalized-free factorization and additive multi-trace dimensions; interactions first appear through connected correlators, anomalous dimensions, and operator mixing at subleading orders. The expansion is controlled only after the $N$-scaling of every operator and coupling, the saddle, and any nonuniform kinematic limit are stated.

Required background. Free and generalized-free theories supplies factorized correlators and double-trace families. Large-N limits and normalizations fixes vector-model counting. Helpful background. Vector models and auxiliary-field saddles develops the path integral, while large-N factorization and master fields compares vector and matrix counting.

Normalization and factorization

Normalize the fundamental vector so that

$$\langle \phi_i(x)\phi_j(0)\rangle =\frac{\delta_{ij}}{(x^2)^{\Delta_\phi}}.$$

A normalized singlet bilinear has the schematic form

$$\mathcal O(x)=\frac{1}{\sqrt{N}}:\!\phi_i\phi_i\!:(x),$$

with an additional $N$-independent factor chosen to make $\langle\mathcal O\mathcal O\rangle=(x^2)^{-\Delta_\mathcal O}$. A connected $k$-point correlator of such bilinears then scales as

$$\langle\mathcal O_1\cdots\mathcal O_k\rangle_{\mathrm{conn}} =O\!\left(N^{1-k/2}\right).$$

Thus a normalized singlet three-point coefficient is generically $O(N^{-1/2})$, and its connected four-point function is $O(N^{-1})$. The disconnected four-point function is $O(1)$ and has the generalized-free decomposition into double-trace operators $[\mathcal O\mathcal O]_{n,\ell}$ with

$$\Delta_{n,\ell}=2\Delta_\mathcal O+2n+\ell+O(N^{-1}).$$

The terminology “single trace” and “multi-trace” is useful by analogy, but in a vector model a singlet bilinear is the elementary invariant composite; it is not a matrix trace. Large-$N$ factorization alone also does not imply a sparse spectrum or a large gap. The separation between factorization and the extra large-gap assumption is emphasized in Fitzpatrick and Kaplan 2012, §§2–3.

The critical O(N) saddle

For a quartic vector model, scale the coupling as

$$S=\int d^dx\left[ \frac12(\partial\phi_i)^2+\frac{\lambda}{4N}(\phi_i\phi_i)^2 \right].$$

Formally introduce an auxiliary field $\sigma$ through

$$S[\phi,\sigma]=\int d^dx\left[ \frac12(\partial\phi_i)^2+\frac12\sigma\phi_i\phi_i -\frac{N}{4\lambda}\sigma^2 \right].$$

The Gaussian contour for $\sigma$ is chosen so that integrating it out reproduces the stable positive quartic interaction; the displayed real saddle is reached by the corresponding contour deformation. Integrating out the $N$ vector components gives

$$S_{\mathrm{eff}}[\sigma] =\frac N2\operatorname{Tr}\log(-\partial^2+\sigma) -\frac{N}{4\lambda}\int d^dx\,\sigma^2,$$

so saddle fluctuations are explicitly weighted by $N$. The conformal critical saddle is selected by tuning the relevant mass deformation and solving the gap equation with the chosen regulator. Other saddles can describe massive or symmetry-broken phases and must not be silently substituted for the critical one.

At the interacting critical point for $2<d<4$, the Hubbard–Stratonovich field replaces the free scalar bilinear as a primary whose leading dimension is $2$. In $d=3$, the first corrections are

$$\begin{aligned} \Delta_\phi&=\frac12+\frac{4}{3\pi^2N}+O(N^{-2}),\\ \Delta_\sigma&=2-\frac{32}{3\pi^2N}+O(N^{-2}). \end{aligned}$$

Equivalently, $\eta=2\Delta_\phi-(d-2)=8/(3\pi^2N)+O(N^{-2})$. These signs and factors provide a sensitive test of whether $\eta$, $\Delta_\phi$, and the thermal exponent have been translated consistently. The auxiliary-field derivation and general-$d$ critical exponents are reviewed in Moshe and Zinn-Justin 2003, §§2.1–2.5 and 3.1–3.3.

The same diagrammatic counting generates the first corrections to OPE data. Exchange of $\sigma$ produces an $O(1/N)$ connected four-point function of the fundamentals; decomposing it yields anomalous dimensions and OPE-coefficient corrections for singlet, symmetric-traceless, and antisymmetric bilinears. At a fixed spin and fixed excitation number this is an ordinary asymptotic expansion. Taking spin, excitation number, spacetime dimension, or a nearly degenerate operator difference to scale with $N$ can reorganize the series.

Mixing and higher-spin currents

At $N=\infty$, bilinear currents of all even spins in the singlet sector saturate the appropriate conservation bound. At finite $N$, only the exact stress tensor and currents of exact global symmetries remain conserved; the other currents acquire $O(1/N)$ anomalous dimensions and recombine with their divergence operators. The weakly broken higher-spin page develops this mechanism.

Multi-trace degeneracies are common. If $\mathcal O_A$ share the same $N=\infty$ dimension and quantum numbers, their first corrections are eigenvalues of the full $O(1/N)$ mixing matrix computed with the leading two-point metric. Assigning a correction to a suggestive composite expression before diagonalization is basis-dependent.

The following comparison table records the calculational claim, error scale, and failure test for this page’s large-$N$ sector.

Target	Dimension or sector	Parameter	Observable	Computed order	Error estimate	Independent check	Evidence basis	Known failure
Critical $O(N)$ fundamental	$d=3$, fixed kinematics	$1/N$	$\Delta_\phi=\tfrac12+\tfrac{4}{3\pi^2N}$	through $1/N$	$O(N^{-2})$ before any nonuniform limit	compare $\eta=2\Delta_\phi-1$ and fixed-dimension bootstrap data	critical auxiliary-field saddle and $1/N$ diagrams	using small $N$ without higher orders or resummation
Critical $O(N)$ singlet	$d=3$, scalar singlet sector	$1/N$	$\Delta_\sigma=2-\tfrac{32}{3\pi^2N}$	through $1/N$	$O(N^{-2})$ plus mixing where degenerate	compare the thermal exponent and gap-equation normalization	critical saddle and auxiliary-field two-point function	confusing $\sigma$ with the free $:\!\phi^2\!:$ primary
Normalized singlet bilinears	vector model at fixed $k$	$1/\sqrt N$ per normalized extra boundary insertion	connected $k$-point scaling $N^{1-k/2}$	leading index counting	diagram- and channel-dependent subleading powers	explicit index-loop count for $k=2,3,4$	Wick and saddle counting	inconsistent operator normalization
Double-trace tower	fixed $n,\ell$ as $N\to\infty$	$1/N$	$2\Delta_\mathcal O+2n+\ell+O(1/N)$	leading spectrum	anomalous dimension and OPE correction at the next order	crossing of the connected four-point function	generalized-free factorization plus connected exchange	a large-spin or large-$n$ limit nonuniform in $1/N$
Nearly conserved higher-spin currents	fixed $s>2$ in the singlet tower	$1/N$	$\gamma_s=O(1/N)$	parametric leading order	model-dependent coefficient and mixing	norm of the nonconservation descendant	multiplet recombination plus large-$N$ factorization	treating approximate conservation as an exact symmetry

Limits on the conclusion

Saddle test. Check stability, the auxiliary-field contour, and the tuned relevant deformation. A formal stationary point need not define the desired CFT.

Counting test. Draw the index loops for a representative two-, three-, and four-point graph after normalizing the external operators. If the announced power of $N$ changes, the entire OPE hierarchy must be revised.

Nonuniformity test. Repeat the estimate with spin, excitation number, and $d-2$ or $4-d$ explicit. Terms such as $\ell/N$ or $1/[N(d-2)]$ signal a double-scaling problem.

Finite-N test. Compare successive $1/N$ orders or an independent method. The mere existence of a formal expansion does not bound its remainder at a particular small $N$.

Two tests must remain distinct: a truncation comparison probes numerical stability, while a sparse-spectrum consistency test adds an assumption not implied by vector factorization. Large-$N$ factorization alone is not evidence for a large gap.

Exercises

Derive the $N^{1-k/2}$ scaling of a connected $k$-point function of normalized bilinears.

Solution

A connected Wick contraction of $k$ unnormalized bilinears has one closed vector-index loop and is therefore $O(N)$. Each normalized bilinear contributes $N^{-1/2}$, giving $N\,N^{-k/2}=N^{1-k/2}$.

At what parametric spin can an omitted correction proportional to $\ell/N^2$ compete with a retained term proportional to $1/N$?

Solution

Their ratio is $\ell/N$. They compete when $\ell=O(N)$, so the fixed-spin $1/N$ expansion is not uniform in that regime.

References

• Fitzpatrick, A. L., and Kaplan, J. (2012), “AdS field theory from conformal field theory,” Journal of High Energy Physics 2012(10), 032. doi:10.1007/JHEP10(2012)032. Open PDF
• Moshe, M., and Zinn-Justin, J. (2003), “Quantum field theory in the large $N$ limit: A review,” Physics Reports 385, 69–228. doi:10.1016/S0370-1573(03)00112-3. Open PDF