Large-N Factorization and Classical Bulk Scaling

Large-$N$ factorization is a statement about the cumulants of a fixed set of suitably normalized single-trace observables. Once nonzero one-point functions have been removed, their leading full correlators become Wick-like and their connected three- and four-point functions are suppressed. This is the counting expected of weak bulk vertices, but it neither determines the light spectrum nor supplies a locality scale. Those conclusions require independent spectral and dynamical tests.

Required background. Observable and Regime Matrix for Quantum Gravity fixes which boundary and bulk observables are being compared.

Helpful background. Large-N CFT Data and Vector Models supplies the boundary organization. Tensor Large N and Melonic Dominance shows that large-$N$ counting need not be planar.

If index counting or normalization conventions are unfamiliar, first review Large-N Limits, Normalizations, and Orders of Limits and Large-N Factorization and Master-Field Claims. Volume VII owns those general methods; this page asks what they do and do not license about a bulk description.

Reading path. First fix centering and normalization, then check the matrix powers in an exactly soluble Gaussian example. The final sections translate that hierarchy into conditional bulk loop counting and use a vector model to show why factorization alone stops short of local Einstein gravity.

Centered operators and connected cumulants

A connected correlator is a cumulant: it is the part left after subtracting every product associated with a nontrivial partition of the insertions. Let $\mathcal O_i$ be single-trace operators, and first remove their one-point functions,

$$\widetilde{\mathcal O}_i =\mathcal O_i-\langle\mathcal O_i\rangle.$$

For nonidentity primaries in the conformal vacuum on flat space, the one-point functions already vanish. Centering must instead be done explicitly in a thermal state, in curved space, in the presence of a boundary or defect, or whenever a scalar condensate is allowed.

Now fix the two-point normalization so that

$$G_{ij}(x) \equiv \langle\widetilde{\mathcal O}_i(x)\widetilde{\mathcal O}_j(0)\rangle_{\mathrm c} =\mathcal O(N^0).$$

In a matrix-like genus expansion, with the operator set, coupling, dimensions, number of insertions, and kinematics held fixed, the planar hierarchy is

$$\left\langle \widetilde{\mathcal O}_{i_1}\cdots\widetilde{\mathcal O}_{i_n} \right\rangle_{\mathrm c} =\mathcal O\!\left(N^{2-n}\right), \qquad n\geq 2.$$

The genus-counting page owns the full topology derivation; here only its planar result is needed. The essential large-$N$ property is the suppression of connected single-trace correlators relative to their disconnected products Aharony et al. 2000, § 1.2, eqs. (1.5)–(1.8), pp. 10–15, PDF and Heemskerk et al. 2009, § 2.4, pp. 7–8, and § 3.2, pp. 10–11, PDF. In particular,

$$\langle\widetilde{\mathcal O}_1\widetilde{\mathcal O}_2\widetilde{\mathcal O}_3\rangle_{\mathrm c} =\mathcal O(N^{-1}), \qquad \langle\widetilde{\mathcal O}_1\widetilde{\mathcal O}_2 \widetilde{\mathcal O}_3\widetilde{\mathcal O}_4\rangle_{\mathrm c} =\mathcal O(N^{-2}).$$

The cumulant expansion then gives the full centered four-point function,

$$\begin{aligned} \left\langle\widetilde{\mathcal O}_1\widetilde{\mathcal O}_2 \widetilde{\mathcal O}_3\widetilde{\mathcal O}_4\right\rangle ={}&G_{12}G_{34}+G_{13}G_{24}+G_{14}G_{23}\\ &+\left\langle\widetilde{\mathcal O}_1\widetilde{\mathcal O}_2 \widetilde{\mathcal O}_3\widetilde{\mathcal O}_4\right\rangle_{\mathrm c}. \end{aligned}$$

Thus the correction to pairwise factorization is $\mathcal O(N^{-2})$. Without centering, this formula is incomplete: it must also contain products involving one-point functions, including mean-times-connected-three-point terms. “Factorization” never means that those terms may be silently discarded.

An explicit matrix normalization

The phrase “canonically normalized adjoint field” does not fix the power of $N$ carried by a trace; the action convention must be stated. Let $M(x)$ be an $N\times N$ Hermitian matrix field, with $\operatorname{Tr}$ taken over its color indices. A clean convention is

$$S=N\int d^dx\,\operatorname{Tr}\!\left[ \frac12(\partial M)^2+V(M) \right],$$

with couplings in $V$ held fixed in the chosen large-$N$ limit. The double-line propagator then carries $1/N$. A connected planar graph with $n$ fixed-length trace insertions has $n$ boundaries and scales as $N^{2-n}$ ’t Hooft 1974, §§ 2–3, pp. 462–466, PDF. For a renormalized fixed-length single-trace operator whose mixing with operators of the same quantum numbers has already been resolved, write

$$X_k=\bigl[\operatorname{Tr}(M^k)\bigr]_{\mathrm R}, \qquad \mathcal O_k=Z_k^{-1/2}\bigl(X_k-\langle X_k\rangle\bigr),$$

where the remaining $N$-independent finite factor $Z_k$ can be chosen so that the connected two-point function of $\mathcal O_k$ is order one. No additional factor of $N$ is required in this convention. The next page treats the separate problem of resolving single-trace, multi-trace, and collective-field mixing.

This operator is not the expectation-normalized trace often used for master-field statements. If

$$\tau_k=\frac{X_k}{N} =\frac1N\bigl[\operatorname{Tr}(M^k)\bigr]_{\mathrm R},$$

then $\langle\tau_k\rangle=\mathcal O(1)$ but its connected variance is $\mathcal O(N^{-2})$. The order-one fluctuation underlying $\mathcal O_k$ is

$$N\bigl(\tau_k-\langle\tau_k\rangle\bigr) =X_k-\langle X_k\rangle,$$

followed by the $N$-independent factor $Z_k^{-1/2}$. The two conventions answer different questions: $\tau_k$ has a deterministic large-$N$ limit, while $\mathcal O_k$ resolves the order-one fluctuation that can be matched to a canonically normalized bulk field.

The comparison is easiest to audit in one place. The powers below apply to fixed $k$ in the displayed matrix convention and to connected correlators within the same fixed large-$N$ regime.

Three trace normalizations

Convention	Definition	Leading connected scales	What it is useful for
Renormalized uncentered trace	Xk = [Tr(Mk)]R	Mean typically O(N); n-point cumulant O(N2−n) for n ≥ 2	Direct double-line counting
Expectation-normalized trace	τk = Xk/N	Mean O(1); n-point cumulant O(N2−2n)	Deterministic or master-field limit
Centered unit-two-point fluctuation	𝒪k = Zk−1/2(Xk − ⟨Xk⟩)	Mean 0; two-point O(1); n-point cumulant O(N2−n)	Comparison with a canonically normalized bulk field

The zero-dimensional Gaussian model, where no UV composite-operator renormalization is needed, makes the normalization visible without diagrammatic shorthand:

$$S=\frac N2\operatorname{Tr}M^2, \qquad \langle M_{ij}M_{kl}\rangle =\frac1N\delta_{il}\delta_{jk}.$$

A Hermitian matrix has $N$ real diagonal components and $N(N-1)$ real off-diagonal components. With the displayed Gaussian weight, $\sqrt N M_{ii}$ and the variables $\sqrt{2N}\,\operatorname{Re}M_{ij}$, $\sqrt{2N}\,\operatorname{Im}M_{ij}$ for $i<j$ are independent standard normals. Consequently $NX$ is chi-squared with $N^2$ real degrees of freedom for $X=\operatorname{Tr}M^2$, and Wick contraction gives

$$\langle X\rangle=N, \qquad \left\langle(X-N)^2\right\rangle_{\mathrm c}=2.$$

Hence $\mathcal O=(X-N)/\sqrt2$ has unit two-point function. Its third and fourth cumulants are $2\sqrt2/N$ and $12/N^2$, respectively, exactly exhibiting the normalized hierarchy.

Another common convention rescales the microscopic field or includes inverse-coupling factors so that $\widehat X=N X$ has a connected two-point function of order $N^2$. The order-one centered fluctuation is then

$$\frac{\widehat X-\langle\widehat X\rangle}{N} =X-\langle X\rangle,$$

followed by the same finite normalization, $1/\sqrt2$ in the Gaussian example. The unrescaled powers differ, but the normalized connected correlators and their physical hierarchy agree.

Conditional bulk interpretation

For a canonically normalized bulk field $\varphi$ in $D=d+1$ dimensions, an interaction

$$S_{\mathrm{bulk}}\supset\int d^{d+1}x\,\sqrt{\lvert g\rvert}\left( \frac{g_3}{3!}\varphi^3+ \frac{g_4}{4!}\varphi^4\right)$$

is most cleanly compared at fixed AdS radius $L$ through the dimensionless couplings

$$\lambda_3=g_3L^{(5-d)/2}, \qquad \lambda_4=g_4L^{3-d}.$$

The normalized matrix hierarchy is reproduced when $\lambda_3=\mathcal O(N^{-1})$, $\lambda_4=\mathcal O(N^{-2})$, and, more generally, every canonically normalized $m$-field vertex obeys $\lambda_m=\mathcal O(N^{2-m})$. For a connected interaction diagram with at least one counted vertex, let $V_m$ be the number of such vertices, $I$ the number of internal lines, $n$ the number of external lines, and $\ell_{\mathrm{loop}}=I-\sum_mV_m+1$ the loop number. Then

$$\prod_m\lambda_m^{V_m} =\mathcal O\!\left(N^{\sum_m(2-m)V_m}\right) =\mathcal O\!\left(N^{2-n-2\ell_{\mathrm{loop}}}\right),$$

where $\sum_m mV_m=n+2I$. Nontrivial tree-level interaction diagrams therefore scale as $N^{2-n}$, and each additional bulk loop costs $1/N^2$; the canonically normalized free two-point function is the separately fixed $n=2$ base case. This counting presupposes a bulk dictionary and canonically normalized propagators; factorization alone does not construct either one.

The loop statement assumes a fixed number of light species and uniformly bounded sums over intermediate states. A family with $N_{\mathrm{sp}}(N)$ comparable fields can instead produce an effective loop parameter $N_{\mathrm{sp}}(N)/N^2$. The connected-correlator page develops this species test.

Even at fixed species, the inference requires identifiable light operators and a perturbative map to bulk fields. Factorization alone does not say that the fields are spectrally isolated, that their interactions admit a derivative expansion, or that a metric is the only low-spin mediator. A perturbative $1/N$ expansion, a finite low-dimension Fock sector, and a locality condition such as polynomial Mellin boundedness are logically separate inputs Fitzpatrick and Kaplan 2013, §§ 1 and 2.4, pp. 1–4 and 14–15, PDF.

Vector-model counterexample

Large-$N$ vector models also factorize, but with different powers. In the free-vector counting that makes the contrast explicit, let repeated vector indices $a,b=1,\ldots,N$ be summed and take the renormalized centered bilinear singlet

$$j(x)=\frac1{\sqrt N}\left[\phi^a(x)\phi^a(x) -\left\langle\phi^b(x)\phi^b(x)\right\rangle\right],$$

the leading connected Wick graph has one closed vector-index loop. Its factor of $N$ and the $N^{-1/2}$ from each insertion give

$$\left\langle j(x_1)\cdots j(x_n)\right\rangle_{\mathrm c} =\mathcal O\!\left(N^{1-n/2}\right),$$

so the connected three- and four-point functions scale as $N^{-1/2}$ and $N^{-1}$. They still vanish as $N\to\infty$, even though they do not follow the matrix hierarchy.

In the free $O(N)$ model just used, the singlet sector contains one exactly conserved current for every even spin, and Klebanov and Polyakov conjecture the corresponding minimal bosonic AdS$_4$ higher-spin description Klebanov and Polyakov 2002, § 2, eqs. (3)–(5), p. 3, PDF. In the critical model, the higher-spin currents are conserved only at leading order as $N\to\infty$ and acquire $\mathcal O(1/N)$ anomalous dimensions at finite $N$ Klebanov and Polyakov 2002, § 2, pp. 4–5, PDF. A theorem sharpens the exact-conservation obstruction: in a unitary three-dimensional CFT with the usual CFT properties, finite stress-tensor two-point coefficient, finitely many primaries below any fixed dimension, and a unique stress tensor, one exactly conserved current of spin greater than two implies an infinite conserved tower Maldacena and Zhiboedov 2013, Introduction, pp. 2–3, and § 5.2, pp. 15–17, PDF. The higher-spin-gap page develops what must change before an Einstein regime is possible.

Sending $N\to\infty$ suppresses connected interactions among the fixed, normalized collective operators in both examples; it does not make the microscopic QFT weakly coupled. The strongest common conclusion is classical or weakly coupled collective behavior in a specified sector. Einstein-like locality requires additional sparsity and a parametrically large higher-spin gap.

Orders of limits and evidence ceiling

The $N$ powers should be checked after fixing two-point normalization and holding ’t Hooft couplings, operator dimensions, the number of insertions, and kinematics fixed. Operators whose length grows with $N$, exponentially late times, or a number of species that grows with $N$ define different scaled limits and can defeat the displayed hierarchy.

Factorization licenses weakly coupled collective behavior for the tested sector. It does not establish a sparse spectrum, a derivative expansion, a metric description, or a nonperturbative dictionary. Volumes VII and IX develop the large-$N$ method and the CFT data. The immediate next page, Single-Trace, Multi-Trace, and Collective-Field Organization for Bulk Criteria, tests when a normalized boundary operator can be assigned a stable one-particle interpretation.

Common pitfalls

Normalizing without centering. Dividing by a power of $N$ fixes the two-point scale; subtracting $\langle\mathcal O\rangle$ removes the one-point cumulant. These are different operations, and the simple pairwise four-point formula needs both the declared normalization and vanishing one-point functions.

Treating suppressed as absent. At finite $N$, connected correlators are small rather than zero. The $1/N$ and $1/N^2$ terms carry bulk interactions, anomalous dimensions, and loop corrections.

Promoting large $N$ to Einstein gravity. Factorization controls a quantum-interaction hierarchy. Spectral sparsity, a higher-spin gap, strong coupling, and a locality window remain independent checks.

Exercises

Why centering matters

Write $\mathcal O_i=m_i+\delta\mathcal O_i$ with $\langle\delta\mathcal O_i\rangle=0$. Explain why the full four-point function cannot be written only as a sum of pairwise two-point functions plus the connected four-point function when some $m_i$ are nonzero.

Solution: expand the uncentered four-point function

Expanding the product produces terms such as

$$m_1\left\langle \delta\mathcal O_2\delta\mathcal O_3\delta\mathcal O_4 \right\rangle_{\mathrm c},$$

as well as $m_im_j\langle\delta\mathcal O_k\delta\mathcal O_l\rangle$, products of four means, and the analogous permutations. The mean-times-two-point terms do occur inside pairings of full two-point functions, but those pairings count the product of four means three times. In terms of full uncentered two-point functions, the exact identity is

$$\begin{aligned} \langle\mathcal O_1\mathcal O_2\mathcal O_3\mathcal O_4\rangle ={}&\sum_{(ij)(k\ell)} \langle\mathcal O_i\mathcal O_j\rangle \langle\mathcal O_k\mathcal O_\ell\rangle +\kappa_{1234}\\ &+\sum_{i=1}^{4}m_i\,\kappa^{(3)}_{\widehat i} -2m_1m_2m_3m_4, \end{aligned}$$

where $\kappa^{(3)}_{\widehat i}$ denotes the connected three-point cumulant with entry $i$ omitted. Replacing every operator by $\widetilde{\mathcal O}_i=\delta\mathcal O_i$ sets all $m_i$ to zero and leaves the three centered pairings plus the connected four-point function.

Gaussian matrix check

For the Gaussian Hermitian matrix model above, $NX$ has a chi-squared distribution with $N^2$ real degrees of freedom. Use $\kappa_n(\chi_k^2)=2^{n-1}(n-1)!k$ to derive the first four cumulants of $X=\operatorname{Tr}M^2$, and then normalize $\mathcal O=(X-N)/\sqrt2$.

Solution: compute the Gaussian cumulants

Because $X=\chi_{N^2}^2/N$,

$$\kappa_n(X)=2^{n-1}(n-1)!N^{2-n}.$$

Thus

$$\kappa_1=N, \qquad \kappa_2=2, \qquad \kappa_3=\frac8N, \qquad \kappa_4=\frac{48}{N^2}.$$

Centering changes only $\kappa_1$, and division by $\sqrt2$ gives

$$\langle\mathcal O^2\rangle_{\mathrm c}=1, \qquad \langle\mathcal O^3\rangle_{\mathrm c}=\frac{2\sqrt2}{N}, \qquad \langle\mathcal O^4\rangle_{\mathrm c}=\frac{12}{N^2}.$$

The example realizes the same $N^{2-n}$ hierarchy as the connected planar diagrams.

Diagnose an unfamiliar large-N sector

Suppose a family of unit-normalized scalar singlets $J$ obeys

$$\langle JJJ\rangle_{\mathrm c}=\mathcal O(N^{-1/2}), \qquad \langle JJJJ\rangle_{\mathrm c}=\mathcal O(N^{-1}),$$

and a spectral survey finds exactly conserved single-trace currents of every even spin $s\geq2$. Without assuming that you know the microscopic fields, answer three questions: which counting pattern does the correlator hierarchy resemble; what is the strongest bulk claim licensed by all the data; and which missing Einstein-gravity criterion is directly falsified?

Solution: classify the data and set the claim ceiling

The powers match the vector-like hierarchy $\langle J^n\rangle_{\mathrm c}=\mathcal O(N^{1-n/2})$, although scaling alone does not identify the microscopic theory. Their vanishing shows that this singlet sector becomes weakly interacting at large $N$. The conserved tower, however, puts light single-trace states at unbounded spin and therefore directly rules out a parametrically large higher-spin gap. The data license a classical or weakly interacting collective sector, potentially with a higher-spin bulk description; they do not license a local Einstein effective theory.

The chapter overview contains the structure diagram and validity and failure diagram. They are embedded there once so that their shared chapter-level context is not repeated on every article.

For the chapter-wide comparison of assumptions, counterevidence, falsifiers, and claim ceilings, see the claim-domain table.

References

• Aharony, Ofer; Gubser, Steven S.; Maldacena, Juan; Ooguri, Hirosi; and Oz, Yaron. “Large N Field Theories, String Theory and Gravity.” Physics Reports 323, 183–386 (2000). doi:10.1016/S0370-1573(99)00083-6. Open PDF.
• Fitzpatrick, A. Liam, and Jared Kaplan. “AdS Field Theory from Conformal Field Theory.” Journal of High Energy Physics 2013, 054 (2013). doi:10.1007/JHEP02(2013)054. Open PDF.
• Heemskerk, Idse; Penedones, João; Polchinski, Joseph; and Sully, James. “Holography from Conformal Field Theory.” Journal of High Energy Physics 2009, 079 (2009). doi:10.1088/1126-6708/2009/10/079. Open PDF.
• Klebanov, Igor R., and Alexander M. Polyakov. “AdS Dual of the Critical $O(N)$ Vector Model.” Physics Letters B 550, 213–219 (2002). doi:10.1016/S0370-2693(02)02980-5. Open PDF.
• Maldacena, Juan, and Alexander Zhiboedov. “Constraining Conformal Field Theories with a Higher Spin Symmetry.” Journal of Physics A: Mathematical and Theoretical 46, 214011 (2013). doi:10.1088/1751-8113/46/21/214011. Open PDF.
• ’t Hooft, Gerard. “A Planar Diagram Theory for Strong Interactions.” Nuclear Physics B 72, 461–473 (1974). doi:10.1016/0550-3213(74)90154-0. Open PDF.