The large-N limit

Intuition Beginner

Imagine a tiny compass needle at each point of a magnet, but instead of pointing in two or three directions it lives in a space of $N$ directions at once. The number $N$ counts how many independent components the spin has. The large-$N$ idea is to let $N$ grow without bound and ask what the magnet does in that limit.

Why would adding more components make life easier rather than harder? Because the energy depends on an average over all $N$ components. When you average many independent quantities, the result barely wiggles: the law of large numbers tames the fluctuations. A quantity built from one component might jump around, but the same quantity averaged over a thousand components sits almost still at its mean.

So in the large-$N$ limit the field that controls the physics stops fluctuating and freezes at a single best value. Finding that one value is a simple optimisation problem.

Visual Beginner

The picture shows a spin arrow that lives in a high-dimensional space, drawn as a fan of many component axes. As more axes are added, the tip of the averaged arrow settles toward the centre of a sphere and stops jittering.

The point of the image is that more components means less jitter. The averaged length of the spin locks onto one value, and that locked value is exactly what the saddle point computes.

Worked example Beginner

Take $N$ numbers $s_1,s_2,\dots ,s_N$, each drawn at random with average value $0$ and average square value $1$. Build the averaged square length $L=(s_1^2+s_2^2+\cdots +s_N^2)/N$.

For $N=1$ the value of $L$ is just $s_1^2$, which swings between near $0$ and several times $1$ depending on the draw. The spread is large.

For $N=4$ the value $L$ averages four squares, so a high draw in one slot tends to be cancelled by a low draw in another. The typical spread of $L$ around its mean of $1$ shrinks by a factor of about $\sqrt{4}=2$.

For $N=100$ the spread shrinks by about $\sqrt{100}=10$. The averaged length now sits very close to $1$ on almost every draw. What this shows: the relative size of the wiggles falls off like $1/\sqrt{N}$, so as $N$ grows the averaged length is effectively a fixed number. That fixed number is the saddle-point value, and the leftover wiggles of size $1/\sqrt{N}$ are the seed of the $1/N$ corrections studied later.

Check your understanding Beginner

Formal definition Intermediate+

Consider the O(N) vector model: an $N$-component real field $\mathbf{\varphi }(\mathbf{x})=({\varphi }_1,\dots ,{\varphi }_N)$ in $d$ Euclidean dimensions with the O(N)-symmetric weight $e^{-\mathcal{H}[\mathbf{\varphi }]}$,

$$\mathcal{H}[\mathbf{\varphi }]=\int d^{d}x\left[\frac{1}{2}(\nabla \mathbf{\varphi }{)}^2+\frac{1}{2}{r}_0\mathbf{\varphi }\cdot \mathbf{\varphi }+\frac{u}{N}(\mathbf{\varphi }\cdot \mathbf{\varphi }{)}^2\right],$$

where $\mathbf{\varphi }\cdot \mathbf{\varphi }={\sum }_{i=1}^N{\varphi }_i^2$. The factor $1/N$ multiplying the quartic coupling is the scaling that makes the energy per component finite as $N\to \infty$; it is what allows a sensible limit. Conventions follow Itzykson and Drouffe ^{[Itzykson-Drouffe Ch. 6]}.

Introduce a real auxiliary field $\sigma (\mathbf{x})$ by a Hubbard-Stratonovich identity, which rewrites the quartic term as a Gaussian integral over $\sigma$,

$$\exp [-\frac{u}{N}\int (\mathbf{\varphi }\cdot \mathbf{\varphi }{)}^2]=\int \mathcal{D}\sigma \exp [\int (\frac{N}{16u}{\sigma }^2-\frac{1}{2}\sigma \mathbf{\varphi }\cdot \mathbf{\varphi })],$$

up to a constant. The decoupling trades the quartic self-interaction for a field $\sigma$ linearly coupled to $\mathbf{\varphi }\cdot \mathbf{\varphi }$. The $\mathbf{\varphi }$ integral is now Gaussian and can be performed, leaving an effective action $S_{\mathrm{eff}}[\sigma ]$ that is proportional to $N$:

$$S_{\mathrm{eff}}[\sigma ]=N[\frac{1}{2}\mathrm{Tr}\ln (-{\nabla }^2+r_0+\sigma )-\frac{1}{16u}\int {\sigma }^2].$$

Because $S_{\mathrm{eff}}\propto N$, the partition function $\int \mathcal{D}\sigma e^{-S_{\mathrm{eff}}}$ is dominated by the saddle point $\sigma =\bar{\sigma }$ (constant) as $N\to \infty$, with corrections suppressed by powers of $1/N$.

Key theorem with proof Intermediate+

Theorem (gap equation, exact at $N=\infty$). Let $m^2=r_0+\bar{\sigma }$ denote the saddle-point mass. The constant saddle of $S_{\mathrm{eff}}$ satisfies the gap equation

$$m^2=r_0+8u\int \frac{d^{d}q}{(2\pi {)}^d}\frac{1}{q^2+m^2},$$

and this saddle gives the exact free energy density in the limit $N\to \infty$. The critical point is the value of $r_0$ at which $m^2\to 0$.

Proof. The saddle point is the stationary configuration of $S_{\mathrm{eff}}[\sigma ]$. For a constant $\sigma =\bar{\sigma }$ the trace-log evaluates on plane waves: $\frac{1}{2}\mathrm{Tr}\ln (-{\nabla }^2+r_0+\bar{\sigma })=\frac{V}{2}\int \frac{d^{d}q}{(2\pi {)}^d}\ln (q^2+r_0+\bar{\sigma })$, with $V$ the volume. Stationarity $\delta S_{\mathrm{eff}}/\delta \sigma =0$ at constant $\sigma$ reads

$$\frac{1}{2}\int \frac{d^{d}q}{(2\pi {)}^d}\frac{1}{q^2+r_0+\bar{\sigma }}-\frac{1}{8u}\bar{\sigma }=0.$$

Solving for $\bar{\sigma }$ and substituting $m^2=r_0+\bar{\sigma }$ gives $\bar{\sigma }=m^2-r_0=8u\int \frac{d^{d}q}{(2\pi {)}^d}(q^2+m^2{)}^{-1}$, which is the stated gap equation. The exactness in the limit is the content of the Laplace method: since $S_{\mathrm{eff}}=Ns[\sigma ]$ with $s$ independent of $N$, the integral $\int \mathcal{D}\sigma e^{-Ns[\sigma ]}$ concentrates on the minimiser of $s$ with relative corrections $O(1/N)$ from the Gaussian fluctuations of $\sigma$ about $\bar{\sigma }$. Hence the saddle is the leading free energy and is exact at $N=\infty$. $\square$

Bridge. This gap equation builds toward every quantitative result below, and the foundational reason it is exact rather than approximate is that the action is proportional to $N$, so Laplace concentration is rigorous in the limit. This is exactly the spherical-model self-consistency condition of Berlin, Kac, and Stanley, written in field-theory language: the constraint that the averaged length of the spin be fixed is dual to the stationarity of $\sigma$, and the central insight is that the auxiliary field $\sigma$ is the Lagrange multiplier enforcing that constraint. The integral on the right generalises the mean-field self-consistency of 08.02.01 by retaining the full momentum dependence of the propagator, and putting these together shows that the large-$N$ saddle is mean-field theory made exact by the law of large numbers in component space. The leftover Gaussian fluctuation of $\sigma$ appears again in the $1/N$ corrections of the Advanced section, where it generates the loop expansion.

Exercises Intermediate+

Advanced results Master

The exactness at $N=\infty$ is only the start; the real power is that $1/N$ is a controllable expansion parameter. Expanding $\sigma =\bar{\sigma }+\delta \sigma$ and integrating over the Gaussian fluctuations $\delta \sigma$ generates a systematic $1/N$ expansion. Each additional power of $1/N$ corresponds to one more loop of the $\sigma$ propagator, so the $1/N$ expansion is a loop expansion organised not by the coupling strength but by the inverse number of components ^{[Wilson 1973]}. This is the decisive advantage over the $ϵ=4-d$ expansion: the small parameter $1/N$ is available in any fixed dimension $d$, so one can compute exponents directly in $d=3$ rather than extrapolating from near $d=4$.

The critical exponents at leading order are read off from the gap equation and the $\sigma$ propagator. The correlation-length exponent is $\nu =1/(d-2)+O(1/N)$ for $2<d<4$, and the anomalous dimension vanishes at leading order, $\eta =0+O(1/N)$, because the constant saddle shifts only the mass and not the $q^2$ coefficient of the propagator. At next order the $\sigma$ fluctuations supply a momentum-dependent self-energy and give the first nonzero value, $\eta =\frac{8}{N}\frac{\Gamma (d-2)\sin (\pi (d-2)/2)}{\pi \Gamma (d/2-1{)}^2\Gamma (d/2+1)}+\cdots$, which in $d=3$ reduces to $\eta =8/(3{\pi }^{2}N)$ ^{[Brezin-Wallace 1973]}. These $1/N$ exponents agree with the $ϵ$-expansion results where both apply, a stringent cross-check of two unrelated expansions.

The connection to the spherical model of Berlin and Kac is exact, not analogical. Stanley proved in 1968 that the partition function of the spherical model is identical to the $N\to \infty$ limit of the O(N) vector model, component for component, in every dimension ^{[Stanley 1968]}. The spherical constraint ${\sum }_i{s}_i^2=N$ is enforced by a Lagrange multiplier whose stationary value obeys the same gap equation derived above; the auxiliary field $\sigma$ is that multiplier promoted to a fluctuating field. The dimensional dependence is then transparent: for $d\le 2$ the gap integral diverges in the infrared and no ordered phase survives, recovering the absence of long-range order forced by continuous-symmetry fluctuations in 08.02.03; for $2<d<4$ there is a genuine interacting critical point with the exponents above; and for $d>4$ the gap integral is dominated by its ultraviolet end and the exponents lock to their mean-field values.

Synthesis. Putting these together, the large-$N$ limit is the foundational reason a strongly interacting field theory can be solved without perturbing in the coupling: the central insight is that the auxiliary field carries an action proportional to $N$, so its saddle is exact and the leftover fluctuations organise into a loop expansion in $1/N$. This is exactly the mechanism that makes the spherical model of Berlin and Kac a genuine field theory rather than a toy, and the gap equation is dual to the spherical self-consistency constraint under the identification of $\sigma$ with the constraint multiplier. The construction generalises the mean-field saddle of 08.02.01 by keeping the full propagator, and it builds toward the matrix models of 08.14.06, where the same idea returns with $N\to \infty$ taken over matrix indices and the loop expansion reorganised by the genus of surfaces. The bridge across all of these is one principle: a parameter that multiplies the entire action makes the saddle point exact, and the inverse of that parameter is the expansion that restores the fluctuations.

Full proof set Master

Proposition (existence of a critical point for $2<d<4$). For the gap equation $m^2=r_0+8uI(m^2)$ with $I(m^2)=\int \frac{d^{d}q}{(2\pi {)}^d}(q^2+m^2{)}^{-1}$ and a fixed ultraviolet cutoff $\Lambda$, there is a finite critical coupling $r_{0c}$ such that $m^2(r_0)\to 0^{+}$ as $r_0\to r_{0c}^{+}$, and near criticality $r_0-r_{0c}\sim m^{d-2}$, giving $\nu =1/(d-2)$ at leading order in $1/N$.

The function $I(m^2)$ is finite at $m^2=0$ for $d>2$, since the infrared behaviour ${\int }_0{d}^{d}q{q}^{-2}\sim {\int }_0{q}^{d-3}dq$ converges there. Define $r_{0c}=-8uI(0)$, the value at which $m^2=0$ solves the gap equation. Subtracting,

$$r_0-r_{0c}=m^2-8u[I(0)-I(m^2)].$$

The bracket equals $\int \frac{d^{d}q}{(2\pi {)}^d}[q^{-2}-(q^2+m^2{)}^{-1}]=m^2\int \frac{d^{d}q}{(2\pi {)}^d}{q}^{-2}(q^2+m^2{)}^{-1}$. For $2<d<4$ this last integral is infrared-dominated and scales as $m^{d-4}$ by dimensional analysis (rescale $q=mp$), so $I(0)-I(m^2)\sim m^{d-2}$. Therefore $r_0-r_{0c}=m^2-cu{m}^{d-2}$ with $c>0$; for $d<4$ the exponent $d-2<2$, so the $m^{d-2}$ term dominates as $m\to 0$ and $r_0-r_{0c}\sim -cu{m}^{d-2}$ up to sign conventions, i.e. $|t|\sim m^{d-2}$. Inverting gives $m\sim |t{|}^{1/(d-2)}$, hence $\xi =m^{-1}\sim |t{|}^{-1/(d-2)}$ and $\nu =1/(d-2)$. $\square$

Proposition ($\eta =0$ at leading order in $1/N$). In the $N\to \infty$ limit the two-point function of $\mathbf{\varphi }$ is $G(q)=(q^2+m^2{)}^{-1}$ exactly, so the anomalous dimension vanishes at order $N^0$.

At $N=\infty$ the partition function is governed by the constant saddle $\bar{\sigma }$, and the $\mathbf{\varphi }$ field is Gaussian with the shifted mass $m^2=r_0+\bar{\sigma }$. Its propagator is therefore $G(q)=1/(q^2+m^2)$, with the coefficient of $q^2$ equal to one and no momentum-dependent correction. Since the anomalous dimension is defined by the small-$q$ scaling $G(q)\sim q^{-2+\eta }$ at criticality ($m^2=0$), and here $G(q)=q^{-2}$ exactly, one reads $\eta =0$. A nonzero value requires the momentum-dependent self-energy from the $\sigma$ propagator, which is a fluctuation effect entering at order $1/N$; hence $\eta =0+O(1/N)$. $\square$

Connections Master

This unit builds toward the matrix models of 08.14.06, where the large-$N$ limit is taken over matrix indices rather than vector components; the saddle-point-becomes-exact principle is identical, but the loop expansion reorganises into a sum over the genus of surfaces rather than a vector $1/N$ series.

The gap equation derived here generalises the mean-field self-consistency condition of 08.02.01: mean-field theory drops the momentum dependence of the propagator, while the large-$N$ saddle retains the full integral $\int d^{d}q(q^2+m^2{)}^{-1}$, making the large-$N$ limit mean-field theory rendered exact by the law of large numbers over components.

The dimensional structure of the gap equation reproduces the continuous-symmetry physics of 08.04.05 and its companions: for $d\le 2$ the infrared divergence of the gap integral forbids an ordered phase, matching the Mermin-Wagner suppression of long-range order, while for $2<d<4$ the same integral yields the interacting exponents $\nu =1/(d-2)$ and $\eta =O(1/N)$ that the momentum-shell renormalisation group of 08.04.05 obtains independently as a cross-check.

Historical & philosophical context Master

The large-$N$ limit grew from two roots that turned out to be the same tree. On the statistical-mechanics side, Berlin and Kac introduced the spherical model in 1952 as an exactly solvable caricature of a ferromagnet, replacing the hard Ising constraint $s_i=\pm 1$ with the soft global constraint ${\sum }_i{s}_i^2=N$ ^{[Berlin-Kac 1952]}. For sixteen years the spherical model was regarded as a clever but isolated toy, until Stanley proved in 1968 that it is exactly the $N\to \infty$ limit of the O(N) Heisenberg model, so its solvability is a special case of large-$N$ solvability ^{[Stanley 1968]}. On the field-theory side, Wilson and others recognised in the early 1970s that the same limit furnishes a small parameter, $1/N$, that controls a loop expansion in any fixed dimension, freeing critical-exponent calculations from the near-four-dimensional regime of the $ϵ$ expansion ^{[Wilson 1973]}. Brezin and Wallace carried the program to explicit exponents at order $1/N$ ^{[Brezin-Wallace 1973]}. The philosophical lesson is that solvability is rarely an accident of a special model; it is usually the signature of a hidden large parameter that makes a saddle point exact, and identifying that parameter is the substance of "solving" a theory.

Bibliography Master

@book{ItzyksonDrouffe1989,
  author    = {Itzykson, Claude and Drouffe, Jean-Michel},
  title     = {Statistical Field Theory, Volume 1},
  publisher = {Cambridge University Press},
  year      = {1989}
}

@article{Stanley1968Spherical,
  author  = {Stanley, H. Eugene},
  title   = {Spherical Model as the Limit of Infinite Spin Dimensionality},
  journal = {Physical Review},
  volume  = {176},
  year    = {1968},
  pages   = {718--722}
}

@article{BerlinKac1952,
  author  = {Berlin, T. H. and Kac, M.},
  title   = {The Spherical Model of a Ferromagnet},
  journal = {Physical Review},
  volume  = {86},
  year    = {1952},
  pages   = {821--835}
}

@article{Wilson1973LessThanFour,
  author  = {Wilson, Kenneth G.},
  title   = {Quantum Field Theory Models in Less Than Four Dimensions},
  journal = {Physical Review D},
  volume  = {7},
  year    = {1973},
  pages   = {2911--2926}
}

@article{BrezinWallace1973,
  author  = {Br\'ezin, Edouard and Wallace, David J.},
  title   = {Critical Behavior of a Classical Heisenberg Ferromagnet with Many Degrees of Freedom},
  journal = {Physical Review B},
  volume  = {7},
  year    = {1973},
  pages   = {1967--1974}
}