Large-N expansions and saddle points

Consider a real integral with weight $\exp[-Nf(x)]$ and a unique minimum $x_*$ satisfying positive curvature $f''(x_*)$. Expand the exponent to quadratic order. A fluctuation becomes appreciably suppressed when $Nf''(x_*)(x-x_*)^2$ is of order one, which gives width proportional to $N^{-1/2}$, rather than $N^{-1}$. The integration measure contributes the same square-root factor to the leading normalization. A large prefactor does not mean every quantity starts with an inverse-first-power correction.

For $f(x)=x^2/2+\lambda x^4/4$ with nonnegative fixed $\lambda$, the only minimum is at zero and the integral converges. Substitute $y=\sqrt N x$. The quadratic weight becomes a unit Gaussian, while the quartic interaction is suppressed by $\lambda/N$. This explicitly identifies the expansion parameter for the integral. It is a pedagogical saddle model, not a claim that a field theory with infinitely many momentum modes reduces to one ordinary variable.

Expand the last exponential to first order and integrate using the unit Gaussian fourth moment, which is three. This produces a relative correction $-3\lambda/(4N)$. For fixed positive $\lambda$, further terms organize an asymptotic expansion as $N$ becomes large. The positive integrand does not require every truncated correction to be positive; the negative first correction simply records that the quartic term suppresses the tails compared with a Gaussian of the same curvature.

In an $O(N)$ vector field theory, a collective auxiliary field can acquire an effective action proportional to the number of components. Its saddle gives a self-consistent leading description, and fluctuations generate inverse-component corrections. Gauge theories with many colors have a different combinatorics: holding the appropriate 't Hooft coupling fixed organizes diagrams by topology, often in powers of inverse color number squared for suitable normalized observables. Calling both limits large N does not make their leading diagrams or correction powers identical.

If curvature vanishes, the Gaussian width estimate fails and a higher-order term determines the scaling. If multiple saddles compete, discarding all but one may miss a phase transition or exponentially small contributions. Taking the large-component limit before a large-volume, long-time, or critical limit can also change the answer. A reliable calculation states which limits are ordered and checks whether the nominally suppressed corrections become enhanced in the observable of interest.

Large N is therefore a method with a domain, rather than a universal solution to strong coupling. Compare successive orders, check symmetries and spectral constraints, and use independent numerical or exact results when available. Applying a controlled expansion at N near one can still yield useful estimates, but its asymptotic motivation alone provides no numerical guarantee there. Establishing error control for such extrapolations is part of the research problem, not a detail hidden by a smooth leading-order curve.