Large-N Dynamics

Large $N$ converts some strongly coupled quantum theories into controlled saddle or graph expansions, but only after the index family, field and action normalization, held coupling, normalized observable, state, and order of limits have been fixed. This chapter develops vector auxiliary fields, matrix eigenvalue densities, ribbon topology, factorization, gauge-theory organization, volume equivalence, tensor melons, and the critical regimes where a fixed $1/N$ expansion fails.

Helpful background. Asymptotic scales, remainders, and uniformity supplies the distinction between pointwise and uniform expansions. Diagrammatics and symmetry factors supplies the graph construction whose index factors are reorganized here.

Enter this chapter

The phrase “large $N$” does not identify one approximation. Three questions must be answered before any calculation:

1. What grows? A vector has one index, a matrix or adjoint field has two, and a rank-three tensor has three with a specified covariance and invariant.
2. What remains fixed? The interaction is scaled so propagators, vertices, and index loops compete. In gauge theory this is the renormalized ’t Hooft coupling, not $g_{\mathrm{YM}}$ itself.
3. Which limit is meant? Volume, continuum, infrared, critical, and $N$ limits may not commute. A sharp $N=\infty$ phase boundary is not a finite-$N$ singularity.

The basic powers already show why the families differ:

$$\begin{array}{c|c|c} \text{family} & \text{declared interaction scaling} & \text{leading free energy}\\ \hline O(N)\ \text{vector} & \lambda(\phi_i\phi_i)^2/(4N) & O(N)\\ \text{matrix or adjoint} & N\operatorname{tr}V,\ \lambda_{\mathrm t}\ \text{fixed} & O(N^2)\\ \text{rank-three tetrahedral tensor} & gN^{-3/2} & O(N^3) \end{array}$$

These powers count degrees of freedom and selected graph families. They are not interchangeable normalizations.

Choose a route by the physical question

Routes through the nine large-N dynamics pages

Question	Start here	Main result	Essential boundary
Which variables and couplings define the limit?	Large-N limits and normalizations	Exact vector, matrix, gauge, and tensor counting conventions	State and order of volume, continuum, infrared, and N limits are part of the claim
How does a vector theory become a saddle?	Vector auxiliary-field saddles	Renormalized O(N) sigma-model gap equation and auxiliary kernel	The leading saddle is not the exact finite-N vacuum
How do matrix eigenvalues encode phases?	Matrix eigenvalues and loop equations	Normalized one-cut and two-cut quartic densities	Positivity, W(z) ∼ 1/z, filling fractions, and critical limit order select the solution
Why do matrix and gauge graphs organize by genus?	Double-line topology	NV−E+F = N2−2h and boundary counting	Euler counting is not by itself a string theory
When do invariant observables factorize?	Factorization and master fields	Connected normalized traces vanish in a selected clustering state	Factorization does not imply a unique ordinary classical configuration
What is genuinely string-like in a planar gauge theory?	Planar gauge organization	Handles, Wilson-loop boundaries, and long-flux-tube expansion	Confinement and a complete gauge/string dictionary require additional dynamics
When can a reduced volume reproduce an extended theory?	Volume reduction and equivalence	Neutral-sector loop-equation matching	Center and projection symmetries must remain unbroken on the scaling path
Why do tensors select melons rather than planar graphs?	Tensor melonic dominance	Colored face counting and a closed leading two-point equation	The scaling and leading family are invariant- and covariance-specific
When does fixed-order 1/N reasoning fail?	Corrections and double scaling	The quartic N2/3(r−rc) scaling window	Critical, infrared, and exponentially small sectors require new control data

A coherent reading path

Begin with Large-N Limits, Normalizations, and Orders of Limits. Its figure and semantic comparison fix every convention reused later.

Then follow the branch matching the degrees of freedom:

• Vector: Vector Models, Auxiliary Fields, and Large-N Saddles derives dimensional transmutation and the quadratic auxiliary kernel.
• Matrix: Matrix Eigenvalue Saddles, Loop Equations, and Phase Transitions solves an equilibrium density, then Double-Line Counting and the Topological Expansion derives the graph topology.
• Gauge: read double-line counting first, then Large-N Factorization and Master-Field Claims, Planar Gauge Dynamics and String-Like Organization, and Volume Reduction, Large-N Equivalence, and Their Hypotheses.
• Tensor: Tensor Large N and Melonic Dominance starts again from the declared covariance and invariant instead of borrowing a matrix genus.

Finish with Subleading Corrections, Double Scaling, and Nonuniform Limits. It tests where every earlier leading statement can lose uniformity.

The chapter’s control statement

Every large-$N$ conclusion should be readable in the form

$$\begin{aligned} &\text{index family and action normalization}\\ &+\ \text{held renormalized couplings and normalized operators}\\ &+\ \text{regulator, state, boundary conditions, and limit order}\\ &+\ \text{leading saddle or graph family and an error test}\\ &\Longrightarrow\text{one bounded observable statement}. \end{aligned}$$

This prevents several common overextensions:

• a vector cactus expansion is not a matrix genus expansion;
• a formal one-cut resolvent is not physical until its density is normalized and nonnegative;
• $N^{2-2h-b}$ does not prove a string dual;
• factorization in one phase does not apply to an unselected mixture;
• volume independence does not cover center-charged observables;
• melonic dominance in one tensor invariant is not universal;
• a fixed $1/N$ series does not control a joint critical or infrared limit;
• absence at every algebraic order does not exclude $e^{-cN}$ effects.

Review the chapter

1. A matrix action is written without an overall $N$, but its graphs are assigned $N^{2-2h}$. What must be checked?

Solution

Derive the propagator and vertex powers from the written quadratic and interaction terms. Without the overall $N$, both are $O(1)$ for the same field normalization, so a graph scales as $N^F$, not $N^{V-E+F}$. Topology alone cannot restore the missing factors.

2. A normalized invariant has variance $O(N^{-2})$ in each of two phases, but the equal phase mixture has values $\pm v$. Does the mixture factorize?

Solution

No. Its mean is zero and its second moment tends to $v^2$, so its variance is $O(1)$. A phase-selecting source and an explicit volume, source-removal, and $N$ order are required.

3. A quartic matrix model is tuned so $N^{2/3}(r-r_c)$ is fixed. May one use a fixed one-cut genus truncation?

Solution

No. This is the support-merging critical window, where nominally subleading orders are nonuniform and infinitely many must be resummed into the double-scaled description.