Loop Vertex Expansion for Convergent Field Theories

• Loop Vertex Expansion is a method that reorganizes perturbative series into convergent tree sums using intermediate field representations and the BKAR forest formula.
• It employs techniques such as the Hubbard–Stratonovich transformation and resolvent bounds to control amplitudes in matrix, tensor, and field-theoretic models.
• The approach extends to multiscale renormalization and variational settings, proving analyticity and Borel summability of the perturbation series.

Loop Vertex Expansion (LVE) is a constructive reorganization of perturbation theory designed to compute connected Schwinger functions and free energies as convergent sums over trees, rather than as divergent sums over Feynman graphs. It combines an intermediate field representation with a forest formula and a replica trick, and it was developed for local and non-local interactions, including vector, matrix, tensor, and renormalized field-theoretic models. In the standard quartic setting, the interaction is rewritten through an auxiliary field, the original fields are integrated out, and the logarithm of the partition function is converted into a sum over connected trees whose amplitudes are controlled by resolvent bounds (Gurau et al., 2013, Sazonov, 2023).

1. Constructive definition and standard architecture

The standard LVE uses three ingredients: an intermediate field representation, a replica trick, and the Brydges–Kennedy–Abdesselam–Rivasseau forest formula. A central point of the formalism is that these are canonical combinatorial tools and do not require space-time dependent lattices; in the multiscale setting, this is one of the reasons the method is presented as independent of the space-time geometry (Gurau et al., 2013).

For quartic interactions, the basic transformation is the Hubbard–Stratonovich-type identity

$$e^{-\lambda \phi^4/2} = \int e^{-\sigma^2/2}\, e^{\,i\sqrt{\lambda}\,\sigma\,\phi^2}\, d\sigma.$$

In matrix models, one similarly introduces an auxiliary Hermitian matrix field $A$, integrates out the original matrix variables, and obtains a non-polynomial effective action containing a $\Tr\log$ term and resolvents of the form

$$\left(1-i\sqrt{\frac{\lambda}{a^2 N}}\,A\right)^{-1}.$$

A typical quartic complex matrix model is

$Z_{N}(\lambda)=\int dM\, \exp\!\left\{-\Tr(MM^\dagger)-\frac{\lambda}{2N}\Tr\big[(MM^\dagger)^2\big]\right\},$

with free energy $F[\lambda,N]=-\frac{1}{N^2}\log Z_N(\lambda)$ (Sazonov, 2023).

The standard constructive workflow is: expand the interaction in a Taylor series, replicate the intermediate field into $n$ copies, apply the BKAR forest formula, and convert the logarithm of the partition function into a sum over connected trees. In this form, connected quantities are represented by tree amplitudes with ordered products of resolvents around faces and corners. The standard outcome is a convergent tree expansion, rather than a graph expansion, and this is the core constructive meaning of LVE (Sazonov, 2023, Rivasseau, 2 Jun 2026).

2. What the expansion actually resums

A common misconception is that LVE merely replaces Feynman graphs by trees within a fixed perturbative order. The letter “How are Feynman graphs resumed by the Loop Vertex Expansion?” states the opposite: each LVE tree collects pieces of ordinary Feynman graphs from many different perturbative orders, and the decisive regrouping occurs only after an intermediate-field extension and a graph collapse procedure (Rivasseau et al., 2010).

The starting point is the usual forest-formula weight attached to a spanning tree $T\subset G$ of a connected graph $G$,

$$w(G,T)=\int_0^1 \prod_{\ell\in T} dw_\ell \prod_{\ell\notin T} x_\ell(\{w\}),$$

with barycentric identity

$A$0

This gives a naive repacking

$A$1

For bosonic theories, however, this naive rearrangement is not sufficient, because the cancellations needed for convergence occur between different perturbative orders, not only at fixed order (Rivasseau et al., 2010).

The genuine LVE mechanism begins by decomposing each quartic vertex into one of three pairings of its four half-lines. An order-$A$2 labeled $A$3 vacuum graph therefore has $A$4 labeled “3-body extensions.” Ordinary lines in the extended graph are then collapsed into bold vertices, called loop vertices, and the forest formula is applied to the collapsed graph rather than to the original Feynman graph. The resulting amplitudes are indexed by spanning trees $A$5 of collapsed graphs $A$6, and each tree amplitude

$A$7

resums an infinite family of ordinary graphs. In this sense, LVE is a tree expansion of dressed loop vertices, not of bare Feynman vertices (Rivasseau et al., 2010).

3. Analyticity domains, resolvent bounds, and summability

In quartic matrix and tensor settings, the standard analyticity domain of LVE is a cardioid-like region. For the quartic matrix model, the domain recalled in the variational LVE work is

$A$8

equivalently on the Riemann surface of $A$9,

$\Tr\log$0

Inside this domain, the free energy and cumulants are analytic, and LVE provides a constructive proof of Borel summability of the perturbation series (Sazonov, 2023).

The appearance of the cardioid is tied to the standard resolvent estimate. Writing the variational parameter as

$\Tr\log$1

one obtains

$\Tr\log$2

This bound gives uniform control of tree amplitudes, but only in a domain that shrinks near the branch cut of $\Tr\log$3 (Sazonov, 2023).

The same constructive logic appears in zero-dimensional higher-order models. For $\Tr\log$4 in zero dimension, the free energy $\Tr\log$5 is shown to be Borel–Le Roy summable of order $\Tr\log$6, and the paper develops the $\Tr\log$7 case in detail as the model example. The general conclusion is that the LVE extends beyond quartic interactions and matches the correct summability order for stable even potentials (Rivasseau et al., 2010).

4. Generalizations beyond the standard quartic form

One line of generalization replaces the quartic intermediate-field paradigm by a loop vertex representation built from higher-order combinatorics. For the zero-dimensional $\Tr\log$8 model,

$\Tr\log$9

the key statement is that the important feature to extend the loop vertex expansion is not to use an intermediate field representation, but rather to force integration of exactly one particular field per vertex of the initial action. The resulting loop-vertex action is expressed through Fuss–Catalan generating functions,

$$\left(1-i\sqrt{\frac{\lambda}{a^2 N}}\,A\right)^{-1}.$$0

and the main analytic estimate is

$$\left(1-i\sqrt{\frac{\lambda}{a^2 N}}\,A\right)^{-1}.$$1

This yields an absolutely convergent tree expansion for $$\left(1-i\sqrt{\frac{\lambda}{a^2 N}}\,A\right)^{-1}.$$2 in a pacman domain $$\left(1-i\sqrt{\frac{\lambda}{a^2 N}}\,A\right)^{-1}.$$3 (Rivasseau, 2017).

A second line of generalization is the inductive realization of LVE. In the one-dimensional $$\left(1-i\sqrt{\frac{\lambda}{a^2 N}}\,A\right)^{-1}.$$4 theory, the construction starts from an intermediate-field representation

$$\left(1-i\sqrt{\frac{\lambda}{a^2 N}}\,A\right)^{-1}.$$5

derives a Polchinski-type equation for $$\left(1-i\sqrt{\frac{\lambda}{a^2 N}}\,A\right)^{-1}.$$6, rewrites it as a functional integral equation, and solves it recursively. The paper emphasizes that this avoids explicit use of the BKAR forest formula and avoids resolvent expansion, while Catalan numbers and the coefficients $$\left(1-i\sqrt{\frac{\lambda}{a^2 N}}\,A\right)^{-1}.$$7 encode the recursive tree structure. For $$\left(1-i\sqrt{\frac{\lambda}{a^2 N}}\,A\right)^{-1}.$$8 and $$\left(1-i\sqrt{\frac{\lambda}{a^2 N}}\,A\right)^{-1}.$$9, the two-point function has an absolutely convergent expansion and exponential decay (Zhao, 2018).

5. Multiscale LVE and renormalization

Ordinary LVE works well for ultraviolet-convergent theories or for a single renormalization-group slice, but renormalization requires scale analysis. The multiscale loop vertex expansion (MLVE) enlarges the formalism by combining multiscale decomposition, intermediate-field representation, forest formulas, and a Mayer-type mechanism that imposes a hard-core constraint on slice labels within each block. In the vector-type quartic model with propagator mimicking the power counting of $Z_{N}(\lambda)=\int dM\, \exp\!\left\{-\Tr(MM^\dagger)-\frac{\lambda}{2N}\Tr\big[(MM^\dagger)^2\big]\right\},$0, the paper states that an ordinary LVE would fail to treat even this simplest superrenormalizable model, whereas the MLVE performs the ultraviolet limit and proves analyticity in the Borel summability domain (Gurau et al., 2013).

In the corrected construction of $Z_{N}(\lambda)=\int dM\, \exp\!\left\{-\Tr(MM^\dagger)-\frac{\lambda}{2N}\Tr\big[(MM^\dagger)^2\big]\right\},$1, the renormalized partition function is written with Wick ordering,

$Z_{N}(\lambda)=\int dM\, \exp\!\left\{-\Tr(MM^\dagger)-\frac{\lambda}{2N}\Tr\big[(MM^\dagger)^2\big]\right\},$2

and the intermediate-field representation introduces the resolvent

$Z_{N}(\lambda)=\int dM\, \exp\!\left\{-\Tr(MM^\dagger)-\frac{\lambda}{2N}\Tr\big[(MM^\dagger)^2\big]\right\},$3

The paper identifies an important error in the earlier cleaning expansion and replaces it with a slice-testing expansion that marks propagators and tadpole counterterms scale by scale, followed by a two-level jungle expansion. The resulting free energy is absolutely convergent uniformly in the ultraviolet cutoff for

$Z_{N}(\lambda)=\int dM\, \exp\!\left\{-\Tr(MM^\dagger)-\frac{\lambda}{2N}\Tr\big[(MM^\dagger)^2\big]\right\},$4

and the ultraviolet limit is analytic in the cardioid domain and is the Borel sum of its perturbative expansion (Rivasseau et al., 2014).

These renormalized constructions were then applied in several directions. The two-dimensional Euclidean $Z_{N}(\lambda)=\int dM\, \exp\!\left\{-\Tr(MM^\dagger)-\frac{\lambda}{2N}\Tr\big[(MM^\dagger)^2\big]\right\},$5 quantum field theory was constructed using LVE, reproducing results of standard constructive theory such as the Borel summability of the Schwinger functions in the coupling constant (Rivasseau et al., 2011). In tensor field theory, the multiscale loop vertex expansion was used to construct cumulants up to a finite order in the quartic $Z_{N}(\lambda)=\int dM\, \exp\!\left\{-\Tr(MM^\dagger)-\frac{\lambda}{2N}\Tr\big[(MM^\dagger)^2\big]\right\},$6 model and to prove analyticity and Borel summability of the cumulants up to finite order (Rivasseau, 2022). In noncommutative field theory, the two-dimensional Grosse–Wulkenhaar model on the Moyal plane was constructed with multiscale loop vertex expansions; the paper treats renormalization with this tool, adapts Nelson’s argument, proves Borel summability of the perturbation series, and states that this is the first non-commutative quantum field theory model to be built in a non-perturbative sense (Wang, 2011).

6. Variational LVE, cumulants, and matrix-model topology

The variational loop vertex expansion (VLVE) modifies the standard construction by choosing the initial approximation to depend on the coupling constant. In the quartic matrix model, the unperturbed quadratic part is replaced by

$Z_{N}(\lambda)=\int dM\, \exp\!\left\{-\Tr(MM^\dagger)-\frac{\lambda}{2N}\Tr\big[(MM^\dagger)^2\big]\right\},$7

and the partition function is rewritten as

$Z_{N}(\lambda)=\int dM\, \exp\!\left\{-\Tr(MM^\dagger)-\frac{\lambda}{2N}\Tr\big[(MM^\dagger)^2\big]\right\},$8

The same LVE machinery—Taylor expansion, replicas, forest formula, and tree expansion—is then applied. The main result is that the free energy $Z_{N}(\lambda)=\int dM\, \exp\!\left\{-\Tr(MM^\dagger)-\frac{\lambda}{2N}\Tr\big[(MM^\dagger)^2\big]\right\},$9 is analytic uniformly in $F[\lambda,N]=-\frac{1}{N^2}\log Z_N(\lambda)$0 for

$F[\lambda,N]=-\frac{1}{N^2}\log Z_N(\lambda)$1

which extends analyticity across the usual branch cut region and to arbitrarily large couplings (Sazonov, 2023).

The same variational mechanism has been extended from free energy to cumulants. For the bounded-rank quartic complex matrix model, the ordinary cumulants and scalar cumulants are represented as convergent sums over LVE trees with cilia,

$F[\lambda,N]=-\frac{1}{N^2}\log Z_N(\lambda)$2

and the scalar cumulants encode the nontrivial coefficients in the Weingarten/topological decomposition. The paper proves analyticity uniformly in $F[\lambda,N]=-\frac{1}{N^2}\log Z_N(\lambda)$3 in the sector

$F[\lambda,N]=-\frac{1}{N^2}\log Z_N(\lambda)$4

under the source norm condition

$F[\lambda,N]=-\frac{1}{N^2}\log Z_N(\lambda)$5

with Nevanlinna–Sokal-type remainder estimate

$F[\lambda,N]=-\frac{1}{N^2}\log Z_N(\lambda)$6

The same work also establishes a topological expansion

$F[\lambda,N]=-\frac{1}{N^2}\log Z_N(\lambda)$7

which matches the usual large-$F[\lambda,N]=-\frac{1}{N^2}\log Z_N(\lambda)$8 ribbon-graph structure and the Weingarten calculus picture (Rivasseau, 2 Jun 2026).

Taken together, these developments show that LVE is no longer restricted to weak-coupling free energies of ultraviolet-convergent quartic models. This suggests a broader constructive framework in which tree-based resummation, multiscale renormalization, variational choice of background, cumulant control, and $F[\lambda,N]=-\frac{1}{N^2}\log Z_N(\lambda)$9-topological organization are treated within a common analytic scheme (Sazonov, 2023, Rivasseau, 2 Jun 2026).