Combinatorial Summation of Feynman Diagrams

• Combinatorial summation is defined as a collection of techniques in QFT that replace exhaustive graph-by-graph perturbative enumeration with structured algebraic and combinatorial formalisms.
• Matrix and permutation methods in φ⁴ theory convert derivative assignments to row–column magic squares, enabling systematic computation of diagram multiplicities and symmetry factors.
• Advanced resummation approaches, including worldline master formulas and dynamic programming schemes, enhance convergence and reveal asymptotic behaviors in diagrammatic expansions.

Searching arXiv for recent and foundational papers relevant to combinatorial summation of Feynman diagrams. Searching for the 2018 $\phi^4$ matrix approach paper and related work on symmetry factors and resummation. Combinatorial summation of Feynman diagrams denotes a family of methods that replace direct graph-by-graph perturbative enumeration with algebraic, graph-theoretic, or combinatorial formalisms that generate complete diagram classes, multiplicities, and symmetry factors in a structured way. In one important realization for vacuum diagrams in $\phi^4$ theory, the perturbative expansion of the generating functional can be rewritten in terms of row–column magic squares, permutation actions, and adjacency matrices, so that summing over Feynman graphs becomes equivalent to summing over equivalence classes of combinatorial matrices (Castro et al., 2018). More broadly, the subject encompasses exact counting of connected graphs, determinant- and dynamic-programming-based summation schemes in many-body theory, worldline master formulas that package many diagrams into single parameter integrals, and constructive resummations that reorganize divergent graph series into convergent tree-indexed expansions (Shi, 16 Jul 2025, Brouder et al., 2013, Ahmadiniaz et al., 2022, Rivasseau et al., 2013).

1. Perturbative graph summation as a combinatorial problem

In perturbative quantum field theory, diagrammatic expansions originate in the Dyson or generating-functional expansion, where repeated application of Wick’s theorem produces large sets of contraction patterns. Different contractions may represent the same topological graph and therefore must be grouped with the correct multiplicity or, equivalently, weighted by the inverse symmetry factor. In $\phi^4$ theory, the Euclidean generating functional can be written as

$$Z(J)=\exp\left[-\mathcal{V}\left(\frac{\delta}{\delta J}\right)\right]\exp\left(\frac{1}{2}J\Delta J\right),$$

with

$$\mathcal{V}\Big(\frac{\delta}{\delta J}\Big) = \frac{g}{4!}\int \mathrm{d}^{d}x\,\frac{\delta^{4}}{\delta J^{4}(x)}.$$

At order $g^m$, one obtains $4m$ functional derivatives acting on $(J\Delta J)^{2m}$, and every surviving contraction at $J\to 0$ corresponds to a vacuum Feynman diagram (Castro et al., 2018).

This basic observation admits several reorganizations. One may count contractions directly, use automorphism groups, convert the problem into recurrence relations for weighted graph counts, or encode entire diagram classes through determinants, path integrals, or auxiliary algebraic structures. A recurring theme is that the “sum over diagrams” is rarely treated as an unstructured set: it is organized by graph isomorphism, connectedness, irreducibility, topology, or permutation orbits (Brouder et al., 2013, Lin et al., 2018).

A standard source of confusion is the distinction between counting Wick contractions and counting diagram topologies. These are not identical. The raw perturbative expansion counts contraction patterns, whereas the diagrammatic expansion groups together all contractions leading to the same graph and divides by the appropriate symmetry factor. In graph-theoretic language, this factor is the size of the automorphism group of the graph, and the inverse factor is the weight with which the topology contributes (Hue et al., 2010).

2. Matrix and permutation methods in $\phi^4$ vacuum theory

A particularly explicit combinatorial summation scheme is the matrix approach for vacuum $\phi^4$0 diagrams. Its starting point is the observation that at perturbative order $\phi^4$1, the action of $\phi^4$2 derivatives on the $\phi^4$3 source factors in $\phi^4$4 is isomorphic to a constrained balls-in-boxes assignment problem. This yields an $\phi^4$5 matrix $\phi^4$6 whose entries record how many derivatives associated with vertex $\phi^4$7 act on source factors associated with vertex label $\phi^4$8, subject to

$\phi^4$9

Thus $\phi^4$0 is a row–column magic square with nonnegative integer entries and row and column sums equal to $\phi^4$1 (Castro et al., 2018).

Each row of such an RC-magic square specifies the arguments of a basic four-leg object

$\phi^4$2

and the explicit derivative evaluation gives

$\phi^4$3

Repeated arguments generate the corresponding local contractions, such as

$\phi^4$4

These objects are represented by adjacency-matrix blocks, and products of $\phi^4$5’s are translated into sums of full adjacency matrices through a formal map $\phi^4$6 satisfying

$\phi^4$7

In this way, an RC-magic square is converted into a product of local blocks, then into a sum of complete adjacency matrices for vacuum graphs, with coefficients equal to graph multiplicities (Castro et al., 2018).

The internal multiplicity associated with a fixed matrix $\phi^4$8 is

$\phi^4$9

which counts the number of derivative assignments within each row. Summing over all RC-magic squares and then merging adjacency matrices that are equivalent under vertex relabeling yields the total multiplicity $$Z(J)=\exp\left[-\mathcal{V}\left(\frac{\delta}{\delta J}\right)\right]\exp\left(\frac{1}{2}J\Delta J\right),$$0 of each vacuum graph $$Z(J)=\exp\left[-\mathcal{V}\left(\frac{\delta}{\delta J}\right)\right]\exp\left(\frac{1}{2}J\Delta J\right),$$1 (Castro et al., 2018).

This construction turns “summation of Feynman diagrams” into summation over combinatorial matrices and permutation orbits. The total number of RC-magic squares grows quickly,

$$Z(J)=\exp\left[-\mathcal{V}\left(\frac{\delta}{\delta J}\right)\right]\exp\left(\frac{1}{2}J\Delta J\right),$$2

for $$Z(J)=\exp\left[-\mathcal{V}\left(\frac{\delta}{\delta J}\right)\right]\exp\left(\frac{1}{2}J\Delta J\right),$$3, but the action of row and column permutations dramatically reduces the effective search space. The numbers of equivalence classes are

$$Z(J)=\exp\left[-\mathcal{V}\left(\frac{\delta}{\delta J}\right)\right]\exp\left(\frac{1}{2}J\Delta J\right),$$4

and Burnside’s lemma gives the class count from the $$Z(J)=\exp\left[-\mathcal{V}\left(\frac{\delta}{\delta J}\right)\right]\exp\left(\frac{1}{2}J\Delta J\right),$$5 action on the set of RC-magic squares (Castro et al., 2018).

3. Multiplicity, symmetry factors, and automorphisms

A central issue in any combinatorial summation is the relation between multiplicities in the generating-functional expansion and symmetry factors in the graph expansion. In the $$Z(J)=\exp\left[-\mathcal{V}\left(\frac{\delta}{\delta J}\right)\right]\exp\left(\frac{1}{2}J\Delta J\right),$$6 matrix approach, once the total multiplicity $$Z(J)=\exp\left[-\mathcal{V}\left(\frac{\delta}{\delta J}\right)\right]\exp\left(\frac{1}{2}J\Delta J\right),$$7 of a diagram at order $$Z(J)=\exp\left[-\mathcal{V}\left(\frac{\delta}{\delta J}\right)\right]\exp\left(\frac{1}{2}J\Delta J\right),$$8 is known, one recovers the standard inverse symmetry factor through

$$Z(J)=\exp\left[-\mathcal{V}\left(\frac{\delta}{\delta J}\right)\right]\exp\left(\frac{1}{2}J\Delta J\right),$$9

This matches the usual graph-theoretic statement that $$\mathcal{V}\Big(\frac{\delta}{\delta J}\Big) = \frac{g}{4!}\int \mathrm{d}^{d}x\,\frac{\delta^{4}}{\delta J^{4}(x)}.$$0 (Castro et al., 2018).

A general formula for symmetry factors across field theories is

$$\mathcal{V}\Big(\frac{\delta}{\delta J}\Big) = \frac{g}{4!}\int \mathrm{d}^{d}x\,\frac{\delta^{4}}{\delta J^{4}(x)}.$$1

where $$\mathcal{V}\Big(\frac{\delta}{\delta J}\Big) = \frac{g}{4!}\int \mathrm{d}^{d}x\,\frac{\delta^{4}}{\delta J^{4}(x)}.$$2 is the number of vertex permutations preserving the topology, $$\mathcal{V}\Big(\frac{\delta}{\delta J}\Big) = \frac{g}{4!}\int \mathrm{d}^{d}x\,\frac{\delta^{4}}{\delta J^{4}(x)}.$$3 counts self-conjugate bubbles, $$\mathcal{V}\Big(\frac{\delta}{\delta J}\Big) = \frac{g}{4!}\int \mathrm{d}^{d}x\,\frac{\delta^{4}}{\delta J^{4}(x)}.$$4 counts vertices with two identical bubbles, and $$\mathcal{V}\Big(\frac{\delta}{\delta J}\Big) = \frac{g}{4!}\int \mathrm{d}^{d}x\,\frac{\delta^{4}}{\delta J^{4}(x)}.$$5 counts sets of $$\mathcal{V}\Big(\frac{\delta}{\delta J}\Big) = \frac{g}{4!}\int \mathrm{d}^{d}x\,\frac{\delta^{4}}{\delta J^{4}(x)}.$$6 identical lines connecting the same pair of vertices (Hue et al., 2010). This formula was derived in a way that does not depend on field spin, only on whether fields are self-conjugate or non self-conjugate, so the scalar-theory combinatorics extends to QED, QCD, and related theories after the appropriate identification of field classes (Hue et al., 2010).

One consequence is that external connected diagrams built only from vertices with three different fields have symmetry factor $$\mathcal{V}\Big(\frac{\delta}{\delta J}\Big) = \frac{g}{4!}\int \mathrm{d}^{d}x\,\frac{\delta^{4}}{\delta J^{4}(x)}.$$7, as in spinor QED. Another is that distinct directed diagrams may collapse to the same undirected schematic topology, in which case the effective inverse symmetry factor is the sum of the individual inverse symmetry factors,

$$\mathcal{V}\Big(\frac{\delta}{\delta J}\Big) = \frac{g}{4!}\int \mathrm{d}^{d}x\,\frac{\delta^{4}}{\delta J^{4}(x)}.$$8

This matters whenever combinatorial summation is performed on “equivalent” undirected classes rather than fully directed graphs (Hue et al., 2010).

The emphasis on automorphisms also clarifies a common misconception: symmetry factors are not ad hoc normalizations appended at the end of a computation. They are the precise quotient between the raw count of Wick-contraction realizations and the number of distinct graph topologies contributing to the same integrand. Any successful combinatorial summation method must reproduce them, either explicitly through automorphism groups or implicitly through an equivalent combinatorial mechanism (Hue et al., 2010, Castro et al., 2018).

4. Exact counting and asymptotic enumeration of connected graphs

A second line of work studies diagram summation through exact counting recurrences and generating functions. One recurrence arises from connected multigraphs with $$\mathcal{V}\Big(\frac{\delta}{\delta J}\Big) = \frac{g}{4!}\int \mathrm{d}^{d}x\,\frac{\delta^{4}}{\delta J^{4}(x)}.$$9 vertices and cyclomatic number $g^m$0, weighted by inverse automorphism order: $g^m$1 These quantities satisfy

$g^m$2

with $g^m$3 and vanishing outside the physical range (Brouder et al., 2013). The first term corresponds to adding an edge that raises the loop number, while the double sum corresponds to joining two smaller connected graphs by a bridge edge (Brouder et al., 2013).

After normalization,

$g^m$4

the recurrence becomes polynomially weighted and more tractable. For fixed $g^m$5, the generating functions are rational; for fixed $g^m$6, one recovers classical graph sequences. In particular,

$g^m$7

encodes tree-level connected graphs, while

$g^m$8

is identified with OEIS A001865, the connected unicyclic multigraph count (Brouder et al., 2013). This shows that exact weighted enumeration of connected Feynman graphs is naturally linked to classical combinatorial sequences (Brouder et al., 2013).

A more specialized zero-dimensional QFT framework gives exact formulas for connected Wick-contraction counts with $g^m$9 external legs at order $4m$0. There, the total number of contractions is

$4m$1

and connected counts are extracted through logarithms and composition sums. The connected diagram count with external-leg normalization is

$4m$2

with explicit formulas for arbitrary $4m$3 in terms of compositions and generating-function inversions (Castro et al., 2020). This approach is purely combinatorial but useful for asymptotic analysis, since the coefficients display the factorial growth characteristic of perturbation theory (Castro et al., 2020).

A further asymptotic development uses zero-dimensional path integrals represented as generalized hyperelliptic curves. For an action

$4m$4

the formal integral

$4m$5

encodes weighted sums over all Feynman multigraphs, and the asymptotics of its coefficients are obtained from the dominant singularities of the inverse function defined by

$4m$6

The resulting all-order asymptotic expansion shows that disconnected, connected, 1PI, and skeleton graph counts all exhibit factorial growth governed by the dominant saddle points of the zero-dimensional action (Borinsky, 2017). This suggests a direct link between renormalized asymptotic enumeration and resurgence-type structures (Borinsky, 2017).

5. Determinantal, dynamic-programming, and many-body summation schemes

In many-body physics, combinatorial summation often aims not merely to count graphs but to evaluate the total contribution of all connected diagrams at fixed perturbative order. The combinatorial summation (CoS) framework does this by returning from determinant expansions to the underlying permutation combinatorics. The determinant is organized in terms of cycle covers and Mahajan–Vinay clow sequences, and dynamic programming is used to sum them with constraints enforcing connectedness and model-specific structure (Shi, 16 Jul 2025).

The original CoS scheme tracks visited interaction structures and enforces connectivity directly in the DP state. For pure contact interactions, a three-state interaction mask per interaction gives complexity

$4m$7

while a more general visited-set construction gives

$4m$8

for cases where the full visited-set data are required (Shi, 16 Jul 2025). The 2025 extension generalizes this to models with $4m$9 internal interaction types, reducing the effective exponential base from $(J\Delta J)^{2m}$0 to $(J\Delta J)^{2m}$1, with overall cost

$(J\Delta J)^{2m}$2

The key combinatorial compression is that each interaction carries a single mask state chosen from

$(J\Delta J)^{2m}$3

rather than $(J\Delta J)^{2m}$4 independent three-state masks (Shi, 16 Jul 2025).

This framework is relevant to models such as the SU(4) Hubbard model on a honeycomb lattice, spinless fermions with frustrated interactions, and SU(2) Hubbard models on Kagome lattices (Shi, 16 Jul 2025). Its significance for combinatorial summation lies in the fact that it deterministically sums all connected bare-$(J\Delta J)^{2m}$5 diagrams of a given order while preserving cancellations that would be difficult to capture by naive stochastic sampling (Shi, 16 Jul 2025).

A related but conceptually distinct development appears in the combinatorial analysis of bold diagrams in many-body perturbation theory. There, the self-energy is expressed as a sum over 1PI diagrams, and every such diagram admits a unique decomposition into a 2PI skeleton together with maximal Green-function insertions. This yields the bold expansion

$(J\Delta J)^{2m}$6

where bold lines carry the full Green function rather than the bare propagator (Lin et al., 2018). The nontrivial combinatorial statement is that the redundancy of reconstructing a given bare diagram from its skeleton plus insertions is exactly compensated by the relation between the symmetry factor of the full diagram and those of the skeleton and insertions (Lin et al., 2018).

This bold framework underlies systematic approximations such as Hartree–Fock, GF2, and GW, all of which can be interpreted as truncations or partial resummations of the skeleton expansion (Lin et al., 2018). A plausible implication is that bold-diagram expansions are best understood not as heuristic reorganizations but as exact combinatorial resummations whose validity depends on the uniqueness of skeleton decompositions. The same work also shows why an analogous bold expansion does not exist for the free energy: closed connected diagrams do not admit unique skeleton decompositions (Lin et al., 2018).

6. Resummation, worldline master formulas, and constructive tree expansions

Another major branch of the subject does not enumerate individual graph topologies first, but packages many graphs into a single analytic object. In the worldline formalism, the one-loop $(J\Delta J)^{2m}$7-photon amplitude in scalar QED is represented by a master integral over proper time and insertion positions,

$(J\Delta J)^{2m}$8

where the integration over all $(J\Delta J)^{2m}$9 automatically sums all permutations of external legs along the loop (Ahmadiniaz et al., 2022). At higher loops, worldline sewing combines families of topologically distinct Feynman diagrams into common parameter integrals, so “summing diagrams” becomes an exercise in evaluating global moduli integrals rather than adding individual graph contributions (Ahmadiniaz et al., 2022).

The same formalism supports nonperturbative gauge-transformation identities. In particular, the generalized Landau–Khalatnikov–Fradkin relation for $J\to 0$0-point fermionic correlators is derived directly at the worldline path-integral level, effectively summing all longitudinal-photon insertions into a compact multiplicative endpoint factor (Ahmadiniaz et al., 2022). This illustrates a broader principle: once the diagrammatic combinatorics are encoded in path-integral insertion variables, many distinct graph families can be resummed algebraically (Ahmadiniaz et al., 2022).

Constructive quantum field theory provides a more radical resummation. Instead of summing over connected Feynman graphs directly,

$J\to 0$1

one distributes each graph amplitude over its spanning trees using universal rational weights

$J\to 0$2

where $J\to 0$3 counts the Hepp sectors in which $J\to 0$4 is the leading tree under Kruskal’s algorithm (Rivasseau et al., 2013). This yields the exact identity

$J\to 0$5

Applied naively to ordinary bosonic Feynman graphs this does not suffice for convergence, but after passing to intermediate-field representations and collapsing ordinary cycles to “loop vertices”, one obtains the Loop Vertex Expansion, a convergent tree-indexed series whose sum reproduces the Borel sum of the original perturbative expansion (Rivasseau et al., 2013).

This constructive resummation highlights a distinct meaning of “combinatorial summation”: not merely grouping equivalent contractions or topologies, but reindexing the entire perturbative series by trees rather than graphs, using weights derived from sector combinatorics and positivity properties of associated matrices (Rivasseau et al., 2013).

7. Extensions, special geometries, and contemporary directions

Recent work extends combinatorial summation into several less traditional directions. In the single-polaron problem, all self-energy diagrams of a given order can be generated iteratively from noncrossing diagrams encoded by Dyck paths and Stieltjes–Rogers polynomials, with the Ward–Takahashi identity used to generate vertex corrections from lower-order self-energy diagrams. The total number of self-energy diagrams at order $J\to 0$6 is

$J\to 0$7

the number of irreducible diagrams is given by a weighted Dyck-path sum, and the noncrossing subclass is counted by Catalan numbers (Miškić et al., 27 May 2025). This organization is designed to improve sign cancellation in approximation-free numerical summation (Miškić et al., 27 May 2025).

In scattering-amplitude geometry, ordinary planar Feynman diagrams at $J\to 0$8 are generalized to planar collections and planar matrices of metric trees for higher $J\to 0$9 biadjoint amplitudes. For $\phi^4$0 one sums over planar collections, and for $\phi^4$1 over planar matrices. Enumeration is already substantial: for $\phi^4$2, $\phi^4$3, and $\phi^4$4 there are $\phi^4$5, $\phi^4$6, and $\phi^4$7 planar collections, while for $\phi^4$8 and $\phi^4$9 there are $\phi^4$00 and $\phi^4$01 planar matrices (Cachazo et al., 2019). Here combinatorial summation is intertwined with tropical Grassmannians, mutation graphs, and positive geometries, so the “diagram” being summed is itself a higher-dimensional array of tree data (Cachazo et al., 2019).

There are also attempts to perform all-orders summation for specific gauge-theory observables directly at the functional level. One example rewrites the dressed photon propagator in QED using Fradkin functionals and Gaussian linkage operators, thereby summing broad classes of vacuum-polarization graphs and extracting leading-logarithmic behavior of $\phi^4$02 in increasingly refined approximations (Fried et al., 2010). Although model-dependent, this illustrates a functional version of combinatorial summation where graph classes are generated by operator algebra rather than explicit graph enumeration (Fried et al., 2010).

Finally, the connection between Feynman diagrams and string worldsheets has recently been made explicit in a half-BPS sector of $\phi^4$03 SYM. There, each two-matrix-model ribbon graph maps to a Strebel graph and hence to a point in moduli space, while the same graph encodes a Belyi map to the target sphere. The sum over Feynman diagrams becomes a sum over discrete lattice points in string moduli space, valid to all orders in the $\phi^4$04 expansion (Gopakumar et al., 2024). This suggests a broader interpretation of combinatorial summation as a reorganization by worldsheet topology and moduli rather than by graphs alone (Gopakumar et al., 2024).

A final contemporary development concerns renormalization with multi-indices. Here “pre-Feynman diagrams” are encoded as monomials in variables $\phi^4$05 recording vertex valences, and a multi-index Hopf algebra with extraction–contraction coproduct and twisted antipode reproduces BPHZ renormalization while acting directly on the expansion of the measure. In the $\phi^4$06 example, the renormalization map transforms the formal exponential of the bare interaction into the exponential of a renormalized interaction with shifted couplings, giving a direct combinatorial route from renormalized diagram expansion to renormalized measure (Bruned et al., 14 Jan 2025). This suggests a bridge between Connes–Kreimer-type renormalization and the multi-index structures familiar from singular SPDE theory (Bruned et al., 14 Jan 2025).