Some quantum field theories with extra symmetry, called supersymmetric theories, have a remarkable property. Their infinite-dimensional path integrals can be evaluated exactly, because they collapse to a finite sum of residues of a rational function of a few complex variables. This collapse is called supersymmetric localization. Physicists use it routinely, but nobody has asked how expensive that final sum is. In particular, nobody has asked which property of the theory's input data decides whether the sum is cheap or explodes.
We propose that the answer is already known in another field. The input data of an abelian gauge theory is an integer charge matrix: one row per matter field, one column per gauge direction. In polyhedral combinatorics, the subdeterminants of an integer matrix (determinants of its square submatrices) are known to control path lengths in linear programming, the "thickness" of polytopes, and lattice-point counts. The note's claim is that the same subdeterminants control the number of poles and flux sectors in localization formulas. It then turns that claim into an algorithm with a provable cost bound and tests it.
Core Claim
The setting is the twisted index of three-dimensional $\mathcal{N}=2$ gauge theories on a circle times a Riemann surface $\Sigma_g$ of genus $g$. The twisted index is an integer-valued count of supersymmetric ground states. For such theories, the Coulomb-branch localization formula is a Jeffrey–Kirwan (JK) residue. The JK residue is a rule, depending on an auxiliary vector $\eta$, that decides which intersection points of the singular hyperplanes of the integrand are kept and with what sign. The formula is summed over magnetic fluxes $m \in \mathbb{Z}^r$ (The 3d twisted index and wall-crossing; The twisted index and topological saddles).
The singular hyperplanes come from the charge vectors $Q_i \in \mathbb{Z}^r$. Here $r$ is the gauge rank and the variables live on a complex torus. So the equations $x^{Q_i}=y_i$ for a basis $S$ of $r$ charges have exactly $|\det Q_S|$ solutions. This gives a direct count of JK-selected poles,
$$ P_\eta(Q) \;=\; \sum_{S:\ \eta \in \mathrm{cone}(Q_S)} |\det Q_S| \;\le\; \sum_{S \text{ basis}} |\det Q_S| \;=\; M_Q(1,1), $$
where $M_Q$ is the arithmetic Tutte polynomial of the charge configuration. The final equality is the volume specialization of the zonotope Ehrhart identity proved in Ehrhart polynomial and arithmetic Tutte polynomial. That work gives this specialization for general integer vector configurations; we reuse it and do not re-derive it.
The flux sum is a second enumeration problem. The contributing fluxes, after the standard geometric-series truncation, are lattice points in a polytope whose constraint matrix is again $Q$. We therefore claim that the number of flux sectors is governed by facet widths, which Enumerating Integer Points in Polytopes with Bounded Subdeterminants bounds by subdeterminants of $Q$.
Two consequences follow. First, if the largest subdeterminant of $Q$ is bounded, the index is computable by the JK route in time polynomial in the number of fields for fixed rank. The totally unimodular case, where every subdeterminant is $0$ or $\pm1$ as for quiver incidence matrices, reduces to counting bases in a cone. Second, wall-crossing, the jump of the index as a real parameter crosses a wall, becomes a difference of such cone-restricted sums across a wall of the chamber decomposition. This gives a combinatorial reading of the abelian wall-crossing formula of The 3d twisted index and wall-crossing.
Relation to Prior Work
The localization papers supply the object and the correctness targets.
- The 3d twisted index and wall-crossing derives abelian wall-crossing.
- The twisted index and topological saddles shows that residues "at infinity" arising from nonzero Chern–Simons levels are topological saddles. It proves that vortex plus topological saddles reproduce the JK answer for $U(1)$.
Neither paper asks what the evaluation costs. Both leave non-abelian groups open because pole structures become uncontrolled. Our work does not change their physics. It measures and bounds their formula.
Two discrete-geometry papers supply the method.
- Solving Totally Unimodular LPs with the Shadow Vertex Algorithm shows that unimodularity bounds the vertex–facet condition number by $n$, giving strongly polynomial pivot counts.
- Enumerating Integer Points in Polytopes with Bounded Subdeterminants bounds facet widths by subdeterminants. Its main theorem needs every $n\times n$ minor to be nonzero.
That last hypothesis matters. Physical charge matrices routinely have dependent charges, for example an adjoint field or two fields of equal charge. For such matrices some minors vanish and the hyperplane arrangement is degenerate. We treat this constraint directly: the degenerate case is exactly where JK needs perturbation, and exactly where the enumeration theorem is silent.
On Counting Integral Points in a Convex Rational Polytope is the closest methodological relative. It evaluates multidimensional residues iteratively, one variable at a time, and states that it could not bound the cost because poles multiply at each step. We supply the missing bound in the setting where the pole set is a determinantal count. We also compare our cone-restricted enumeration against its iterated scheme.
Other connecting work does not touch this question:
- On exact correlation functions in SU(N) N = 2 superconformal QCD links localization to computational simplification through Toda decoupling of differential equations, not residue combinatorics.
- Multimatrix models and tri-Sasaki Einstein spaces sums over quiver spanning trees, but at large $N$ and for sphere partition functions.
- Scattering Amplitudes from Intersection Theory localizes amplitudes on hyperplane arrangements without any subdeterminant analysis.
Research Program
Abelian theorem. For $U(1)^r$ with charge matrix $Q$, largest absolute subdeterminant $\Delta$, $n$ chirals and genus $g$, we will prove a bound on the JK evaluation cost. The pole count is bounded by $P_\eta(Q)$. The flux-sector count is bounded through facet width by a function of $\Delta$, $r$ and the Chern–Simons levels. We will state the bound explicitly and show it is tight on families.
Enumeration algorithm. The algorithm walks the chambers of the arrangement using shadow-vertex-style pivots. Each basis is visited only if its cone contains $\eta$. The $|\det Q_S|$ torsion poles per basis are then enumerated by Smith normal form. Unimodularity makes each basis contribute one pole, which is where strongly polynomial behaviour comes from.
Degenerate case. We will handle zero minors by a symbolic perturbation of $\eta$ and of the fugacities. Crucially, we will prove that the pole count is then bounded by the arithmetic Tutte data of the matroid of $Q$ rather than by a count of generic perturbed poles. This is the step that the non-zero-minor hypothesis in the enumeration paper blocks. It is also the step most likely to fail.
Wall-crossing and non-abelian extension. We will express wall-crossing as a signed sum over bases crossing a wall and check it against the abelian formula. We will then extend to $U(N)$ and quivers by restricting to the Cartan subalgebra. There, vector multiplets contribute numerators but not poles, so matter weights alone define the arrangement. Adjoint matter makes the arrangement degenerate, which is precisely why the degenerate step matters.
Conjecture. We will test numerically whether $P_\eta(Q)$ itself equals a specialization of an arithmetic Tutte-type polynomial refined by the chamber of $\eta$. We state this as a conjecture, not a claim.
Baselines
Naive reference. Enumerate every $r$-subset of charges, solve for its poles, and apply the JK sign rule, looping over a generous flux box. This is correct by construction, is the implementation most practitioners write, and sets the floor any bound must beat. It will share our code for residue extraction, so only pole and flux selection differ.
Iterated residues. One-variable-at-a-time residues in the style of On Counting Integral Points in a Convex Rational Polytope. It is the strongest general-purpose exact method for multidimensional rational residue sums, and its pole blow-up is the documented bottleneck we claim to control. We will run it with the same exact-arithmetic backend and the best variable ordering found by searching orderings within the same time budget.
Higgs-branch sum. The Higgs-branch evaluation of the index as a sum of vortex saddles plus topological saddles, following The twisted index and topological saddles. It is the physically preferred competitor and is exact where available, currently for $U(1)$. We will implement it for the abelian families where the paper's classification applies and compare on identical theories.
No reported runtime or pole-count result was found on Lacuna for any of these methods. The only quantitative correctness standard in the file is exact agreement of the JK and saddle sums in the $U(1)$ case.
Ablation. We will isolate the key ingredient by running our chamber walk with the subdeterminant-based flux truncation replaced by a fixed generous flux box. This separates the value of the polytope bound from the value of cone-restricted pole selection.
Success criteria. Beating the baselines means two things. The integer index must be identical on every instance. Pole counts and wall-clock time must grow polynomially in $n$ at fixed $\Delta$ and $r$, where iterated residues grow faster, and must track the predicted $P_\eta(Q)$ within a constant factor.
Why not just…?
Why not just use iterated one-variable residues, which already work? They do work, and they are our main baseline. However, the authors of On Counting Integral Points in a Convex Rational Polytope report that the number of poles grows at each step with no bound they could prove. Intermediate poles can be spurious, cancelling only in the final sum. If our experiments show that a good variable ordering keeps iterated residues within a constant of $P_\eta(Q)$, the algorithmic claim weakens to a purely analytic bound. We would report that honestly.
Why not just evaluate on the Higgs branch? For $U(1)$ it is exact and often short, per The twisted index and topological saddles. But topological-saddle moduli spaces are unclassified for non-abelian groups, which is exactly where cost matters. A related real threat is the Bethe-vacuum form of the index, whose cost scales with the number of vacua. It is not in our file, so we cannot cite it here. We will implement it for abelian families and report where it wins, which is likely at small $\Delta$ with large genus.
Why not algebraic partial fractions via Gröbner bases or the Nullstellensatz? On Counting Integral Points in a Convex Rational Polytope tried the Nullstellensatz route and found it feasible only for few constraints, because its linear systems explode. We include it at small $n$ to confirm that crossover and do not pursue it further.
Why not Monte Carlo or machine learning? Randomized volume estimation reaches below 2% mean error up to 100 dimensions in Practical Polytope Volume Approximation. Classifiers reach above 90% on quiver duality in Quiver mutations, Seiberg duality, and machine learning. Both, however, estimate or guess an integer whose exact value and jumps are the physics. Neither can certify an index or a wall-crossing.
Proposed Experiments
No benchmark for localization cost exists in the file, so we build one.
Correctness anchors. These are the $U(1)$ theories with Chern–Simons levels from The twisted index and topological saddles. They also test the abelian wall-crossing formula of The 3d twisted index and wall-crossing.
Controlled families. These are $U(1)^r$ theories with random integer charge matrices at fixed largest subdeterminant $\Delta \in \{1,2,4,8\}$, for $r$ up to 6 and $n$ up to 40. The families include totally unimodular incidence matrices of random quivers and necklace quivers from Multimatrix models and tri-Sasaki Einstein spaces.
Degenerate families. These have repeated and adjoint-like charges, which zero out minors.
Non-abelian set. This set covers $U(2)$ and $U(3)$ with fundamentals and adjoints, restricted to the Cartan.
Metrics. We record exact index agreement across all methods, poles actually evaluated versus predicted $P_\eta(Q)$ and versus $M_Q(1,1)$, flux sectors visited versus the facet-width bound, and wall-clock time at matched hardware.
Decisive checks. The first check is whether runtime at fixed $\Delta$ scales polynomially in $n$ while iterated residues do not. Failure here falsifies the main claim. The second check is whether degenerate charge matrices stay within the Tutte-data bound. Failure here tells us the degenerate step needs a new idea. The third check is whether the wall-crossing jump computed as a signed basis sum matches the known abelian formula on every instance.
References
- The 3d twisted index and wall-crossing
- The twisted index and topological saddles
- Ehrhart polynomial and arithmetic Tutte polynomial
- Enumerating Integer Points in Polytopes with Bounded Subdeterminants
- Solving Totally Unimodular LPs with the Shadow Vertex Algorithm
- On Counting Integral Points in a Convex Rational Polytope
- On exact correlation functions in SU(N) N = 2 $$ \mathcal{N}=2 $$ superconformal QCD
- Multimatrix models and tri-Sasaki Einstein spaces
- Scattering Amplitudes from Intersection Theory
- Practical Polytope Volume Approximation
- Quiver mutations, Seiberg duality, and machine learning