When physicists propose a new theory of gravity, or a new way to couple a field to spacetime curvature, the first question is whether it is sick. There are two main kinds of sickness. The first is a tachyonic instability: a mode with negative effective mass-squared that grows exponentially, like a ball resting on top of a hill. The second is an Ostrogradsky ghost: a mode with negative kinetic energy, which appears when equations of motion contain more than two time derivatives and lets the vacuum decay without bound. Papers establish these verdicts one theory at a time, using a patchwork of checks. These include counting propagating modes around flat space, reading off signs of kinetic and mass terms, estimating the energy at which modes become strongly coupled, perturbing on curved backgrounds, and inspecting the poles of scattering amplitudes. Nobody has asked which of these checks are actually needed, or which pathologies no cheap check can tell apart.
This note proposes treating that question as a sensor-placement problem from combinatorics. The possible pathologies of a theory family play the role of network faults. The diagnostic checks play the role of sensors. A minimal battery of checks that gives every pathology pattern a unique signature is a variant of an identifying code, a set of sensors whose readings pin down exactly which node failed. The central object is the twin: two pathology hypotheses that every available check answers identically, so no external sensor can separate them. The concrete claim is that twins in this problem correspond to known physical degeneracies. The clearest cases are the shared strong-coupling scale of tensor and longitudinal modes in non-minimally coupled Proca theory, and the Proca and Kalb-Ramond duality. It follows that separating twins requires a specific, and more expensive, class of check. Once the problem is stated this way, it can be computed and falsified.
Core Claim
Fix a family of theories, for example a Proca vector with couplings $\alpha R A^2$ and $\beta R_{\mu\nu}A^\mu A^\nu$. Each point in its parameter space carries some pattern of pathologies, possibly several at once:
- a ghost in a given spin sector,
- a tachyonic or runaway mode on de Sitter space,
- gradient instabilities,
- strong coupling of a given mode at a given scale.
Each diagnostic check returns a discrete answer. Examples are "sign of the spin-1 kinetic coefficient," "does the flat-space strong-coupling scale scale as $(\beta/M_{\rm pl}^2 m^2)^{1/3}$," "negative-frequency modes on FLRW," and "negative residue or negative-mass pole in the tree four-point amplitude."
These pieces form a bipartite graph between pathology hypotheses and checks. A discriminating battery is a set of checks $T$ that gives every hypothesis $h$ a distinct response set $N(h)\cap T$. If a pair of hypotheses is joined to exactly the same checks, it is a twin pair, and no battery drawn from that check class can separate it.
The claim has two parts. First, for the theory families in the file, the minimal battery is small but must include at least one curved-background check. Flat-space and amplitude-level checks alone leave twins. Second, those twins are exactly the pairs the literature already treats as physically entangled:
- In Unveiling the Inconsistency of the Proca Theory with Nonminimal Coupling to Gravity, Hell shows that the Ricci-tensor coupling makes tensor modes and the longitudinal vector mode strongly coupled at the same scale. Any scale-based check therefore sees them as twins.
- In On the duality of massive Kalb-Ramond and Proca fields, Hell shows that a massive Kalb-Ramond field and a quartic Proca field share one strong-coupling scale, carried by different modes in each picture.
A secondary claim concerns field redefinitions. Hell's disformal transformation $g_{\mu\nu}\to g_{\mu\nu}-(\beta/M_{\rm pl}^2)A_\mu A_\nu$ removes the tensor strong coupling and the runaway modes. In this language, it splits a twin pair: after the redefinition, a cheap check separates the hypotheses. That gives a concrete test for the open problem the file highlights, namely whether an analogous redefinition exists for the Kalb-Ramond two-form of On the Kalb-Ramond field with non-minimal coupling to gravity. Any candidate redefinition can be checked for whether it splits the corresponding twins and whether it introduces new higher-time-derivative (Ostrogradsky) hypotheses.
Research Program
Theory families. We build a symbolic catalogue of five families, chosen because each has a published verdict that can serve as ground truth:
- non-minimal Proca, from Hell's paper above and from On the cosmological degrees of freedom of Proca field with non-minimal coupling to gravity;
- non-minimal Kalb-Ramond, from Hell and Obata;
- New General Relativity, whose parameter space Revisiting stability in new general relativity finds generically plagued by vector-mode and ghost or tachyonic instabilities, with only a few healthy branches;
- conformal gravity under different boundary conditions, from On the ghost problem of conformal gravity, where Neumann conditions plus an extra condition in de Sitter, or "alternative" conditions in Minkowski, remove the ghost classically;
- curved-space Galileons and DBI scalars, from Symmetries for Galileons and DBI scalars on curved space, where keeping the nonlinear shift symmetries on de Sitter forces a trade-off between ghosts and tachyons.
Parameter cells and hypotheses. Each family's parameter space is divided into cells by the sign changes of the relevant coefficients. Each cell, combined with each background (Minkowski, de Sitter, FLRW), generates hypotheses. A hypothesis is a set of up to $k$ simultaneously sick modes, matching the $k$-fault setting.
Checks. We implement roughly a dozen checks, grouped into four tiers by cost:
- flat-space spin-projector sign checks;
- strong-coupling scale extraction from cubic and quartic vertices;
- linear perturbation on de Sitter or FLRW backgrounds;
- amplitude-pole checks.
The amplitude tier is where the vertex-operator direction plays a supporting role. For field-theory families we compute residues with ordinary Feynman rules. For two calibration cases we use worldline and worldsheet vertex-operator results. Tree-level 11D supergravity amplitudes from the pure spinor worldline supplies a BRST- and gauge-invariant four-point amplitude that must trigger no pathology hypothesis; this is a negative control. It was made possible by the non-minimal ghost-number-zero construction of The 11D pure spinor ghost number zero vertex operator. The bosonic string amplitudes built from coherent and DDF vertex operators in Scattering, absorption and emission of highly excited strings contain the $m^2=-1/\alpha'$ tachyon pole and must trigger it; this is a positive control.
These calibrations also test a conceptual point. The BRST ghosts of the pure spinor formalism are gauge-fixing variables, not Ostrogradsky ghosts. A well-built amplitude check must not mistake one for the other.
Solving for batteries. Given the incidence graph, we search for minimum-cost discriminating batteries, with cost weighted by each check's tier. Candidate sensors are restricted to the check side of the bipartite graph. The Padoa-definability encoding of Solving the Identifying Code Set Problem with Grouped Independent Support handles this naturally, because it looks for an independent support over a designated variable subset. We then report three things: the twin classes under each check tier, the minimal battery, and whether candidate field redefinitions change the twin structure.
Baselines
The naive reference point is to run every check on every parameter cell. By construction this is always correct whenever any battery is correct, so it sets the accuracy ceiling. The comparison is therefore about cost, and about the structural information exhaustive checking does not give: it reports verdicts but never says which checks were redundant or which hypotheses are twins. The proposal beats it if a battery of at most half the checks, weighted by tier cost, reproduces the exhaustive verdicts on every cell.
The strongest existing method for the combinatorial core is gismo, from Solving the Identifying Code Set Problem with Grouped Independent Support. On its 50 undirected networks benchmark it scaled to 21,363 nodes, where integer linear programming failed beyond 494 nodes. It reached a median speedup of up to 520x, with sensor sets typically within 10% of the optimum. On the physics-diagnosis task itself, no reported result was found on Lacuna. We therefore compare against the standard integer linear program on the same incidence graphs, under the same time limit and machine. Our adapted gismo counts as an improvement only if it returns batteries of equal or smaller cost on the larger $k=2$ New General Relativity graphs where the integer program times out. At $k=1$ we expect the two to tie, and we will report that.
The physics gold standard is a full Hamiltonian constraint analysis per cell, which counts degrees of freedom and determines their signs. We treat it as an oracle for ground truth on flat backgrounds wherever the literature did not already supply the verdict. Its cost is the denominator for our cost savings.
The key ablation removes the curved-background tier. If the twin classes do not change, the claim that separating twins needs curved checks is false. A second ablation removes the amplitude tier, to see whether amplitude checks ever provide unique discrimination or are always redundant with the spin-projector signs.
Why not just…?
Why not just run the full Hamiltonian analysis for every theory, as papers already do? For a handful of theories this is feasible, and it is our oracle. It does not scale, though. The New General Relativity analysis in Revisiting stability in new general relativity was restricted to linearized Minkowski perturbations, so curved-background runaways, the pathology Hell found for Proca, lie outside it. The Hamiltonian count also cannot report strong-coupling twins, because those are properties of interaction vertices, not of the free theory. If the oracle alone matched every published verdict, including curved-background ones, at comparable cost, the proposal would reduce to bookkeeping. The tier ablation is designed to show whether that is the case.
Why not just use an integer program or greedy set cover and skip the definability machinery? At small sizes this may be enough, and we expect a tie at $k=1$. The reason to go further comes from Solving the Identifying Code Set Problem with Grouped Independent Support, which reports that integer-program constraints grow as $O(|V|^k)$ for $k$ simultaneous faults. Multi-pathology hypotheses in New General Relativity push $|V|$ into that regime. If the integer program still finishes on our largest graphs, we will say so and drop the solver claim. The twin-structure results do not depend on which solver is used.
Why not just note that the pathologies are known to be degenerate and stop there? Hell's papers already show that the modes share a scale. What they do not show is that no check in the flat-space class can split them, or which cheaper check a disformal redefinition makes decisive. Those are negative, "no check exists" statements. They become provable here as twin-freeness failures of an explicit incidence graph. The same statements also tell anyone searching for a Kalb-Ramond analogue of the disformal fix exactly what that fix must achieve.
Decisive Checks
The first test is calibration. Every published verdict across the five families must be reproduced, the pure spinor 11D supergravity amplitude must flag nothing, and the bosonic string amplitude must flag its tachyon. Any miss means a check is mis-specified.
The second test is the core prediction. The Proca tensor/longitudinal pair and the Proca/Kalb-Ramond dual pair should be twins under every flat-space and scale check, and they should be split by the de Sitter runaway check. A cheap flat-space check that separates them refutes the claim.
The third test concerns redefinitions. After applying Hell's disformal transformation, the Proca twin pair should become separable at the flat-space tier. We will then test candidate two-form redefinitions for the same effect, and in each case count any new higher-time-derivative hypotheses the redefinition introduces.
The last check comes from the New General Relativity scan with up to two simultaneous pathologies. There we report battery size, cost against exhaustive checking and the oracle, and solver scaling against the integer program.
References
- Unveiling the Inconsistency of the Proca Theory with Nonminimal Coupling to Gravity
- On the duality of massive Kalb-Ramond and Proca fields
- On the Kalb-Ramond field with non-minimal coupling to gravity
- On the cosmological degrees of freedom of Proca field with non-minimal coupling to gravity
- Revisiting stability in new general relativity
- On the ghost problem of conformal gravity
- Symmetries for Galileons and DBI scalars on curved space
- Tree-level 11D supergravity amplitudes from the pure spinor worldline
- The 11D pure spinor ghost number zero vertex operator
- Scattering, absorption and emission of highly excited strings
- Solving the Identifying Code Set Problem with Grouped Independent Support