Data-Defined Stability Boundaries for Slowly Unstable Galileon Cosmologies

Some candidate theories of dark energy contain a flaw that is real but may be slow. These are scalar fields whose small mass is protected by a symmetry, and when they are written on an expanding universe they develop a tachyonic instability: a perturbation that grows exponentially, like a ball balanced on a hilltop. For curved-space Galileons derived from branes, the growth time can be comparable to the age of the universe, and the authors of Galileons on cosmological backgrounds call such models "phenomenologically viable." The wider stability literature instead treats runaway modes as fatal. So today, whether a model is benign or catastrophic is settled by argument, not by measurement.

This note proposes to settle it with data. Whether a slow instability matters depends on two things: how fast it grows, which the theory's couplings fix, and how large the seed perturbation was, which the theory leaves free. We treat the seed amplitude as a nuisance parameter, meaning an unknown the fit must account for without it being the quantity of interest. We then fit both the couplings and the seed with an unbinned extended maximum likelihood fit. That is a statistical method that scores every observed object individually against a continuous prediction and also scores how many objects were seen. The result replaces a verbal viability criterion with a confidence region whose edge is set by the data. A second, statistical claim sits alongside it: near this edge, the usual shortcut for turning a likelihood into error bars (Wilks' theorem) should undercover, so the method must be calibrated explicitly.

Core Claim

The Galileon theories of New Class of Effective Field Theories from Embedded Branes and Symmetries for Galileons and DBI scalars on curved space describe our 4D spacetime as a sheet (a brane) bending inside a 5D space. Built this way, their equations of motion are second order by construction. This means they avoid the Ostrogradsky instability: the negative-energy "ghost" states that generically appear when a theory's equations contain higher time derivatives.

Ghost-freedom is therefore a hard condition that does not depend on time: a kinetic coefficient must be positive. This matches how the literature treats ghosts as non-negotiable. On the ghost problem of conformal gravity shows that ghosts in conformal gravity survive even the favourable Maldacena boundary conditions. Revisiting stability in new general relativity discards unstable branches outright.

The tachyonic condition is different in kind. Its severity is a growth factor, and that factor only acquires meaning relative to an observation window and a seed amplitude. Our claim is that this asymmetry should shape the statistical model. Ghost conditions enter as a hard boundary on the parameters. Tachyonic growth enters the prediction itself: the growing mode is evolved from an unknown seed amplitude $\epsilon$ and changes the expansion history the data see.

Concretely, for couplings $\theta$ of the five brane Lagrangians $\mathcal{L}_1$ to $\mathcal{L}_5$, we define a growth factor:

$$ G(\theta) = \exp\!\int_{t_i}^{t_0} \sqrt{\max\!\big(0,\,-m_{\rm eff}^2(\theta,t)\big)}\,dt . $$

Here $m_{\rm eff}^2(\theta,t)$ is the effective mass-squared of the field perturbation computed on the minisuperspace background (the reduction to a field that depends only on time, following Galileons on cosmological backgrounds). The times $t_i$ and $t_0$ are the start of the fitted epoch and today. The combination $\epsilon\, G(\theta)$ is the perturbation amplitude today.

Where $\epsilon\, G(\theta)$ exceeds a fixed fraction of the background field, linear theory no longer applies and the model makes no prediction. We treat that region as a second, data-independent excluded zone, not as a region the fit disfavours.

The novelty is not "fit Galileons to data" in the abstract. It is three things together:

  • the stability of the model becomes a jointly inferred quantity in $(\theta,\epsilon)$;
  • the boundary between benign and catastrophic is reported as a confidence region;
  • the statistical behaviour of that region near the excluded zones is calibrated rather than assumed.

Research Program

Theory layer. We take the FRW (expanding-universe) Galileon construction on power-law backgrounds $a(t)\sim t^\alpha$ from Galileons on cosmological backgrounds. We start with the case where only the lowest-order terms are switched on, because the $\mathcal{L}_4$ and $\mathcal{L}_5$ terms are known to be dense. We then extend the terms order by order.

For each $\theta$, the code computes three things:

  • the ghost-free kinetic condition;
  • the history of $m_{\rm eff}^2$;
  • the background expansion, with the growing mode added from seed $\epsilon$.

The mass terms whose coefficients are fixed by the bulk geometry, $m^2\propto R$, are kept exactly as derived. They are what distinguish these theories from a generic quintessence field with a hand-chosen potential.

Observable layer. The natural event-level data are catalogues of individual distance indicators. The survey rate of such objects depends on comoving volume, which depends on the expansion history. This is where the extended term of the likelihood earns its place: the total number of objects, and not only their distribution in redshift and distance, carries information about $\theta$.

The likelihood has the form $\mathcal{L}=e^{-\nu(\theta,\epsilon)}\prod_i \nu(\theta,\epsilon)\,p(z_i,d_i\mid\theta,\epsilon)$. Here $\nu$ is the expected number of objects, and $p$ is the continuous density of each object's redshift $z_i$ and measured distance $d_i$, with the instrument response folded in. This follows the practice of building resolution and efficiency directly into continuous densities in Yan et al.. Mixing in contaminating populations follows the signal-plus-background structure used in Mitchell et al..

Inference layer. We construct confidence regions in the couplings $\theta$ by the Neyman method with toy Monte Carlo. This means building each interval directly from thousands of simulated datasets rather than from a formula that assumes large samples. The seed $\epsilon$ is profiled, and both excluded zones are imposed.

Two ingredients come from existing work:

  • Using the ordering of likelihood ratios to handle a physical boundary follows the problem faced in low-count searches such as the CLEO $\tau\to\mu\gamma$ search. There, 1.8 fitted signal events at 1.0σ were turned into a 90% confidence-level limit.
  • Marginalising nuisance parameters event by event follows BASiL. That work also documents cut-based methods returning unphysical negative signal rates.

Neither paper faced a boundary whose location depends on a nuisance parameter. Here it does: the edge of the linear-validity zone depends on $\epsilon$.

Scope, stated honestly. The research file contains no cosmological distance catalogue and no reported Galileon fit. Every event-level dataset it lists probes Standard Model or conventional astrophysical physics, and none is sensitive to the dark-energy sector. We will therefore not repurpose, for example, PandaX-4T or LHAASO data for a question they cannot answer. The deliverable is the method, fully validated on simulated catalogues. Application to a real catalogue is a clearly separated final step.

Baselines

Theory-only criterion (the "do nothing" reference). The naive reference is the current state of the art: declare a model viable if its tachyonic growth time exceeds the age of the universe, as argued in Galileons on cosmological backgrounds. We run it by evaluating $G(\theta)$ on the same parameter grid and marking points with $G$ below $e$ as viable.

Our method beats it if two things hold on simulations. First, the data-driven region must disagree with this criterion on a measurable fraction of parameter space. Second, the disagreement must track the injected seed: models the criterion calls viable are excluded when $\epsilon$ is large, and some it calls non-viable are allowed when $\epsilon$ is small. If the two regions coincide for all realistic seeds, the proposal's central premise fails.

Binned Poisson fit with a hard stability cut. The strongest competitor is the standard cosmology workflow. It bins the catalogue in redshift and distance, fits a binned Poisson likelihood, and imposes both ghost and tachyon conditions as a hard truncation of the prior. It is strong because it is simple and robust, and in low dimensions binning often loses little.

For fairness, it gets the same simulated catalogues, the same theory code and the same compute. We give it a bin-width optimisation on held-out simulations, so that no information loss is attributed to poor binning. For comparison, in a different domain, unbinned event weighting gave a 15–20% precision gain at 10% signal-to-noise and nearly halved some spectral error bars (D'Amico et al.). No reported result was found on Lacuna for any fit of curved Galileon parameters, binned or unbinned, so there is no published number in our domain to beat.

We count the comparison as won if two things hold. The unbinned fit must shrink the 68% region in $\theta$ by at least 15% at fixed coverage. And the binned-with-hard-cut method must show coverage below nominal near the tachyonic boundary, while ours stays nominal.

Ablation that isolates the key ingredient. The central ablation keeps everything unbinned. It only swaps our treatment of the tachyon, where growth enters the prediction with a free seed, for a hard cut at a fixed growth factor. If the hard-cut version matches ours in both coverage and region shape, then treating the seed as a nuisance adds nothing, and the idea reduces to a better-binned standard fit. A second ablation drops the extended Poisson term, to measure how much constraining power comes from the object count alone.

Why not just…?

Why not just bin the data and use a chi-square fit? For one or two observables with many objects per bin, a binned fit may lose little, and we expect this to be a real threat at the high-statistics end. The argument for unbinned fitting is strongest in two situations. One is when counts are sparse at high redshift, where the growing mode acts. The other is when selection efficiency varies continuously within a bin. The CLEO comparison shows the general pattern: an unbinned fit kept 15.2% signal efficiency where a counting method needed tight cuts that discarded signal (Ahmed et al.). The optimised-binning baseline tests this directly. If binning matches us across catalogue sizes, we will say so and keep only the stability-boundary contribution.

Why not just impose the stability inequalities as a hard prior and move on? This is the most serious alternative, because it is what a careful Bayesian would do. Its weakness is that the tachyonic "inequality" is not a property of $\theta$ alone. A hard cut at growth factor $G_{\max}$ silently fixes the seed, so a single number decides the benign-versus-catastrophic question that the literature disputes. The ablation above settles whether this matters numerically. We commit to reporting a null result if it does not.

Why not just argue viability from screening or field redefinitions, with no data? Theory alone has not closed these questions. The $c_s=0$ mode (a perturbation that does not propagate) found in non-minimally coupled Proca cosmology is left to a hoped-for screening mechanism (De Felice and Hell). Disformal field redefinitions cure Ricci-tensor couplings but have no local form for Riemann couplings (Obata). Stable spherical DBI Galileons are forced to be superluminal (Goon et al.). For slow tachyons specifically, "is the growth observable?" is an empirical question by definition. We deliberately exclude the non-linear regime rather than claim to resolve it.

Proposed Experiments

Coverage study. The first decisive check is coverage. We generate 10⁴ simulated catalogues at several true points $(\theta,\epsilon)$ placed on, near and far from the tachyonic and linear-validity boundaries. We then measure how often each method's 68% and 95% regions contain the truth. The prediction is that Wilks-based intervals and hard-cut intervals undercover near the boundaries, while the Neyman construction does not.

Discrimination test. The second check asks whether data can separate a benign tachyon from a stable model. We inject catalogues from a model with growth time near the age of the universe and fit both hypotheses. We report the catalogue size at which the growing mode is detected at 3σ, as a function of $\epsilon$. This curve is the concrete, quantitative answer to the dispute between Galileons on cosmological backgrounds and the stricter stability literature.

Robustness test. The third check tests robustness to the theory truncation. We repeat the discrimination test as $\mathcal{L}_4$ and then $\mathcal{L}_5$ are switched on, to see whether conclusions drawn from the low-order model survive.

The proposal is falsified if the ablation shows no difference from a hard cut, or if the detection threshold lies far beyond any plausible catalogue size. In either case, the honest conclusion is that slow Galileon tachyons are observationally inert, and the viability debate is purely theoretical.