When Averaging Lies: Disorder-Driven Ghost and Tachyon Instabilities in Random Media

When a wave travels through a weakly random medium, its average behaves as if the medium were absorbing it. The scattered energy is not lost. It moves into incoherent, rapidly fluctuating components that the average no longer sees. Mathematicians have made this precise for stable waves. Field theorists, meanwhile, worry about two kinds of unstable fields. The first is the tachyon, a field with negative mass-squared that grows exponentially. The second is the ghost, a mode with negative kinetic energy. Ghosts appear whenever a theory has more than two time derivatives; this result is Ostrogradsky's theorem. This note asks what happens when such unstable fields propagate through a weakly fluctuating background. That is a realistic situation: a curvature coupling $\xi R$ turns any fluctuation of the spacetime curvature $R$ into a random potential.

Our claim is that the standard tool gives the wrong stability verdict exactly where it matters. That tool is homogenization of the averaged field, which yields an "effective complex potential" whose imaginary part looks like damping. The averaged field can decay while the field's energy grows exponentially. For ghosts this happens generically: scattering between positive- and negative-energy modes conserves only an indefinite quantity, so both populations can grow together. We propose to derive the correct description, which is a transport (kinetic) equation for wave energy with a sign structure inherited from the ghost. We will compute the stability thresholds it predicts and check them against large Monte Carlo ensembles.

Core Claim

Two homogenized descriptions exist for waves in weak random media. The first is the mean field, which obeys a damped deterministic equation, as in The random Schrödinger equation: homogenization in time-dependent potentials and The 1D Schrödinger equation with a spacetime white noise: the average wave function. The second is the energy density, which obeys a linear Boltzmann equation; the same Ryzhik paper and Gaussian fluctuations from random Schrödinger equation show how the Wigner transform links waves to this transport picture. For stable waves the two descriptions agree that nothing blows up. We claim they disagree in two settings.

The first setting is ghost fields, such as a fourth-order scalar of Pais–Uhlenbeck type, which has a positive-energy branch $s=+1$ and a negative-energy branch $s=-1$. Let $W_s(k)$ be the energy density on branch $s$ at wavenumber $k$, and let $\tau = \varepsilon^2 t$ be kinetic time for a random potential of strength $\varepsilon$. We conjecture the limiting equation

$$\partial_\tau W_s(k) = 2\mu_s(k)\,W_s(k) + \sum_{s'=\pm}\int \sigma_{ss'}(k,k')\,\big[W_{s'}(k') - s s'\,W_s(k)\big]\,dk' .$$

Here $\sigma_{ss'}\ge 0$ is the scattering kernel set by the potential's space-time power spectrum, and $\mu_s(k)$ is a rescaled tachyonic growth rate that is nonzero only in the near-critical regime described below. The factor $ss'$ is the new ingredient. It makes $\sum_s s\int W_s$ conserved, while the positive total $\sum_s \int W_s$ increases at a rate proportional to the inter-branch cross-section. One can verify both properties directly from the symmetry of $\sigma$. The averaged field still decays at the usual damping rate.

The second setting is near-critical tachyons, with $m^2_{\rm eff} = -\varepsilon^2\mu$. Here tachyonic growth and scattering act on the same $O(\varepsilon^{-2})$ time scale. Scattering moves energy out of the unstable long-wavelength band into stable short-wavelength modes, but some of it scatters back in. The true stability threshold is where the top eigenvalue of the kinetic operator crosses zero. That threshold sits at a larger $\mu$ than the mean-field threshold (growth rate equals damping rate). This leaves a finite window of "false stability" whose width we will compute.

The novelty, as far as the compiled literature shows, lies in three things. No existing work spans homogenization and these instabilities. No kinetic limit for indefinite-energy fields exists. And no one has stated the mean-field versus energy disagreement as a computable threshold gap.

Research Program

The first part is analytic. We will extend the Duhamel-expansion argument of Ryzhik's time-dependent random Schrödinger paper to a two-branch Klein–Gordon system with an indefinite conserved form. The main technical change is the combinatorics: the pairing estimates currently rely on positivity of the conserved norm, and will have to be replaced by bounds that survive exponential growth on times of order $\varepsilon^{-2}$. We will use time-dependent random potentials. That avoids Anderson localization, which the research file flags as a failure mode of static weak disorder, and it matches fluctuating curvature. Pair creation of positive and ghost quanta then appears in $\sigma_{+-}$ at the potential's power spectrum evaluated at the frequency sum. The derivation will either confirm the sign structure in the equation above or show where it fails.

The second part is spectral. For the tachyonic case we will compute the top eigenvalue of the kinetic operator as a function of $\mu$ and the scattering rate. From that we extract the false-stability window and its scaling. We will also test whether the real part of the effective potential shifts $m^2_{\rm eff}$ enough to move the threshold.

The third part connects the results to physics. In Symmetries for Galileons and DBI scalars on curved space and New Class of Effective Field Theories from Embedded Branes, Goon, Hinterbichler and Trodden build ghost-free scalar theories from brane embeddings. On de Sitter space these theories have $m^2\propto R$ tachyons, which the authors tolerate only if the tachyon's growth time exceeds the age of the universe. We will ask how a stochastic curvature background changes that criterion. In On the ghost problem of conformal gravity, Hell, Lüst and Zoupanos remove ghosts with boundary conditions that select a ghost-free subspace. Our framework predicts that weak disorder repopulates that subspace through $\sigma_{+-}$, giving a finite lifetime to the classical boundary-condition cure. This is a concrete, testable limitation of that method.

Baselines

The naive reference is the deterministic linear stability analysis of the unperturbed background. In this analysis only the sign of $m^2_{\rm eff}$ and the presence of a ghost branch matter, which is the criterion used by Goon et al. We will run it on exactly the same parameter grid as every other method. It is the default that physicists apply, and beating it means correctly predicting the stability changes that disorder causes and it misses.

The strongest existing method is the homogenized mean-field equation with an effective complex potential. We will compute that potential with the Ryzhik construction, or with the renormalized white-noise version of Gu and Komorowski when the noise is white in time. This is the most developed rigorous tool for averaged waves in random media, and it is the one a careful reader would apply. It receives the same random-potential statistics and no free parameters. We beat it if, on at least the near-critical tachyon grid and all ghost runs, our kinetic thresholds match Monte Carlo energy growth while the mean-field predictions do not.

No reported result exists for this benchmark: the compiled file reports no reported result found on Lacuna for random-media tachyon or ghost stability. The gravity papers explicitly contain no numerical data. The only nearby numbers are dissipative-ODE results. For example, high-order HMM₂ reached error $2.3\times10^{-9}$ in 0.09 s, versus 5.44 s for a coupled solver, on a linear stiff system (High Order Numerical Homogenization for Dissipative Ordinary Differential Equations). These numbers show what reduced models achieve when fast modes decay, which is precisely the condition our systems violate.

The ground truth is direct Monte Carlo simulation of the full stochastic field equation, which we describe under the experiments below.

The key ablation is a "sign-blind" kinetic equation, in which $ss'$ is replaced by $+1$ so the ghost branch is treated as ordinary positive energy. This ablation isolates whether the indefinite sign structure, and not merely the presence of a second branch or of transport physics, is what predicts the growth.

Why not just…?

Why not just trust the effective complex potential, since its imaginary part already damps the field? That damping measures loss of coherence, not loss of energy. The research file states this directly for the stable case, and The 1D Schrödinger equation with a spacetime white noise: the average wave function shows that the coherent field carries an exponentially shrinking fraction of the energy. With a ghost branch, the discarded incoherent part is exactly where the growth occurs. The mean-field baseline compared against Monte Carlo energy curves will show this, or refute us.

Why not just run Monte Carlo and skip the theory? This is a real threat for one-dimensional and two-dimensional toy models, because simulation is the ground truth there. Simulation cannot reach three dimensions over cosmological numbers of oscillation periods. It also cannot reach $\varepsilon\to 0$, where the relevant times scale as $\varepsilon^{-2}$. And it does not yield a threshold formula. If the kinetic equation fails to predict the simulated thresholds within ensemble error, the theory has no value and we will say so.

Why not just treat each mode as a randomly driven parametric oscillator, whose second moments are already known to grow? For a single mode, that is the right answer, and it is a serious competitor. Our claim is about the spatially extended problem. There, scattering between wavenumbers either rescues the tachyonic band (by draining energy) or dooms the ghost (by pair conversion), and a single-mode treatment captures neither. We will include a mode-by-mode stochastic oscillator prediction as an extra baseline. If it matches Monte Carlo as well as the kinetic equation does, we will narrow the claim to the ghost case.

Why not just remove the instability by a field redefinition, boundary conditions, or reduction to a slow manifold? Each of these fails in the setting we study.

Proposed Experiments

The ground-truth solver is a split-step spectral integrator for one- and two-dimensional systems. It covers a tachyonic Klein–Gordon field and a fourth-order Pais–Uhlenbeck scalar, each with a time-dependent Gaussian random potential $\varepsilon V(t,x)$ standing in for $\xi\,\delta R$. We will run ensembles of 10³–10⁴ realizations across $\varepsilon\in[0.02,0.2]$, the near-critical parameter $\mu$, and the potential's correlation length and time. Solver accuracy will be checked on stable cases against the known Gaussian-fluctuation and Rayleigh-intensity limits from Gaussian fluctuations from random Schrödinger equation. We will report the following metrics:

  • the growth rate of the mean field;
  • the growth rate of the mean energy;
  • the per-branch Wigner spectra;
  • an energy-spectrum error that weights wavenumbers, in the spirit of the frequency-space error metric of PDEBench.

Three checks decide the outcome. First, in the tachyonic case, there must be a measurable window of $\mu$ in which the simulated mean decays while the energy grows. The edges of that window must match the kinetic eigenvalue crossing to within ensemble error and scale with $\varepsilon$ as predicted. Second, in the ghost case, the energy growth rate must scale with the inter-branch cross-section $\sigma_{+-}$, which we vary through the potential's power spectrum, while the sign-blind ablation predicts bounded energy. Third, simulations initialized in the ghost-free subspace selected by boundary conditions must show ghost-branch population appearing at the rate $\sigma_{+-}$ predicts. If the first two checks fail, the kinetic sign structure is wrong. If only the tachyonic window is absent, the work reduces to a statement about ghosts, which is still useful.