Fidelity Catastrophe: Mechanisms & Implications

• Fidelity catastrophe is a regime where system stability breaks down due to small perturbations, manifesting as vanishing overlaps or proxy misalignments across disciplines.
• It involves mechanisms like multifractal intensity correlations in quantum many-body physics, nonperturbative impurity effects in quasiperiodic systems, and objective misspecification in AI safety.
• Understanding fidelity catastrophe offers actionable insights into mitigating extreme sensitivity in systems ranging from condensed matter to stochastic evolutionary dynamics.

Searching arXiv for papers using the term and adjacent usages. “Fidelity catastrophe” is a cross-disciplinary term used for qualitatively different phenomena that share a common structural motif: a system becomes catastrophically sensitive when a notion of fidelity collapses, broadens into an extreme distribution, or is shown to be insufficiently aligned with the quantity of real interest. In condensed-matter and many-body quantum physics, the term is closely tied to orthogonality catastrophe, where a local perturbation drives the overlap between many-body states to vanish in the thermodynamic limit, with especially sharp behavior at criticality and in open-system variants (Kettemann, 2016, Tonielli et al., 2018, Chen et al., 2021). In quasiperiodic systems, the same language is used for anomalous fidelity decay in the Aubry-André family, including a “statistical exponential orthogonality catastrophe” in the localized phase (Vahedi et al., 2022). In AI safety, “fidelity catastrophe” denotes catastrophe caused by insufficient fidelity of a proxy objective to the true reward under strong optimization pressure (Marklund et al., 16 Mar 2026). In stochastic evolutionary dynamics, it denotes a low-mutation instability in the stochastic Eigen model, distinct from the classic high-mutation error catastrophe (Crosato et al., 18 Aug 2025). Across these usages, fidelity catastrophe names a regime in which a small perturbation, finite mismatch, or state-dependent fluctuation produces a qualitatively disproportionate response.

1. Term, scope, and recurrent structure

The term does not have a single universal definition across fields. In quantum many-body settings, fidelity is an overlap quantity, such as $F=|\langle \Psi_0|\Psi_{0,\mathrm{imp}\rangle|$ for impurity problems in free-fermion systems (Vahedi et al., 2022), the overlap $F$ between the ground state of a Fermi liquid and the one of the same system with an added potential impurity (Kettemann, 2016), or the adiabatic fidelity $$\mathcal{F}(\lambda)=|\langle \Phi_\lambda \,|\, \Psi_\lambda\rangle|^2$$ in driven systems (Chen et al., 2021). In these settings, catastrophe refers to asymptotically vanishing overlap, unusually strong scaling, or an extreme disparity between typical and mean behavior.

In AI safety, fidelity refers to the relation between a proxy reward/objective $\hat r$ and the true reward $r^*$. The failure mode is defined as catastrophe caused by insufficient fidelity of the proxy reward/objective to the true reward, when combined with strong optimization pressure and advanced capability (Marklund et al., 16 Mar 2026). In stochastic evolutionary theory, fidelity refers to replication fidelity $\hat\mu=\mu_{ii}$, and the catastrophe occurs when rates of mutation fall beneath a threshold, yielding noise-induced multistability and stochastic switching among short-lived regimes of effectively clonal behavior (Crosato et al., 18 Aug 2025).

A plausible implication is that the phrase is best understood operationally rather than semantically: it marks regimes where a fidelity-like quantity ceases to behave perturbatively. The concrete mechanism differs by domain—multifractality, gapless infrared correlations, proxy misspecification, or multiplicative demographic noise—but the mathematical and conceptual role is similar.

2. Orthogonality catastrophe and critical disorder

A central use of the concept arises in the study of the Anderson orthogonality catastrophe at the Anderson metal-insulator transition (AMIT). For a noninteracting Fermi liquid ground state $|\psi\rangle$ perturbed by a short-ranged static impurity of strength $\lambda$ at position ${\bf x}$, the overlap is

$\chi = F^2 = |\langle \psi|\psi'\rangle|^2,$

and Anderson’s bound gives

$F$0

with the Anderson integral

$F$1

Thus the fidelity is controlled by local wavefunction intensities and their correlations across the Fermi energy (Kettemann, 2016).

Near the AMIT, the key input is multifractality. The intensity correlation function

$F$2

behaves at criticality as

$F$3

with $F$4 and $F$5 (Kettemann, 2016). At the transition itself,

$F$6

so $F$7, and therefore the typical fidelity decays exponentially with system size,

$F$8

This is the paper’s main result: at criticality the orthogonality catastrophe is not merely algebraic but exponentially enhanced by multifractal correlations (Kettemann, 2016).

Away from criticality, the behavior is phase-dependent. On the metallic side, the paper recovers Anderson’s power-law result,

$F$9

with a power that increases as the Fermi energy approaches the mobility edge,

$$\mathcal{F}(\lambda)=|\langle \Phi_\lambda \,|\, \Psi_\lambda\rangle|^2$$0

On the insulating side,

$$\mathcal{F}(\lambda)=|\langle \Phi_\lambda \,|\, \Psi_\lambda\rangle|^2$$1

independent of total system size for $$\mathcal{F}(\lambda)=|\langle \Phi_\lambda \,|\, \Psi_\lambda\rangle|^2$$2, so the fidelity saturates to a constant (Kettemann, 2016).

A major conceptual point is the separation between typical and mean fidelity. The typical quantity

$$\mathcal{F}(\lambda)=|\langle \Phi_\lambda \,|\, \Psi_\lambda\rangle|^2$$3

follows the phase-dependent scaling above, but $$\mathcal{F}(\lambda)=|\langle \Phi_\lambda \,|\, \Psi_\lambda\rangle|^2$$4 is log-normally distributed with a width diverging at the AMIT, and the mean fidelity

$$\mathcal{F}(\lambda)=|\langle \Phi_\lambda \,|\, \Psi_\lambda\rangle|^2$$5

tends to one at the AMIT even while $$\mathcal{F}(\lambda)=|\langle \Phi_\lambda \,|\, \Psi_\lambda\rangle|^2$$6 (Kettemann, 2016). The physical explanation given is the presence of anomalously low-intensity regions, or local pseudogaps, where the impurity has almost no effect. This makes the AMIT version of fidelity catastrophe fundamentally distributional rather than reducible to a single exponent.

3. Quasiperiodic chains and the Aubry-André family

The Aubry-André (AA) and extended Aubry-André (EAA) models provide a second major setting in which fidelity catastrophe language is used. The overlap between unperturbed and impurity-perturbed many-body ground states is again the central object, and because the model is noninteracting the fidelity can be written exactly as a determinant of overlaps of the occupied one-particle states and is bounded by the Anderson integral

$$\mathcal{F}(\lambda)=|\langle \Phi_\lambda \,|\, \Psi_\lambda\rangle|^2$$7

(Vahedi et al., 2022).

The paper explicitly compares a single-site impurity ($$\mathcal{F}(\lambda)=|\langle \Phi_\lambda \,|\, \Psi_\lambda\rangle|^2$$8) with an extended impurity ($$\mathcal{F}(\lambda)=|\langle \Phi_\lambda \,|\, \Psi_\lambda\rangle|^2$$9). For a local impurity of strength $\hat r$0,

$\hat r$1

The AA model has three regimes as $\hat r$2 is tuned: extended/metallic phase for $\hat r$3, critical point / critical phase at $\hat r$4, and localized/insulating phase for $\hat r$5. At the critical point, all eigenstates are multifractal and the average level spacing scales as $\hat r$6, with $\hat r$7, while $\hat r$8 and hence $\hat r$9 in $r^*$0 (Vahedi et al., 2022).

The generic critical prediction is

$r^*$1

which at the AA critical point becomes a very strong decay because $r^*$2 and $r^*$3. However, the numerics for a weak single-site impurity do not show this behavior. The paper finds power-law decay in the metallic phase,

$r^*$4

and also power-law decay in the critical phase, with the critical decay faster than in the metallic phase but still not exponential (Vahedi et al., 2022).

The explanation given is that the simple two-point intensity-correlation approximation is incomplete for a local impurity in the critical AA model. The impurity strongly modifies the local wavefunction intensity at the impurity site, and for multifractal critical states this introduces nonperturbative corrections and multipoint correlations. The exact Green’s-function expression for the perturbed intensity shows that naive replacement of perturbed states by unperturbed ones misses these corrections, and the additional correlations weaken the infrared divergence of $r^*$5, reducing it from the expected critical exponential scaling to a much weaker, effectively logarithmic growth (Vahedi et al., 2022).

The localized phase exhibits a different mechanism, termed a statistical exponential orthogonality catastrophe. For strong localization, a single impurity mainly shifts the energy of the localized state at its site. At fixed particle number, this can change which localized orbitals are occupied, producing a Bernoulli-like fidelity distribution with $r^*$6 if the occupation pattern does not change and $r^*$7 if a level is pushed across the Fermi energy. As system size grows, the probability of mismatch accumulates in a way that gives an exponential suppression of the typical fidelity (Vahedi et al., 2022). The mechanism is therefore statistical rather than the usual metallic continuum-coupling mechanism.

For an extended impurity, the picture changes again. In the critical AA phase, the fidelity becomes more strongly suppressed as $r^*$8 grows, and for sufficiently large $r^*$9 the data are consistent with an exponential decay in $\hat\mu=\mu_{ii}$0. In the EAA model with a mobility edge, extended impurities and especially stronger $\hat\mu=\mu_{ii}$1 show clear signs of exponential orthogonality catastrophe at the mobility edge. The same paper also considers a parametric perturbation $\hat\mu=\mu_{ii}$2, for which both average and typical fidelities show exponential decay at the critical point, in agreement with the estimate $\hat\mu=\mu_{ii}$3 (Vahedi et al., 2022).

4. Driven and dissipative many-body systems

In driven many-body systems, fidelity catastrophe appears in adiabatic form. The relevant quantity is the adiabatic fidelity

$\hat\mu=\mu_{ii}$4

where $\hat\mu=\mu_{ii}$5 is the actual driven state and $\hat\mu=\mu_{ii}$6 is the instantaneous ground state (Chen et al., 2021). The paper “Bounds on quantum adiabaticity in driven many-body systems from generalized orthogonality catastrophe and quantum speed limit” introduces the generalized orthogonality catastrophe

$\hat\mu=\mu_{ii}$7

and notes that in a broad class of systems

$\hat\mu=\mu_{ii}$8

with $\hat\mu=\mu_{ii}$9 growing with system size in many models, so $|\psi\rangle$0 can become exponentially small in $|\psi\rangle$1 (Chen et al., 2021).

The paper combines generalized orthogonality catastrophe with a quantum speed limit,

$|\psi\rangle$2

to derive improved inequalities for estimating adiabatic fidelity: $|\psi\rangle$3 and

$|\psi\rangle$4

The second bound is nearly sharp when the system size is large, as illustrated using the driven Rice-Mele model, where

$|\psi\rangle$5

(Chen et al., 2021). In this usage, fidelity catastrophe is not a separate named phase transition but a large-system adiabatic fragility induced jointly by orthogonality catastrophe and finite-speed evolution.

Open-system dynamics provides another variant. In “Orthogonality catastrophe in dissipative quantum many body systems,” the fidelity is

$|\psi\rangle$6

and the main result is the universal long-time scaling

$|\psi\rangle$7

when the system supports long-range correlations (Tonielli et al., 2018). The exponential term $|\psi\rangle$8 signals environmental decoherence, while the algebraic factor $|\psi\rangle$9 is the orthogonality-catastrophe-like contribution generated by infrared singularities in a second-order cumulant expansion suited for Liouvillian dynamics.

For the critical transverse-field Ising chain with local dephasing $\lambda$0, the paper finds

$\lambda$1

and interprets the positive algebraic factor as a critical slowing down of decoherence (Tonielli et al., 2018). The same structure is substantiated for XY and XX chains and for the two-dimensional Bose gas deep in the superfluid phase with local particle heating. This suggests that fidelity catastrophe in open many-body settings is less about vanishing static overlap than about universal, correlation-controlled suppression of return probability under local dissipation.

5. Fidelity catastrophe in AI alignment and consequentialist optimization

In AI safety, “fidelity catastrophe” has a distinct meaning. The paper “Consequentialist Objectives and Catastrophe” defines it as catastrophe caused by insufficient fidelity of the proxy reward/objective to the true reward, when combined with strong optimization pressure and advanced capability (Marklund et al., 16 Mar 2026). The underlying claim is that a fixed consequentialist objective becomes dangerous not because the system is clumsy, but because a highly capable optimizer can exploit even tiny misspecifications in the objective and drive the world into outcomes much worse than what would happen under simple, uninformed behavior.

The formal setup distinguishes the true reward $\lambda$2 from the proxy $\lambda$3, with the executed policy

$\lambda$4

Catastrophe is defined relative to the contemporary value

$\lambda$5

which is the best performance achievable by an uninformed policy, and the primordial value

$\lambda$6

A safety threshold $\lambda$7 is chosen, and performance below $\lambda$8 is called catastrophic (Marklund et al., 16 Mar 2026).

The paper’s central lower bound states that if $\lambda$9, ${\bf x}$0 is iid, and ${\bf x}$1, then

${\bf x}$2

Here attainability is

${\bf x}$3

The theorem’s interpretation is that avoiding catastrophe requires the proxy reward ${\bf x}$4 to encode many bits about the true reward ${\bf x}$5, and that the required information can be enormous when the safety threshold is high relative to primordial value and outcomes are highly attainable (Marklund et al., 16 Mar 2026).

A major point of emphasis is that simple or random behavior is safe, while catastrophic risk arises due to extraordinary competence rather than incompetence. The mitigation proposed is capability constraint. The paper defines a regularized policy distribution

${\bf x}$6

and argues that for any bit budget ${\bf x}$7, if there is variation among optimal uninformed policies, then there exists a proxy ${\bf x}$8 with ${\bf x}$9 and some $\chi = F^2 = |\langle \psi|\psi'\rangle|^2,$0 such that $\chi = F^2 = |\langle \psi|\psi'\rangle|^2,$1 (Marklund et al., 16 Mar 2026). In this literature, fidelity catastrophe is therefore a theorem-backed alignment failure mode rather than an overlap phenomenon.

6. Stochastic evolutionary dynamics and the low-mutation catastrophe

The stochastic Eigen model introduces yet another meaning. The classic Eigen model is synonymous with error catastrophe: when mutation rates are sufficiently high, the genetic variant with the largest replication rate does not occupy the largest fraction of the total population. The 2025 paper adds a distinct phenomenon, the fidelity catastrophe, which occurs at sufficiently low mutation / high replication fidelity once finite-population noise is treated properly (Crosato et al., 18 Aug 2025).

The stochastic model has $\chi = F^2 = |\langle \psi|\psi'\rangle|^2,$2 variants with population counts $\chi = F^2 = |\langle \psi|\psi'\rangle|^2,$3, replication rates $\chi = F^2 = |\langle \psi|\psi'\rangle|^2,$4, death rates $\chi = F^2 = |\langle \psi|\psi'\rangle|^2,$5, and mutation probabilities $\chi = F^2 = |\langle \psi|\psi'\rangle|^2,$6. In the deterministic $\chi = F^2 = |\langle \psi|\psi'\rangle|^2,$7 limit, the population fractions evolve toward the dominant eigenvector of

$\chi = F^2 = |\langle \psi|\psi'\rangle|^2,$8

The fidelity catastrophe appears only in the stochastic version, through state-dependent fluctuations. The system-size expansion yields a reduced Fokker-Planck equation

$\chi = F^2 = |\langle \psi|\psi'\rangle|^2,$9

with noise-induced term

$F$00

The total effective force is written as

$F$01

(Crosato et al., 18 Aug 2025).

The mechanism is noise-induced multistability. In mixed states near the center of the simplex, fluctuations are large; near a corner, if one variant is nearly fixed, the stochastic effects are much weaker. This gradient in noise strength creates an effective drift toward corners or toward the center depending on the replication fidelity $F$02. The resulting stationary distribution can become multimodal, and the population stochastically switches between temporary states where one variant is effectively clonal (Crosato et al., 18 Aug 2025).

The critical fidelity is obtained from the point where the unstable fixed point enters the simplex: $F$03 or equivalently

$F$04

For $F$05, the stationary distribution becomes multimodal and clonal switching occurs; for $F$06, the distribution remains unimodal and centered near the deterministic quasispecies state (Crosato et al., 18 Aug 2025).

The paper emphasizes the regime $F$07, with $F$08 the length of the genome, so the number of possible variants can be far larger than the population size. Since $F$09, large genotype spaces make the phenomenon more likely. The asymptotic relation

$F$10

implies that when $F$11, the fidelity and error thresholds collide, leaving only a vanishingly small interval of mutation rates for which the model is neither in the fidelity- nor error-catastrophe regimes (Crosato et al., 18 Aug 2025). In this field, fidelity catastrophe is therefore a low-mutation instability complementary to the classic high-mutation error catastrophe.

7. Common themes, distinctions, and recurrent misconceptions

The strongest commonality across these literatures is not a shared microscopic mechanism but a shared asymptotic logic. In the AMIT problem, multifractal intensity correlations make the typical fidelity collapse as

$F$12

at criticality (Kettemann, 2016). In the AA family, critical multifractality alone does not guarantee that a weak single-site impurity will realize the expected exponential catastrophe, because nonperturbative impurity-induced corrections and multipoint correlations can soften the divergence (Vahedi et al., 2022). In driven systems, generalized orthogonality catastrophe and quantum speed limit jointly bound adiabatic tracking, making large systems fragile even under slow protocols (Chen et al., 2021). In open systems, local dissipation produces a return-probability decay

$F$13

rather than a simple exponential, showing that critical correlations can reshape decoherence (Tonielli et al., 2018). In AI alignment, capability amplifies proxy misspecification into catastrophe, so fidelity catastrophe is fundamentally about optimization against an insufficiently informative objective (Marklund et al., 16 Mar 2026). In stochastic evolutionary theory, excessive replication fidelity destabilizes quasispecies behavior because multiplicative finite-$F$14 noise induces clonal multistability (Crosato et al., 18 Aug 2025).

Several misconceptions are directly contradicted by the cited work. One is that “catastrophe” always means the mean overlap vanishes. At the AMIT, the typical fidelity converges to zero exponentially fast with system size while the mean fidelity converges to one (Kettemann, 2016). Another is that critical multifractality automatically yields exponential orthogonality catastrophe for any local impurity; the AA results show that a weak single-site impurity can display only power-law decay in the critical phase (Vahedi et al., 2022). A further misconception is that catastrophe necessarily arises from noise, irrationality, or incompetence. In the AI setting, the central claim is the opposite: catastrophic risk arises due to extraordinary competence rather than incompetence (Marklund et al., 16 Mar 2026).

This suggests a useful unifying interpretation. Fidelity catastrophe occurs when the object that mediates stability—state overlap, adiabatic tracking, return probability, proxy-objective fidelity, or replication fidelity—enters a regime in which rare events, strong optimization, long-range correlations, or multiplicative fluctuations dominate the response. The resulting behavior is not merely large in magnitude; it is qualitatively discontinuous with naive perturbative expectations.