Collapse Propagation Model

• Collapse Propagation Model is a family of dynamical frameworks that describe how a localized collapse event triggers a chain reaction across interconnected system components.
• The models employ explicit state variables, propagation operators, and threshold mechanisms to capture dynamics in settings such as structural engineering, quantum systems, deep learning, and financial networks.
• Practical insights reveal that parameters like compaction ratios, data freshness, and residual scaling govern whether collapse is amplified, arrested, or reversed in these systems.

Searching arXiv for relevant papers on “collapse propagation model” and related uses of the term across domains. “Collapse Propagation Model” denotes a class of dynamical descriptions in which collapse is treated as a process that advances through a structured system rather than as an isolated event. In the literature, the term and closely related formulations appear in progressive structural collapse, airway reopening in collapsed elastic networks, objective wavefunction collapse, hierarchical polymer collapse, neural and representational collapse in deep networks, recursive model collapse in generative learning, world-model collapse in long-horizon agents, and cascading financial failures [(Beck, 2008); (Li et al., 2022); (Okon et al., 2018); (Bunin et al., 2015); (Súkeník et al., 2023); (Khelifa et al., 18 Feb 2026); (Song et al., 30 Jun 2026); (Stella et al., 2023)]. Across these settings, the state variable may be an avalanche front, a branching finger, a density matrix, a layerwise feature ensemble, a divergence sequence, or a network equity vector, but the central question is similar: what local mechanism initiates collapse, how does it couple to neighboring degrees of freedom, and under what conditions is propagation amplified, arrested, or redirected.

1. Scope and domain-specific meanings

The expression does not name a single universal formalism. It instead designates a family of models in which collapse has an explicit propagation law. In engineering, the propagating object can be a collapse front or an air finger; in quantum foundations, a collapse operator or chaining process; in machine learning, a layerwise degeneration of representations or a recursive drift of distributions; and in network science, a threshold-driven contagion of insolvency [(Beck, 2008); (Li et al., 2022); (Okon et al., 2018); (Saada et al., 2024); (Khelifa et al., 18 Feb 2026); (Stella et al., 2023)].

Domain	State variable	Propagation law or mechanism
Progressive structural collapse	$x_2(t)$, the avalanche front	Front motion opposed by $R(x_2)$ and modified by compaction (Beck, 2008)
Objective quantum collapse	$|\psi(t)\rangle$, $\rho(t)$, or chained DoFs	CSL-type stochastic suppression or per-chaining event probability $1/\Sigma$ (Okon et al., 2018, Li, 24 Feb 2026)
Deep representation collapse	$H_l$, $\boldsymbol{\Sigma}_\ell$, $\rho^\ell$	Layerwise DNC/PFC or spectral-gap-driven rank collapse (Súkeník et al., 2023, Saada et al., 2024, Noci et al., 2022, Wang et al., 2024)
Recursive generative collapse	$D_i=\chi^2(\hat p^i\|data)$	Discounted accumulation of score errors across generations (Khelifa et al., 18 Feb 2026)
World-model collapse in agents	$R_t$, $R(x_2)$0, $R(x_2)$1	Transition from accurate world state to corrupted planning (Song et al., 30 Jun 2026)
Financial cascades	$R(x_2)$2	Threshold-triggered losses transmitted by cross-holdings (Stella et al., 2023)

This breadth matters because identical vocabulary masks different ontologies. In some literatures collapse is a physical event; in others it is a failure mode, a spectral degeneracy, or a recursive bias amplification. What unifies them is not substance but architecture: an initially localized disturbance becomes system-wide through explicit coupling rules.

2. Recurrent mathematical architecture

Several representative propagation laws make the shared structure visible. In progressive building collapse, the avalanche front obeys

$R(x_2)$3

with compaction ratio $R(x_2)$4, resistive force $R(x_2)$5, and front coordinate $R(x_2)$6 (Beck, 2008). In financial contagion, the update law is

$R(x_2)$7

where $R(x_2)$8 is the cross-holdings matrix, $R(x_2)$9 is the primitive-asset contribution, and $|\psi(t)\rangle$0 encodes threshold-triggered failure costs (Stella et al., 2023). In recursive diffusion-model training, the collapse-propagation approximation is

$|\psi(t)\rangle$1

with $|\psi(t)\rangle$2 the fresh-data fraction and $|\psi(t)\rangle$3 the score-error energy (Khelifa et al., 18 Feb 2026). In transformer spectral analysis, the width-collapse law is

$|\psi(t)\rangle$4

for any fixed layer $|\psi(t)\rangle$5, where $|\psi(t)\rangle$6 is stable rank and $|\psi(t)\rangle$7 is context length (Saada et al., 2024).

Taken together, these equations suggest a recurring decomposition into four ingredients. First, there is a state variable whose evolution is tracked explicitly. Second, there is a propagation operator, such as a mechanical resistance law, a network matrix, a stochastic collapse operator, or a layerwise linearization. Third, there is a nonlinearity or threshold mechanism, such as compaction, failure costs, spectral outliers, or branch switching. Fourth, there is a criterion for irreversible collapse: front advance, decoherence, rank-one degeneration, positive divergence floor, or node failure. This suggests that “collapse propagation model” is best understood as a structural template rather than a domain-specific theorem.

3. Physical, engineering, and material realizations

In progressive structural collapse, Beck’s one-dimensional continuum model divides the building into a rigid top block, a compacted section, and an intact lower section. Mass conservation yields the kinematic relation $|\psi(t)\rangle$8, so the avalanche front moves faster than the top block. A central correction to Bažant and Verdure is that proper treatment of compaction amplifies the resistive term to $|\psi(t)\rangle$9, implying that compaction increases effective resistance and slows the avalanche rather than accelerating it (Beck, 2008). The same paper further argues that an avalanche cannot supply the energy required for a $\rho(t)$0 heat wave, because the associated dimensionless resistance $\rho(t)$1 is two orders of magnitude larger than the apparent structural resistance $\rho(t)$2; this leads to the conclusion that any major weakening wave must precede, not be driven by, the avalanche (Beck, 2008).

In airway reopening, the collapse-propagation problem is inverted: an air finger advances through a liquid-filled elastic Y-bifurcation and reopens a collapsed network. The main experimental control parameter is the initial area ratio $\rho(t)$3, with low, moderate, and high collapse reported as $\rho(t)$4 and $\rho(t)$5 in the main channel. Steady propagation in the parent channel is typically lost near the bifurcation and later recovered in the daughter channels, but at high collapse experimentally indistinguishable parent-channel fingers can lead to multiple reopening states in the daughters. The relaxation time $\rho(t)$6 and relaxation distance $\rho(t)$7 are largest when viscous and surface-tension forces dominate and decrease toward a plateau when elastic and surface-tension forces balance, implying that practical branching networks may be too short for steady propagation to remain established (Li et al., 2022).

In polymer collapse, the coalescence model replaces a rapidly collapsing chain by infalling spherical blobs that coalesce on contact. Distances between monomers are assigned at their first coalescence and are not subsequently equilibrated. The model reproduces the dependence of distance on contour separation $\rho(t)$8 observed in molecular dynamics and explains the slow approach to the expected $\rho(t)$9 scaling by a wide distribution of blob sizes (Bunin et al., 2015). Here propagation is hierarchical: local blobs merge into larger blobs, and the history of coalescence fixes an ultrametric geometry.

A related threshold-based realization appears in cosmological structure formation. In the excursion-set approach, collapse propagates from initial Lagrangian overdensities through a barrier condition $1/\Sigma$0. The standard ellipsoidal-collapse barrier depends on tidal ellipticity $1/\Sigma$1 and prolateness $1/\Sigma$2, but simulation results show that protohalo triaxiality must also be included. The modified ECE model, which treats collapsing perturbations as initially ellipsoidal rather than spherical, provides a more accurate description of the measured minimum overdensities of recently collapsed objects (Ludlow et al., 2011). This suggests that in threshold-propagation problems, geometric state variables can be as important as scalar load variables.

4. Objective-collapse and relativistic quantum models

In quantum foundations, collapse propagation models specify how superpositions are dynamically suppressed. Okon and Sebastián propose a modification of CSL in which the collapse operator is not mass density but maximally integrated information, $1/\Sigma$3. The stochastic solution retains the CSL form,

$1/\Sigma$4

so superpositions of different $1/\Sigma$5-eigenstates are exponentially suppressed as $1/\Sigma$6 (Okon et al., 2018). The model treats consciousness as a physical property represented by a Hermitian operator and explicitly rejects the slogan that consciousness, as something extra-physical, “causes” collapse. A major technical limitation is that no practical construction of $1/\Sigma$7 is currently known (Okon et al., 2018).

A different proposal ties collapse to a new quantum correlation called chaining. In the state-chaining model, each appearance of one or several new chainings along a degree of freedom has collapse probability $1/\Sigma$8, and a chain of $1/\Sigma$9 such steps yields cumulative event probability

$H_l$0

Macroscopic measuring devices and Schrödinger-cat scenarios become classical because they realize long chaining cascades, whereas a mirror can remain interference-compatible because it entangles without chaining (Li, 24 Feb 2026). The paper reports lower bounds such as $H_l$1 from interference-related considerations (Li, 24 Feb 2026).

Relativistic formulations sharpen the notion of propagation in spacetime. In the scalar-field CSL model, the density matrix obeys

$H_l$2

and a superposition of two coherent clumps is driven toward field eigenstates whose eigenvalues are close to the clump profiles $H_l$3 (Bedingham et al., 2019). The model exhibits sensible collapse behavior but also unavoidable particle production, so it is not phenomenologically viable (Bedingham et al., 2019). Pearle’s later relativistic dynamical collapse model instead uses the spacetime particle-number density $H_l$4, smeared to $H_l$5, and obtains

$H_l$6

with the crucial property that the vacuum is not excited (Pearle, 2014). For a clump in two locations, the collapse rate matches nonrelativistic CSL when $H_l$7 and $H_l$8, but that choice causes unacceptable mass change over cosmological time; taking $H_l$9 to be the size of the universe and $\boldsymbol{\Sigma}_\ell$0 preserves satisfactory collapse behavior while keeping mass change acceptably small (Pearle, 2014).

These quantum models differ on ontology, but they share a precise propagation question: what operator, correlation structure, or spacetime density carries the instability from a microscopic superposition into a macroscopically definite branch?

5. Representation collapse in deep networks

In deep learning, collapse propagation refers to the way simple class geometry or representational degeneration moves from one layer to another. The deep unconstrained features model (DUFM) formalizes deep neural collapse in a nonlinear multi-layer setting. For binary classification, balanced data, ReLU activations, and MSE with layerwise Frobenius regularization, the $\boldsymbol{\Sigma}_\ell$1-DUFM optimum exhibits DNC1–DNC3 through the depth of the head whenever

$\boldsymbol{\Sigma}_\ell$2

where DNC1 is within-class collapse, DNC2 is orthogonal class means, and DNC3 is weight–mean alignment (Súkeník et al., 2023). The analysis is static rather than dynamical, but it supplies an exact global-optimum explanation for why collapse can propagate backward from the last layer to earlier layers.

The transformer literature identifies a different propagation mechanism. A spectral analysis of random softmax attention shows an outlier singular value at $\boldsymbol{\Sigma}_\ell$3 and bulk singular values of order $\boldsymbol{\Sigma}_\ell$4, creating a spectral gap. This yields a new pathology, rank collapse in width, with

$\boldsymbol{\Sigma}_\ell$5

for fixed $\boldsymbol{\Sigma}_\ell$6, meaning that as context length grows, the token covariance becomes effectively rank-one even without taking depth to infinity (Saada et al., 2024). The same spectral gap also exacerbates vanishing and exploding gradients, and removing the outlier eigenvalue(s) is proposed as a practical mitigation (Saada et al., 2024). A related signal-propagation analysis proves that rank collapse hinders training by causing the gradients of the queries and keys to vanish at initialization, while value gradients remain nonzero; the proposed fix is depth-dependent residual scaling, $\boldsymbol{\Sigma}_\ell$7 and $\boldsymbol{\Sigma}_\ell$8, which keeps correlations bounded away from complete collapse (Noci et al., 2022).

For ResNets, the proposed generalization is Progressive Feedforward Collapse. The degree of collapse at layer $\boldsymbol{\Sigma}_\ell$9 is quantified by

$\rho^\ell$0

by the simplex-ETF distance $\rho^\ell$1, and by the NCC accuracy $\rho^\ell$2; the conjecture is that collapse increases during forward propagation, and the abstract states that the metrics of PFC indeed monotonically decrease across depth on various datasets (Wang et al., 2024). This is a shift from endpoint NC to a layerwise collapse trajectory.

A common misconception in this literature is that collapse is only a last-layer or depth-asymptotic artifact. The cited analyses argue otherwise: collapse can be a global optimum of a deep surrogate model, a width effect of softmax attention, or a progressively measurable forward-pass phenomenon (Súkeník et al., 2023, Saada et al., 2024, Wang et al., 2024).

6. Recursive, agentic, and systemic collapse

When models are retrained on their own outputs, collapse propagation becomes explicitly intergenerational. In score-based diffusion models, each generation is trained on the mixture

$\rho^\ell$3

and the relevant quantities are $\rho^\ell$4 and $\rho^\ell$5. The theory shows that $\rho^\ell$6 in the perturbative regime and that the accumulated divergence behaves like a discounted sum of squared score errors with discount factor $\rho^\ell$7 (Khelifa et al., 18 Feb 2026). Persistent score-error floors imply persistent drift, whereas sufficiently accurate score estimation and sufficiently large fresh-data fractions yield bounded divergence (Khelifa et al., 18 Feb 2026). Here propagation is recursive: errors are partially forgotten and partially inherited.

In long-horizon language agents, world-model collapse is studied as a phase transition. The order parameter is

$\rho^\ell$8

where $\rho^\ell$9 is state cardinality and $D_i=\chi^2(\hat p^i\|data)$0 is dependency density. The reported phase diagram has a solved plateau, a narrow transition band, and a collapse floor, and the mechanistic signature is that world-state fidelity fails before action validity: $D_i=\chi^2(\hat p^i\|data)$1 on failed episodes (Song et al., 30 Jun 2026). This is an important corrective to the view that agent collapse is merely bad action selection. The evidence indicates that the agent acts from a corrupted world model, and stronger models translate the critical boundary without removing the qualitative transition (Song et al., 30 Jun 2026).

In the global financial system, collapse propagation is a thresholded network dynamical system. The equity vector follows

$D_i=\chi^2(\hat p^i\|data)$2

with $D_i=\chi^2(\hat p^i\|data)$3 a nonnegative Schur matrix of cross-holdings, $D_i=\chi^2(\hat p^i\|data)$4 primitive-asset value, and $D_i=\chi^2(\hat p^i\|data)$5 failure costs (Stella et al., 2023). The paper proves existence and uniqueness results for equilibria, gives explicit formulas such as

$D_i=\chi^2(\hat p^i\|data)$6

and introduces sign-space iteration,

$D_i=\chi^2(\hat p^i\|data)$7

to compute attractive equilibria and identify failure sets (Stella et al., 2023). This is among the clearest threshold-network instantiations of collapse propagation: the propagation channel is $D_i=\chi^2(\hat p^i\|data)$8, the local instability is a threshold crossing, and the eventual state is an equilibrium sign pattern.

These recursive and networked models suggest a broader interpretation. Collapse propagation often has a memory term, a damping term, and a threshold or spectral condition separating benign from runaway regimes. In diffusion models the memory factor is $D_i=\chi^2(\hat p^i\|data)$9; in agents it is the local error-amplification regime near a critical boundary; in finance it is the monotone iteration of failure sets under cross-holdings (Khelifa et al., 18 Feb 2026, Song et al., 30 Jun 2026, Stella et al., 2023).

7. Conceptual themes and unresolved issues

Several conceptual themes recur. First, collapse is rarely purely local. Structural avalanche fronts, elastic reopening fingers, and financial insolvencies advance through explicit couplings; quantum models use commuting collapse generators, integrated-information operators, or chaining graphs; deep-network models use layerwise mappings or spectral concentration [(Beck, 2008); (Li et al., 2022); (Okon et al., 2018); (Li, 24 Feb 2026); (Saada et al., 2024)]. Second, collapse is often controlled by a small set of effective parameters: compaction ratio $R_t$0, fresh-data fraction $R_t$1, collapse rate $R_t$2 or $R_t$3, chaining constant $R_t$4, state cardinality $R_t$5, dependency density $R_t$6, or residual-scaling coefficients $R_t$7 [(Beck, 2008); (Khelifa et al., 18 Feb 2026); (Pearle, 2014); (Li, 24 Feb 2026); (Song et al., 30 Jun 2026); (Noci et al., 2022)]. Third, propagation can be slowed or even reversed by terms that initially appear secondary: compaction increases effective resistance in structural collapse, fresh data geometrically discounts old errors in diffusion retraining, and residual scaling prevents query/key gradient extinction in transformers [(Beck, 2008); (Khelifa et al., 18 Feb 2026); (Noci et al., 2022)].

Open problems are domain-specific but structurally parallel. Engineering models remain highly reduced and often one-dimensional [(Beck, 2008); (Li et al., 2022)]. Quantum models still face unresolved operator-construction problems, relativistic extensions, or energy-production pathologies (Okon et al., 2018, Bedingham et al., 2019). Deep-learning theories are frequently limited to initialization, binary classification, or simplified heads (Súkeník et al., 2023, Saada et al., 2024). Agentic collapse work currently relies on synthetic environments with exact gold state (Song et al., 30 Jun 2026). Financial models provide explicit equilibria, but real calibration of $R_t$8, $R_t$9, and $R(x_2)$00 remains a separate empirical problem (Stella et al., 2023).

What these literatures jointly establish is not a single doctrine but a stable research pattern: collapse becomes analytically tractable when it is modeled as a propagating process with identifiable state variables, couplings, and thresholds. In that sense, a collapse propagation model is less a specific equation than a methodological commitment to describing collapse as transmission, recursion, and geometry rather than as an instantaneous endpoint.