Safe Particle Flow: Theory and Applications

• Safe Particle Flow is a framework that enforces safety at the density level by modifying particle drifts via barrier function constraints.
• It translates safety requirements on the evolving density into drift corrections, ensuring forward invariance and exponential stability across applications.
• The method offers computational tractability and improved accuracy compared to traditional pointwise constraint approaches in variational inference and filtering.

Searching arXiv for papers on safe particle flow and related particle-flow formulations. arxiv_search(query="\"safe particle flow\" variational inference control barrier function", max_results=10, sort_by="relevance") Searching arXiv for related particle flow filtering and stochastic particle flow work. arxiv_search(query="particle flow filtering stochastic importance sampling invertible particle flow arXiv", max_results=10, sort_by="relevance") Searching arXiv for particle-physics uses of “safe” particle/energy flow representations. arxiv_search(query="energy flow networks infrared collinear safety particle flow arXiv", max_results=10, sort_by="relevance") Safe particle flow most specifically denotes a control-barrier-function formulation for constrained variational inference in which safety is defined at the level of the variational density rather than only at the level of individual particles. In that formulation, a particle drift is modified so that the evolving density remains supported on a prescribed safe set, with theoretical guarantees of constraint satisfaction and forward invariance (Yi et al., 12 Sep 2025). More broadly, adjacent literatures use closely related language to describe particle-flow methods that preserve other forms of admissibility, including unbiasedness and finite-time stability in Bayesian filtering, invertibility for importance-corrected proposal transport, reduced stiffness via optimized diffusion, energy conservation in detector reconstruction, and infrared-and-collinear safety in jet representations (Dai et al., 2021, Li et al., 2016, Zhang et al., 2024, Kobylianskii et al., 27 Aug 2025, Komiske et al., 2018).

1. Terminology and scope

In the variational-inference literature, safe particle flow is a principled method for enforcing equality and inequality constraints on a variational density by introducing a barrier functional on the space of probability density functions and translating density-level safety conditions into constraints on the particle drift via the Liouville equation (Yi et al., 12 Sep 2025). In filtering and transport, the same phrase or closely related language is also used for particle-flow constructions that avoid particle degeneracy, preserve consistency, or stabilize numerical integration (Melo et al., 2015, Dai et al., 2021). In collider physics, “particle flow” usually refers instead to reconstruction or representation of particles from detector signals, and safety then means physical admissibility, such as energy conservation, robust handling of overlapping showers, or infrared-and-collinear safety (Kobylianskii et al., 27 Aug 2025, Komiske et al., 2018).

Context	Safety notion	Representative papers
Constrained variational inference	Density support remains in a safe set	(Yi et al., 12 Sep 2025)
Bayesian filtering	Unbiasedness, consistency, stability, invertibility	(Dai et al., 2021, Li et al., 2016, Melo et al., 2015, Zhang et al., 2024, Servadio, 2 May 2025)
Collider reconstruction and ML	Energy conservation, soft assignment, IRC safety	(Kobylianskii et al., 27 Aug 2025, Komiske et al., 2018, Elagin et al., 2012)

This terminological spread is important because the underlying mathematical objects differ. In constrained variational inference, the central object is a time-evolving density $q(x;t)$; in stochastic particle-flow filtering it is a homotopy from prior to posterior; in collider applications it is either detector-object assignment or a permutation-invariant observable defined on sets of particles. A common thread is that “safe” does not mean generic robustness in an informal sense, but a formally specified invariance or admissibility property.

2. Density-level formulation in constrained variational inference

The constrained variational-inference formulation starts from the problem of approximating a Bayesian posterior $p(x\mid z)$ with a variational density $q(x)$ while enforcing support constraints. The safe set is written as

$$\mathcal{S}=\bigcap_{i\in\mathcal{I}} \mathcal{S}_i,\qquad \mathcal{S}_i=\{x\mid g_i(x)\ge 0\},$$

and the intended support condition is that $q(x)=0$ for all $x\notin\mathcal{S}$ (Yi et al., 12 Sep 2025). The paper’s central move is to encode this requirement by a barrier functional on the space of densities rather than by a pointwise barrier only in state space: $$h_i(q)=-\int_{\mathcal{X}\setminus \mathcal{S}_i} g_i(x)\,q(x)\,dx.$$ The feasible set of densities is correspondingly written as

$$\mathcal{Q}_s=\left\{q\;\middle|\; \int_{\mathcal{X}\setminus \mathcal{S}} q(x)\,dx = 0\right\}.$$

A key lemma states that the zero level set $h(q)=\mathbf{0}$ corresponds exactly to densities supported on $\mathcal{S}$ (Yi et al., 12 Sep 2025). This is the conceptual distinction between safe particle flow and earlier approaches that, as stated in the paper, often enforce constraints only on particles but not on the entire density. The formulation therefore makes support preservation a property of the variational approximation itself.

Forward invariance is imposed through a differential inequality on the barrier functional,

$p(x\mid z)$0

with $p(x\mid z)$1 (Yi et al., 12 Sep 2025). The paper states that this condition guarantees that the support of $p(x\mid z)$2 remains in $p(x\mid z)$3 for all $p(x\mid z)$4. It also states exponential stability of infeasibility, so the safety requirement is not merely asymptotic but dynamically enforced along the transport.

3. Liouville dynamics, control barrier functions, and drift correction

The density-level condition becomes operational through the Liouville equation. Particles evolve according to

$p(x\mid z)$5

and the induced density evolution is

$p(x\mid z)$6

(Yi et al., 12 Sep 2025). This connects the time derivative of the variational density directly to the particle drift and makes it possible to derive control-barrier-function constraints on the drift itself.

The main theorem stated in the paper is that if, for each constraint $p(x\mid z)$7,

$p(x\mid z)$8

then the feasible density set is forward-invariant and exponentially stable (Yi et al., 12 Sep 2025). A local, particle-wise sufficient condition is then introduced in standard CBF form: $p(x\mid z)$9 for all $q(x)$0, with $q(x)$1 (Yi et al., 12 Sep 2025).

The drift is synthesized by starting from an unconstrained drift $q(x)$2, for example a Stein drift, and adding a minimal correction $q(x)$3. The correction is obtained by solving a quadratic program for each particle: $q(x)$4 subject to

$q(x)$5

The safe particle drift is then

$q(x)$6

(Yi et al., 12 Sep 2025). According to the paper, this yields a computationally tractable method with forward invariance, exponential stability, and feasibility of the control input under mild regularity conditions.

The reported numerical simulations compare safe particle flow with projected particle flow and unconstrained flows under both equality and inequality constraints. The paper states that the projected method sometimes failed to satisfy all constraints, especially when combining equality and inequality types, whereas safe particle flow enforced both constraints rigorously at the density level and achieved accurate posterior approximation with low KL divergence (Yi et al., 12 Sep 2025). This suggests that the decisive innovation is not merely adding a barrier to particle trajectories, but lifting the safety condition to the density manifold and then projecting it back to drift space.

4. Relation to stochastic particle-flow filtering

The broader particle-flow filtering literature supplies the background against which safe particle flow is most naturally understood. A parameterized family of stochastic particle flow filters derives flows driven by a linear combination of prior information and measurement likelihood information, proves unbiasedness under linear measurement and Gaussian distributions, establishes consistency, and develops finite-time stability, finite-time contractive stability, and finite-time stochastic stability (Dai et al., 2021). In that framework, the particle flow SDE is

$q(x)$7

and a Lyapunov analysis shows that if $q(x)$8, the error contracts exponentially (Dai et al., 2021). The practical implication drawn in the paper is that choosing $q(x)$9 is beneficial for safe implementation because diffusion suppresses numerical error growth.

Stochastic Particle Flow recasts the update as a Langevin diffusion whose stationary solution is the posterior,

$$\mathcal{S}=\bigcap_{i\in\mathcal{I}} \mathcal{S}_i,\qquad \mathcal{S}_i=\{x\mid g_i(x)\ge 0\},$$0

and uses equally weighted samples associated with local solutions of the Fokker–Planck equation to produce a Gaussian-mixture approximation of the filtering density (Melo et al., 2015). The paper emphasizes the absence of weight degeneracy, maintenance of high ESS in challenging settings, and convergence in total variation under suitable conditions. In this line of work, safety is closely tied to avoiding weight collapse and maintaining a representative sample cloud in high dimension.

Later work on importance sampling with stochastic particle flow introduces a diffusion term into the particle-flow ODE, formulates a flow-induced proposal density after migration of a Gaussian mixture model, and optimizes the diffusion matrix to improve the accuracy–computational complexity tradeoff (Zhang et al., 2024). The stated objective

$$\mathcal{S}=\bigcap_{i\in\mathcal{I}} \mathcal{S}_i,\qquad \mathcal{S}_i=\{x\mid g_i(x)\ge 0\},$$1

balances stiffness reduction against excess diffusion, and the paper reports reduced stiffness of the ODE together with improved estimating accuracy in a highly nonlinear 3-D source localization scenario (Zhang et al., 2024). Here, safety refers primarily to numerically reliable transport through stiff regions.

A distinct but complementary strand embeds deterministic particle flows inside standard particle filtering by enforcing invertibility of the flow map. The resulting change-of-variables proposal density yields computationally efficient weight updates while preserving the usual particle-filter convergence guarantees (Li et al., 2016). The proposal takes the form

$$\mathcal{S}=\bigcap_{i\in\mathcal{I}} \mathcal{S}_i,\qquad \mathcal{S}_i=\{x\mid g_i(x)\ge 0\},$$2

so safety is achieved by principled importance correction rather than by hard state constraints (Li et al., 2016).

High-order particle-flow filters derived from differential algebra and Taylor expansion series push this stability perspective further. They construct drift and diffusion as polynomials, introduce DAPFFv1-$$\mathcal{S}=\bigcap_{i\in\mathcal{I}} \mathcal{S}_i,\qquad \mathcal{S}_i=\{x\mid g_i(x)\ge 0\},$$3 and DAPFFv2-$$\mathcal{S}=\bigcap_{i\in\mathcal{I}} \mathcal{S}_i,\qquad \mathcal{S}_i=\{x\mid g_i(x)\ge 0\},$$4, and report that high-order terms improve posterior fidelity relative to Gromov flow and the “exact” flow, with the paper explicitly linking these improvements to stability and safety in nonlinear regimes (Servadio, 2 May 2025). This suggests a broad taxonomy in which safe particle flow can mean barrier-certified support preservation, but can also denote stable and accurate transport under discretization, curvature, or stiffness constraints.

5. Safety notions in particle-physics reconstruction and representation

In high-energy physics, particle-flow terminology predates the variational-inference usage and refers to detector-level reconstruction. A probabilistic particle flow algorithm for high-occupancy environments was proposed to resolve overlapping calorimeter energy depositions of spatially close particles using a statistically consistent probabilistic procedure. The abstract states that the method is nearly free of ad-hoc corrections, improves energy resolution, and provides new handles such as the uncertainty of the jet energy on an event-by-event basis and the estimate of the probability of a given particle hypothesis for a given detector response; when applied to hadronic jets from tau decays in the CDF-II detector at Fermilab, it demonstrated reliable and robust performance (Elagin et al., 2012). In this context, safety means reliable reconstruction under overlap and occupancy.

Recent transformer-based particle-flow reconstruction sharpens the physical-constraint interpretation. GLOW combines incidence matrix supervision from HGPflow with a MaskFormer architecture, using a 6-layer self-attention transformer encoder and a 4-layer masked cross-attention transformer decoder (Kobylianskii et al., 27 Aug 2025). The incidence matrix $$\mathcal{S}=\bigcap_{i\in\mathcal{I}} \mathcal{S}_i,\qquad \mathcal{S}_i=\{x\mid g_i(x)\ge 0\},$$5 represents the fraction of the energy from detector object $$\mathcal{S}=\bigcap_{i\in\mathcal{I}} \mathcal{S}_i,\qquad \mathcal{S}_i=\{x\mid g_i(x)\ge 0\},$$6 attributed to particle $$\mathcal{S}=\bigcap_{i\in\mathcal{I}} \mathcal{S}_i,\qquad \mathcal{S}_i=\{x\mid g_i(x)\ge 0\},$$7, and the details state that each detector object’s total assigned energy cannot exceed its measurement, directly embedding energy conservation (Kobylianskii et al., 27 Aug 2025). The model also permits fractional, overlapping assignments of cluster energy to several particles, which the paper presents as a more robust treatment of overlapping showers than hard one-to-one assignment.

A different notion of safety appears in machine-learned jet observables. Komiske, Metodiev, and Thaler introduced Energy Flow Networks and Particle Flow Networks by adapting Deep Sets to particle jets. EFNs are written as

$$\mathcal{S}=\bigcap_{i\in\mathcal{I}} \mathcal{S}_i,\qquad \mathcal{S}_i=\{x\mid g_i(x)\ge 0\},$$8

while PFNs use the general form

$$\mathcal{S}=\bigcap_{i\in\mathcal{I}} \mathcal{S}_i,\qquad \mathcal{S}_i=\{x\mid g_i(x)\ge 0\},$$9

(Komiske et al., 2018). The paper states that EFNs respect infrared and collinear safety by construction, because the latent representation is an energy-weighted linear sum over purely geometric inputs. It also states that PFNs outperform EFNs on quark/gluon discrimination and that including particle identification information improves classification performance, indicating that IRC-unsafe information is helpful for that task (Komiske et al., 2018). This is a precise example of a trade-off between formal safety and representational freedom.

These collider uses are mathematically distinct from safe particle flow in variational inference. Nonetheless, they share a structural pattern: safety is encoded by construction—through energy-conserving assignments, statistically consistent overlap resolution, or a representation class restricted to IRC-safe observables—rather than added as a post hoc correction.

6. Conceptual distinctions, misconceptions, and open issues

A common misconception is that constraining particle trajectories is equivalent to constraining the evolving approximation. The constrained variational-inference formulation explicitly rejects this equivalence by stating that existing methods often enforce constraints only on particles but not on the entire density $q(x)=0$0 (Yi et al., 12 Sep 2025). The defining feature of safe particle flow in that setting is therefore density-level support invariance, not merely pointwise feasibility of sample paths.

A second misconception is that deterministic flow is inherently more correct than stochastic flow. The filtering literature does not support such a blanket conclusion. Stochastic particle flow is introduced precisely to address failures of deterministic flows in capturing higher moments and in maintaining stability in nonlinear, high-dimensional settings (Melo et al., 2015). Diffusion optimization is later used to reduce stiffness and improve the accuracy–computational complexity tradeoff (Zhang et al., 2024), and the parameterized family of stochastic flows proves finite-time stability properties not available from a purely heuristic drift construction (Dai et al., 2021).

A third misconception is that “safe particle flow” designates a single algorithm across all disciplines. The literature instead points to several non-equivalent safety criteria. In constrained variational inference, safety is support preservation of $q(x)=0$1; in Bayesian filtering, it is often unbiasedness, consistency, invertibility, or stable transport; in collider reconstruction, it is energy consistency and robust shared assignment; in jet learning, it is infrared-and-collinear safety (Yi et al., 12 Sep 2025, Li et al., 2016, Kobylianskii et al., 27 Aug 2025, Komiske et al., 2018).

The current literature also reveals concrete limitations. The proofs of unbiasedness and correct covariance in the unified stochastic-flow family are stated under linear measurements and Gaussian prior and likelihood distributions (Dai et al., 2021). In the differential-algebra formulation, DAPFFv2-$q(x)=0$2 is limited to maximum feasible order $q(x)=0$3 because higher orders are nullified by evaluation at zero deviation (Servadio, 2 May 2025). In jet learning, PFNs outperform EFNs on quark/gluon discrimination, which shows that safety by construction can exclude useful task-relevant information (Komiske et al., 2018). A plausible implication is that future work will continue to balance formal invariance guarantees against approximation power, computational tractability, and the degree to which physically meaningful structure should be hard-coded rather than learned.

Across these domains, safe particle flow is best understood not as a monolithic technique but as a design principle for particle-based transport: specify an admissibility criterion at the appropriate level—density support, proposal correctness, numerical stability, conservation law, or IRC invariance—and construct the flow so that the criterion is enforced by the transport itself rather than by ex post repair.