Time-Invariant Evolutionary Paths

• Time-Invariant Evolutionary Path Characteristic is the concept of conserved cumulative quantities along evolutionary trajectories that remain stable despite changes in timing or ordering.
• It spans multiple disciplines, applying from molecular representation learning with chemically valid edit paths to weak-selection dynamics in evolutionary game theory.
• Recent research demonstrates that invariance under time reversal, shifts, or asymptotic limits provides a robust framework for analyzing evolutionary processes.

Searching arXiv for the cited papers and the phrase to ground the article in current records. A time-invariant evolutionary path characteristic is a property of an evolutionary or evolution-like process that remains invariant with respect to an admissible notion of temporal progression, trajectory parameterization, or path realization. Across the literature, the phrase does not denote a single universal construct. Instead, it appears in several technically distinct settings: molecular representation learning, stochastic evolutionary game dynamics, Lévy-coded branching processes, path-dependent stochastic analysis, multiscale stochastic evolution equations, and lineage processes under changing environments. In each case, the central issue is whether some cumulative quantity associated with an evolutionary path is independent of path ordering, invariant under time reversal, stable under time shifts, or asymptotically reduced to a lower-dimensional invariant dynamics. The most explicit recent formulation in machine learning is PCEvo, where the relevant invariant object is the predicted endpoint property change accumulated along chemically feasible edit paths between molecules (Li et al., 27 Jan 2026).

1. Conceptual scope and meanings across fields

The expression has field-dependent meanings. In molecular representation learning, PCEvo treats the transformation between molecular graphs as a virtual evolutionary path and imposes that the net property change between two molecules should be independent of which chemically valid edit ordering is used, so that the learned mapping is path-invariant rather than strictly time-based (Li et al., 27 Jan 2026). In stochastic evolutionary game theory, the ratio of the path probability density under selection plus drift to that under neutrality is defined as an evolutionary path characteristic that depends on the endpoints of the path but not on elapsed time, and whose expectation is non-decreasing under the stated weak-selection conditions (Yu et al., 15 Aug 2025).

In branching-process theory, the relevant invariant is a reversal symmetry: a whole excursion of a spectrally positive Lévy process above its past infimum is invariant in law under a transformation built from reversing the pre-maximum and post-maximum pieces, and this induces time-reversal invariance at extinction for some branching processes (Felipe et al., 2016). In path-dependent stochastic analysis, time-invariant evolutionary path behavior is encoded by an evolutionary semigroup on path space, where dynamics seen from time $t$ are obtained by shifting the path to time $0$, applying an expectation operator, and shifting back (Denk et al., 2 Jul 2025). In slow–fast stochastic systems, a related idea appears in asymptotic form: after fast transients decay, solutions are exponentially attracted to a random invariant manifold, and in the singular limit this converges to a slow manifold governing long-time dynamics (Fu et al., 2012). In evolutionary lineage models under gradual environmental change, the present-day ancestral path, when reversed in time, converges to a homogeneous Ornstein–Uhlenbeck process in the stationary regime, yielding a stationary ancestral path law rather than a static state (Calvez et al., 2021).

These usages are not interchangeable. A necessary distinction is between path invariance, time-reversal invariance, time-shift covariance, stationary reversed-path laws, and asymptotic reduction to invariant manifolds. A plausible implication is that the phrase is best understood as a family resemblance term rather than a single definition.

2. Path invariance in molecular representation learning

The most direct use of the concept in contemporary AI for science appears in "PCEvo: Path-Consistent Molecular Representation via Virtual Evolutionary" (Li et al., 27 Jan 2026). The method starts from a standard molecular dataset

$\mathcal{D}=\{(G_i,y_i)\}_{i=1}^N,$

and converts it into an evolutionary dataset containing triples $(G_s,G_t,\tau)$, where $G_s$ is a source molecule, $G_t$ is a target molecule, and $\tau\in\mathcal{T}(G_s,G_t)$ is a chemically feasible edit path from source to target (Li et al., 27 Jan 2026). To avoid unrealistic pairings, the source is chosen from the target’s structural neighborhood using Tanimoto similarity on fingerprints: $$\mathcal{N}(G_t) = \text{Top}_\text{K} \left( \{ G_s \in \mathcal{G}_{src} \mid G_s \neq G_t \}, f_{Tanimoto} \right).$$

For each selected pair, the method solves a maximum common subgraph-like alignment problem to map atoms between the two graphs. The symmetric difference is decomposed into a minimal unordered edit set

$\mathcal{S}=\{o_1,\dots,o_M\},$

consisting of elementary operations such as $\text{ADD/REMOVE/REPLACE\_ATOM}$ and $0$0. Because these operations are not order-independent chemically, PCEvo constructs a dependency DAG

$0$1

encoding topological prerequisites such as “an atom must exist before a bond can be added” and “a node must be detached before it is removed.” Any topological sort of this DAG is a valid virtual path, and the method samples up to $0$2 distinct valid paths for the same pair, making the framework explicitly multi-path (Li et al., 27 Jan 2026).

The invariance claim is formulated at the level of cumulative prediction. For a path

$0$3

the shared encoder $0$4 maps each intermediate graph to a latent vector $0$5. Local changes are defined by

$0$6

where $0$7 is a nonlinear delta predictor. The total predicted change along the path is

$0$8

Supervision uses the endpoint label difference through

$0$9

while absolute endpoint supervision is retained via

$\mathcal{D}=\{(G_i,y_i)\}_{i=1}^N,$0

If multiple valid paths $\mathcal{D}=\{(G_i,y_i)\}_{i=1}^N,$1 connect the same endpoints, the model enforces

$\mathcal{D}=\{(G_i,y_i)\}_{i=1}^N,$2

This is the paper’s precise sense of a time-invariant evolutionary path characteristic: the molecule is not assumed to evolve in real time, but the mapping from structural evolution to property change is path-invariant under chemically feasible edit sequences (Li et al., 27 Jan 2026). The invariance is therefore conditional rather than absolute. It holds only over the constrained family $\mathcal{D}=\{(G_i,y_i)\}_{i=1}^N,$3, not over arbitrary graph sequences; it assumes that endpoint property differences can be decomposed into local increments that are learnable and approximately additive along the path; and its theoretical analysis uses a weak-dependence or mixing assumption with an effective sample size $\mathcal{D}=\{(G_i,y_i)\}_{i=1}^N,$4 (Li et al., 27 Jan 2026). The paper further reports that simply increasing the number of paths without $\mathcal{D}=\{(G_i,y_i)\}_{i=1}^N,$5 does not reliably help, whereas enabling path consistency yields stable gains when multiple paths are available, which indicates that invariance is learned from diversity among valid orderings rather than from path multiplicity alone (Li et al., 27 Jan 2026).

3. Endpoint-density ratios under weak selection

A formally different but terminologically close construct is introduced in "The invariance and non-decreasing expectation of an evolutionary path characteristic under weak selection" (Yu et al., 15 Aug 2025). The setting is a finite population of fixed size $\mathcal{D}=\{(G_i,y_i)\}_{i=1}^N,$6 with two phenotypes $\mathcal{D}=\{(G_i,y_i)\}_{i=1}^N,$7 and $\mathcal{D}=\{(G_i,y_i)\}_{i=1}^N,$8, frequencies

$\mathcal{D}=\{(G_i,y_i)\}_{i=1}^N,$9

and a $(G_s,G_t,\tau)$0 payoff matrix

$(G_s,G_t,\tau)$1

Expected payoffs are

$(G_s,G_t,\tau)$2

with population mean payoff

$(G_s,G_t,\tau)$3

The paper defines

$(G_s,G_t,\tau)$4

when $(G_s,G_t,\tau)$5, so that $(G_s,G_t,\tau)$6 is the unique interior singularity where $(G_s,G_t,\tau)$7 (Yu et al., 15 Aug 2025).

Fitnesses under weak selection are

$(G_s,G_t,\tau)$8

with mean fitness

$(G_s,G_t,\tau)$9

The time derivative of mean fitness decomposes as

$G_s$0

where

$G_s$1

The point of departure from Fisher-type arguments is that in stochastic, frequency-dependent dynamics, mean fitness need not be monotone (Yu et al., 15 Aug 2025).

Using a Moran-process diffusion approximation, the transition coefficients are

$G_s$2

with drift and diffusion

$G_s$3

The Fokker–Planck equation for $G_s$4 is

$G_s$5

while under neutrality the density $G_s$6 satisfies

$G_s$7

Under weak selection,

$G_s$8

The evolutionary path characteristic is then defined as the ratio

$G_s$9

Its defining feature is that it depends on the endpoints $G_t$0, not on the elapsed time $G_t$1 (Yu et al., 15 Aug 2025). The paper interprets $G_t$2 as a path more favored under selection than under drift, $G_t$3 as less favored, and $G_t$4 as indistinguishable from neutrality. Its expected value is

$G_t$5

Under the boundary conditions

$G_t$6

the derivative becomes

$G_t$7

or equivalently

$G_t$8

with positive integrals $G_t$9. The paper concludes that under neutrality,

$\tau\in\mathcal{T}(G_s,G_t)$0

and under weak selection with $\tau\in\mathcal{T}(G_s,G_t)$1, especially for large populations $\tau\in\mathcal{T}(G_s,G_t)$2,

$\tau\in\mathcal{T}(G_s,G_t)$3

Thus the expectation of the time-invariant ratio does not decline over time (Yu et al., 15 Aug 2025).

A plausible implication is that this construct is not a property of individual realized paths alone, but of the relation between selected and neutral path-density ensembles. The paper presents it explicitly as an alternative to mean fitness as a measure of evolutionary progression in stochastic frequency-dependent systems (Yu et al., 15 Aug 2025).

4. Time-reversal invariance in branching genealogies

A further major meaning of a time-invariant evolutionary path characteristic appears in "Branching processes seen from their extinction time via path decompositions of reflected Lévy processes" (Felipe et al., 2016). The probabilistic setting is a spectrally positive Lévy process $\tau\in\mathcal{T}(G_s,G_t)$4 with no negative jumps, in the (sub)critical regime $\tau\in\mathcal{T}(G_s,G_t)$5, so that $\tau\in\mathcal{T}(G_s,G_t)$6 does not drift to $\tau\in\mathcal{T}(G_s,G_t)$7. The reflected process is

$\tau\in\mathcal{T}(G_s,G_t)$8

The excursion process of $\tau\in\mathcal{T}(G_s,G_t)$9 away from $$\mathcal{N}(G_t) = \text{Top}_\text{K} \left( \{ G_s \in \mathcal{G}_{src} \mid G_s \neq G_t \}, f_{Tanimoto} \right).$$0 is studied under Itô excursion theory, and for a generic excursion $$\mathcal{N}(G_t) = \text{Top}_\text{K} \left( \{ G_s \in \mathcal{G}_{src} \mid G_s \neq G_t \}, f_{Tanimoto} \right).$$1, the time of its maximum is

$$\mathcal{N}(G_t) = \text{Top}_\text{K} \left( \{ G_s \in \mathcal{G}_{src} \mid G_s \neq G_t \}, f_{Tanimoto} \right).$$2

The path decomposition separates pre-supremum and post-supremum subpaths,

$$\mathcal{N}(G_t) = \text{Top}_\text{K} \left( \{ G_s \in \mathcal{G}_{src} \mid G_s \neq G_t \}, f_{Tanimoto} \right).$$3

and the key transformation is space-time reversal

$$\mathcal{N}(G_t) = \text{Top}_\text{K} \left( \{ G_s \in \mathcal{G}_{src} \mid G_s \neq G_t \}, f_{Tanimoto} \right).$$4

The main theorem states that, under the excursion measure $$\mathcal{N}(G_t) = \text{Top}_\text{K} \left( \{ G_s \in \mathcal{G}_{src} \mid G_s \neq G_t \}, f_{Tanimoto} \right).$$5, reversing both the pre- and post-maximum pieces and concatenating them leaves the law of the excursion unchanged: $$\mathcal{N}(G_t) = \text{Top}_\text{K} \left( \{ G_s \in \mathcal{G}_{src} \mid G_s \neq G_t \}, f_{Tanimoto} \right).$$6 where

$$\mathcal{N}(G_t) = \text{Top}_\text{K} \left( \{ G_s \in \mathcal{G}_{src} \mid G_s \neq G_t \}, f_{Tanimoto} \right).$$7

More specifically, both components are separately invariant: $$\mathcal{N}(G_t) = \text{Top}_\text{K} \left( \{ G_s \in \mathcal{G}_{src} \mid G_s \neq G_t \}, f_{Tanimoto} \right).$$8 and

$$\mathcal{N}(G_t) = \text{Top}_\text{K} \left( \{ G_s \in \mathcal{G}_{src} \mid G_s \neq G_t \}, f_{Tanimoto} \right).$$9

The significance for evolutionary theory comes from the coding of branching genealogies by these excursions. In the finite-variation case, the excursion acts as the contour process of a splitting tree and its local time process is a Crump–Mode–Jagers branching process; in the infinite-variation case, it codes a continuum genealogy. Because the excursion is invariant under the reversal transform, the paper derives that the (sub)critical CMJ branching process and the excursion away from $\mathcal{S}=\{o_1,\dots,o_M\},$0 of the critical Feller diffusion are invariant under time reversal from their extinction time (Felipe et al., 2016). The local time process also satisfies

$\mathcal{S}=\{o_1,\dots,o_M\},$1

which the paper interprets as a profile symmetry in height.

Here the characteristic is not an endpoint-density ratio or additive path functional, but a symmetry of the law of genealogical trajectories when viewed from extinction. This suggests a distinct notion of invariance: the process has no preferred temporal orientation once conditioned and re-centered in the specified manner.

5. Time shifts, semigroups, and path-dependent analytic evolution

In "Martingales and Path-Dependent PDEs via Evolutionary Semigroups" (Denk et al., 2 Jul 2025), the relevant notion is not evolutionary biology but path-dependent stochastic evolution. The authors work on a path space $\mathcal{S}=\{o_1,\dots,o_M\},$2 of continuous trajectories $\mathcal{S}=\{o_1,\dots,o_M\},$3 state space, where the past is encoded by $\mathcal{S}=\{o_1,\dots,o_M\},$4, and define the shift

$\mathcal{S}=\{o_1,\dots,o_M\},$5

An evolutionary semigroup $\mathcal{S}=\{o_1,\dots,o_M\},$6 on past-path space satisfies

$\mathcal{S}=\{o_1,\dots,o_M\},$7

or equivalently arises from a homogeneous expectation operator $\mathcal{S}=\{o_1,\dots,o_M\},$8 via

$\mathcal{S}=\{o_1,\dots,o_M\},$9

The paper identifies this as the analytic encoding of time-invariant evolutionary path behavior (Denk et al., 2 Jul 2025).

The central theorem establishes an equivalence between a compensated martingale and a mild final value problem. For adapted $\text{ADD/REMOVE/REPLACE\_ATOM}$0 and $\text{ADD/REMOVE/REPLACE\_ATOM}$1, with

$\text{ADD/REMOVE/REPLACE\_ATOM}$2

the shifted quantities

$\text{ADD/REMOVE/REPLACE\_ATOM}$3

satisfy

$\text{ADD/REMOVE/REPLACE\_ATOM}$4

if and only if

$\text{ADD/REMOVE/REPLACE\_ATOM}$5

In that case, $\text{ADD/REMOVE/REPLACE\_ATOM}$6 is a mild solution of

$\text{ADD/REMOVE/REPLACE\_ATOM}$7

where $\text{ADD/REMOVE/REPLACE\_ATOM}$8 is the full generator of the evolutionary semigroup (Denk et al., 2 Jul 2025).

A key further notion is the $\text{ADD/REMOVE/REPLACE\_ATOM}$9-derivative

$0$00

The paper proves that

$0$01

if and only if there exists an adapted $0$02 such that

$0$03

is an $0$04-martingale, in which case

$0$05

For the stopping expectation operator $0$06, this derivative coincides with Dupire’s time derivative (Denk et al., 2 Jul 2025).

This use of “time-invariant evolutionary path behavior” differs from both PCEvo and the weak-selection ratio. It is a shift-covariant analytic formalism on path space. A plausible implication is that “evolutionary” here refers to semigroup evolution rather than biological evolution, but the underlying structural idea remains an invariance under change of temporal reference frame.

6. Asymptotic invariance, stationary lineages, and counterexamples to universality

Not all relevant work supports a single invariant quantity. In "Slow Manifolds for Multi-Time-Scale Stochastic Evolutionary Systems" (Fu et al., 2012), the authors consider a coupled fast–slow stochastic system

$0$07

with assumptions including

$0$08

After random transformation of the fast variable, the system generates a cocycle $0$09, and for sufficiently small $0$10 there exists a Lipschitz random invariant manifold

$0$11

with invariance

$0$12

It also has exponential tracking: $0$13 As $0$14, the manifold converges to a slow manifold

$0$15

The paper explicitly interprets this as a sense in which the dynamics become effectively time-invariant or asymptotically time-invariant after elimination of the fast motion (Fu et al., 2012).

A related but more lineage-specific stationary path law appears in "Dynamics of lineages in adaptation to a gradual environmental change" (Calvez et al., 2021). In the moving frame, the macroscopic density satisfies

$0$16

and under

$0$17

there exists a unique positive stationary solution

$0$18

For a uniformly sampled individual at time $0$19, the backward-time version of the limiting spine process is asymptotically the homogeneous Ornstein–Uhlenbeck process

$0$20

The paper interprets this as a stationary ancestral path: in the stationary regime, the distribution of reversed ancestral trajectories does not depend on the observation time $0$21 except through a stationary OU law (Calvez et al., 2021).

By contrast, "Time and Knowability in Evolutionary Processes" argues against any general expectation of a time-invariant path characteristic in biological evolution (Sober et al., 2013). Under the Markov Chain Convergence Theorem, for a process with constant transition probabilities, irreducibility, aperiodicity, and the Markov property,

$0$22

as the time separating past and present approaches infinity. The Data Processing Inequality further implies that in a causal chain $0$23,

$0$24

Within a Moran-process analysis, the paper shows that evidence about ancestral states depends strongly on the form of selection and on mutation, and concludes that information about the past generally decays rather than remaining invariant (Sober et al., 2013). This serves as an important corrective: invariance results are model-specific, not universal.

7. Comparative synthesis and methodological significance

The literature supports several non-equivalent formalizations of a time-invariant evolutionary path characteristic.

Setting	Invariant object	Sense of invariance
Molecular representation learning	$0$25 constrained to match $0$26	Path-independence across chemically feasible edit orderings (Li et al., 27 Jan 2026)
Weak-selection evolutionary dynamics	$0$27	Endpoint-dependent, time-independent density ratio (Yu et al., 15 Aug 2025)
Lévy-coded branching processes	Excursion and local-time laws	Time reversal from the maximum or extinction time (Felipe et al., 2016)
Path-dependent stochastic analysis	Evolutionary semigroup $0$28 on shifted paths	Invariance under time shift in path space (Denk et al., 2 Jul 2025)
Slow–fast stochastic systems	Random invariant manifold and slow-manifold limit	Asymptotic elimination of fast time dependence (Fu et al., 2012)
Adapting lineages in moving environments	Reversed ancestral path law	Stationary OU limit in backward time (Calvez et al., 2021)

Several common structural themes recur. First, invariance typically requires an explicitly defined admissible class of paths or transformations. In PCEvo, it is the constrained family $0$29 of chemically feasible edit paths (Li et al., 27 Jan 2026). In the weak-selection theory, it is the diffusion-approximation path ensemble relative to neutrality (Yu et al., 15 Aug 2025). In the Lévy excursion setting, it is a reversal transformation around the excursion maximum (Felipe et al., 2016). Second, invariance is usually attached to a cumulative or encoded object rather than instantaneous state values: accumulated property change, path-density ratios, excursion laws, shifted semigroup evolution, or reduced manifold dynamics. Third, the validity of the invariant description is conditional on structural assumptions such as weak selection, stationarity, spectral gap, chemical feasibility, or specific reflection and excursion constructions.

A common misconception is to read “time-invariant” as implying literal absence of temporal structure. The surveyed works point in the opposite direction. PCEvo is explicitly “time-like” in that it uses sequences of intermediate molecular states, but the learned quantity is invariant to admissible edit ordering (Li et al., 27 Jan 2026). The weak-selection ratio is derived from a full time-dependent Fokker–Planck and path-integral formalism, yet the resulting characteristic depends only on endpoints (Yu et al., 15 Aug 2025). The Lévy and lineage results concern path laws seen backward from distinguished events such as the maximum or extinction, not static systems (Felipe et al., 2016, Calvez et al., 2021). Time invariance therefore usually means invariance of a derived path functional under a specified temporal transformation, not the absence of dynamics.

Taken together, these works indicate that the phrase time-invariant evolutionary path characteristic names a broad research motif: identifying quantities attached to evolution-like trajectories that remain stable under admissible reorderings, reversals, shifts, or singular limits. The most concrete recent instantiation is path-consistent molecular representation learning, where this principle becomes a trainable inductive bias for few-shot property prediction (Li et al., 27 Jan 2026). More broadly, the motif provides a technical language for expressing when an evolving system possesses a conserved, symmetric, or reference-frame-independent description of cumulative change.