Stochastic Cloning (SC): An Overview

• Stochastic Cloning (SC) is a set of techniques that replicate states, trajectories, particles, or lineages to manage uncertainties in simulation, filtering, and branching models.
• In simulation and delayed-state filtering, SC redistributes computational budgets or augments states to accurately capture transition dynamics and cross-time correlations.
• In rare-event simulation and clonal dynamics, SC reduces variance and enables precise estimation of large-deviation observables and clone-size distributions.

Searching arXiv for relevant papers on stochastic cloning across simulation, estimation, and rare-event methods. arxiv_search(query="stochastic cloning", max_results=10, sort_by="relevance") arxiv_search(query="(Huang et al., 2019) SPSC stochastic cloning", max_results=5, sort_by="relevance") Stochastic Cloning (SC) is not a single universally standardized method; rather, the term denotes several technically distinct constructions that share a common operation—replication of states, trajectories, particles, or lineages under stochastic dynamics. In discrete-time stochastic simulation, SC is the cloning step in SPSC (Simulation, Partitioning, Selection, Cloning), where selected trajectories are replicated to redistribute a fixed simulation budget across representative regions of state space. In robotics and navigation, SC denotes augmented-state filtering with clones of prior states so delayed-state measurements such as odometry or time-differenced carrier phase (TDCP) can be processed consistently. In rare-event simulation, cloning algorithms are continuous-time interacting-particle approximations of normalized Feynman–Kac dynamics for estimating large-deviation quantities. In stochastic clonal dynamics, the term is used for exact, time-independent treatments of clone-size distributions in Markovian branching processes [(Huang et al., 2019); (Mina et al., 28 Aug 2025); (Angeli et al., 2018); (Angeli et al., 2019); (A et al., 2013)].

1. Terminological scope and shared structure

The technical meaning of SC depends on the modeling context. What is cloned may be an entire simulation trajectory, a delayed filter state, a particle in a mean-field ensemble, or a biological lineage state indexed by clone size. The mathematical purpose also changes: budget redistribution in simulation, preservation of cross-time covariance in filtering, variance reduction in rare-event sampling, or exact combinatorial characterization of branching statistics.

Domain	Meaning of SC	Primary role
Discrete-time simulation	Cloning selected trajectories (“delegates”)	Redistribute fixed replications across discovered regions
Robotics and navigation	Augment state with clones of past states	Handle delayed-state correlated measurements
Rare-event simulation	Clone fitter particles and kill less-fit particles	Approximate Feynman–Kac normalized dynamics
Clonal branching models	Exact stochastic clone-size framework	Derive clone-size and mutation-count distributions

A plausible unifying interpretation is that SC makes branching explicit at an intermediate representation level. In SPSC, this branching occurs between time intervals of a black-box simulator; in Kalman filtering, it occurs in the state representation through retained past states; in rare-event simulation, it occurs through selection in an interacting particle system; in branching-process biology, it is the object of analysis itself rather than an algorithmic device.

2. SC as the cloning step in SPSC for discrete-time stochastic simulations

In SPSC, SC is the fourth step of a coordinated execution policy—Simulation, Partitioning, Selection, and Cloning—designed to estimate probabilities of solution sets in discrete-time stochastic simulations when the number of replications is limited. The observable process is a random vector $\boldsymbol{X}_t$, and the probability of interest at final time $T$ is

$$\mathbb{P}(\boldsymbol{X}_T \in \mathcal{S})=\theta_{\mathcal{S}}.$$

The classical Monte Carlo estimator under $n$ independent replications is the relative frequency

$\hat{\theta}_{\mathcal{S}}=\frac{1}{n}\sum_{i=1}^n \mathds{1}_{\mathcal{S}}(\boldsymbol{X}_T^{(i)}).$

SPSC instead decomposes the target probability across intermediate times and partitions of the observable space by the law of total probability, then estimates empirical transition probabilities between partition cells over a coarser time grid $\{t_{(0)}=0,t_{(1)},\dots,t_{(m)}=T\}$ (Huang et al., 2019).

Operationally, the interval $[0,T]$ is split into $m$ pieces, with $m=5$ and $t_{(i)}=i\cdot T/5$ in the reported experiments. For each interval, $T$0 replications are simulated from the current states to the next checkpoint. The $T$1 observable vectors at that checkpoint are partitioned by an unsupervised clustering algorithm; the implementation uses k-means with a preset number of clusters $T$2. From each cluster, the trajectory whose state is nearest to the centroid in Euclidean distance is selected as a delegate. SC then restores the authorized total number of trajectories $T$3 by cloning delegates: each delegate receives $T$4 clones, and if $T$5, the remaining $T$6 clones are assigned by randomly selecting delegates for one additional copy each. No weights are assigned, and all clones are treated equally. The stochastic aspect comes from subsequent evolution: clones are identical at the cloning time but evolve independently afterward, so they fan out around the selected states.

The transition estimator is based on clone counts. If $T$7 cloned replications originate in partition cell $T$8 at time $T$9, and $$\mathbb{P}(\boldsymbol{X}_T \in \mathcal{S})=\theta_{\mathcal{S}}.$$0 of them reach cell $$\mathbb{P}(\boldsymbol{X}_T \in \mathcal{S})=\theta_{\mathcal{S}}.$$1 at time $$\mathbb{P}(\boldsymbol{X}_T \in \mathcal{S})=\theta_{\mathcal{S}}.$$2, then

$$\mathbb{P}(\boldsymbol{X}_T \in \mathcal{S})=\theta_{\mathcal{S}}.$$3

The overall estimator is a sum of products of such empirical conditional probabilities over all cell sequences ending in the target set. The paper does not claim unbiasedness under arbitrary partitioning and cloning strategies, and it explicitly notes that the estimator depends on how partitions are formed, that poorly chosen partitions can bias transition estimates, and that the estimator’s consistency is not proven.

The reported case study uses a Wolf–Sheep Predation NetLogo variant with vegetation, prey, and predator agents, discrete time, and population-count observables. At $$\mathbb{P}(\boldsymbol{X}_T \in \mathcal{S})=\theta_{\mathcal{S}}.$$4, reference probabilities from 30,000 Monte Carlo runs are approximately $$\mathbb{P}(\boldsymbol{X}_T \in \mathcal{S})=\theta_{\mathcal{S}}.$$5 for extinction of both prey and predators, $$\mathbb{P}(\boldsymbol{X}_T \in \mathcal{S})=\theta_{\mathcal{S}}.$$6 for predator extinction with prey survival, and $$\mathbb{P}(\boldsymbol{X}_T \in \mathcal{S})=\theta_{\mathcal{S}}.$$7 for stable coexistence. Under the limited budget $$\mathbb{P}(\boldsymbol{X}_T \in \mathcal{S})=\theta_{\mathcal{S}}.$$8, repeated 1000 times, SPSC improves detection rates relative to Monte Carlo: for “both extinct,” $$\mathbb{P}(\boldsymbol{X}_T \in \mathcal{S})=\theta_{\mathcal{S}}.$$9 versus $n$0; for “predators extinct,” $n$1 versus $n$2; for “all survive,” both $n$3; and for detecting all three solutions in the same run, $n$4 versus $n$5. Using absolute error and aggregate relative error, Wilcoxon–Mann–Whitney tests at $n$6 show smaller errors for SPSC with extremely small $n$7-values, including $n$8 for the aggregate comparison. The interpretation given is that partitioning plus equitable delegate cloning improves coverage of the solution space under a tight simulation budget.

Historically, this use of cloning is situated against prior cloning in parallel simulation, but the implementation is intentionally lightweight: selected delegates are copied to maintain $n$9, with no importance weights and no model-specific rare-event levels. The method is black-box in the sense that it requires only short-horizon simulation, an observable-space distance, clustering, and the ability to copy and resume simulator states.

3. SC in delayed-state filtering and navigation

In robotics and navigation, SC is a state-augmentation technique for measurements that depend on both current and prior states. The canonical measurement model is

$\hat{\theta}_{\mathcal{S}}=\frac{1}{n}\sum_{i=1}^n \mathds{1}_{\mathcal{S}}(\boldsymbol{X}_T^{(i)}).$0

with process model

$\hat{\theta}_{\mathcal{S}}=\frac{1}{n}\sum_{i=1}^n \mathds{1}_{\mathcal{S}}(\boldsymbol{X}_T^{(i)}).$1

After linearization,

$\hat{\theta}_{\mathcal{S}}=\frac{1}{n}\sum_{i=1}^n \mathds{1}_{\mathcal{S}}(\boldsymbol{X}_T^{(i)}).$2

A single-time-state Kalman filter assumes measurements depend only on the current state; when that assumption is violated, the residual is correlated with prior state errors. SC resolves this by augmenting the state with cloned prior states, for example

$\hat{\theta}_{\mathcal{S}}=\frac{1}{n}\sum_{i=1}^n \mathds{1}_{\mathcal{S}}(\boldsymbol{X}_T^{(i)}).$3

and using the standard KF or EKF update in the augmented space. The off-diagonal covariance blocks $\hat{\theta}_{\mathcal{S}}=\frac{1}{n}\sum_{i=1}^n \mathds{1}_{\mathcal{S}}(\boldsymbol{X}_T^{(i)}).$4 are the critical terms encoding current–delayed cross-correlation. A properly derived delayed-state Kalman filter (DSKF), however, yields exactly the same posterior mean and covariance as SC without explicit augmentation; this equivalence is established algebraically in block form for linear and EKF-linearized models, directly contradicting the misconception that Kalman filters cannot handle correlated delayed-state measurements (Mina et al., 28 Aug 2025).

The innovation covariance in the non-augmented DSKF can be written directly as

$\hat{\theta}_{\mathcal{S}}=\frac{1}{n}\sum_{i=1}^n \mathds{1}_{\mathcal{S}}(\boldsymbol{X}_T^{(i)}).$5

with gains

$\hat{\theta}_{\mathcal{S}}=\frac{1}{n}\sum_{i=1}^n \mathds{1}_{\mathcal{S}}(\boldsymbol{X}_T^{(i)}).$6

$\hat{\theta}_{\mathcal{S}}=\frac{1}{n}\sum_{i=1}^n \mathds{1}_{\mathcal{S}}(\boldsymbol{X}_T^{(i)}).$7

These are precisely the block components of the augmented SC gain. The equivalence holds under linear-Gaussian models or EKF linearization at the same point, correct maintenance of cross-covariance blocks, proper handling of residual–process-noise correlation, and the invertible-state-transition condition used in the DSKF derivation.

The computational trade-off is explicit:

Approach	Arithmetic complexity	Memory
SC	$\hat{\theta}_{\mathcal{S}}=\frac{1}{n}\sum_{i=1}^n \mathds{1}_{\mathcal{S}}(\boldsymbol{X}_T^{(i)}).$8	$\hat{\theta}_{\mathcal{S}}=\frac{1}{n}\sum_{i=1}^n \mathds{1}_{\mathcal{S}}(\boldsymbol{X}_T^{(i)}).$9
DSKF	$\{t_{(0)}=0,t_{(1)},\dots,t_{(m)}=T\}$0	$\{t_{(0)}=0,t_{(1)},\dots,t_{(m)}=T\}$1

These expressions show that DSKF reduces flops per iteration relative to SC. In memory, DSKF tends to be favorable when state dimension $\{t_{(0)}=0,t_{(1)},\dots,t_{(m)}=T\}$2 is large relative to measurement dimension $\{t_{(0)}=0,t_{(1)},\dots,t_{(m)}=T\}$3, while SC can be favorable when $\{t_{(0)}=0,t_{(1)},\dots,t_{(m)}=T\}$4 is large relative to $\{t_{(0)}=0,t_{(1)},\dots,t_{(m)}=T\}$5. Both formulations generalize to sliding windows, multiple delayed measurements, time-varying delays, and EKF/UKF settings, provided the joint moments or cross-covariances are maintained correctly.

A concrete aerospace instantiation is the stochastic-cloning UD filter for lunar GNSS-based orbit and clock estimation with TDCP measurements. There, the live state is

$\{t_{(0)}=0,t_{(1)},\dots,t_{(m)}=T\}$6

and the augmented state with one clone is

$\{t_{(0)}=0,t_{(1)},\dots,t_{(m)}=T\}$7

TDCP depends on consecutive epochs, so SC provides the required past-state handle, while UD factorization,

$\{t_{(0)}=0,t_{(1)},\dots,t_{(m)}=T\}$8

preserves symmetry and positive semi-definiteness by construction and avoids square roots. The framework processes all TDCP measurements first, then current-state-only ionosphere-free pseudorange measurements, and uses a delayed-state smoother because standard RTS smoothing does not account for future measurements that directly constrain past states. In 50 Monte Carlo runs over 180 hours, combining ionosphere-free pseudorange and TDCP yields RMS position SISE of approximately $\{t_{(0)}=0,t_{(1)},\dots,t_{(m)}=T\}$9 and RMS velocity SISE of approximately $[0,T]$0 in filtering, while smoothing gives approximately $[0,T]$1 position and $[0,T]$2 velocity RMS; pseudorange-only gives approximately $[0,T]$3 RMS position and $[0,T]$4 RMS velocity (Iiyama et al., 23 Jan 2026).

4. SC in continuous-time rare-event simulation and large deviations

In statistical physics and rare-event analysis, SC denotes continuous-time cloning algorithms for estimating large-deviation quantities of stochastic dynamics. The underlying process is a continuous-time Markov jump process with generator

$[0,T]$5

For additive path functionals,

$[0,T]$6

the tilted generator is

$[0,T]$7

It is split as

$[0,T]$8

where $[0,T]$9 is the mutation part with modified rates $m$0, and $m$1 is the potential. The Feynman–Kac semigroup

$m$2

is not probability conserving; normalizing it yields a non-linear evolution for $m$3, represented by a McKean process. Cloning algorithms are particle approximations of this non-linear process rather than simulations of the original process conditioned on a rare event. The distinction from a Doob $m$4-transform is explicit in the literature (Angeli et al., 2018).

The empirical approximation uses an ensemble of $m$5 clones with empirical measure

$m$6

and the standard estimator of the scaled cumulant generating function is

$m$7

Selection may be represented in several equivalent ways: clone-or-kill with threshold $m$8, pure fitness-increase selection

$m$9

or a residual-sampling analogue. The threshold formulation interprets killing as replacement of less-fit states by samples from $m=5$0 and cloning as copying fitter states into uniformly chosen destinations. The choice of representation affects total selection rate and hence variance.

The continuous-time cloning generator combines mutation-plus-cloning with killing. Each particle has a mutation-plus-cloning clock at rate $m=5$1, and the expected number of offspring per mutation event is

$m=5$2

A two-point offspring distribution on adjacent integers minimizes the second moment $m=5$3 and therefore sharpens the error bounds. The cloning-factor estimator updates multiplicatively at selection events and satisfies

$m=5$4

so that

$m=5$5

is an alternative estimator.

The convergence theory is rigorous and uniform in time. For generic particle approximations,

$m=5$6

$m=5$7

For cloning algorithms, the constants depend additionally on uniform bounds on $m=5$8 and the modified exit rates, and on $m=5$9, the offspring-distribution second moment. A parallel formulation gives finite-time error terms of order $t_{(i)}=i\cdot T/5$0 or $t_{(i)}=i\cdot T/5$1 when a burn-in fraction $t_{(i)}=i\cdot T/5$2 is used (Angeli et al., 2018, Angeli et al., 2019).

The literature also makes a strong methodological connection to sequential Monte Carlo and particle filters. Martingale and carré du champ identities yield the characteristic $t_{(i)}=i\cdot T/5$3 scaling in empirical-measure fluctuations, and adaptive choices of the normalization constant $t_{(i)}=i\cdot T/5$4, such as $t_{(i)}=i\cdot T/5$5 or the median of $t_{(i)}=i\cdot T/5$6, reduce total selection rate. In lattice-gas examples, especially the inclusion process, pairwise fitness-increase selection (“filter2”) is reported to yield smaller selection rates and estimator variance than standard cloning with $t_{(i)}=i\cdot T/5$7 while preserving the mean in the convergent regime.

5. SC in stochastic clonal dynamics and clone-size distributions

In stochastic clonal dynamics, SC is a time-independent framework for exact analysis of clone-size distributions in Markovian branching processes. The core epithelial progenitor model has dividing basal cells $t_{(i)}=i\cdot T/5$8 with three division outcomes: $t_{(i)}=i\cdot T/5$9 with probability $T$00, $T$01 with probability $T$02, and $T$03 with probability $T$04. Under homeostasis, $T$05 on average across the population. Conditioning on clone size $T$06 rather than clonal age makes division timing irrelevant for the distribution of proliferating versus differentiating progeny. If $T$07 is the probability that a clone of total size $T$08 contains $T$09 proliferating cells, then each division increments $T$10 while $T$11 performs a nearest-neighbor random walk with up, flat, and down steps of probabilities $T$12, $T$13, and $T$14, with absorbing boundary at $T$15 (A et al., 2013).

This discrete structure maps exactly to weighted Motzkin lattice paths. Up steps $T$16 correspond to $T$17, flat steps $T$18 to $T$19, and down steps $T$20 to $T$21. Paths start at $T$22, must stay at $T$23, and absorption at $T$24 corresponds to a fully differentiated clone. The probability generating functions are given explicitly: $T$25

$T$26

$T$27

For fully differentiated clones,

$T$28

The gambler’s ruin formulation gives the eventual full-differentiation probability

$T$29

Thus neutrality or subcriticality implies certain eventual differentiation, whereas a supercritical proliferative tilt leaves a nonzero probability of indefinite survival of proliferators.

The same time-independent strategy yields an exact alternative formulation of the Luria–Delbrück problem. In the neutral pure-birth setting, a non-mutant division produces one mutant daughter with probability $T$30 and two non-mutants with probability $T$31. If $T$32 is the probability of observing $T$33 mutant cells when total population size is $T$34, then

$T$35

with $T$36. The framework reproduces the heavy-tailed mutant-count distributions associated with early mutation events and extends to non-neutral selection through relative fitness $T$37. It also gives exact results for the number of clones, $T$38, for clone-size distributions of sub-clones, and for mutation counts in randomly selected clones.

The principal advantage of this framework is exactness under size conditioning: it avoids explicit modeling of cell-cycle time distributions and exposes combinatorial structure via Motzkin and Dyck-type counting. Its limitation is equally clear: age and timing information are suppressed, so if cycle-length heterogeneity or time-resolved dynamics are essential, age-dependent models remain necessary.

6. Trade-offs, misconceptions, and comparative interpretation

Across the literatures, SC trades representational simplicity against statistical or computational control. In SPSC, cloning is simple and model-agnostic, but the estimator depends on partition quality, selection rule, and cloning allocation. The baseline uses uniform cloning across clusters; the authors explicitly suggest future schemes that allocate more clones to small clusters and fewer to large ones in order to counterbalance class imbalance. They also note that no weights are tracked, so bias can be introduced when selection or cloning interacts unfavorably with the dynamics, especially under mis-specified partitions or atypical delegates (Huang et al., 2019).

In delayed-state filtering, the central misconception is that Kalman filters are inherently unable to process correlated delayed-state measurements. The DSKF result shows that the issue is not capability but formulation: SC handles the problem by augmentation, while DSKF handles the same cross-time covariance blocks directly. SC remains attractive when augmented-state code paths, clone stacking, or sparsity-exploiting implementations fit existing tooling, whereas DSKF is preferable when state dimension is large and arithmetic cost is critical (Mina et al., 28 Aug 2025).

In rare-event simulation, a different misconception concerns conditioning. The non-linear McKean dynamics approximated by cloning algorithms arise from Feynman–Kac normalization, not from exact conditioning of the original process on the rare event. Efficiency depends strongly on reducing selection-induced correlations. Adaptive normalization, pairwise fitness-increase selection, and minimal-variance offspring distributions are therefore not merely implementation details; they control genealogical degeneracy and estimator variance (Angeli et al., 2018).

The branching-process usage emphasizes another contrast. There, SC is not primarily a numerical resampling mechanism but an exact stochastic framework indexed by clone size. This suggests that the phrase “stochastic cloning” names either an algorithmic operator or the stochastic object being analyzed, depending on whether replication is enacted computationally or represented analytically (A et al., 2013).

Taken together, these uses show that SC is best understood as a family of replication-based constructions rather than a single method. The common theme is controlled duplication under uncertainty; the differences lie in what is duplicated, how cross-correlation is retained, whether weights are present, and whether the objective is estimation of terminal probabilities, consistent assimilation of delayed measurements, approximation of large-deviation observables, or exact characterization of clone-size statistics.