Spatial Ancestral Chains: Theory & Applications

• Spatial Ancestral Chains (SAC) are explicitly defined lineage chains augmented with spatial coordinates, providing a framework for tracking ancestry in diverse spatial models.
• They are applied in individual-based speciation, locally regulated population models, and hierarchical scene reconstruction, linking genealogy with spatial dynamics.
• SAC methodologies offer practical insights into coalescence phenomena, diffusive scaling, and efficient ancestral data composition, enhancing both evolutionary analysis and computational imaging.

Searching arXiv for the cited SAC-related papers and adjacent work to ground the article in current arXiv records. I'll look up the arXiv entries directly to confirm the core papers and neighboring literature. {"query":"(Costa et al., 2017) Registering the evolutionary history in individual-based models of speciation", "max_results": 5} Spatial Ancestral Chains (SAC) denote ancestor-tracing structures in which lineage is represented together with an explicit spatial embedding through time. In spatially explicit, individual-based simulations of speciation, a SAC for an extant individual is the ordered sequence of ancestor–descendant links, augmented by spatial coordinates at each generation, tracing the lineage back to a founder (Costa et al., 2017). In spatial population models with local regulation, SAC refers to the genealogical path of a sampled individual through space-time, represented as a random walk in a dynamic random environment, with collections of such paths forming coalescing systems (Birkner et al., 2019). In monocular dynamic reconstruction, the same term is used in a different sense: a Spatial Ancestral Chain is the ancestor sequence of nodes in a temporal partition tree, queried recursively to compose spatial dynamics across hierarchy levels (Wang et al., 12 Feb 2026). The shared structure is ancestry organized as a chain and coupled to spatial state, but the mathematical objects, observables, and inferential roles differ across these literatures.

1. Core definitions and formal structure

In the spatially explicit speciation setting, for individual $i$ extant at generation $t^*$, the ancestral chain is

$$A_i = \{ i_{t^*}, p(i,t^*), p(p(i,t^*),t^*-1), \dots, \mathrm{founder} \},$$

and the spatial SAC augments each node with coordinates $(x_{\cdot}(\tau), y_{\cdot}(\tau))$ for every generation $\tau$ on the path (Costa et al., 2017). A single-lineage SAC is therefore a backward-time lineage record with explicit space-time localization. For two sampled individuals, SAC also supports a spatial coalescent representation: one traces both lineages back to the first common ancestor and extracts the two spatial paths from the extant samples to their MRCA (Costa et al., 2017).

In locally regulated spatial population models, SAC is operationalized as the spatial position of the ancestor $n$ generations back. The central construction is a process $X=(X_n)_{n\ge 0}$, where $X_n$ is the spatial location of the ancestor $n$ generations in the past of a sampled individual (Birkner et al., 2019). Conditional on the environment, $X$ is a time-inhomogeneous Markov chain. Multiple SACs then form a coalescing system of random walks in the same environment, and in one spatial dimension the diffusive scaling limit of the full collection is the Brownian web (Birkner et al., 2019).

A precursor to this viewpoint appears in the branching Markov chain framework, where the lineage of a typical generation-$t^*$0 individual is represented by a non-homogeneous Markov chain $t^*$1 with transition kernel

$t^*$2

This is a Doob $t^*$3-transform of the one-step mean kernel and functions as the ancestral chain of a typical descendant under a size-biased change of measure (Bansaye, 2013). In homogeneous branching random walks, this chain reduces to the underlying spatial random walk (Bansaye, 2013). This suggests that SAC in later spatial-population work can be viewed as a concrete, spatially embedded specialization of a broader ancestral-chain formalism.

In WorldTree, SAC is defined on tree nodes rather than biological individuals. For a temporal partition tree node $t^*$4, the chain

$t^*$5

collects ancestors along the root-to-node path, and the rendered representation at time $t^*$6 is the union of the current node’s deformed Gaussians with those of all ancestors (Wang et al., 12 Feb 2026). Here “ancestral” refers to hierarchy within a learned decomposition rather than genealogy in a population.

2. SAC in spatially explicit individual-based speciation models

The speciation framework is a spatially explicit, individual-based model with constant population size $t^*$7, haploid genomes of length $t^*$8 binary loci, sex labels in $t^*$9, and coordinates $$A_i = \{ i_{t^*}, p(i,t^*), p(p(i,t^*),t^*-1), \dots, \mathrm{founder} \},$$0 on a periodic $$A_i = \{ i_{t^*}, p(i,t^*), p(p(i,t^*),t^*-1), \dots, \mathrm{founder} \},$$1 lattice (Costa et al., 2017). Distances are computed with the torus metric

$$A_i = \{ i_{t^*}, p(i,t^*), p(p(i,t^*),t^*-1), \dots, \mathrm{founder} \},$$2

Mating is local and compatibility-constrained: individual $$A_i = \{ i_{t^*}, p(i,t^*), p(p(i,t^*),t^*-1), \dots, \mathrm{founder} \},$$3 can mate only with partners $$A_i = \{ i_{t^*}, p(i,t^*), p(p(i,t^*),t^*-1), \dots, \mathrm{founder} \},$$4 satisfying $$A_i = \{ i_{t^*}, p(i,t^*), p(p(i,t^*),t^*-1), \dots, \mathrm{founder} \},$$5 and genetic compatibility $$A_i = \{ i_{t^*}, p(i,t^*), p(p(i,t^*),t^*-1), \dots, \mathrm{founder} \},$$6, where $$A_i = \{ i_{t^*}, p(i,t^*), p(p(i,t^*),t^*-1), \dots, \mathrm{founder} \},$$7 is the Hamming distance between binary genotypes (Costa et al., 2017). Species at generation $$A_i = \{ i_{t^*}, p(i,t^*), p(p(i,t^*),t^*-1), \dots, \mathrm{founder} \},$$8 are defined as connected components of the genetic compatibility graph $$A_i = \{ i_{t^*}, p(i,t^*), p(p(i,t^*),t^*-1), \dots, \mathrm{founder} \},$$9, so conspecific individuals need not be directly compatible if they are connected through compatible intermediates (Costa et al., 2017).

Within this model, SAC sits between two bookkeeping schemes. The Most Recent Common Ancestor Time matrix (MRCAT) stores pairwise times to MRCA among extant individuals and is driven by parent pointers $(x_{\cdot}(\tau), y_{\cdot}(\tau))$0, while the Sequential Speciation and Extinction Events algorithm (SSEE) records forward-time species-level branching and extinction and yields the true species phylogeny (Costa et al., 2017). SAC links these resolutions by using MRCAT to extract individual-level spatial lineages and coalescent paths, and by using SSEE to attach species-level spatial attributes such as ranges or centroids through time to branches traversed by those lineages (Costa et al., 2017).

For asexual models, the MRCAT recurrence is

$(x_{\cdot}(\tau), y_{\cdot}(\tau))$1

with $(x_{\cdot}(\tau), y_{\cdot}(\tau))$2 and $(x_{\cdot}(\tau), y_{\cdot}(\tau))$3 for $(x_{\cdot}(\tau), y_{\cdot}(\tau))$4. For sexual models, maternal and paternal matrices evolve separately,

$(x_{\cdot}(\tau), y_{\cdot}(\tau))$5

and a combined-lineage version may use

$(x_{\cdot}(\tau), y_{\cdot}(\tau))$6

Coordinates $(x_{\cdot}(\tau), y_{\cdot}(\tau))$7 are stored per generation for SAC reconstruction (Costa et al., 2017).

The lineage reconstruction itself is direct. Starting from a focal individual at time $(x_{\cdot}(\tau), y_{\cdot}(\tau))$8, one follows parent pointers backward, storing identifiers, generation stamps, and coordinates:

$(x_{\cdot}(\tau), y_{\cdot}(\tau))$9

For two individuals, one can recover the MRCA from $\tau$0, trace both parent chains until the first common ancestor, and extract the two coordinate sequences down to the coalescence time (Costa et al., 2017).

The model demonstrations use $\tau$1, $\tau$2, $\tau$3, $\tau$4, $\tau$5, $\tau$6, and $\tau$7 (Costa et al., 2017). Across 50 simulations at $\tau$8 and $\tau$9 generations, trees from MRCAT were consistently closer to the true SSEE phylogeny than distance-based trees by Robinson–Foulds topology and $n$0 branch-length distributions, and the advantage increased at $n$1 (Costa et al., 2017). Because SAC is reconstructed from parent pointers rather than from extant genetic distances alone, this result supports its use for fine-scale spatial genealogy in simulations.

3. SAC in locally regulated spatial population models

In spatial population models with local regulation, SAC is not primarily a storage device but a stochastic process induced by forward demographic dynamics. Two prototypical cases are emphasized: a discrete-time contact process interpreted through oriented percolation, and the logistic branching random walk (LBRW) (Birkner et al., 2019).

For the contact process, the environment is an i.i.d. Bernoulli field $n$2 on $n$3. A site is open if $n$4 and closed otherwise. Occupancy evolves by open nearest-neighbor propagation, and local regulation arises because an inhabitable site chooses a single parent uniformly among occupied neighbors, so competition intensifies when many candidate parents are present (Birkner et al., 2019). In the stationary regime for $n$5, the time-reversed indicator field $n$6 iff there is an infinite directed open path starting at $n$7 serves as the dynamic environment for the ancestral walk (Birkner et al., 2019). On the event $n$8, the ancestral lineage from the origin moves by

$n$9

for $X=(X_n)_{n\ge 0}$0, where $X=(X_n)_{n\ge 0}$1 is the oriented percolation backbone (Birkner et al., 2019).

For the LBRW, $X=(X_n)_{n\ge 0}$2 is the number of individuals at $X=(X_n)_{n\ge 0}$3 in generation $X=(X_n)_{n\ge 0}$4, and each individual at $X=(X_n)_{n\ge 0}$5 produces offspring with Poisson mean

$X=(X_n)_{n\ge 0}$6

followed by migration according to a symmetric, aperiodic finite-range kernel $X=(X_n)_{n\ge 0}$7 (Birkner et al., 2019). Local regulation is explicit in the competition term $X=(X_n)_{n\ge 0}$8, which depresses reproduction in crowded neighborhoods. In the stationary regime, the ancestral lineage of a uniformly sampled individual at $X=(X_n)_{n\ge 0}$9 has transition kernel

$X_n$0

The environment seen by the lineage is therefore the time-reversed stationary population field (Birkner et al., 2019).

A central technical contribution is the regeneration construction. In the contact-process case, a local path exploration with additional randomness yields regeneration times

$X_n$1

with increments $X_n$2 and $X_n$3 such that $X_n$4 are i.i.d., $X_n$5 is symmetrically distributed, and both $X_n$6 and $X_n$7 have exponential tails (Birkner et al., 2019). In the LBRW, regeneration is obtained through coarse-grained good blocks, percolation comparison, strong coupling, and regeneration cones (Birkner et al., 2019).

These renewal structures imply diffusive scaling. For the contact process, conditioning on $X_n$8,

$X_n$9

for almost every environment, and both annealed and quenched central limit theorems hold with a centered, non-degenerate, isotropic Gaussian limit and variance

$n$0

For the LBRW, under the small-competition regime ensuring a unique non-trivial stationary law, one has

$n$1

and an annealed CLT with a non-degenerate Gaussian limit (Birkner et al., 2019). The interpretation is that SAC behaves, at large scales, like a centered random walk whose effective variance encodes local regulation and environmental fluctuations.

Collections of SACs exhibit coalescence phenomena. In one spatial dimension, the entire set of ancestral paths on the oriented percolation backbone converges under diffusive scaling to the Brownian web, the continuum object of coalescing Brownian motions (Birkner et al., 2019). This is a system-level statement: the SAC ensemble, not only individual chains, has a universal scaling limit.

4. Pairwise coalescence, isolation by distance, and effective-size interpretations

The pairwise analysis of SAC is developed explicitly for two-dimensional LBRW lineages (Birkner et al., 2024). Two individuals are sampled at time $n$2, one from site $n$3 and one from site $n$4, conditional on occupancy. Their ancestral locations are $n$5 and $n$6, and the coalescence time is

$n$7

Meeting time $n$8 satisfies $n$9 (Birkner et al., 2024).

The main asymptotic statement is a two-dimensional tail law. Under $X$0, small competition parameters, and stationarity,

$X$1

An analogous result holds first for $X$2, and finer analysis shows that $X$3 is negligible at scale $X$4 (Birkner et al., 2024). The result agrees with classical two-dimensional random-walk hitting asymptotics and with the suitably scaled stepping stone model (Birkner et al., 2024).

The methodological core is a joint regeneration construction for the two lineages. Space-time is coarse-grained into boxes, regeneration times $X$5 are defined for the pair, and the increments have polynomial tail control:

$X$6

A coupling to independent environments produces approximation bounds of order $X$7 for the first regeneration increment, and a separation lemma gives exponential control on the time needed for the pair to separate to radius $X$8 (Birkner et al., 2024). Hitting probabilities for spheres then reduce the large-distance genealogy to near-harmonic behavior of $X$9 in two dimensions.

The same framework links pairwise SAC behavior to identity by descent. If the mutation rate is scaled as $t^*$00, then

$t^*$01

which matches the corresponding scaling of Malécot’s continuous-space approximation and the stepping stone model (Birkner et al., 2024). This supports the replacement of fluctuating local sizes by fixed effective sizes for large-scale inference in locally regulated populations (Birkner et al., 2024). In the terminology of SAC, pairwise links in the ancestral chain network have a predictable large-distance decay, and regeneration blocks provide the tractable decomposition that makes this asymptotic regime analyzable.

The spatially explicit speciation model gives related, but simulation-based, spatial diagnostics. Range overlap is

$t^*$02

and isolation by distance can be quantified through a Mantel correlation between geographic and genetic distance matrices (Costa et al., 2017). Centroid trajectories $t^*$03 align species movement with lineage chains and permit allopatric-versus-sympatric interpretation of branches traversed by SAC (Costa et al., 2017). These are not asymptotic theorems, but they place SAC in a macroevolutionary spatial-signature analysis.

5. SAC in hierarchical dynamic scene reconstruction

In WorldTree, SAC is part of a monocular dynamic reconstruction framework that combines a Temporal Partition Tree (TPT) with a complementary spatial hierarchy (Wang et al., 12 Feb 2026). The dynamic scene is represented by motion bases $t^*$04 and canonical Gaussians $t^*$05, with deformation driven by a KNN motion graph and blended $t^*$06 motions via Dual Quaternion Blending (Wang et al., 12 Feb 2026). TPT recursively partitions the time interval and assigns each node a subset of motion bases and Gaussians. A child focuses on a narrower temporal interval and inherits only a subset of its parent’s Gaussian primitives (Wang et al., 12 Feb 2026).

SAC supplies the missing spatial context created by this inheritance truncation. For node $t^*$07, the ancestral chain $t^*$08 contains the node’s ancestors in TPT, and the deformed Gaussian set rendered at time $t^*$09 is

$t^*$10

Thus a node’s render is the union of its own deformed Gaussians and the deformed Gaussians from all ancestors on the chain (Wang et al., 12 Feb 2026). SAC therefore performs composition rather than parameter fusion: ancestors remain separately optimized, and their contributions are aggregated only at render time (Wang et al., 12 Feb 2026).

The paper frames this as “hierarchical spatial composition and modeling specialization” and “hierarchical motion decoupling” (Wang et al., 12 Feb 2026). Each ancestor has its own motion bases and Gaussians, optimized only on its own interval, while SAC recursively queries these ancestors to provide complementary spatial dynamics that were pruned during inheritance (Wang et al., 12 Feb 2026). During breadth-first TPT construction, chains propagate by appending the current node to the parent chain:

$t^*$11

Nodes at the same depth have the same ancestor-chain initialization but are optimized independently, with frozen ancestors, enabling parallel depth-wise optimization (Wang et al., 12 Feb 2026).

WorldTree reports that SAC improves reconstruction quality beyond TPT alone. On NVIDIA-LS, enabling TPT on top of a baseline brings LPIPS from $t^*$12 to $t^*$13 and mAVGE from $t^*$14 to $t^*$15; adding SAC further improves LPIPS to $t^*$16 and mAVGE to $t^*$17 (Wang et al., 12 Feb 2026). With BA and SW enabled, the final metrics are PSNR $t^*$18, SSIM $t^*$19, LPIPS $t^*$20, AVGE $t^*$21, mPSNR $t^*$22, mSSIM $t^*$23, and mAVGE $t^*$24 (Wang et al., 12 Feb 2026). On DyCheck, the reported mLPIPS is $t^*$25 versus MoSca’s $t^*$26, an improvement of $t^*$27 relative to the second-best method (Wang et al., 12 Feb 2026).

Ablations emphasize chain length. With TPT depth fixed, the full chain yields mPSNR $t^*$28, mSSIM $t^*$29, and LPIPS $t^*$30, whereas progressively reduced chains yield mPSNR $t^*$31, mSSIM $t^*$32, and LPIPS $t^*$33 (Wang et al., 12 Feb 2026). This indicates that the ancestor composition itself, rather than merely hierarchical splitting, is essential in that setting.

6. Algorithms, complexity, limitations, and cross-domain significance

The algorithmic profile of SAC depends sharply on domain. In the speciation model, MRCAT requires storing pairwise genealogical times over extant individuals, with $t^*$34 update time and $t^*$35 storage per generation; full parent-pointer storage for SAC reconstruction is $t^*$36 in memory, while checkpointing reduces this to $t^*$37 at an $t^*$38 extraction penalty per query (Costa et al., 2017). Sparse representations are generally not useful for $t^*$39 because most entries are nonzero, so scaling strategies rely instead on downsampling, windowing, and checkpointing (Costa et al., 2017). SSEE preserves extinct species branches, but individual-level SACs for extinct individuals require archive storage if they are to remain queryable (Costa et al., 2017).

In the local-regulation literature, the main limitation is analytic rather than storage-related. The environments are non-elliptic and, in the contact-process case, not Markovian in a local sense, so general random-walk-in-dynamic-random-environment results do not directly apply (Birkner et al., 2019). Model-specific regeneration constructions are therefore essential. Open problems include characterizing $t^*$40 as $t^*$41 for the contact process, establishing a quenched CLT for the LBRW, and understanding higher-dimensional scaling limits of coalescing systems, since no Brownian-web analogue exists in $t^*$42 where Brownian motions do not meet (Birkner et al., 2019). In two-dimensional LBRW, the coalescence asymptotics require stationarity, small competition, and finite-range dispersal and competition; the results identify the decay profile but not the prefactors needed for direct numerical prediction (Birkner et al., 2024).

In WorldTree, SAC adds a union over ancestors during rendering, so rasterization cost grows with the number of Gaussian primitives included from ancestral nodes (Wang et al., 12 Feb 2026). Practical efficiency is maintained by limiting tree depth and inherited Gaussian counts. The experiments use TPT depth $t^*$43, inherited Gaussian counts of $t^*$44 at layer $t^*$45 and $t^*$46 at layer $t^*$47, BA for $t^*$48 steps, Static Warm-up for $t^*$49 steps at the root, and layerwise optimization schedules of $t^*$50 steps for NVIDIA-LS and $t^*$51 for DyCheck (Wang et al., 12 Feb 2026). The method depends on pretrained depth, tracking, and flow priors; failures are reported on small, fast-moving regions and under pervasive scene-wide non-rigid deformation (Wang et al., 12 Feb 2026).

Across domains, SAC consistently serves as an ancestry-conditioned spatial representation, but its formal role changes. In branching and population models, SAC is a stochastic lineage object used to derive many-to-one formulas, coalescence laws, CLTs, and scaling limits [(Bansaye, 2013); (Birkner et al., 2019); (Birkner et al., 2024)]. In speciation simulations, SAC is a reconstructible data structure that connects individual genealogy, space, and species-level phylogeny (Costa et al., 2017). In dynamic reconstruction, SAC is a render-time composition mechanism over hierarchical ancestors that restores spatial support without re-entangling optimization (Wang et al., 12 Feb 2026). This suggests that the unifying content of the term is not a single biological or computational ontology, but a common design principle: spatial information becomes tractable when ancestry is organized as an explicit chain and queried across scales.