Future Causal Completion

• Future Causal Completion is a framework for reconstructing or forecasting admissible future states by applying causal constraints across geometric, generative, and counterfactual inference settings.
• It employs methods like Lorentzian geometry, latent-space light-cone restrictions, and low-rank matrix/tensor completion to delimit the set of plausible future outcomes.
• This approach enables precise future forecasting and imputation while addressing challenges related to topology, semantic validity, and domain-specific causal assumptions.

Searching arXiv for recent and relevant papers on future causal completion, causal completion, and future counterfactual forecasting. Future Causal Completion denotes a family of problems and formalisms concerned with completing, constraining, or forecasting causally admissible futures from partial present or past information. The term is used in several technically distinct literatures. In geometric and Lorentzian settings, it refers to the attachment or characterization of future ideal points by means of causal structure, typically via indecomposable past sets or related boundary constructions (Smaï, 2023, Hau et al., 2022, Müller, 2019, Hau et al., 29 Sep 2025). In machine learning and generative modeling, it refers to restricting generated future states to those that are causally reachable from observed states, for example by light-cone constraints in latent space (Vlontzos et al., 2020). In causal inference, panel data, and tensor completion, it refers to recovering missing future or post-treatment counterfactual outcomes under low-rank or structured assumptions (Athey et al., 2017, Auerbach et al., 2022, Deb et al., 9 Nov 2025, Luo et al., 12 Mar 2026, Zhou et al., 17 Mar 2026). A related but distinct strand treats inaccessible future observations as privileged supervision during training for strictly causal predictors at deployment (Li et al., 1 Jul 2026). Across these domains, the unifying object is a future that is not merely unknown but only partially admissible: the task is to specify, recover, or forecast the subset of future states compatible with causal structure, causal dynamics, or causal estimands.

1. Conceptual scope and major meanings

The phrase has no single universal definition, but the available literature supports a stable taxonomy. One meaning is geometric future completion, where a spacetime is enlarged by adding future ideal points derived intrinsically from causal relations rather than from conformal compactification (Smaï, 2023, Müller, 2019). Another is causally constrained future generation, where a model generates future frames or states only from regions of latent space deemed causally reachable from observations (Vlontzos et al., 2020). A third is counterfactual future completion in causal inference, where the missing future untreated or treated outcomes required for causal effect estimation are imputed under structured assumptions such as low rank, tensor factorization, or latent dynamics (Athey et al., 2017, Auerbach et al., 2022, Deb et al., 9 Nov 2025). A fourth is future-assisted causal learning, where future information appears only during training and is removed at inference, so that the deployed system remains strictly causal (Li et al., 1 Jul 2026).

These meanings are connected by a common asymmetry between observed histories and unobserved futures. In each case, the future is not treated as an arbitrary continuation. The admissible continuation is constrained either by chronology and causality, by a latent propagation law, by the structure of potential outcomes, or by a train/test separation that prohibits future access at deployment. This suggests a useful unifying description: future causal completion is the problem of reconstructing or characterizing future states under explicit causal admissibility conditions. That description is interpretive, but it matches the formal role played by future cones, TIPs, g-computation formulas, low-rank panels, and dynamic factor forecasts across the cited work (Vlontzos et al., 2020, Smaï, 2023, Forastiere et al., 2024, Deb et al., 9 Nov 2025).

A common misconception is that all such work concerns future prediction in the same sense. The literature distinguishes at least three different targets. In Lorentzian geometry, the target is a completed causal boundary rather than a forecasted trajectory (Smaï, 2023, Hau et al., 2022). In generative modeling, the target is a plausible future state or frame constrained by geometry, not necessarily the true realized future (Vlontzos et al., 2020). In causal inference, the target is typically a future or post-treatment counterfactual potential outcome, or an average causal contrast built from such outcomes, rather than a physically evolving state (Athey et al., 2017, Forastiere et al., 2024).

2. Geometric and Lorentzian formulations

In the Geroch–Kronheimer–Penrose tradition, future causal completion is built from indecomposable past sets. For a globally hyperbolic spacetime $M$, the future completion $\hat M$ is the set of IPs, decomposing into proper indecomposable past sets $I^-(p)$ for ordinary points and terminal indecomposable past sets $I^-(\gamma)$ for future-inextendible timelike or causal curves (Smaï, 2023, Hau et al., 2022). In this framework, future ideal points are TIPs, and the future causal boundary consists precisely of these TIPs (Smaï, 2023).

A major structural result is available in the simply connected or developable, globally hyperbolic, conformally flat setting without conjugate points. There, the causal completion is a topological manifold with boundary homeomorphic to $S\times [0,1]$, where $S$ is a Cauchy hypersurface (Smaï, 2023). In an enveloping-space representation

$$\Omega=\{(x,t)\in U\times \mathbb{R}: f^-(x)<t<f^+(x)\},$$

the future causal boundary is homeomorphic to the graph of $f^+$, with each future ideal point identified with $I^-(p)\cap \Omega$ for a unique $p\in \operatorname{graph}(f^+)$ (Smaï, 2023). The graphs of $\hat M$0 are achronal, and each future-inextendible timelike curve meets the future graph exactly once (Smaï, 2023). This yields an unusually concrete future completion, replacing abstract equivalence classes by explicit future endpoints.

A related line of work studies the topology placed on the future completion. On $\hat M$1, two topologies are frequently compared: the chronological topology and the stronger topology $\hat M$2. The latter is metrizable and, in the causally continuous setting, is the coarsest causally continuous topology (Müller, 2019). The difference in convergence between the two topologies is characterized by the fact that $\hat M$3-limits in the larger past-set space may be decomposable, whereas the weaker topology on $\hat M$4 only records maximal indecomposable components (Müller, 2019). This makes $\hat M$5 especially relevant when one wants a metric-compatible completion rather than only an order-theoretic one.

The synthetic extension of this program equips future causal completions with the structure of Lorentzian pre-length spaces. For a globally hyperbolic spacetime, the future completion $\hat M$6 can be endowed with a metric $\hat M$7, a chronological relation $\hat M$8, a causal order $\hat M$9 given by inclusion of indecomposable past sets, and a time separation $I^-(p)$0, yielding $I^-(p)$1 as a Lorentzian pre-length space (Hau et al., 2022). For globally hyperbolic generalized Robertson–Walker spacetimes, this synthetic picture can be sharpened further: the future causal completion is a globally hyperbolic Lorentzian pre-length space provided the chronological topology is Hausdorff in the infinite-integral regime, and unconditionally in the finite-integral regime (Hau et al., 29 Sep 2025). In the finite-integral case, the future boundary behaves like a terminal slice $I^-(p)$2; in the infinite-integral case, it becomes a null-cone-type boundary governed by Busemann data (Hau et al., 29 Sep 2025).

Low-regularity geometry reveals that even the notion of chronological future can become unstable. For merely continuous Lorentzian metrics, the chronological future defined using locally Lipschitz curves may fail to be open, may differ from the one defined using piecewise $I^-(p)$3 timelike curves, and may exhibit interior or exterior bubbling (Grant et al., 2019). This matters directly for future completion because classical constructions of future sets and ideal endpoints assume openness, push-up, and agreement across curve classes. The low-regularity results therefore identify conditions under which future causal completion remains canonical and conditions under which the very notion of “future set” becomes ambiguous (Grant et al., 2019).

3. Causally admissible future generation in latent space

A distinct but conceptually parallel formulation appears in latent generative modeling. “Causal future prediction in a Minkowski space-time” treats future prediction as latent-space continuation under a Lorentzian geometry (Vlontzos et al., 2020). An encoded sample is interpreted as an event $I^-(p)$4 in a $I^-(p)$5-dimensional Minkowski latent space with metric

$I^-(p)$6

or its $I^-(p)$7-dimensional analogue (Vlontzos et al., 2020). The Minkowski squared interval between $I^-(p)$8 and $I^-(p)$9 is

$I^-(\gamma)$0

with timelike, spacelike, and lightlike separations defined by the sign of this quantity (Vlontzos et al., 2020).

In this framework, a candidate future is causally admissible when it lies in the future light cone of the observed state. The implied admissibility condition is

$I^-(\gamma)$1

with equality on the light-cone boundary (Vlontzos et al., 2020). Proper time along a path is real only when the path remains timelike or lightlike, yielding the latent finite-propagation constraint $I^-(\gamma)$2 (Vlontzos et al., 2020). This turns future completion into a constrained latent search problem rather than unconstrained extrapolation.

For multiple observations $I^-(\gamma)$3, each induces a future light cone $I^-(\gamma)$4, and the candidate future set is their intersection

$I^-(\gamma)$5

At target time $I^-(\gamma)$6, valid completions lie in

$I^-(\gamma)$7

This intersection formalizes the idea that a future must be reachable from all observed past states simultaneously (Vlontzos et al., 2020). The method is architecture-agnostic in the sense that the causal mechanism is an inference-time wrapper around a pretrained encoder-decoder latent model. In the reported experiments, the base latent model is the Poincaré VAE of Mathieu et al., mapped to Minkowski coordinates through an orthochronous diffeomorphism, with a $I^-(\gamma)$8-dimensional Minkowski embedding reported as best (Vlontzos et al., 2020).

The empirical demonstrations are proof-of-concept rather than benchmark-heavy. On a custom Moving MNIST setup, 100,000 latent samples drawn from a wrapped Gaussian and filtered by a single cone yielded acceptance rates of $I^-(\gamma)$9 samples at $S\times [0,1]$0, $S\times [0,1]$1 at $S\times [0,1]$2, and $S\times [0,1]$3 at $S\times [0,1]$4, indicating that shorter horizons constrain the future more strongly (Vlontzos et al., 2020). Intersecting cones from multiple observations produces multiple plausible futures rather than a unique deterministic next frame (Vlontzos et al., 2020). A failure mode is digit transmutation, such as a $S\times [0,1]$5 becoming a $S\times [0,1]$6, which remains causally admissible in the latent geometry but semantically undesirable (Vlontzos et al., 2020). This clarifies the scope of the guarantees: they are geometric causal guarantees, not semantic identity guarantees or structural-causal-model guarantees (Vlontzos et al., 2020).

4. Future completion in causal inference, panels, and tensors

In causal inference, future causal completion typically means imputing missing potential outcomes that lie in the future or post-treatment region of a panel or tensor. In matrix completion for panel data, the untreated potential-outcome matrix $S\times [0,1]$7 is partially observed on untreated unit-time cells and missing on treated cells. The goal is to reconstruct those missing untreated outcomes so that treatment effects can be estimated (Athey et al., 2017). The proposed estimator is nuclear-norm-regularized matrix completion with unregularized unit and time fixed effects: $S\times [0,1]$8 with completed untreated counterfactuals

$S\times [0,1]$9

for treated cells (Athey et al., 2017). This reframes future counterfactual completion as structured missing-data recovery.

Tensor completion generalizes the same logic to multivariate longitudinal outcomes. The data are arranged into a tensor $S$0, where the third mode indexes outcomes (Auerbach et al., 2022). The convex estimator penalizes nuclear norms of unfoldings, and in the application the first unfolding is used: $S$1 Under the stated low-rank shared-factor conditions, the non-asymptotic error rate improves from roughly $S$2 in the single-outcome matrix case to roughly $S$3 up to logarithmic factors when $S$4 (Auerbach et al., 2022). This shows that auxiliary outcomes can improve future or post-treatment counterfactual completion by enlarging the effective information content.

Several later extensions make the completion problem more structured. Spatial causal tensor completion jointly models multiple exposures and multiple outcomes while approximating unmeasured spatial confounders with graph-Laplacian eigenvectors (Zhou et al., 17 Mar 2026). The potential outcomes tensor $S$5 is given a low-rank Tucker representation

$S$6

with the spatial factor decomposed as $S$7, where $S$8 are low-frequency graph Laplacian eigenvectors (Zhou et al., 17 Mar 2026). The method combines tensor completion, generalized propensity weighting, and spectral adjustment to address unmeasured spatial confounding (Zhou et al., 17 Mar 2026). This suggests that future causal completion is increasingly treated as a structured representation problem rather than a purely algebraic imputation problem.

When treatments are multiple and sparse, Mixed Synthetic Nearest Neighbors extends causal matrix completion under MNAR observation by borrowing across treatment levels (Luo et al., 12 Mar 2026). The key assumption is that latent row factors are shared across treatments, which makes the reconstruction coefficients treatment-invariant even when the final target entry remains treatment-specific (Luo et al., 12 Mar 2026). Under MCAR, the expected number of usable anchors can improve by a factor

$S$9

relative to treatment-isolated SNN, which is especially valuable when the target treatment level is rare (Luo et al., 12 Mar 2026).

A separate line addresses genuinely future horizons rather than only missing post-treatment cells. FOCUS forecasts future counterfactuals in panel data by first estimating a treatment-specific low-rank factor model and then forecasting the latent factors forward under a stable VAR(1): $$\Omega=\{(x,t)\in U\times \mathbb{R}: f^-(x)<t<f^+(x)\},$$0 For a fixed treatment state $$\Omega=\{(x,t)\in U\times \mathbb{R}: f^-(x)<t<f^+(x)\},$$1, the future forecast target is

$$\Omega=\{(x,t)\in U\times \mathbb{R}: f^-(x)<t<f^+(x)\},$$2

estimated by

$$\Omega=\{(x,t)\in U\times \mathbb{R}: f^-(x)<t<f^+(x)\},$$3

(Deb et al., 9 Nov 2025). Under the stated conditions, the error obeys

$$\Omega=\{(x,t)\in U\times \mathbb{R}: f^-(x)<t<f^+(x)\},$$4

with asymptotic normality also established (Deb et al., 9 Nov 2025). This is one of the clearest direct formulations of future causal completion as future counterfactual forecasting rather than contemporaneous completion.

5. Forecasting future interventions and future-privileged supervision

A formal causal-inference treatment of future interventions asks when effects estimated in the past can be transported to future implementation periods. The relevant estimand is the future average treatment effect on a designated future target set: $$\Omega=\{(x,t)\in U\times \mathbb{R}: f^-(x)<t<f^+(x)\},$$5 Identification requires not only sequential randomization in the observed data but also temporal transportability assumptions that equate the distribution of potential outcomes across observed and future windows conditional on relevant treatment histories and measured modifiers (Forastiere et al., 2024). The core transported conditional mean equality is

$$\Omega=\{(x,t)\in U\times \mathbb{R}: f^-(x)<t<f^+(x)\},$$6

which allows future causal effects to be reconstructed from forecast or scenario-specified future pre-treatment modifier histories $$\Omega=\{(x,t)\in U\times \mathbb{R}: f^-(x)<t<f^+(x)\},$$7 (Forastiere et al., 2024).

A key reduction result shows that, under the paper’s temporal transportability assumptions, it suffices to condition on pre-treatment modifiers before the carry-over period: $$\Omega=\{(x,t)\in U\times \mathbb{R}: f^-(x)<t<f^+(x)\},$$8 (Forastiere et al., 2024). This is technically important because it reduces future causal completion to two subproblems: forecasting the relevant future pre-treatment context and transporting the conditional causal response surface from the past (Forastiere et al., 2024). A plausible implication is that future causal completion in longitudinal policy settings is as much a transportability problem as a missing-data problem.

A different use of the future appears in future-privileged supervision for strictly causal models. In egocentric gaze estimation, ECOGaze introduces a future-aware branch available only during training and a strictly causal branch used at test time (Li et al., 1 Jul 2026). The deployed model satisfies

$$\Omega=\{(x,t)\in U\times \mathbb{R}: f^-(x)<t<f^+(x)\},$$9

while during training the future-aware branch can attend to $f^+$0 for a tunable horizon $f^+$1 (Li et al., 1 Jul 2026). The supervision combines a ground-truth KL term and a future-privileged supervision KL term with stop-gradient on the future branch (Li et al., 1 Jul 2026). Across EGTEA Gaze+ and Ego4D, the gains are non-monotonic in $f^+$2: optimal performance occurs around $f^+$3, corresponding to roughly $f^+$4–$f^+$5 seconds on EGTEA and $f^+$6, about $f^+$7 seconds, on Ego4D (Li et al., 1 Jul 2026). This is not future completion at inference time, but it shows that near-future information can sharpen the representation of the present causal state during training (Li et al., 1 Jul 2026).

6. Limits, controversies, and cross-domain synthesis

A first recurring limitation is that many “causal” guarantees are not causal in the structural-intervention sense. The Minkowski latent framework provides geometric reachability guarantees, not identification of intervention effects from observational data (Vlontzos et al., 2020). Likewise, future-privileged supervision improves a causal predictor but does not produce a future completion mechanism at test time (Li et al., 1 Jul 2026). In causal inference, low-rank or tensor completion recovers counterfactuals only under latent-factor and overlap assumptions, not from causal structure alone (Athey et al., 2017, Auerbach et al., 2022).

A second limitation concerns topology and regularity. In low-regularity Lorentzian geometry, even the chronological future may fail to be open and may depend on the curve class used (Grant et al., 2019). This complicates any future completion built on future sets or boundary hypersurfaces. The literature therefore suggests that causally plain settings, Hausdorff chronological topologies, or CLT-type metric refinements are not technical conveniences but structural requirements for a robust completion theory (Grant et al., 2019, Müller, 2019, Hau et al., 29 Sep 2025).

A third limitation is semantic or domain mismatch. In latent generative prediction, a future can be causally reachable yet semantically wrong, as in digit transmutation (Vlontzos et al., 2020). In panel completion, a recovered future counterfactual may be statistically regularized but causally invalid if treatment timing depends on idiosyncratic shocks not captured by the latent structure (Athey et al., 2017). In temporal transportability, identification fails if future effect modifiers drift outside observed support or if future intervention implementation changes materially (Forastiere et al., 2024).

Across domains, however, the same abstract pattern recurs. There is always an observed prefix or past object, a latent or geometric admissible region, and a rule for restricting future candidates. In Lorentzian geometry, the admissible future is encoded by TIPs, future cones, Busemann classes, or boundary graphs (Smaï, 2023, Hau et al., 29 Sep 2025). In generative models, it is encoded by cone intersections in latent Minkowski space (Vlontzos et al., 2020). In causal inference, it is encoded by the support of observed histories, low-rank spans, treatment-specific panels, or transported g-formulas (Athey et al., 2017, Forastiere et al., 2024, Deb et al., 9 Nov 2025). This suggests that future causal completion is best understood not as one method class but as a recurring formal problem: specifying which futures remain admissible once causal structure is imposed.

A plausible implication is that future work will continue to hybridize these ingredients. The existing literature already points toward structure-aware causal completion: tensor methods augmented with spatial spectra (Zhou et al., 17 Mar 2026), multi-treatment local completion under MNAR (Luo et al., 12 Mar 2026), and dynamic factor forecasting beyond static low-rank imputation (Deb et al., 9 Nov 2025). In geometric settings, synthetic Lorentzian structures on causal completions suggest a route to completion theory beyond smooth manifolds, though low-regularity pathologies remain a constraint (Hau et al., 2022, Grant et al., 2019). The term “Future Causal Completion” is therefore most useful when understood as an umbrella for methods that turn unconstrained extrapolation into constrained causal extension, while keeping clear which notion of causality—geometric, topological, generative, potential-outcome, or strictly non-anticipatory—is actually in force.