acceptodds
Under review as a conference paper at ICLR 2027

DeCaf: Decomposed Subspaces for Longitudinal Treatment Effect Estimation

Abstract

Estimating how a patient’s outcomes might change under different treatment plans over time is a central need in personalized medicine. Longitudinal observational data, especially from electronic health records (EHR), can help address this by capturing how clinical decisions and patient states evolve across repeated visits. In real-world settings, however, treatment practices and data collection processes often vary across environments (e.g., across hospitals, cohorts or over time). As a result, signals that reflect the patient's underlying clinical state and treatment response, i.e., the factors through which interventions causally influence outcomes, can be entangled with environment-specific artifacts: systematic, setting-induced patterns (e.g., differences in treatment protocols and documentation practices), making multi-step counterfactual forecasting unreliable. To address this challenge, we propose DeCaf, a framework that learns a structured decomposed representation by separating a patient's longitudinal history into two latent subspaces: a outcome relevant subspace that encodes stable, intervention-relevant factors, and an environment/nuisance subspace that captures variation associated with treatment practice, cohort, and measurement processes. We then combine this representation with a trajectory generator to produce counterfactual trajectories under alternative treatment sequences, enabling coherent and stable forecasts over time.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
causal inference

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SDC6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies treatment effect estimation and multi-step counterfactual prediction under policy and environment shifts. It proposes DECAF, a two-stage framework that first separates patient history into outcome-related and environment-related subspaces, then freezes the encoder and uses conditional diffusion with contrastive regularization to predict future states in the outcome-related subspace. Experiments on a tumor growth simulator, semi-synthetic MIMIC-III, and real MIMIC-IV show lower average errors than baselines in most cross-environment prediction tasks.

Critical strengths

The paper has a clear motivation and combines environment separation with multi-step counterfactual prediction. It separates outcome-related and environment-related information, then predicts future states in the outcome-related subspace. This design may reduce the accumulation of environment-specific bias over multiple steps.

The experiments are systematic, with several baselines on simulated, semi-synthetic, and real clinical data. Ablation studies examine different generative models and loss terms. The authors also clearly state the theoretical assumptions and the lack of counterfactual ground truth in real data, helping readers understand the limits of the results.

Critical weaknesses

The first stage uses outcome supervision only for one-step prediction, then freezes the encoder. It is unclear whether the learned representation retains the information needed for longer-term predictions. Since the environment and patient condition may be related, removing environment information could also remove useful clinical information that one-step supervision may not preserve.

Evidence for multi-step counterfactual prediction remains limited, mainly coming from the tumor simulator. Semi-synthetic MIMIC-III evaluates only one-step treatment effects, while real MIMIC-IV evaluates only factual prediction. Some long-horizon gains are also small compared with the variation across random seeds, so their stability needs further validation.

Questions for authors
  1. The first stage uses outcome supervision only for one-step prediction. What evidence shows that the frozen representation retains the information needed for multi-step prediction?

  2. Could the model match positive pairs mainly through their shared environment vector, without sufficiently constraining the generated counterfactual states?

  3. How do DECAF and the baselines differ in their use of environment labels and tuning budgets? How can DECAF be used when training labels for environments are unavailable, or when the environment changes within a patient’s trajectory?

  4. Which data-generating mechanisms and parameters differ across the eight semi-synthetic MIMIC-III environments? Does this setup satisfy the theory’s assumptions about stable mechanisms across environments?

6 comments

We thank the reviewer for the careful and constructive review.


Question 1. Information retained for multi-step prediction

Response: Thank you for raising this concern. Multi-step prediction (Eq. 3) applies the one-step transition KK under each treatment and then the one-step head mt,a(ztc)m_{t,a}(\mathbf{z}^c_t) of Theorem 1:

E[yt+τaˉ∣Ht]=∫mt+τ−1,at+τ−1(zt+τ−1)∏s=tt+τ−2Ks,as(dzs+1∣zs),zt=ztc\displaystyle \mathbb{E}[y_{t+\tau}^{\bar a}\mid\mathbb H_t]=\int m_{t+\tau-1,a_{t+\tau-1}}(z_{t+\tau-1})\prod_{s=t}^{t+\tau-2}K_{s,a_s}(dz_{s+1}\mid z_s),\qquad z_t=\mathbf{z}^c_t

This needs two conditions: a correct one-step head, and ztc\mathbf{z}^c_t carrying everything in Ht\mathbb H_t that matters for the next state (transition sufficiency, Eq. 2). One-step supervision does not guarantee the second, so Section 5 states it separately. When both hold, with head error ϵy\epsilon_y, head bound BB and per-step transition error ϵK\epsilon_K (total variation), the rollout error satisfies

∣μτ−μ^τ∣≤ϵy+2B(τ−1) ϵK\displaystyle \lvert\mu_\tau-\hat\mu_\tau\rvert\le\epsilon_y+2B(\tau-1)\,\epsilon_K

so it grows at most linearly with the horizon; we will add the bound and proof to App. H. Whether training meets these conditions is an empirical question, which the experiments below address.

First, we examined whether one-step supervision is the limiting factor. If one-step supervision discarded information needed for rollout, multi-step targets should improve the rollout. We tested two variants under the Table 10 protocol (Tumor Growth nRMSE, four-policy mean over three seeds; Δ\Delta: per-seed difference to DeCaf, mean ± SD): in (i), Stage 1 gets extra factual heads that predict yt+τy_{t+\tau} for τ≤5\tau\le5 under the realized treatments, encoder frozen; in (ii), the encoder is fine-tuned with Stage 2, keeping the Stage-1 outcome loss.

Stage-1 supervisionτ=1\tau=1τ=3\tau=3τ=5\tau=5Δ\Delta at τ=5\tau=5
DeCaf (one-step, frozen)1.3711.5251.880—
(i) + multi-step heads, frozen1.4131.4971.833−0.047 ± 0.025
(ii) fine-tuned with Stage 21.3541.5771.958+0.078 ± 0.031

Fine-tuning the encoder with Stage 2 (without the separation losses) helps a little at one step but hurts at three and five. Multi-step heads cost a little at one step and help a little at five. The frozen one-step encoder retains most of what the rollout needs; we will include the multi-step-head variant in the revision.

Second, we examined how much the history knows about the future beyond ztc\mathbf{z}^c_t. If ztc\mathbf{z}^c_t carries everything in Ht\mathbb H_t that matters for the future, the rollout error cannot be predicted from Ht\mathbb H_t any better than from ztc\mathbf{z}^c_t itself. For each τ\tau we predict the rollout error yt+τ−μ^τ(ztc)y_{t+\tau}-\hat\mu_\tau(\mathbf{z}^c_t) on factual trajectories from ztc\mathbf{z}^c_t, from Ht\mathbb H_t, and from zte\mathbf{z}^e_t (R2R^2 of the residual, means over three seeds). The zc\mathbf{z}^c row measures fitting error of the model itself; the excess of the Ht\mathbb H_t row over it is the information in the history that ztc\mathbf{z}^c_t misses; the ze\mathbf{z}^e row answers the second part of W1.

Residual predicted fromτ=1\tau=1τ=2\tau=2τ=3\tau=3τ=4\tau=4τ=5\tau=5
zc\mathbf{z}^c, Tumor Growth0.030.040.050.070.08
Full history, Tumor Growth0.040.070.060.110.09
ze\mathbf{z}^e, Tumor Growth0.020.02−0.010.030.02
zc\mathbf{z}^c, MIMIC-IV0.050.120.140.180.21
Full history, MIMIC-IV0.080.170.190.220.27
ze\mathbf{z}^e, MIMIC-IV0.020.040.030.060.04

Using the full history instead of ztc\mathbf{z}^c_t improves residual R2R^2 by only 0.01–0.04 on Tumor Growth and 0.03–0.06 on MIMIC-IV. Most of the predictable residual is fitting error that ztc\mathbf{z}^c_t itself predicts, not information missing from it. The probes using zte\mathbf{z}^e_t alone reach residual R2R^2 of at most 0.03 and 0.06, respectively.

Third, the multi-step results already in the paper point the same way. zc\mathbf{z}^c keeps 97.9% of the full-history next-outcome R2R^2 on MIMIC-IV (Table 12); on Tumor Growth the four-policy mean error grows by 0.13–0.19 per step from τ=2\tau=2 to 55 without compounding (Table 11), and at τ=5\tau=5 DeCaf has the lowest or near-lowest mean in all four policies (Table 9).

In summary, the frozen encoder performs well in multi-step prediction. Adding multi-step supervision gives only a small improvement, while fine-tuning the encoder does not improve the longer-horizon results.

Question 2. Matching through the shared environment vector

Response: The generated state z~t+1c\tilde{\mathbf{z}}^c_{t+1} comes from the diffusion sampler given (ztc,at′)(\mathbf{z}^c_t,a_t'), and zte\mathbf{z}^e_t enters only the projection heads, not the sampler. If the pair loss acted only through zte\mathbf{z}^e_t, removing the loss would not change the rollout, but it does (Table 10: τ=3\tau=3 1.525 to 1.923, τ=5\tau=5 1.880 to 1.997).

We first retrained Stage 2 without zte\mathbf{z}^e_t in either projection head (variant B in our reply to Reviewer oC7g). The error increased from 1.525 to 1.594 at τ=3\tau=3 and from 1.880 to 1.948 at τ=5\tau=5. We also tested the full model by shuffling the generated state, zte\mathbf{z}^e_t, or both across patients while keeping the projection heads fixed. The retrieval accuracies are shown below:

Generated sideRetrieval accuracy (%)
Original54.3
Generated state permuted7.8
ze\mathbf{z}^e permuted35.6
Both permuted1.6

Most of the matching comes from the generated state; zte\mathbf{z}^e_t acts as context.


Question 3. Environment labels; tuning; missing or changing labels

Response: All methods are trained in the same setting. DeCaf uses the environment label only in the Stage-1 losses (Eq. 7), one per trajectory (policy in Tumor Growth, SCM environment in MIMIC-III, care unit in MIMIC-IV); Stage 2 and inference do not use it. As this input is specific to DeCaf, we checked whether it explains the improvement: variant A (reply to Reviewer oC7g) uses the same label through the adversarial objective, without the ze\mathbf{z}^e branch, and is worse at every horizon (1.492 / 1.727 / 2.083 vs 1.371 / 1.525 / 1.880). The label alone does not explain the improvement.

On tuning efforts, we used the same configuration for DeCaf across all three benchmarks, without dataset-specific tuning (App. A.4), the baselines use the defaults of Melnychuk et al. (2022) (App. F), and every method is early-stopped on the same source validation split, with no OOD data used for selection (App. D.1).

We use one environment label per trajectory during training, such as the care unit in our MIMIC-IV experiments. If labels are unavailable, one option would be to infer environments from the data. If the environment changes within a trajectory, the classifiers in Eq. (7) could use a different label at each time step. Both settings are beyond the scope of this paper, but we appreciate the reviewer raising them. They are important in real world settings, for transfer across hospitals and for patients who move between units. We will discuss them in Sec. 7 as limitations and future work.


Question 4. MIMIC-III environments; stable-mechanism assumption

Response: Thank you for the careful reading. The eight environments use the same SCM equations and observed covariates; they differ in patients and in the parameter values θe\theta_e (autoregressive and covariate weights, treatment-effect intercept and slope, noise scale).

Theorem 5 states a sufficient condition for a shared predictor on zc\mathbf{z}^c to be invariant across environments. In the Tumor Growth dataset, environments differ only in the treatment-assignment policy, while the simulator's response mechanism remains fixed (App. D.2). The semi-synthetic MIMIC-III dataset tests a different setting: the held-out environments 4–7 use different SCM parameter values, introducing changes in the outcome mechanism itself. DeCaf still has the lowest OOD √PEHE there (0.314 vs 0.434, Table 3). DeCaf improves under both shifts: a policy shift with a fixed response mechanism, and a change in the outcome mechanism itself.


Weakness 1. One-step supervision; loss of information

Please see our response to Q1.


Weakness 2. Multi-step evidence

Response: We extended the semi-synthetic MIMIC-III benchmark to treatment plans of length τ=1,…,5\tau=1,\dots,5 under the counterfactual protocol (App. D.2): the SCM is iterated under the plan with the observed covariates held fixed, Stage 2 is trained on the source environments, and we report nRMSEcf(τ)\mathrm{nRMSE}_{\mathrm{cf}}(\tau) (mean ± SD, three seeds).

nRMSEcf(τ)\mathrm{nRMSE}_{\mathrm{cf}}(\tau), MIMIC-IIIτ=1\tau=1τ=3\tau=3τ=5\tau=5
CDM0.452 ± 0.0230.729 ± 0.0351.495 ± 0.052
DR-CRN0.535 ± 0.0240.795 ± 0.0461.764 ± 0.083
G-Transformer0.492 ± 0.0620.824 ± 0.0581.964 ± 0.091
DeCaf0.394 ± 0.0310.712 ± 0.0411.376 ± 0.042

DeCaf has the lowest mean error at all three reported horizons. The difference from CDM is small at τ=3\tau=3 and larger at τ=5\tau=5, where DeCaf reduces the mean error by 8.0%. These results provide additional multi-step counterfactual evidence beyond the tumor simulator.

We thank the reviewer and would be glad to discuss any remaining concerns

Dear Reviewer,

Thank you again for taking the time to review our paper and share your feedback. We have included additional experiments and analyses in our response. When you have a chance, we would appreciate your thoughts and would be happy to discuss any remaining questions.

Reviewer_LTALOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose a two-stage framework (DeCaf) for estimating individualized treatment effects, aiming to separate out causal and environmental variables into two subspaces. The proposed framework is first trained to learn the latent space representation for the environment for pre-treatment history. Then, the causal subspace is trained using diffusion objectives. Finally, the treatment effects are then estimated with generation.

Results: The authors compare their model against 9 different other causal learning approaches in three different environments 1) the Tumor Growth simulator 2) semi-synthetic MIMIC-III and 3) a MIMIC-IV ventilation cohort.

Discussion: The authors find noticeable improvements in all three of the tasks presented, though some have some natural overlap with other baselines. The authors measure their performance in a number of conventional treatment effect measures, as well as some derivative ones.

Critical strengths

The idea of separating out the latent space into an environmental effect and background effect for counterfactual generation seems to be novel.

The authors beat a majority of the baselines in all settings tested in terms of the selected metrics for each task.

Critical weaknesses

I appreciate the authors appropriately disclosed their use of LLM writing, however it still suffers from a decent amount of mannered prose and some interwoven thoughts, and numbers displayed all throughout the paper. I think the writing is generally better for LLM writing, but it needs some improvements.

What is the backbone for the patient history representation? That seems to have been glossed over. Is the backbone the same for each of these models?

The authors use three random seeds to generate estimates of variance, which is good. Some of these estimates of variance do overlap with other models, but that is not that surprising given the nature of treatment effect estimation.

Questions for authors

I don’t see code referenced anywhere, at least in the main text. Will this be provided now or at publication?

2 comments

We thank the reviewer for the comments on the writing and the experiments.


Question 1. Code

Response: The code is available at https://anonymous.4open.science/r/decaf-code-D846/.


Weakness 1. Writing

Response: We appreciate the feedback and will work on making the writing clearer and more concise.


Weakness 2. Backbone

Response: Thank you for this question. DeCaf uses a temporal backbone: a standard two-layer GRU with hidden size 128, followed by two linear projections to the branches (dc,de)=(64,32)(d_c,d_e)=(64,32) (App. A.4); the denoiser is a three-layer MLP with hidden size 512. Every DeCaf variant in Tables 2, 10, 11 and 13 uses this same backbone. The baselines use their own published backbones (App. F), several of them more complex: LSTMs (CRN, RMSN), Transformers (Causal Transformer, G-Transformer) and neural CDEs (TE-CDE), all on the same splits and protocol.

We kept the backbone simple, because the contribution of DeCaf is the decomposition, not the sequence model: it separates the part of the history that drives treatment response (zc\mathbf{z}^c) from the environment-related part (ze\mathbf{z}^e) and rolls out only in zc\mathbf{z}^c. A plain GRU also makes the comparison against the Transformer and CDE baselines conservative. Table 11 isolates this directly. The unsplit row uses the same GRU without the decomposition and rolls out in its single 128-d state; averaged over the four policies it scores 1.475 at τ=1\tau=1 and 2.175 at τ=5\tau=5, in the range of the stronger baselines in Table 9 (CTD-NKO 1.558 / 1.991, CDM 1.665 / 2.381, CRN 1.728 / 2.186, CT 1.861 / 2.349). With the same GRU, rolling out in zc\mathbf{z}^c gives 1.371 / 1.880. The improvement comes from the decomposition.


Weakness 3. Seeds and variance

Response: We use three random seeds (10, 101, 1010) and report mean ± SD across them. Some comparisons are close; on these benchmarks the differences between methods are small in absolute terms, so overlapping intervals are expected. For this reason we also report relative gains, which are substantial on the main benchmarks: on MIMIC-III, OOD PEHE\sqrt{\mathrm{PEHE}} drops from 0.434 (CDM) to 0.314 and OOD ATE error from 0.213 (DR-CRN) to 0.157 (Table 3), and on Tumor Growth DeCaf has the lowest mean in 13 of the 15 OOD settings.

We hope this answers the reviewer's questions, and we are happy to clarify anything further.

Dear Reviewer,

Thank you again for recognizing the novelty of our approach and your suggestions. We have clarified the backbone architecture and shared the code in our response. We would appreciate any further feedback you may have.

Reviewer_oC7gOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes DECAF for longitudinal counterfactual prediction under environment and policy shift. It learns an outcome-relevant latent subspace and an environment-related subspace, then uses treatment-conditional latent diffusion for multi-step rollout. Tumor Growth and semi-synthetic MIMIC-III provide oracle counterfactuals, and MIMIC-IV evaluates factual SpO2 forecasting. The results show the model outperforms baselines.

Critical strengths
  1. The paper proposed a decomposed longitudinal representation learning objective that separates outcome-relevant signals from environment-related variation and specifies the identification and sufficiency conditions required for treatment-effect estimation under shift.

  2. The paper proposed a treatment-conditional diffusion model that performs multi-step counterfactual forecasting in the outcome-relevant latent space while keeping the environment branch fixed.

  3. The authors conducted extensive experiments on both real-world and synthetic datasets, covering policy shift, environment shift, and factual clinical forecasting across Tumor Growth, semi-synthetic MIMIC-III, and MIMIC-IV. The reported results show that DECAF generally outperforms the baselines, supporting the effectiveness of the proposed method.

Critical weaknesses
  1. The role of the environment branch is not isolated by the current ablations. Table 2 compares rollout in z_c, an "unsplit z," and z_e, but the construction and parameterization of the unsplit representation are not specified. Table 10 removes the environment-separation loss, not the z_e variable itself. No matched experiment removes z_e while holding the encoder, diffusion model, parameter budget, and training schedule fixed. It therefore remains unclear whether z_e improves accuracy or mainly acts as a training signal.
  2. Code, split-generation scripts, and per-seed results are not available at the review stage.
Questions for authors

Can the authors provide a variable-level diagnostic showing which measured variables are represented in z_c versus z_e? Feature attribution, supervised probes for individual variables, and reconstruction tests could make the proposed separation directly testable and clinically interpretable.

4 comments

We thank the reviewer for the constructive comments.


Question 1. Variable-level diagnostic of z^c vs z^e

Response: Yes; thank you for the suggestion. On Tumor Growth the simulator state and the environment are known, so it is clear which branch should carry each variable. With the encoder frozen, we fit a small MLP of the same architecture and tuning budget from ztc\mathbf{z}^c_t, from zte\mathbf{z}^e_t and from their concatenation to predict each variable below (source-patient splits, three seeds); as the patient parameters and the policy are only partly identifiable from a finite history, a full-history predictor is the reference.

Target variableMetriczc\mathbf{z}^cze\mathbf{z}^e[zc;ze][\mathbf{z}^c;\mathbf{z}^e]Full history (ref.)
Tumor volumeR2R^20.900.560.930.96
Residual chemotherapy exposureR2R^20.680.280.730.78
Response parameter ρ\rhoR2R^20.350.140.390.42
Chemo sensitivity βc\beta_cR2R^20.310.150.340.39
Radio sensitivity αr\alpha_rR2R^20.250.120.290.31
Radio sensitivity βr\beta_rR2R^20.170.110.210.27
Recent mean diameter (policy input)R2R^20.840.720.890.91
Assignment probability (patient's policy)R2R^20.540.630.790.82
Environment label (policy γ\gamma; chance 1/3)bal. acc.0.430.780.810.86

The first six variables determine the treatment response and come mainly from zc\mathbf{z}^c, close to the full-history reference (lower for the radiotherapy parameters, which are harder to identify from a history), and much less from ze\mathbf{z}^e. The last two define the environment: the label is recovered from ze\mathbf{z}^e at 0.78 (reference 0.86) but from zc\mathbf{z}^c only at 0.43 (chance 0.33); the assignment probability is recovered better from ze\mathbf{z}^e (0.63 vs 0.54) and only the concatenation reaches the reference, as it depends on both state and policy. Each branch mainly carries its intended group; some overlap is expected since they share the GRU state.


Weakness 1. Isolating the environment branch; the unsplit representation

Response: Thank you for suggesting this comparison. ze\mathbf{z}^e is not used at inference; the rollout uses only ztc\mathbf{z}^c_t and the treatment plan (Alg. 3). It enters training in two places: the Stage-1 environment classifier on ze\mathbf{z}^e (Eq. 7), which keeps environment information in the GRU state that the adversary on zc\mathbf{z}^c would otherwise remove, and the Stage-2 pair projections (Eq. 15).

The unsplit model (App. E.2) is a Stage 1 with the same GRU and a single 128-d representation, trained with the outcome loss only, followed by the same Stage 2; it removes the decomposition training as a whole (five-step error 2.175 vs 1.880, Table 2).

To isolate ze\mathbf{z}^e we ran the matched experiment the reviewer describes (GRU, diffusion model, splits, seeds and schedule fixed as in Table 10; both stages retrained): A removes ze\mathbf{z}^e (branch, classifier and pair-loss input), keeping the adversarial, outcome and sufficiency losses on zc\mathbf{z}^c; A′ is A with the outcome and projection heads widened to match the parameter count (dcd_c stays 64); B keeps ze\mathbf{z}^e in Stage 1 but drops it from the Stage-2 projections. Last column: per-seed difference to Full at τ=5\tau=5, mean ± SD.

Variantze\mathbf{z}^e in Stage 1ze\mathbf{z}^e in pair lossParams (S1/S2)τ=1\tau=1τ=3\tau=3τ=5\tau=5Δ\Delta at τ=5\tau=5
Full DeCafyesyes0.20M / 0.81M1.3711.5251.880—
Unsplit (Table 11)nono0.22M / 0.86M1.4751.6182.175+0.295 ± 0.047
A: w/o ze\mathbf{z}^enono0.195M / 0.80M1.4921.7272.083+0.203 ± 0.034
A′: A, capacity-matchednono0.20M / 0.81M1.4511.6632.025+0.145 ± 0.052
B: ze\mathbf{z}^e in Stage 1 onlyyesno0.20M / 0.81M1.3711.5941.948+0.068 ± 0.021

B changes only Stage 2, so its one-step error equals the full model's; removing the branch changes the parameter count by only about 2%, so A is close to capacity-matched. Removing ze\mathbf{z}^e raises error at every horizon; A′ recovers only part of it, so the branch matters beyond its parameter count. B lies between A′ and the full model at longer horizons, so both roles contribute: the Stage-1 classifier (A′ vs Full at τ=1\tau=1) and the pair-loss input (B vs Full at longer horizons; see Reviewer SDC6, Q2).


Weakness 2. Code availability

Response: The code is available at https://anonymous.4open.science/r/decaf-code-D846/.

Dear Reviewer,

Thank you again for your helpful suggestions and positive feedback. We have included additional experiments and analyses in our response and would appreciate any further comments you may have.

Thank the authors for the response. My concerns are addressed. I will keep my positive rating score.

We’re glad they addressed your concerns, and we appreciate your feedback and positive assessment of our work! If there are specific remaining concerns that we could address during the discussion period, we would be happy to provide further clarification or results.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.