Proximal Inference: Methods & Applications

• Proximal inference is a family of inferential constructions that uses auxiliary objects such as bridge functions and proximal operators to recover target quantities hidden by unmeasured confounding or computational challenges.
• It has wide applications in causal inference, penalized estimation, and Bayesian computation, enabling robust identification and efficient estimation through innovative techniques like doubly robust estimators and Moreau–Yosida smoothing.
• Recent advancements integrate proximal methods with neural network-based optimizations and stochastic trajectory updates, enhancing performance in synthetic control, regularization, and non-smooth optimization tasks.

Searching arXiv for papers on “proximal inference” to ground the article in current literature. Proximal inference denotes a family of inferential constructions in which an auxiliary object—most commonly a bridge function, a proximal operator, or a KL-regularized projection—is used to recover a target quantity that is otherwise hidden, unstable, or computationally inconvenient. In recent arXiv literature, the term appears most prominently in causal inference with unmeasured confounding, where proxy variables identify counterfactual means through outcome and treatment bridges, but it also appears in stochastic trajectory optimization through reverse-KL and proximal-KL updates, in penalized estimation through proximal maps applied to an initial estimator, and in Bayesian computation through Moreau–Yosida smoothing (Cui et al., 2020, Yu et al., 8 May 2026, Quaini et al., 2022, Zhou et al., 2022).

1. Mathematical forms and recurrent constructions

In proximal causal inference, the central objects are bridge functions. The outcome bridge $h(w,a,x)$ satisfies

$$E\bigl[Y\mid Z,A=a,X=x\bigr] = \int h(w,a,x)\,dF(w\mid Z,A=a,X=x),$$

while the treatment bridge $q(z,a,x)$ satisfies

$$E\bigl[q(Z,a,x)\mid W,A=a,X=x\bigr] = \frac1{f(A=a\mid W,X)}.$$

These Fredholm equations of the first kind are the basic identification devices for the average treatment effect and related functionals (Cui et al., 2020).

In proximal estimation, the defining object is the proximal operator under a possibly singular inner product $J$: $\prox_{\lambda P,J}(\bar\theta) = \arg\min_{\theta\in\mathbb R^p} \Big\{ \tfrac12(\theta-\bar\theta)^\top J(\theta-\bar\theta)+\lambda P(\theta) \Big\},$ with the resulting estimator written as $\hat\theta=\prox_{\lambda P,J}(\bar\theta)$. This framework is used to analyze ridge, lasso, elastic net, group-lasso, and related penalized procedures within a single convex-analytic template (Quaini et al., 2022).

In proximal MCMC, the non-smooth regularizer $\phi$ is replaced by its Moreau–Yosida envelope

$$\phi_\gamma(x) = \min_{u\in\mathbb R^p} \Bigl\{\phi(u)+\tfrac1{2\gamma}\|x-u\|^2\Bigr\},$$

whose gradient is

$$\nabla\phi_\gamma(x) = \tfrac1\gamma\bigl(x-\mathrm{prox}_{\gamma\phi}(x)\bigr).$$

This converts a non-differentiable posterior into a smooth approximation amenable to Langevin and Hamiltonian Monte Carlo updates (Zhou et al., 2022).

In stochastic trajectory optimization, PISTO defines the iterate by minimizing a reverse-KL objective regularized by a proximal KL term between successive Gaussian proposals: $$E\bigl[Y\mid Z,A=a,X=x\bigr] = \int h(w,a,x)\,dF(w\mid Z,A=a,X=x),$$0 The proximal term is explicitly interpreted as a trust-region penalty in KL geometry (Yu et al., 8 May 2026).

A plausible unifying theme is that each formulation inserts an auxiliary projection, smoothing, or bridge relation between the observed problem and an otherwise inaccessible target.

2. Core proximal causal inference

The canonical proximal causal setup observes i.i.d. draws of $$E\bigl[Y\mid Z,A=a,X=x\bigr] = \int h(w,a,x)\,dF(w\mid Z,A=a,X=x),$$1, where $$E\bigl[Y\mid Z,A=a,X=x\bigr] = \int h(w,a,x)\,dF(w\mid Z,A=a,X=x),$$2 is treatment, $$E\bigl[Y\mid Z,A=a,X=x\bigr] = \int h(w,a,x)\,dF(w\mid Z,A=a,X=x),$$3 is outcome, $$E\bigl[Y\mid Z,A=a,X=x\bigr] = \int h(w,a,x)\,dF(w\mid Z,A=a,X=x),$$4 is an unobserved confounder, $$E\bigl[Y\mid Z,A=a,X=x\bigr] = \int h(w,a,x)\,dF(w\mid Z,A=a,X=x),$$5 is a treatment-inducing proxy, and $$E\bigl[Y\mid Z,A=a,X=x\bigr] = \int h(w,a,x)\,dF(w\mid Z,A=a,X=x),$$6 is an outcome-inducing proxy. Instead of the usual exchangeability condition $$E\bigl[Y\mid Z,A=a,X=x\bigr] = \int h(w,a,x)\,dF(w\mid Z,A=a,X=x),$$7, proximal causal inference assumes

$$E\bigl[Y\mid Z,A=a,X=x\bigr] = \int h(w,a,x)\,dF(w\mid Z,A=a,X=x),$$8

together with positivity and completeness conditions ensuring that integration against the proxy distributions is injective (Cui et al., 2020).

Under these assumptions, identification proceeds through two complementary representations. Any solution of the outcome-bridge equation yields the proximal g-formula

$$E\bigl[Y\mid Z,A=a,X=x\bigr] = \int h(w,a,x)\,dF(w\mid Z,A=a,X=x),$$9

so that the average treatment effect $q(z,a,x)$0 is obtained by integrating $q(z,a,x)$1 over $q(z,a,x)$2. Any solution of the treatment-bridge equation yields the weighted representation

$q(z,a,x)$3

and therefore

$q(z,a,x)$4

(Cui et al., 2020).

The role of completeness is decisive. In the bridge-function formulation, completeness guarantees uniqueness of the bridge and therefore point identification. When completeness holds, the proxy pair $q(z,a,x)$5 effectively replaces direct measurement of $q(z,a,x)$6; when it fails, the same proxy structure may still support partial identification rather than point identification, as discussed below (Ghassami et al., 2023).

3. Estimation, efficiency, and machine learning in proximal causal inference

Semiparametric proximal causal inference characterizes an efficient influence function for the average treatment effect: $q(z,a,x)$7 From this influence function one obtains a proximal doubly-robust estimator,

$q(z,a,x)$8

which is consistent if either the outcome-bridge model or the treatment-bridge model is correctly specified, and locally efficient when both are correct (Cui et al., 2020).

This semiparametric program extends to complex longitudinal studies. For two treatment occasions, proximal causal inference under a marginal structural mean model introduces time-indexed proxies $q(z,a,x)$9 and $$E\bigl[q(Z,a,x)\mid W,A=a,X=x\bigr] = \frac1{f(A=a\mid W,X)}.$$0, outcome bridges $$E\bigl[q(Z,a,x)\mid W,A=a,X=x\bigr] = \frac1{f(A=a\mid W,X)}.$$1, treatment bridges $$E\bigl[q(Z,a,x)\mid W,A=a,X=x\bigr] = \frac1{f(A=a\mid W,X)}.$$2, and a class of regular and asymptotically linear estimators indexed by efficient scores. The resulting proximal doubly-robust estimating equation remains unbiased if either the outcome-bridge models or the treatment-bridge models are correctly specified (Ying et al., 2021).

For right-censored time-to-event outcomes, proximal inverse probability-weighted and proximal doubly robust estimators identify

$$E\bigl[q(Z,a,x)\mid W,A=a,X=x\bigr] = \frac1{f(A=a\mid W,X)}.$$3

and

$$E\bigl[q(Z,a,x)\mid W,A=a,X=x\bigr] = \frac1{f(A=a\mid W,X)}.$$4

Both estimators are shown to be uniformly consistent and asymptotically normal, with the doubly robust version remaining valid under $$E\bigl[q(Z,a,x)\mid W,A=a,X=x\bigr] = \frac1{f(A=a\mid W,X)}.$$5 (Ying et al., 2022).

Several computational strategies have been developed for the ill-posed bridge equations. Neural Maximum Moment Restriction rewrites the conditional bridge restriction as an infinite family of unconditional moments and minimizes the squared worst-case moment over $$E\bigl[q(Z,a,x)\mid W,A=a,X=x\bigr] = \frac1{f(A=a\mid W,X)}.$$6 in an RKHS. Empirically, NMMR-U achieved the lowest causal MSE at all sample sizes in the demand benchmark, and NMMR-V gave by far the best performance as $$E\bigl[q(Z,a,x)\mid W,A=a,X=x\bigr] = \frac1{f(A=a\mid W,X)}.$$7 increased in the dSprite benchmark; the method also scales favorably because it trains a single neural network using mini-batch SGD (Kompa et al., 2022).

An alternative is regression-based PCI, which replaces direct solution of integral equations by two-stage generalized linear models. In the identity–identity, log–log, and logit–logit cases, observed-data regressions on suitable transformations of $$E\bigl[q(Z,a,x)\mid W,A=a,X=x\bigr] = \frac1{f(A=a\mid W,X)}.$$8 recover the causal parameter $$E\bigl[q(Z,a,x)\mid W,A=a,X=x\bigr] = \frac1{f(A=a\mid W,X)}.$$9, making PCI implementable with standard GLM software. In the SUPPORT application, the proximal 2SLS estimate for days lived was $J$0 with $J$1 CI $J$2, and the proximal estimate for 30-day survival was log-OR $J$3 with $J$4 CI $J$5 (Liu et al., 2024).

4. Extensions, robustness, and proxy design

Proximal causal inference has been extended far beyond the baseline point-treatment setting. In synthetic control, donor outcomes and surrogates can be treated as proxy systems for latent factors, leading to GMM estimators for synthetic-control weights and ATT; one paper shows that, if $J$6, only post-treatment data suffice for identification, while another frames the post-treatment treated-minus-synthetic-control contrast as a time series and uses excluded donors as proxies of latent confounders (Liu et al., 2023, Shi et al., 2021).

The same bridge logic has been adapted to hidden mediators and hidden outcomes. For hidden mediators, proximal identification yields formulas for $J$7, hidden front-door identification, and the population intervention indirect effect, together with multiply-robust influence-function-based estimators (Ghassami et al., 2021). For hidden outcomes, three proxies $J$8 together with conditional independence, completeness, distinctness, and label-uniqueness identify the full-data law $J$9; influence-function-based estimators are then shown to be multiply robust and asymptotically normal (Guo et al., 11 May 2026). Modified treatment policies have likewise been incorporated through bridge representations

$\prox_{\lambda P,J}(\bar\theta) = \arg\min_{\theta\in\mathbb R^p} \Big\{ \tfrac12(\theta-\bar\theta)^\top J(\theta-\bar\theta)+\lambda P(\theta) \Big\},$0

together with cross-fitted debiased machine learning estimators (Olivas-Martinez et al., 12 Dec 2025).

A major line of work concerns robustness to weak, invalid, or unconventional proxies. When completeness fails, proximal methods need not collapse completely: sharp bounds on causal effects can be obtained without identifying a bridge function, and the resulting non-smooth bounds can be approximated by LogSumExp smoothing for bootstrap inference (Ghassami et al., 2023). When many candidate treatment proxies are available and some may be invalid, fortified proximal causal inference assumes only that at least $\prox_{\lambda P,J}(\bar\theta) = \arg\min_{\theta\in\mathbb R^p} \Big\{ \tfrac12(\theta-\bar\theta)^\top J(\theta-\bar\theta)+\lambda P(\theta) \Big\},$1 out of $\prox_{\lambda P,J}(\bar\theta) = \arg\min_{\theta\in\mathbb R^p} \Big\{ \tfrac12(\theta-\bar\theta)^\top J(\theta-\bar\theta)+\lambda P(\theta) \Big\},$2 proxies are valid, introduces the fortified function class

$\prox_{\lambda P,J}(\bar\theta) = \arg\min_{\theta\in\mathbb R^p} \Big\{ \tfrac12(\theta-\bar\theta)^\top J(\theta-\bar\theta)+\lambda P(\theta) \Big\},$3

and constructs the multiply robust, locally efficient estimator fPMR (Yu et al., 16 Jun 2025). Under a canonical proximal linear structural equations model, adaptive proximal causal inference instead uses a majority-valid-TCP rule, LASSO-based median estimation, and an adaptive LASSO refinement that is root-$\prox_{\lambda P,J}(\bar\theta) = \arg\min_{\theta\in\mathbb R^p} \Big\{ \tfrac12(\theta-\bar\theta)^\top J(\theta-\bar\theta)+\lambda P(\theta) \Big\},$4 consistent and oracle-efficient under the stated regularity conditions (Rakshit et al., 25 Jul 2025).

Proxy construction itself has become a subject of methodological analysis. With text data, two instances of pre-treatment text are split into $\prox_{\lambda P,J}(\bar\theta) = \arg\min_{\theta\in\mathbb R^p} \Big\{ \tfrac12(\theta-\bar\theta)^\top J(\theta-\bar\theta)+\lambda P(\theta) \Big\},$5 and $\prox_{\lambda P,J}(\bar\theta) = \arg\min_{\theta\in\mathbb R^p} \Big\{ \tfrac12(\theta-\bar\theta)^\top J(\theta-\bar\theta)+\lambda P(\theta) \Big\},$6, passed through two zero-shot models to produce $\prox_{\lambda P,J}(\bar\theta) = \arg\min_{\theta\in\mathbb R^p} \Big\{ \tfrac12(\theta-\bar\theta)^\top J(\theta-\bar\theta)+\lambda P(\theta) \Big\},$7 and $\prox_{\lambda P,J}(\bar\theta) = \arg\min_{\theta\in\mathbb R^p} \Big\{ \tfrac12(\theta-\bar\theta)^\top J(\theta-\bar\theta)+\lambda P(\theta) \Big\},$8, and filtered by an observed conditional odds-ratio heuristic $\prox_{\lambda P,J}(\bar\theta) = \arg\min_{\theta\in\mathbb R^p} \Big\{ \tfrac12(\theta-\bar\theta)^\top J(\theta-\bar\theta)+\lambda P(\theta) \Big\},$9; under the paper’s assumptions this design satisfies the proximal identification conditions while several naive alternatives do not (Chen et al., 2024). For systems with feedback, regression-based bidirectional proximal causal inference identifies two causal directions in a linear SEM with unmeasured confounding and derives Bi-TSLS estimators plus sensitivity formulas indexed by $\hat\theta=\prox_{\lambda P,J}(\bar\theta)$0 and $\hat\theta=\prox_{\lambda P,J}(\bar\theta)$1 (Min et al., 18 Jul 2025). A plausible implication of this literature is that “proxy validity” is no longer treated as a binary prerequisite alone, but increasingly as an estimand-adjacent modeling problem involving robustness, diagnostics, and sensitivity analysis (Ringlein et al., 30 Dec 2025).

5. Proximal inference in stochastic trajectory optimization

In stochastic trajectory optimization, PISTO recasts STOMP within a variational-inference framework. Starting from the STOMP objective

$\hat\theta=\prox_{\lambda P,J}(\bar\theta)$2

the paper shows that

$\hat\theta=\prox_{\lambda P,J}(\bar\theta)$3

where

$\hat\theta=\prox_{\lambda P,J}(\bar\theta)$4

Thus STOMP implicitly minimizes the KL divergence from a Boltzmann trajectory distribution (Yu et al., 8 May 2026).

PISTO stabilizes this update by adding a proximal KL penalty to the previous proposal. The proximal objective admits an equivalent surrogate target

$\hat\theta=\prox_{\lambda P,J}(\bar\theta)$5

which can be written as

$\hat\theta=\prox_{\lambda P,J}(\bar\theta)$6

with

$\hat\theta=\prox_{\lambda P,J}(\bar\theta)$7

Reverse-KL moment matching then yields the closed-form update

$\hat\theta=\prox_{\lambda P,J}(\bar\theta)$8

Because direct sampling from $\hat\theta=\prox_{\lambda P,J}(\bar\theta)$9 is intractable, the method uses importance sampling from $\phi$0, normalized weights $\phi$1, and the Monte Carlo approximation

$\phi$2

(Yu et al., 8 May 2026).

The second KL term is explicitly interpreted as enforcing a controlled neighborhood of the previous iterate in KL geometry, preventing overly large jumps that can destabilize Monte Carlo variational inference. Since the update relies only on sampling trajectories and evaluating $\phi$3, PISTO inherits STOMP’s ability to handle non-differentiable and discontinuous cost functions without modification (Yu et al., 8 May 2026).

Empirically, on robot arm motion planning benchmarks, PISTO achieves an $\phi$4 success rate, outperforming CHOMP ($\phi$5) and STOMP ($\phi$6), while producing shorter, smoother paths at twice the speed of competing stochastic methods. It is further validated on contact-rich MuJoCo locomotion and manipulation tasks, where it consistently outperforms both CEM and MPPI baselines in reward (Yu et al., 8 May 2026).

6. Proximal operators in statistical estimation and Bayesian computation

Outside causal inference, proximal methods provide a general language for regularized estimation. “Proximal Estimation and Inference” defines a penalized estimator as the proximal map of an initial estimator and derives its asymptotic law from three inputs: the limit law of the initial estimator, the limiting penalty subgradient, and the inner product defining the proximal operator. Under the stated conditions,

$\phi$7

and, if $\phi$8 is invertible,

$\phi$9

The same framework yields a general Oracle theorem based on the limit subgradient and covers regular and irregular linear-regression designs, including new ridgeless-type proximal estimators (Quaini et al., 2022).

The proximal perspective also supports Bayesian computation for non-smooth and constrained models. ProxMCMC smooths the posterior by replacing $$\phi_\gamma(x) = \min_{u\in\mathbb R^p} \Bigl\{\phi(u)+\tfrac1{2\gamma}\|x-u\|^2\Bigr\},$$0 with $$\phi_\gamma(x) = \min_{u\in\mathbb R^p} \Bigl\{\phi(u)+\tfrac1{2\gamma}\|x-u\|^2\Bigr\},$$1, samples from

$$\phi_\gamma(x) = \min_{u\in\mathbb R^p} \Bigl\{\phi(u)+\tfrac1{2\gamma}\|x-u\|^2\Bigr\},$$2

and implements either the Unadjusted Langevin Algorithm or Hamiltonian Monte Carlo with gradients

$$\phi_\gamma(x) = \min_{u\in\mathbb R^p} \Bigl\{\phi(u)+\tfrac1{2\gamma}\|x-u\|^2\Bigr\},$$3

The framework extends to adaptive estimation of $$\phi_\gamma(x) = \min_{u\in\mathbb R^p} \Bigl\{\phi(u)+\tfrac1{2\gamma}\|x-u\|^2\Bigr\},$$4, $$\phi_\gamma(x) = \min_{u\in\mathbb R^p} \Bigl\{\phi(u)+\tfrac1{2\gamma}\|x-u\|^2\Bigr\},$$5, and $$\phi_\gamma(x) = \min_{u\in\mathbb R^p} \Bigl\{\phi(u)+\tfrac1{2\gamma}\|x-u\|^2\Bigr\},$$6, and the paper reports applications including lasso, constrained lasso, graphical lasso, matrix completion, and sparse low-rank matrix regression (Zhou et al., 2022).

A distinct Bayesian construction pushes a Gaussian prior through a proximal mapping: $$\phi_\gamma(x) = \min_{u\in\mathbb R^p} \Bigl\{\phi(u)+\tfrac1{2\gamma}\|x-u\|^2\Bigr\},$$7 Because inference is carried out on $$\phi_\gamma(x) = \min_{u\in\mathbb R^p} \Bigl\{\phi(u)+\tfrac1{2\gamma}\|x-u\|^2\Bigr\},$$8, the posterior remains an ordinary density on $$\phi_\gamma(x) = \min_{u\in\mathbb R^p} \Bigl\{\phi(u)+\tfrac1{2\gamma}\|x-u\|^2\Bigr\},$$9, while the induced prior on $$\nabla\phi_\gamma(x) = \tfrac1\gamma\bigl(x-\mathrm{prox}_{\gamma\phi}(x)\bigr).$$0 places mass on lower-dimensional strata generated by the proximal map. This is used to handle varying-dimensional problems such as fused lasso and matrix recovery under unknown rank, with posterior computation by standard HMC and a dynamic flow-network application in which feasibility constraints and sparsity are encoded directly through the prox operator (Xu et al., 2021).

Taken together, these developments show that proximal inference is not a single method but a technical motif. In causal settings it reconstructs counterfactuals from proxy relations; in optimization it regularizes variational updates by KL geometry; in frequentist estimation it interprets penalized procedures as proximal transforms of initial estimators; and in Bayesian computation it smooths or reparameterizes non-smooth posteriors through Moreau–Yosida and proximal maps.