Constraint residuals, graph posteriors, and determinant-corrected full-space targets in Bayesian inverse problems

This paper demonstrates that in finite-dimensional Bayesian inverse problems with equality constraints, sampling via penalized residuals in the full parameter-state space yields a posterior distinct from the reduced-space posterior due to a missing Jacobian determinant factor, and it derives specific determinant corrections required to ensure that zero-noise residual limits correctly recover the graph-lifted reduced posterior.

The Gist

Imagine you are trying to solve a mystery. You have a set of clues (data) and a theory about how the world works (a mathematical model). Your goal is to figure out the true "secret ingredient" (the parameter) that caused the clues you see.

In the world of science, this is called a Bayesian inverse problem. Usually, scientists try to solve this by looking at the "secret ingredient" directly. But sometimes, the math is so hard that they try a different trick: they look at the secret ingredient and the result it produces together, and they just punish any answer where the result doesn't match the rules.

This paper, written by Jonathon Cottom and Emilia Olsson, points out a subtle but dangerous trap in that "different trick." They show that simply punishing the wrong answers isn't enough; you might accidentally punish the right answers too, just because of how you wrote the math.

Here is the breakdown using everyday analogies:

1. The Two Ways to Solve the Puzzle

Imagine you are trying to find the perfect recipe for a cake (the parameter). You know the cake must rise to a specific height (the state equation).

• The "Reduced" Way (The Clean Approach): You assume that for every recipe, there is exactly one height the cake will reach. You calculate that height first, then check if it matches your goal. This is the "gold standard" but can be very slow and computationally expensive.
• The "Full-Space" Way (The Penalty Approach): You write down the recipe and the height together. You tell your computer: "If the height is wrong, give it a big penalty score." You hope that by making the penalty huge, the computer will only keep the recipes where the height is perfect.

2. The Trap: The "Volume" Problem

The authors discovered that the "Full-Space" way has a hidden flaw.

Imagine you are trying to find a needle in a haystack.

• The Problem: If you change the way you measure the "wrongness" of the height (for example, measuring it in inches instead of centimeters, or squaring the error), you change the volume of the space where the "wrong" answers live.
• The Consequence: Even though the "perfect" recipes (the ones where the height is exactly right) are the same in both cases, the probability of picking a specific perfect recipe changes.

The Metaphor:
Think of the "perfect" recipes as a thin, flat sheet of paper floating in 3D space.

• If you use a "naive" penalty (just squaring the error), the math accidentally stretches or squishes the air around that sheet. It makes some parts of the sheet look "thicker" (more likely) and other parts "thinner" (less likely) just because of how you measured the error.
• The result? You end up with a biased list of recipes. You might think a specific cake recipe is the best, not because it fits the data, but because your math accidentally made that spot on the "sheet" look bigger.

3. The Solution: The "Determinant Correction"

The paper provides a fix. It's like adding a specific "volume adjustment" knob to your math.

• The Fix: Before you apply the penalty, you must multiply your math by a specific number (called the determinant of the Jacobian).
• What it does: This number acts like a counter-weight. If your measurement method squished the space, this number puffs it back up. If it stretched the space, this number squeezes it back down.
• The Result: Once you add this correction, the "Full-Space" method gives you the exact same list of best recipes as the "Reduced" (gold standard) method.

4. Why This Matters

The authors aren't saying the "Full-Space" method is bad. In fact, it's very popular because it's often easier to run on computers.

However, they are saying: You cannot just assume that "zero error" equals "correct probability."

• Feasibility vs. Calibration: Getting the error to zero is like making sure you are standing on the right street (Feasibility). But getting the correct probability is like knowing exactly which house on that street you should knock on (Calibration).
• The Warning: If you use advanced computer methods (like ADMM or MCMC) to solve these problems, you must include this "volume correction." If you don't, your computer might be very efficient at finding the right street, but it will be knocking on the wrong doors.

Summary in One Sentence

When using computer tricks to solve complex scientific puzzles by punishing errors, you must add a specific mathematical "volume correction" to ensure you aren't accidentally biasing your results just because of how you measured the error.

The Paper's Core Message:

1. Don't confuse "zero error" with "correct answer."
2. Algebraically equivalent ways of writing an equation can lead to different answers if you don't fix the volume.
3. The Fix: Multiply your penalty by the "Jacobian determinant" (a specific number that accounts for how the math stretches space).
4. The Tool: The authors created a software package called detcorr to help scientists check if they have applied this fix correctly.

Technical Summary

Technical Summary: Constraint Residuals, Graph Posteriors, and Determinant-Corrected Full-Space Targets in Bayesian Inverse Problems

1. Problem Statement

In Bayesian inverse problems constrained by state equations $c(\theta, u) = 0$, practitioners often sample in the full parameter-state space $(\theta, u)$ by penalizing the residual $c(\theta, u)$, rather than eliminating the state $u$ to sample in the reduced space $\theta$ alone. While full-space formulations (using augmented Lagrangians, ADMM, or penalty methods) are computationally advantageous for handling ill-conditioning and leveraging PDE-induced geometry, this paper identifies a fundamental theoretical ambiguity: driving the residual to zero is necessary for feasibility but insufficient to define the correct Bayesian posterior measure.

The authors demonstrate that algebraically equivalent residuals (e.g., $c$ vs. $A(\theta)c$) define the same feasible set but induce different limiting posterior distributions when penalized naively. Specifically, a standard Gaussian penalty on the residual converges to a "zero-noise residual posterior" that differs from the desired "graph-lifted reduced posterior" by a state-Jacobian volume factor.

2. Methodology and Theoretical Framework

The paper operates within the context of finite-dimensional discretizations where the state equation $c(\theta, u) = 0$ has a unique solution $u = G(\theta)$ and a nonsingular state Jacobian $D_u c$.

Key Distinctions:
The authors distinguish three measures often conflated in practice:

1. Reduced Posterior ($\pi_{red}$): The standard Bayesian posterior on $\theta$ obtained by solving $u=G(\theta)$ and evaluating the likelihood.
2. Graph-Lifted Posterior ($\pi_{\Gamma}$): The push-forward of the reduced posterior onto the constraint manifold $\Gamma = \{(\theta, u) : c(\theta, u)=0\}$. This is the target for exact sampling of the reduced problem in full space.
3. Zero-Noise Residual Posterior ($\pi_{res}$): The limit of a full-space formulation where a small-noise likelihood is placed directly on the residual coordinates $c(\theta, u)$.

Theoretical Derivation:
Using a local change of variables from state coordinates to residual coordinates and applying the coarea formula, the authors derive the limiting behavior of the naive penalty posterior:
$$\pi_{\rho}(\theta, u) \propto r(\theta, u) \exp\left(-\frac{\rho}{2} \|c(\theta, u)\|^2\right)$$
As the penalty weight $\rho \to \infty$, the marginal distribution over $\theta$ converges to:
$$\pi_{\theta}^{res}(\theta) \propto r(\theta, G(\theta)) \cdot |\det D_u c(\theta, G(\theta))|^{-1}$$
This result (Theorem 1) shows that the naive penalty introduces an unwanted weighting factor of $|\det D_u c|^{-1}$.

Correction Mechanism:
To recover the graph-lifted reduced posterior from a full-space penalty, the authors propose a determinant correction. The corrected target density is:
$$\tilde{\pi}_{\rho}(\theta, u) \propto r(\theta, u) \cdot |\det D_u c(\theta, u)| \cdot \exp\left(-\frac{\rho}{2} \|c(\theta, u)\|^2\right)$$
This correction cancels the Jacobian volume factor introduced by the change of variables, ensuring the limit matches the reduced posterior. The paper extends this to weighted residuals ($c^T R c$), showing the correction must include an additional $\frac{1}{2}\log \det R(\theta)$ term if the penalty is unnormalized.

3. Key Contributions

The paper makes four specific contributions:

1. Identification of Target Ambiguity: It formally distinguishes between the reduced posterior, its graph lift, and the residual posterior, clarifying that they are distinct measures despite sharing the same feasible set.
2. Penalty Limit Theorem: It proves a finite-dimensional theorem showing that naive penalty formulations converge to a posterior reweighted by the inverse state-Jacobian determinant, provided specific regularity and domination conditions are met.
3. Constructive Corrections: It derives explicit determinant-corrected targets for unweighted, weighted, and rescaled residual penalties that converge to the graph-lifted reduced posterior. It also establishes that hard-constraint invariance (feasibility) is distinct from exact finite-$\rho$ invariance under residual transformations.
4. Sampler-Agnostic Templates and Software: It provides templates for SMC, MCMC, and particle variational methods where augmented-Lagrangian or ADMM steps serve as proposals or preconditioners, while the determinant correction ensures the invariant measure is correct. The authors release the detcorr software package to evaluate these corrections and diagnose the separation between feasibility and calibration.

4. Results and Validation

The paper validates its theoretical claims through:

• Analytic Scalar Example: A simple nonlinear inverse problem ($u=\theta^2$) demonstrates that scaling the residual by a function $a(\theta)$ changes the naive posterior limit (introducing a factor of $a(\theta)^{-q}$) but leaves the feasible set unchanged. The determinant correction recovers the correct posterior.
• PDE-Constrained Benchmark: A one-dimensional elliptic coefficient inverse problem is used to compare three approaches: reduced-space sampling, naive full-space penalty, and determinant-corrected full-space penalty.
  • The naive full-space marginal is shown to be biased (shifted mean and variance) relative to the reduced-space reference.
  • The determinant-corrected full-space marginal matches the reduced-space reference to within numerical tolerance.
  • Diagnostics confirm that while both naive and corrected methods satisfy the constraint ($\|c\| \approx 0$), only the corrected method achieves posterior calibration.

5. Significance and Claims

The paper argues that residual convergence does not imply posterior correctness. The primary significance of the work is the separation of two distinct tasks in constrained Bayesian inference:

1. Feasibility: Driving the residual to zero (often handled by optimization primitives like ADMM or augmented Lagrangians).
2. Posterior Calibration: Ensuring the sampling distribution matches the intended target measure (requiring the determinant correction).

The authors emphasize that augmented-Lagrangian, splitting, and manifold methods are powerful tools for proposal generation, preconditioning, and initialization. However, these algorithms do not automatically define the posterior measure. To obtain exact sampling of the reduced posterior in full space, the target density must be explicitly declared and corrected with the state-Jacobian determinant.

The paper concludes that this distinction is a "discretisation-level warning": in infinite-dimensional settings, the determinant may require renormalization or careful reference measure choices, but the finite-dimensional result serves as a critical diagnostic for computational implementations. The work does not claim to invalidate existing algorithms but rather provides the necessary mathematical "guardrails" to ensure they sample the correct distribution.