acceptodds
Under review as a conference paper at ICLR 2027

When Does Auxiliary Calibration Improve Target Recovery? Nuisance Geometry, Sample Complexity, and Timing

Abstract

Many learning problems combine target-bearing observations with auxiliary measurements that do not observe the target directly, but instead reveal a shared nuisance variable. When, and by how much, can such auxiliary measurements improve target recovery? We study this question in a Gaussian inverse experiment with an acquisition channel that mixes target and nuisance effects and a calibration channel that observes only the nuisance. We show that calibration value is determined not by marginal channel quality, but by the paired geometry relating the nuisance scores in the two channels. The closed relation pairing the two nuisance scores yields an exact resolvent formula for efficient target information, including nonclosed score ranges, and separates calibration-irrecoverable structural loss from finite-budget calibration loss. For unrestricted nuisance, we prove an exact minimax-risk reduction to the efficient Gaussian shift over compact convex target classes; for restricted nuisance classes, we give a nuisance-translation condition under which the reduction remains valid. Under a Hilbert-scale comparison, we derive matching two-resource minimax rates and the critical auxiliary sample size needed to recover the nuisance-free target rate. In parabolic evolution models, calibration timing can change the effective ill-posedness and turn a polynomial auxiliary requirement into an exponential one, while nuisance constraints can change the minimax rate itself. The same geometry also yields target-oriented calibration allocation, which can differ from nuisance-optimal allocation. Experiments support these predictions and illustrate the additional errors-in-variables difficulty of learning the nuisance response.

Primary area:
learning theory
Keywords:
Statistical inverse problems, Calibration, Semiparametric efficiency

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

The paper deals with the problem of target recovery using auxiliary calibration data. Concretely, the goal is to estimate a target xx from an observed channel Y=Ax+Bθ+noiseY = Ax + B\theta + \text{noise}, using a second channel without target, Z=Cθ+noiseZ = C\theta + \text{noise}, where θ\theta is a "shared nuisance parameter". To do so, the authors rely on, using their terms, a paired nuisance geometry and efficient information, and show that, for an unrestricted nuisance, the problem is equivalent in minimax risk to the efficient Gaussian shift. They then derive two-resource minimax rates depending on the number of target-bearing data nn and of auxiliary data mm. They then study the case where auxiliary calibration is collected at different times. Finally, they ask, for a fixed auxiliary calibration budget, where it should be spent.

Critical strengths

First, the overall problem studied, i.e., when auxiliary calibration improves target recovery, is relevant; and the derivations seem correct to me. Second, the theoretical study is exhaustive, from the problem formulation to the minimax analysis and the budget allocation.

Critical weaknesses

The main problem is the formalism: the paper is very hard to read and contains too many undefined notations. The introduction relies on undefined terms from the start ("target-bearing data", "nuisance directions", "closed linear relation"...). This continues in the rest of the paper. Notations are often undefined: the spaces X,Θ,Y,Z\mathcal X, \Theta, \mathcal Y, \mathcal Z in the very first line of theory! A striking example is vv appearing as a dummy variable in eq (2.2), but appears in fact to be the "target direction" which seems to be a key notion to the paper (and this quantity is properly introduced in (Trabs 2015) where the equation comes from). Others : Π\Pi after eq (2.5), the list is not exhaustive.

This problem leads to the second weakness being the clarity of the contribution. The model described in (1.1) is not explained or linked to a concrete application; it is only related to several references listed at line 54 with little explanation. The related-work paragraph begins by citing only papers on the tools used, the first positioning of the work appears with the comparison to "de Hoop et al 2023", and the rest of the paragraph ((attia et al. 2018) and others) is ambiguous about the novelty of the present work. In my view, the contribution is unclear and is poorly put into perspective with the state of the art, even though it seems to be cited.

Questions for authors

Could you state precisely what your results bring compared to the cited works? In particular, the "two-resources rates" seem close to those obtained for deconvolution with an auxiliary sample (Comte & Lacour, 2011) and, the main stated difference with (de Hoop et al., 2023) is the "noncommuting transfer operators". What does your formulation allow that these works do not, and is there a concrete application where the difference matters?

Reviewer_THs3OpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper develops a minimax theory for auxiliary calibration in Gaussian inverse problems, separating irreducible structural loss from finite-sample loss. It derives critical calibration sample sizes and reveals that late calibration in parabolic systems can cause exponential sample complexity, which bounded-norm nuisance constraints can mitigate. It also provides target-oriented resource allocation rules, validated through simulations and fastMRI experiments.

Critical strengths
  1. The paper makes a sharp conceptual point: calibration benefit cannot be judged solely by marginal quality of individual channels. Even with identical Gram operators for nuisance channels, target recovery performance can be completely different, depending on the coupling geometry between nuisance responses of two channels. This is a non‑trivial observation that reshapes how we understand calibration value in inverse problems.

  2. The authors develop rigorous infinite-dimensional inverse-problem theory using operator methods, separating structural and finite-sample losses. They also establish minimax risk bounds and characterize the critical calibration sample size under different nuisance constraints.

Critical weaknesses
  1. All main theorems assume the forward operators AA, nuisance response maps BB and CC are given exactly. In many real‑world applications such as MRI coil‑sensitivity calibration, these operators must be estimated from noisy data. The paper only touches this in limited simulation for learning BB, without extending the main minimax theory to the case of unknown operators. The gap between theory and practice should be discussed more explicitly in main text.

  2. The fastMRI experiment serves as a real‑world demonstration rather than a strict empirical test of the Gaussian inverse‑experiment model. In practical MRI, ACS calibration data is reused for both calibration and image reconstruction, which violates the paper’s clean two‑channel independent‑noise setup. There is a risk readers may misinterpret the fastMRI results as empirical verification of the paper’s central theorems. The authors should highlight this mismatch explicitly and clarify which insights can (and cannot) be transferred to real MRI systems.

  3. The main theoretical results are restricted to linear Gaussian inverse problems. Nonlinear settings are discussed only through local asymptotic normality, which does not guarantee global recovery. The paper should clarify this limitation and discuss the challenges of extending the theory to general nonlinear inverse problems.

Questions for authors

See Weakness.

Reviewer_rZDjOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

Paper introduces and tackles the problem of when can auxiliary measurements help recover the target, when such auxiliary measurements never observe the target itself. Instead, they measure a shared nuisance variable that also affects the target-bearing observations. To achieve this they introduce a paired geometry split into structural loss and a recoverable transfer. The justification is that marginal Grams matrix do not determine problem efficiency (they give a convincing counter-example).

Authors quantify the recoverable transfer using an efficient projection. They then show how this efficient characterization can be expressed as a conditional minimax reduction. They solve it and provide bounds for different conditions and classes to eliminate the nuisance component (to its theoretical limit) and recover the target. They also show how such bounds change with the timing of nuisance signal collection.

Finally they propose an target-oriented design where one minimizes the target uncertainty after nuisance elimination. They show that such target-oriented allocation is optimal: they thus provide an optimal auxiliary measurement design to recover the target. Numerical simulation support their point.

Critical strengths

Paper introduces a formalization and shows (in certain classes with strict assumptions) how one can leverage auxiliary measurements of nuisance to improve target detection, even without ever observing the target. Their solution is elegant, with exact reduction o the efficient Gaussian shift for for unrestricted nuisance that requires almost no structural assumptions.

It provides a strong study of calibration timing, showing how timing qualitatively changes ill-posedness and rates, while deriving matching bounds. Separately, it shows how nuisance constraints can change the rate itself, restoring polynomial late-calibration rates where the unrestricted problem is logarithmic. This adds to the paper width and applicability.

Paper provides a target-oriented allocation with a closed form for the known-response independent-coordinate case. This is a worthy theoretical result. Especially that it follows directly from the same efficient-information geometry.

Critical weaknesses

The paper provides an optimal solution only in the oracle independent-coordinate model. For correlated / non-coordinate channels, authors only claims the efficient-information criterion defines the design. For such cases there is no closed form solution, or an extra response-estimation error. This is not the author's fault, but does limit the generalization and applicability of the paper.

The paper opens up a whole field of research: what if B and C were unknown? Or the relationship non-linear? The theory is limited to certain assumptions. The minimax finding is limited to the cases where the translation condition holds. Again, I do not hold this against authors.

The main limitation is the real-world experiments, which is limited to the appendix. How would one balance the budget (and its n and m components) in a real application?

Questions for authors

How to split the budget between acquisition and calibration in a real application with unknown smoothness, unequal per-unit costs, and calibration measurements that also enter the reconstruction. This falls into future work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.