Suppose a physical calculation gives a limited set of correlation measurements. Many spectra fit them. A reconstruction method returns a smooth-looking curve, and the next calculation uses that curve to choose an action.

A proposal from my recent research planning uses a different formulation: leave the spectrum uncertain and bound the quantity the decision actually uses. Then choose the next correlation measurement by how much it could reduce disagreement about that decision.

This is a developed proposal, not a demonstrated spectral method. Its appeal is that it changes what counts as progress. A better-looking spectrum may be irrelevant; a tighter bound on the decision quantity may be enough.

Preserve the ambiguity

In a schematic inverse problem, observed values satisfy yᵢ ≈ ∫ Kᵢ(ω) ρ(ω) dω. The unknown spectrum ρ is nonnegative and may obey normalization or moment constraints. With limited observations, the admissible set can contain many very different functions.

Suppose the downstream quantity is J(ρ) = ∫ w(ω) ρ(ω) dω. Instead of choosing one reconstruction, compute the smallest and largest J compatible with the observations and assumptions. If both lie on the same side of a decision threshold, the decision is stable within that set even though the spectrum is unresolved.

Conversely, a narrow point estimate for ρ can conceal a wide range of possible J. Regularization may choose a convenient member of the set without showing how much the decision depends on that choice.

The next measurement is a resource allocation

Candidate measurements differ in time, noise, computational fidelity, and cost. The next useful one is not necessarily the measurement that most improves global reconstruction error. It is the one expected to separate the models that imply different decisions.

A proposed acquisition policy would compare expected reduction in the J interval per unit cost. It could choose a new correlator, a time point, a higher-fidelity calculation, or more samples of an existing condition. The policy needs to charge for failed calculations and for the optimization used to choose them.

The uncomfortable part is that expected reduction requires a model of possible outcomes, while a set-based bound need not. That additional distributional assumption should remain visible. A conservative alternative can optimize a worst-case reduction, though that may be too pessimistic to be useful.

A grid is another assumption

A discretized optimization problem only bounds the spectra admitted by its grid and numerical tolerances. A very narrow feature between grid points can invalidate an apparent continuum guarantee. Measurement noise and solver error can widen the feasible set further.

I would keep three claims separate: a bound on the chosen finite representation, a numerical certificate that the optimization was solved accurately, and a justified extension to continuous spectra. The first two do not automatically establish the third.

This distinction is part of the proposal’s research burden. Without it, an attractive interval could give more confidence than an honest point estimate.

Compare decisions under equal budgets

A useful study would supply a downstream threshold or action before acquisition begins. Compare uniform sampling, reconstruction-directed acquisition, an established specialist strategy, and the proposed decision-directed policy at equal total cost. Evaluate whether the chosen action agrees with a high-quality reference, not only whether the recovered spectrum resembles it.

Also include cases where global reconstruction is genuinely necessary and where the decision weight changes after measurement. A policy optimized for one functional may be brittle to a new question. The stored observations should remain reusable so that specializing the acquisition does not require discarding the science.

The ambitious version is a calculation that knows which physical ambiguity it can afford to leave unresolved. The first acceptable result would be much narrower: one inverse problem, one decision functional, and a measured advantage that survives a strong acquisition baseline.

Related notes