How do encoding strength and measurement shots shape double descent in quantum extreme learning machines?

Double descent is a risk curve, not a single number

As model capacity grows, test error first follows the textbook U-shaped curve, spikes sharply at the interpolation threshold — the point where the model has just enough capacity to fit the training data exactly — and then falls again as capacity increases further. This inversion is double descent. The paper treats it as a property of the whole risk curve, not of one trained model, so any claim about it has to specify where on the capacity axis the model sits.

The paper's target is the shape of that curve when features are correlated. For independent features the behavior is well understood; for correlated features the shape is governed by the spectrum of the feature covariance, but earlier analyses typically produced only implicit equations. The paper's first contribution is to close that gap for classical linear regression by characterizing the entire risk curve analytically through two quantities: the condition number of the feature population covariance and a spectral measure of correlation strength.

Two covariance quantities fix the classical risk curve

The analysis assumes the whitened features are sub-Gaussian, which lets the authors control the eigenspectrum of random matrices even when entries are correlated. In the under-parametrized regime this reduces the problem to the Marchenko–Pastur law, giving a closed-form canonical curve governed by the ratio of features to samples, provided a growth condition on the sub-Gaussian norm holds. In the over-parametrized regime the bias–variance decomposition of the minimum-norm interpolator yields a variance term bounded through the condition number of the population covariance.

The practical consequence is that two spectral properties of the feature covariance — how ill-conditioned it is and how strongly its coordinates are correlated — determine where the interpolation peak sits and how the curve behaves beyond it. This is the analytical link the paper then transfers to quantum models.

Why the QELM readout covariance is the right object

For unitary 2-design circuits, the QELM readout is a zero-mean Gaussian process with isotropic population covariance, so the classical result applies with a feature variance of order 2^-N. The readout also admits a finite Fourier decomposition in which the frequency set contains accumulated spectral differences and the coefficients encode the reservoir unitaries, the input state, and the observables.

This Fourier representation is what connects an abstract statistical quantity — the population covariance — to concrete circuit parameters. The paper's central move is to express the QELM population covariance as a weighted Fourier decomposition with cross-frequency rank-one outer products, so that changing a design knob changes the covariance spectrum in a way that can be tracked analytically.

Encoding strength drives the condition number

Writing the covariance as a weighted Fourier sum lets the authors Taylor-expand each weight in the encoding strength s at s = 0. Taking s → 0 reveals a singular limit in which the covariance collapses to a rank-one outer product, and a more careful analysis shows the condition number scales as Ω(s^-2) as s → 0. In other words, weak encoding produces a severely ill-conditioned covariance.

At the other extreme, the authors assume the input density is a K-order Sobolev function and use integration by parts to bound the condition number as s → ∞. The two limits together mean encoding strength is not a neutral preprocessing choice: it directly sets how ill-conditioned the covariance is, and therefore where the risk curve peaks.

Reuploads regularize only when they inject energy

Adding layers changes the Fourier coefficients themselves, not just the frequencies: scaling the input strength rescales spectral differences, but increasing the number of reuploads changes the operators inside the trace. The paper decomposes the updated population covariance into three terms — one over the base frequencies, one over the newly accessible high frequencies, and a symmetric cross term — and absorbs the cross term into a remainder that does not affect the main argument.

The resulting bound shows the condition number decreases with added reuploads if and only if the lowest eigenvalue of the combined base-plus-high-frequency term increases. The authors formalize this as a high-frequency injection δ(ΔL): the new frequencies must put positive energy into the weak eigenspace of the base covariance. If they do, an ill-conditioned covariance can be improved; if they do not, extra layers need not help. The paper states this as a condition, not a general guarantee for arbitrary reservoir unitaries.

Finite shot noise acts like ridge regularization

Measurement shots are usually treated as an estimation cost to be minimized. Here, finite shot noise regularizes the QELM output covariance in the same way the other two knobs do, and the paper derives conditions under which the interpolation threshold remains stable. Finite measurement shot noise can be leveraged to guarantee this stability.

The authors describe the mechanism as suppressing pathological feature correlations, which prevents the interpolation threshold from shifting into the over-parametrized regime. They also note that sufficiently strong finite shot noise reduces correlations enough to restore a single peak, in contrast to the multiple-descent behavior associated with anisotropic covariances that have significantly distinct eigenvalues. Shot budgets therefore trade off estimation accuracy against implicit regularization, and the paper argues they should be tuned with that trade-off in mind when the covariance is poorly conditioned.

What the design recipe does not promise

The paper is explicit that its findings do not guarantee the second descent reaches a lower risk than the under-parametrized regime; that depends on the interplay between the dataset and the model architecture. The over-parametrized bound also relies on a geometric regularity condition on the whitened features that the authors justify for 2-design circuits and, more generally, in a high-dimensional limit under a Gaussian input assumption. Extending that justification to broader non-Gaussian input distributions and to finite, moderate dimensions remains open.

Two further boundaries matter for anyone applying the recipe. The theory does not explain the emergence of second peaks in the over-parametrized regime, leaving the multiple-descent mechanism open. And the reupload analysis identifies when additional layers improve conditioning but does not establish a general, non-perturbative guarantee that the required injection condition holds for arbitrary reservoir unitaries — characterizing which reservoir architectures satisfy it automatically would sharpen the recipe considerably. The theory also addresses ridgeless linear regression on top of a fixed quantum feature map, so whether the same covariance-conditioning perspective extends to trained variational quantum models, to non-linear quantum kernel methods, or to classification losses is left as an open direction.

The priority is conditioning, not raw capacity

The paper's practical message is a priority ordering rather than a single setting. Increase input encoding strength and the number of reuploads as far as circuit depth and coherence budgets allow, and tune the measurement-shot budget to balance estimation cost against implicit regularization. Each of these knobs acts on the same object — the QELM output covariance — so they are not independent levers but different ways of conditioning the same spectrum.

The decision threshold that matters most is the injection condition for reuploads: extra layers improve the condition number only when the newly accessible high frequencies put positive energy into the weak eigenspace of the base covariance. When that condition fails, adding depth is not a remedy. The most useful thing a reader can take from this paper is therefore not that over-parametrization helps, but that the covariance condition number — shaped by encoding strength, reuploads, and shot noise together — is the quantity to steer.

Sources used in this answer