Measurement-Induced Logit Contraction

• Measurement-induced logit contraction is a phenomenon in hybrid quantum neural networks where bounded measurement outputs limit the dynamic range available to the loss function.
• The analysis shows that this contraction compresses class score differences, weakening gradient signals and modifying loss curvature for improved stability and accuracy.
• The introduction of Quantum Measurement Temperature (QMT) rescales measurement outputs before softmax, amplifying gradients and enhancing trainability without altering the quantum circuit.

Searching arXiv for the cited papers and topic to ground the article in current research. Measurement-induced logit contraction denotes a regime in which the mechanism that exposes model outputs to the loss compresses the effective separations between class scores, thereby modifying confidence, curvature, and gradient flow. In the most explicit formulation, hybrid quantum neural networks (QNNs) produce logits as expectation values of bounded observables, so the measurement stage constrains each class score to a narrow interval before softmax–cross-entropy is applied; the resulting loss then operates in a weak-sensitivity regime. Related analyses in cross-entropy training, logit regularization, and class-level logit perturbation describe closely allied effects in which temperature, convex logit penalties, or explicit perturbations contract or expand the logits seen by the loss, altering optimization and generalization dynamics (Mondal et al., 21 Jun 2026, Khanh, 16 Jun 2026, Beck et al., 12 Feb 2026, Li et al., 2022).

1. Conceptual scope

In hybrid QNN classification, measurement-induced logit contraction is a trainability bottleneck arising at the measurement–loss interface. The $i$-th class score is a quantum expectation value

$$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$

and for standard Pauli observables $M_i\in\{X,Y,Z\}$, the operator norm is $1$ with eigenvalues $\pm 1$, so $z_i\in[-1,1]$ for every input $x$ and parameter $\theta$. When these bounded outputs are fed directly into softmax–cross-entropy, class-score separations are intrinsically limited, and the loss cannot realize large confidence gaps even if the circuit is otherwise expressive (Mondal et al., 21 Jun 2026).

A broader reading of the literature shows that the same phrase, or closely related measurement-language, is used to analyze how the loss and normalization stage governs effective logit scale. Under cross-entropy in grokking, increasing temperature contracts $z/T$, desaturates the softmax, and restores gradient signal; under convex logit regularization, the loss pulls logits toward finite per-sample targets; under class-level perturbation, explicit contraction or expansion of logits changes entropy and margins in a targeted way (Khanh, 16 Jun 2026, Beck et al., 12 Feb 2026, Li et al., 2022).

This suggests that measurement-induced logit contraction is best understood not as a single pathology but as a family of output-space phenomena in which the measurement map—bounded observables, temperature scaling, regularization, or perturbation—controls the logit dynamic range available to the loss.

2. Measurement–loss mismatch in hybrid quantum classifiers

The hybrid-QNN formulation is the most concrete and mechanistically specific version of the concept. A variational quantum classifier (VQC) produces one logit per class as a bounded expectation value. Because the bound $z_i\in[-1,1]$ is imposed by quantum mechanics, it is independent of ansatz depth or initialization. The resulting issue is not merely that logits are numerically small, but that the loss surface seen by optimization is compressed at the interface between quantum measurement and classical cross-entropy (Mondal et al., 21 Jun 2026).

The paper formalizes this with a Logit Compression bound. If $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$0 satisfies $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$1, with softmax probabilities

$$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$2

then for one-hot label $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$3 with true class $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$4,

$$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$5

For Pauli-bounded logits $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$6 and $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$7, the maximum achievable confidence is $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$8, and the minimum loss is $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$9. Thus, even an expressive quantum circuit cannot escape the loss’s weak-sensitivity regime when fed bounded logits (Mondal et al., 21 Jun 2026).

The significance of this result is that it isolates a structural mismatch between quantum readout and standard multi-class training. The bottleneck is neither ansatz capacity nor parameter initialization in isolation. It is the fact that physically bounded measurements are used as logits in a loss whose useful operating regime ordinarily assumes a substantially larger logit dynamic range.

3. Sensitivity, gradients, and curvature

The weak-sensitivity regime follows directly from the local differential structure of softmax–cross-entropy. In the standard multi-class setup,

$M_i\in\{X,Y,Z\}$0

and

$M_i\in\{X,Y,Z\}$1

with Jacobian $M_i\in\{X,Y,Z\}$2 (Mondal et al., 21 Jun 2026).

When logits are bounded and close together, $M_i\in\{X,Y,Z\}$3 trends toward uniform, $M_i\in\{X,Y,Z\}$4. The paper identifies three consequences. First, logit differences contract, so $M_i\in\{X,Y,Z\}$5 barely exceeds its competitors. Second, gradients $M_i\in\{X,Y,Z\}$6 are small for most classes and poorly informative. Third, the Hessian with respect to logits,

$M_i\in\{X,Y,Z\}$7

has eigenvalues that concentrate below $M_i\in\{X,Y,Z\}$8, and at uniform $M_i\in\{X,Y,Z\}$9 its largest value is approximately $1$0, reducing second-order sensitivity (Mondal et al., 21 Jun 2026).

This failure mode is distinct from circuit-level barren plateaus. The paper states that the circuit Jacobian $1$1 may have nontrivial entries, yet optimization remains degraded because the loss suppresses those directions after measurement. In that sense, measurement-induced logit contraction is an interface effect: trainability is lost not necessarily inside the circuit, but after the circuit has already produced usable variation in $1$2.

A related cross-entropy analysis in grokking reaches a complementary conclusion from the opposite scale regime. There, the proximal variable controlling delayed generalization is the effective logit scale $1$3, because large logits saturate the softmax and starve gradients. Increasing $1$4 contracts the effective logits, raises entropy, and desaturates the softmax; across a grid of held norms and temperatures, grokking delay collapses onto effective logit scale alone with $1$5 (Khanh, 16 Jun 2026). Taken together, these results show that cross-entropy is highly scale-sensitive at the measurement stage: too little dynamic range produces weak sensitivity, whereas too much produces saturation.

4. Quantum Measurement Temperature

The mitigation proposed for hybrid QNNs is Quantum Measurement Temperature (QMT), a learnable positive scalar $1$6 that rescales measurement outputs before softmax: $1$7 Equivalently, one may learn $1$8 and write $1$9. QMT is trained jointly with classical and quantum parameters and is inserted between the quantum measurement outputs and the classical softmax–cross-entropy loss; it does not alter the quantum ansatz, circuit depth, or measurement operators (Mondal et al., 21 Jun 2026).

The modified softmax and loss are

$\pm 1$0

Temperature scaling expands the effective logit range from $\pm 1$1 to $\pm 1$2, with

$\pm 1$3

As $\pm 1$4, $\pm 1$5 while $\pm 1$6. The mechanism therefore restores an effective logit dynamic range without changing the circuit (Mondal et al., 21 Jun 2026).

The gradient effect is explicit: $\pm 1$7 Hence QMT amplifies parameter-space gradients by $\pm 1$8 while preserving their directions, and increases their variance by $\pm 1$9. The same paper further shows that if $z_i\in[-1,1]$0 is $z_i\in[-1,1]$1-Lipschitz in $z_i\in[-1,1]$2, then $z_i\in[-1,1]$3 is $z_i\in[-1,1]$4-Lipschitz, so loss responsiveness to small logit differences is increased. Despite this amplification, gradients remain controlled: for Pauli-generated rotations and Pauli measurements, $z_i\in[-1,1]$5, and

$z_i\in[-1,1]$6

The intended effect is therefore controlled gradient strengthening rather than uncontrolled explosion (Mondal et al., 21 Jun 2026).

The empirical evidence is extensive. On protein fluorescence microscopy, learned QMT expands effective test-set logit ranges from protein mean $z_i\in[-1,1]$7 to $z_i\in[-1,1]$8, compared with $z_i\in[-1,1]$9 in classical models; on Fashion MNIST, the range expands from $x$0 to $x$1. Test-set mean margins increase from $x$2 to $x$3 on the protein hybrid QNN and from $x$4 to $x$5 on Fashion MNIST. Training stabilizes across five random initializations, with standard deviation reduced from $x$6 to $x$7 in several configurations; gradient variance improves by more than $x$8; and mean test accuracy gains of approximately $x$9–$\theta$0 are reported across datasets, entanglement choices, optimizers, and training strategies. Learned temperatures are typically fractional, with $\theta$1–$\theta$2 depending on setting (Mondal et al., 21 Jun 2026).

Implementation details in the reported workflow are also specific: a classical CNN extracts features $\theta$3; these are encoded into a VQC using $\theta$4 encodings with data re-uploading, $\theta$5 trainable rotations, and CZ or CNOT entanglement; Pauli $\theta$6 is measured on $\theta$7 wires to obtain $\theta$8; QMT applies $\theta$9; parameter updates use Adam with learning rate $z/T$0, cosine annealing with $z/T$1, batch size $z/T$2, and $z/T$3 epochs. Initialization uses $z/T$4, positivity can be enforced with $z/T$5 or $z/T$6, and the paper reports clamping examples such as $z/T$7, while practitioner guidance also recommends $z/T$8 (Mondal et al., 21 Jun 2026).

5. Related formulations in cross-entropy, logit regularization, and perturbation

Several neighboring literatures analyze different mechanisms that contract the logits seen by the loss.

Setting	Mechanism	Reported effect
Hybrid QNNs	Bounded measurement outputs $z/T$9	Weak-sensitivity regime; suppressed gradients
Grokking under CE	Temperature $z_i\in[-1,1]$0 contracts $z_i\in[-1,1]$1	Desaturates softmax; shortens delay
Convex logit regularization	Penalty pulls logits toward finite targets	Logit clustering; FLD alignment under stated conditions
Class-level perturbation	Additive perturbation or $z_i\in[-1,1]$2	Targeted margin contraction or expansion

Under cross-entropy grokking, the central object is the effective logit scale $z_i\in[-1,1]$3 seen by the softmax. The authors show that holding weight norm fixed while varying only output temperature slides the grokking delay across the entire norm-induced range, and matching the effective logit scale back to baseline recovers $z_i\in[-1,1]$4 of the norm-induced delay at modulus $z_i\in[-1,1]$5 and $z_i\in[-1,1]$6 at $z_i\in[-1,1]$7. Under mean-squared error, by contrast, the effective scale is pinned near one by the target and the logit-scale mediation channel is inactive (Khanh, 16 Jun 2026).

Convex logit regularization produces a different but related form of contraction. In binary classification with

$z_i\in[-1,1]$8

strict convexity yields a unique finite minimizer $z_i\in[-1,1]$9 satisfying

$$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$00

or, in penalty form, $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$01. In multiclass settings, the regularizer induces finite per-sample targets $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$02, and optimization contracts logits around these finite targets as tightly as the data geometry allows. For Gaussian data, or whenever logits are sufficiently clustered, minimizing the regularized loss is equivalent to minimizing the coefficient of variation $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$03, which aligns the weight vector with Fisher’s Linear Discriminant. Label smoothing appears as a canonical instance, with target logit gap

$$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$04

This replaces the unregularized tendency toward infinite margins with a finite-target objective (Beck et al., 12 Feb 2026).

Class-level logit perturbation studies contraction and expansion even more explicitly. For softmax cross-entropy,

$$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$05

so the first-order loss change under perturbation $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$06 is $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$07. Positive augmentation aligns $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$08 with $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$09, increasing loss, shrinking margins, and raising entropy; negative augmentation anti-aligns it, decreasing loss and expanding margins. The paper also formalizes a contraction operator $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$10, $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$11, and states that reducing $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$12 flattens probabilities and increases entropy. This is analogous to temperature scaling with $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$13, but the proposed learning scheme applies contraction or expansion at class level through bounded perturbations and reports gains on balanced and long-tail benchmarks, including CIFAR-LT, iNat, VOC-MLT, COCO-MLT, and MS-COCO (Li et al., 2022).

These literatures are united by a common mathematical fact: the output–loss map is not invariant to logit scale under cross-entropy. They differ, however, in which regime is problematic. Hybrid QNNs suffer from insufficient scale after measurement and therefore use QMT to decompress logits. Grokking analyses and some perturbative methods instead treat excessive scale and softmax saturation as the limiting factor, so contraction is used to restore gradient signal.

6. Distinctions, limitations, and open problems

A recurrent misconception is that measurement-induced logit contraction is simply another name for barren plateaus. In the hybrid-QNN formulation this is explicitly false: barren plateaus, expressibility–entanglement-induced gradient collapse, and noise-induced vanishing gradients concern the circuit Jacobian $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$14 inside the circuit, whereas logit contraction degrades trainability after measurement, at the classical loss interface (Mondal et al., 21 Jun 2026).

A second misconception is that QMT is merely post-hoc calibration. Post-hoc temperature scaling calibrates probabilities after training and does not modify optimization dynamics. QMT acts during training, rescales logits before softmax, enlarges loss sensitivity, and changes the gradient magnitude and variance by $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$15 and $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$16, respectively; the paper attributes faster and more stable convergence, larger margins, reduced collapses, and improved final accuracy to this training-time intervention (Mondal et al., 21 Jun 2026).

The limitations are regime-specific. In hybrid QNNs, excessively small $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$17 may worsen calibration, amplify finite-shot noise, or sharpen the softmax too aggressively; the reported safeguards are positivity constraints, $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$18-clamping, weight decay on $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$19, gradient clipping, sufficient shots, and monitoring of expected calibration error. In the grokking setting, numerical precision can imitate or mask saturation effects: a float64 softmax-collapse audit showed that an apparent transition at extreme norm in float32 was spurious, and the authors therefore fit the dose-response law only on the collapse-free range $$z_i(x,\theta)=\langle M_i\rangle=\mathrm{Tr}(\rho(x,\theta)M_i),$$20 (Khanh, 16 Jun 2026).

There is also a terminological boundary. In Bayesian group-sparse categorical logit models, “measurement-induced contraction” refers to posterior contraction driven by likelihood curvature, score concentration, and a data-calibrated prior, rather than contraction of classifier logits themselves. The paper proves contraction rates in prediction and parameter metrics for multi-category logit models under group sparsity, but this is a distinct statistical use of contraction language (Jeong, 2020).

Open questions reported across these works remain largely structural. For hybrid QNNs, they include per-class or per-logit temperatures, adaptive schedules tied to margin or gradient-norm feedback, extensions to regression or structured losses, and systematic study under hardware noise and finite shots. For logit-scale mediation in grokking, larger-scale tasks and LayerNorm-equipped models were not mapped. For convex logit regularization, exact FLD alignment is proved for linear models under Gaussian data or locally quadratic regimes, leaving broader deep end-to-end settings open (Mondal et al., 21 Jun 2026, Khanh, 16 Jun 2026, Beck et al., 12 Feb 2026).

In this sense, measurement-induced logit contraction functions as a unifying lens on how output-space constraints imposed by measurement, normalization, or regularization determine the useful scale of logits presented to the loss. The central technical question is not simply whether logits are large or small, but whether the measurement stage places the loss in a regime where gradients remain informative and optimization is stably coupled to the model’s representational capacity.