Overcoming the Representational Limitations of Amplitude Encoding for Quantum Neural Networks

AI summary

Overview

Research area: Quantum machine learning, specifically quantum neural networks (variational quantum circuits) and the classical-data encoding step that precedes them.

Technical level: Intermediate. The central argument is geometric and can be followed with basic linear algebra, but familiarity with amplitude encoding, qubit registers, and variational training loops helps.

Scope: The paper diagnoses a normalization-induced scale invariance in amplitude encoding, proves it imposes an exact architecture-independent limit on what any downstream quantum circuit can learn, proposes a fixed nonlinear feature map that removes it while keeping logarithmic qubit scaling, and benchmarks the result on three standard classification datasets plus one synthetic dataset.

What This Paper Is About

Amplitude encoding can squeeze a d-dimensional classical vector into roughly log₂ d qubits, but only after dividing by the vector's ℓ₂ norm. That division throws away the vector's length: any input x and any nonzero real multiple λx produce the same quantum state up to an unobservable global phase, so no quantum circuit downstream can ever tell them apart. The paper shows this is not a minor technicality but a hard restriction on the hypothesis class, and it designs a parameter-free preprocessing step that restores the lost scale information without giving up the compact qubit budget.

Key Contributions

1. A precise statement of the bottleneck. The authors formalize normalization-induced scale invariance as an exact, architecture-independent restriction: because 𝒜(λx) = 𝒜(x) for λ > 0 (and up to a global phase for λ < 0), no choice of ansatz, entangling topology, or trainable parameters can recover the distinction. They visualize the consequence as star-shaped decision boundaries that are radially symmetric about the origin and persist throughout training.

2. A general condition for fixing it, plus a proof that nonlinearity alone is not enough. They derive the requirement that a deterministic map f preceding amplitude encoding must satisfy f(x) = f(λx) ⟹ λ = 1 and f(x) = f(y) ⟹ x = y, i.e. the image of f must contain no two distinct points on the same positive ray. They then give a counterexample: h(x) = eˣ is nonlinear, yet h(λx) = λh(x) admits nontrivial solutions x = log λ / (λ − 1) for λ ≠ 1, so scaling ambiguity survives.

3. A fixed feature map that provably removes the ambiguity. The proposed map is Φ(x) = (x, x^∘2, cos(x)) for x ∈ ℝ^d, where x^∘2 is the element-wise square. Any residual scaling solution would have to satisfy both (λy_i)² = λy_i² and cos(λy_i) = λcos(y_i) coordinate-wise, which admits none. It adds no trainable parameters, is defined for all real inputs (raw amplitude encoding is undefined at x = 0), and preserves logarithmic qubit scaling despite tripling the input dimension.

4. A controlled empirical comparison. All quantum models share an identical ansatz with exactly 32 trainable rotation parameters, differing only in encoding, and are compared against a classical feed-forward network across six-fold stratified cross-validation with Holm–Bonferroni-corrected exact McNemar tests.

Main Findings

• Raw amplitude encoding fails on simple 2D geometry. On the three canonical two-dimensional binary problems (linearly separable, concentric circles, interleaving half-moons), a state-vector-simulated VQC with 4 qubits, 12 weights and amplitude encoding trained for 20 epochs produced star-shaped boundaries and Accuracy near 0.5, equivalent to random guessing.

• Angle encoding has its own aliasing problem unless inputs are standardized. Because R_Y(x_j + 2πk)|0⟩ = (−1)^k R_Y(x_j)|0⟩, inputs separated by multiples of 2π encode to physically identical states. Without standardization, the same 4-qubit, 12-weight circuit also sat near 0.5 Accuracy; with standardized inputs it reached around 80%.

• The proposed map restores separability on the toy problems. With Φ(x) = (x, x^∘2, cos(x)) ahead of amplitude encoding, the same setup achieved Accuracy around 90% and well-separated decision boundaries on all three canonical datasets, surpassing angle encoding on the nonlinear ones.

• Large balanced-accuracy gains on three of four benchmarks. Relative to conventional amplitude encoding in the full study: Breast Cancer 0.6991 → 0.8834, Wine 0.7600 → 0.9320, Synthetic 0.5996 → 0.7717. ROC-AUC moved 0.7859 → 0.9692, 0.9031 → 0.9792, and 0.6469 → 0.8571 respectively.

• Digits is the qualified case. Balanced accuracy improved only from 0.2751 to 0.3100, and ROC-AUC and AUPR were slightly lower than conventional amplitude encoding (0.7268 vs 0.7480 and 0.3127 vs 0.3239); all quantum models performed poorly on this ten-class task, with angle encoding at 0.2175 balanced accuracy. The paper states the feature map "substantially improves" on three of the four datasets.

• Statistical significance against conventional amplitude encoding held on all four datasets. Out-of-fold predictions tested with the exact McNemar test and Holm–Bonferroni adjustment gave p_Holm < 0.05 in each case.

• The abstract summarizes the headline gain as 21–44% improvement in classification accuracy over both quantum encoding baselines across the tasks.

• Amplitude encoding generally beat angle encoding among the conventional schemes, with the synthetic dataset the exception (angle 0.6458 vs amplitude 0.5996 balanced accuracy). The authors attribute angle encoding's disadvantage partly to the PCA reduction to at most eight principal components that its qubit cost forced.

• The classical network still led everywhere. Its balanced accuracies were 0.9662 (Breast Cancer), 0.9653 (Wine), 0.9408 (Digits) and 0.8462 (Synthetic). The closest quantum result was Wine at 0.9320 versus 0.9653. The authors explicitly state that demonstrating quantum advantage is not the objective.

• Compression is not the same as fidelity. The concluding argument is that qubit efficiency alone is an insufficient criterion for an encoding, since compression helps only if task-relevant information survives.

Methodology in Plain English

The authors separate two questions that are usually tangled together: what the variational circuit can express, and what the encoding has already made impossible before the circuit runs. They answer the second one analytically, using the geometry of the normalization step. Dividing a vector by its length projects the whole input space onto a sphere, so every point on a ray from the origin collapses to one location. Rotating or entangling qubits afterwards cannot undo a collapse that happened upstream.

From that picture they derive a design rule for a preprocessing map: no two distinct inputs may land on the same ray in the preprocessed space. Testing candidate maps against this rule shows why a single nonlinearity is unreliable, and why pairing squaring with cosine works — the two functions scale in incompatible ways, so no λ ≠ 1 can satisfy both at once.

For the experiments they hold everything except the encoding constant. All three quantum models use the same reverse-linear CNOT ansatz with exactly 32 trainable rotation parameters, trained for 40 epochs with mini-batches of 32 in PennyLane on the lightning.gpu backend, simulated on an NVIDIA Quadro RTX 8000 GPU. That state-vector simulation removes finite-shot uncertainty and hardware noise, so differences can be attributed to encoding. Angle encoding's one-qubit-per-feature cost forced a PCA reduction to at most eight components, while amplitude encoding needed only ⌈log₂ d⌉ qubits with padding to the next power of two. Hyperparameters came from 20 Optuna trials on the validation set inside each fold of a stratified, shuffled six-fold cross-validation, with balanced accuracy as the objective; only the learning rate was tuned for quantum models, while the classical network also had its weight decay tuned. Evaluation used accuracy, balanced accuracy, macro F1, precision, recall, ROC-AUC and AUPR, reported as mean (standard deviation) over the six folds, with one-vs-rest ROC-AUC and macro-averaged AUPR for the multiclass sets.

Why This Matters

Impact on research. The paper converts a known but often glossed-over property of amplitude encoding into a testable design constraint with a formal condition attached. It also supplies a negative result that is easy to reuse: nonlinearity is not sufficient, so any proposed encoding fix needs an argument, not just a nonlinearity. Because the fix adds no trainable parameters and keeps logarithmic scaling, it slots into existing pipelines without changing the model's parameter budget — a meaningful point given how often the alternative in the literature is to bolt a classical neural network in front of the circuit, which the authors note both increases trainable parameters and makes the architecture no longer fully quantum.

Real-world applications (grounded in the tasks the paper evaluates):

• Medical diagnosis from tabular features, such as the 30 cell-nucleus measurements in the Breast Cancer dataset, where radial information about feature magnitude carries diagnostic signal.
• Chemical and agricultural classification from continuous measurements, as in the 13-feature, three-class Wine task.
• Handwritten digit recognition from 8×8 pixel images, a case where the paper's own results show encoding fixes alone are far from sufficient.
• High-dimensional data embedding generally, where the logarithmic qubit budget matters most and where the authors position the map as promising.

Industry relevance. Teams evaluating whether a variational circuit can match a classical baseline on tabular data now have a concrete, cheap preprocessing step to try before concluding the model class is inadequate. The finding that raw amplitude encoding can pin a classifier at chance level regardless of architecture is directly actionable for anyone debugging a quantum model that will not learn.

Future Directions

• Closing the gap to classical models. Even the best quantum result, Wine at 0.9320 balanced accuracy, trails the classical network's 0.9653, and Digits remains near 0.31 for all quantum encodings. The authors state that encoding alone may not resolve current QNN limitations and call for encoding design to be considered alongside circuit architecture and optimization strategy.
• Addressing the multiclass and high-dimensional failure. The Digits result, where the proposed map improved balanced accuracy but slightly reduced ROC-AUC and AUPR relative to plain amplitude encoding, is unexplained and invites investigation into how the tripled dimensionality interacts with many-class problems.
• Testing beyond noiseless simulation. All results come from state-vector simulation with no finite-shot uncertainty and no hardware noise; whether the gains survive shot noise, decoherence, and real device execution is untested here.
• Extending the design rule to other encodings. The condition derived in the paper applies to any deterministic map composed with amplitude encoding, which suggests applying the same ray-collision test to the angle-encoding variants the paper surveys (data reuploading, IQP, QAOA-inspired, Ring, Waterfall), whose costs scale from 𝒪(LN) to 𝒪(LN²).

Target Audience

Quantum machine learning researchers and graduate students working on variational circuits and data encoding, particularly those who benchmark QNNs on tabular or low-dimensional data and need to know which failures are architectural versus encoding-imposed. It is also useful for practitioners building quantum-classical hybrid pipelines who want a parameter-free alternative to prepending a classical network, and for reviewers and theorists interested in formal expressivity limits of encoding schemes. Readers without quantum computing background will follow the geometric argument but may need to look up amplitude encoding and variational circuit basics first.

Authors’ abstract

Quantum Neural Networks (QNNs) offer a framework for learning functions from classical data, but their expressive power depends critically on how inputs are encoded into quantum states. Angle encoding can require resources that scale with the number of features, whereas amplitude encoding represents $d$ features using only $\mathcal{O}(\log_2 d)$ qubits. This compactness comes at a representational cost: normalization makes conventional amplitude encoding invariant to nonzero real rescaling of the input, up to an unobservable global phase. We show that this invariance induces an exact, architecture-independent restriction on the resulting hypothesis class. No downstream quantum circuit can distinguish inputs differing by nonzero real scaling. We develop a general condition under which deterministic preprocessing can remove this ambiguity and show that nonlinearity alone is insufficient. Based on this analysis, we introduce a fixed nonlinear feature map that provably distinguishes all real inputs under subsequent amplitude encoding, without adding trainable parameters and while retaining logarithmic qubit scaling. We evaluate the resulting QNN on three canonical classification benchmarks and a nonlinear synthetic dataset against conventional amplitude and angle encoding, as well as a classical neural network. Across these tasks, the proposed encoding consistently improves classification accuracy over both quantum encoding baselines by 21$-$44 %. Our results establish normalization-induced scale invariance as a concrete representational bottleneck of amplitude encoding and demonstrate that it can be removed while preserving a compact quantum representation, marking a step toward quantum models capable of capturing arbitrary relationships in classical data.