---
title: Angle/Amplitude Encoding in QML
url: https://www.emergentmind.com/topics/angle-amplitude-encoding
type: topic
---

# Angle/Amplitude Encoding in QML

Angle encoding and amplitude encoding are two principal strategies for mapping classical information into computational substrates, most prominently quantum circuits. In contemporary quantum machine learning, angle encoding, also called rotation encoding, maps features into gate parameters such as single-qubit \(R_Y\), \(R_X\), or \(R_Z\) rotations, whereas amplitude encoding stores a normalized classical vector in the amplitudes of a quantum state. The pair defines a recurring trade-off among qubit count, circuit depth, state-preparation cost, expressivity, trainability, and noise robustness: angle-based maps are shallow and parallelizable, while amplitude-based maps offer exponential qubit compression but typically incur substantially higher preparation cost and greater sensitivity to two-qubit errors [2606.05387].

## 1. Formal definitions and canonical constructions

In the canonical angle-encoding construction, a real-valued feature vector \(x\in\mathbb{R}^D\) is embedded by applying one rotation per feature,
\[
U_{\text{angle}(x)} = \bigotimes_{k=1}^{D} R_Y(x_k),
\qquad
|\psi(x)\rangle = U_E(x)|0\rangle^{\otimes D}.
\]
A closely related QCNN formulation writes
\[
|\psi(x)\rangle = \bigotimes_{j=1}^{n} R_Y(\pi x_j)|0\rangle,
\]
while QRNN work uses the product state
\[
\ket{x_t}=\bigotimes_{i=1}^N \left(\cos\!\left(\frac{\hat{x}_{i,t}}{2}\right)\ket{0}+\sin\!\left(\frac{\hat{x}_{i,t}}{2}\right)\ket{1}\right),
\]
with \(\hat{x}_{i,t}\) obtained by MinMax scaling. Across these formulations, the defining feature is locality: each feature controls a rotation on one qubit, and the encoding layer is shallow and naturally parallelizable [2606.05387; 2512.12512; 2508.16784].

Amplitude encoding instead embeds a normalized vector into computational-basis amplitudes. In survey form,
\[
U_{\text{amp}(x)}|0\rangle^n = \frac{1}{\|x\|_2}\sum_{j=0}^{2^n - 1} x_j |j\rangle \equiv |\psi_x\rangle,
\qquad
n=\lceil \log_2 D\rceil.
\]
A related formulation used in hybrid QML is
\[
\ket{\psi} = \sum_{x\in\{0,1\}^{n}} \alpha_x \ket{x},
\qquad
\sum_{x\in\{0,1\}^{n}} |\alpha_x|^2 = 1.
\]
When the feature dimension is not a power of two, unused basis states are assigned zero amplitude. The normalization requirement is intrinsic: amplitude encoding identifies the input with a valid quantum state, so only the direction of the vector is preserved unless additional preprocessing is introduced [2606.05387; 2501.15828; 2508.16784].

Several works refine these canonical forms without changing the underlying distinction. Dense-angle encoding places two features on one qubit through
\[
U_{\text{dense}(x)} = \bigotimes_{k=1}^{\lceil D/2 \rceil} R_Z(x_{2k-1}) R_Y(x_{2k}),
\]
while data re-uploading repeatedly inserts angle-encoded blocks between trainable unitaries,
\[
U_{\text{reu}(x)} = \prod_{\ell=1}^{L} W_\ell(\theta_\ell)\, S(x),
\qquad
S(x)=\bigotimes_{k=1}^D R_Y(x_k).
\]
These variants retain the angle-encoding viewpoint but alter the qubit-efficiency–expressivity balance [2606.05387].

## 2. Resource trade-offs and hardware regimes

The most persistent distinction between angle and amplitude encoding is the contrast between qubit count and preparation cost. In the survey taxonomy, angle encoding uses \(q=D\) qubits, \(g=D\) single-qubit gates, and encoding depth \(d=1\), with zero two-qubit gates in the encoding layer. Dense-angle encoding reduces the qubit count to \(\lceil D/2\rceil\) at depth \(2\). Amplitude encoding compresses to \(q=\lceil\log_2 D\rceil\) qubits, but exact preparation typically has \(g=O(D)\) and \(d=O(D)\), with many two-qubit gates [2606.05387].

| Encoding | Qubits | Encoding-layer cost |
|---|---:|---|
| Angle | \(q=D\) | \(g=D,\ d=1\), no two-qubit encoding gates |
| Dense-angle | \(q=\lceil D/2\rceil\) | depth \(2\) |
| Amplitude | \(q=\lceil\log_2 D\rceil\) | \(g=O(D),\ d=O(D)\), many two-qubit gates |

This qubit-compression versus gate-cost trade-off reappears in application-specific analyses. In an angle-encoding Hadamard test for cosine similarity, amplitude encoding uses \(\mathcal O(\log_2 d)\) qubits and angle encoding \(\mathcal O(d)\) qubits; both have \(\mathcal O(d)\) state-preparation gates in the comparison table, but the circuit depth is \(\mathcal O(\mathrm{poly}(d))\) for amplitude encoding and \(\mathcal O(1)\) for the angle-encoding construction, which relies on parallel elementwise two-qubit subcircuits and classical post-processing. The specific implementation described uses two qubits per component, hence \(2d\) qubits for a \(d\)-dimensional vector [2604.15867].

Noise considerations sharpen the practical distinction. A survey-level fidelity bound under noisy preparation is
\[
F(\rho_{\mathrm{noisy}},|\psi\rangle) \ge (1-p_2)^{g_2}(1-p_1)^{g_1},
\]
with critical error threshold
\[
p^*(E,D,F_{\min}) = 1 - F_{\min}^{1/g_2(E,D)}.
\]
Using \(F_{\min}=0.90\), the reported representative thresholds at \(D=64\) are effectively unconstraining for angle encoding in this model, \(p^*\approx 1.3\times 10^{-3}\) for re-uploading with \(L=5\), \(p^*\approx 1.6\times 10^{-3}\) for IQP, and \(p^*\approx 8\times 10^{-4}\) for amplitude encoding. The survey’s central hardware guideline is that for \(p \ge 10^{-3}\), shallow angle-based encodings consistently outperform amplitude encoding in practice, despite the latter’s exponential qubit advantage [2606.05387].

## 3. Algorithmic constructions and approximation schemes

Angle encoding is often preferred when local structure and shallow depth are paramount. In the angle-encoding Hadamard test for cosine similarity, normalized real vectors \(v,w\) satisfy
\[
Similarity(v,w)=v\cdot w=\sum_{i=1}^d v_i w_i.
\]
Each scalar \(x\in[-1,1]\) is mapped through
\[
\theta_x = 2\arccos(x),
\qquad
U = R_y^\dagger(\theta_x)R_y(\theta_y),
\]
and the Hadamard test yields \(\Re \langle 0|U|0\rangle\). After a first-order approximation of the square-root correction term, the estimator becomes
\[
Similarity(v,w) = \sum_{i=1}^d \Re\langle 0|U_i|0\rangle - d + 1.
\]
The induced bias is defined as
\[
Bias \coloneqq \left(\sum_{i=1}^d \Re \langle 0|U_i|0\rangle \right) - d + 1 - \sum_{i=1}^d v_iw_i,
\]
and is shown to be non-negative. In experiments on random normalized vectors, RMSE decreases and correlation increases with dimension: for \(d=2,4,8,12\), the reported RMSE values are \(0.8012\), \(0.3479\), \(0.1500\), and \(0.0879\), with corresponding correlations \(0.6449\), \(0.7857\), \(0.9301\), and \(0.9642\) [2604.15867].

A different angle-based line of work studies quantum numerical integration through the structure of the angle map
\[
\Theta_g(b)=2\arcsin\!\bigl(\sqrt{g(x_{i(b)})}\bigr),
\qquad b\in\{0,1\}^n.
\]
Here the ancilla amplitude oracle is implemented concretely by angle encoding, since
\[
|0\rangle \mapsto \sqrt{1-g(x_i)}\,|0\rangle+\sqrt{g(x_i)}\,|1\rangle
\]
is prepared by an \(R_Y\) rotation. The hierarchy
\[
\mathcal G_n^{(d)}=\{g:\mathcal X_n\to[0,1]\;:\;\deg(\Theta_g)\le d\}
\]
implies an exact factorization into
\[
\sum_{k=0}^d \binom{n}{k}
\]
multi-controlled \(R_Y\) gates. Combined with QAE, this yields total gate count
\[
O\!\left((\log(1/\varepsilon))^d\,\varepsilon^{-1}\right),
\]
with special case \(d=1\) giving \(O(\varepsilon^{-1}\log(1/\varepsilon))\). The same work proves an unconditional separation for a family with Sobolev regularity \(s<1/2\), where the quantum oracle cost is \(O(1/\varepsilon)\) while deterministic or randomized classical quadrature requires \(\Omega(\varepsilon^{-1/s})\) evaluations [2604.24289].

Amplitude encoding has generated a parallel literature on approximate loading. The Adaptive Interpolating Quantum Transform replaces fixed Fourier truncation with a learned transform-and-truncate workflow:
\[
\mathbf{y}(\mathbf{x}) = U\mathbf{x},
\qquad
\tilde{y}_j(\mathbf{x}) = 
\begin{cases}
y_j(\mathbf{x}), & j \in \mathcal{K}(\mathbf{x}) \\
0, & j \notin \mathcal{K}(\mathbf{x}),
\end{cases}
\]
followed by sparse amplitude encoding and inverse transform. The method preserves \(O(N\log N)\) classical complexity and \(O(n^2)\) quantum gate count, while reporting a \(40\%\) reduction in reconstruction error on financial time-series data and up to \(50\%\) reduction on image datasets at matched sparsity [2603.03803].

Amplitude encoding also appears in quantum algorithms beyond supervised learning. For high-dimensional linear SDEs,
\[
\mathrm{d} \mathbf{X}_t = A(t) \mathbf{X}_t \mathrm{d} t + B(t) \mathrm{d} \mathbf{W}_t,
\]
the solution vector is encoded as
\[
|\mathbf{x}\rangle = \sum_{j=1}^N x_j |j\rangle
\]
up to normalization, and a quantum PRNG circuit is used to amplitude-encode the noise term. Both the Dyson-series-based and Euler–Maruyama-based methods make only \({\rm polylog}(N)\) queries to the PRNG circuit and the block-encodings of \(A\) and \(B\) [2604.24133].

## 4. Empirical behavior in quantum learning architectures

In QRNNs, angle encoding was initially dominant because it is simple and shallow, but comparative results have increasingly highlighted the representational advantages of amplitude encoding when state preparation is made practical. One QRNN study reports that the base amplitude-encoded QRNN generalizes better than the angle-encoded QRNN on Yahoo Finance and Oxford-Man forecasting tasks, with benchmark test MSEs of about \(0.015\) for Angle QRNN and \(0.009\) for Amplitude QRNN, and parameter counts of about \(44\) and \(28\), respectively. The same work shows that adding a pre-normalized magnitude feature improves validation MSE from about \(0.0088\) to \(0.0067\) under MinMax scaling and to about \(0.0056\) under MaxMin scaling, while approximate amplitude preparation through EnQode attains average fidelity about \(0.94\) on smaller state sizes and becomes competitive with exact QSP under IBM Torino noise, with noisy MSEs about \(0.014\) for exact QSP and \(0.013\) for EnQode [2508.16784].

A hybrid quantum neural network for recovery-rate prediction provides a high-dimensional benchmark. With \(256=2^8\) features, amplitude encoding loads the full input into an 8-qubit state, whereas the angle baseline compresses 256 classical inputs to 8 auxiliary outputs and then applies \(R_x(x_i)\) rotations. On a dataset of 1,725 observations and 256 features from 1996 to 2023, the reported best average test RMSEs are \(0.228\) for the amplitude-encoded model, \(0.242\) for the angle-encoded model, and \(0.246\) for a matched FNN baseline [2501.15828].

QCNN results are more regime-dependent. Under depolarizing noise on aggressively downsampled \(4\times 4\) MNIST, angle encoding reaches \(81.250\%\) at low noise and remains comparatively robust as noise increases, while the hybrid phase/angle encoder lags. At \(8\times 8\) MNIST, however, the hybrid scheme reaches \(75.000\%\) at mid noise, whereas angle encoding records \(56.250\%\), indicating a crossover once more feature bandwidth is available. Amplitude-encoded QCNNs are described as sparsely represented in the downsampled grids but strong in lightweight and full-resolution configurations, with training dynamics that closely resemble classical convergence [2512.12512].

Variational-classifier studies reinforce the same point from a different angle: the embedding is a hyperparameter. Under identical model topologies on Wine and Diabetes, one comparative study reports that the difference in accuracy between the best and worst models ranges from \(10\%\) to \(30\%\), with differences reaching up to \(41\%\). On Wine at 10 layers without re-uploading, the best reported model is angle encoding with `RY` at \(97.5\%\), while `RX-RY-RZ` gives \(56.39\%\), a gap of \(41.11\%\). A separate Qiskit VQC study reports that a hybrid amplitude/angle/phase feature map reaches \(0.95\) accuracy with COBYLA and \(0.90\) with L_BFGS_B on a synthetic six-feature classification task, compared with \(0.61\) and \(0.62\) for standard VQC encoding [2508.00768; 2503.14062].

## 5. Temporal, learned, and online extensions

Several recent proposals start from the premise that both standard angle and standard amplitude maps are static encodings. SPATE, or Spiking-Phase Adaptive Temporal Encoding, converts real-valued tabular data into leaky integrate-and-fire spike trains and maps spike statistics into
\[
R_X(\alpha_i),\qquad R_Z(\phi_i),\qquad \mathrm{CRZ}(\lambda\,\beta_{i,k})
\]
operations on feature and temporal qubits. Its motivation is explicit: static angle and amplitude encodings do not directly model intensity/rate, timing/phase, or coarse temporal patterns. Under the encoding-centric evaluation protocol, SPATE reports CKTA \(0.9661\) and Fisher \(7.3651\) on Blobs, compared with CKTA \(0.6322\) and Fisher \(0.7045\) for angle encoding and CKTA \(0.7192\) and Fisher \(1.3296\) for amplitude encoding; on Moons, SPATE reaches CKTA \(0.5057\), while angle and amplitude both give \(0.0145\). The principal counterexample is Circles, where angle encoding is clearly stronger, with CKTA \(0.8176\) versus \(0.2284\) for SPATE [2604.11022].

Online quantum reservoir computing provides a different extension of amplitude encoding. A measurement-enabled protocol implements the partial-trace update of amplitude-encoded QRC through mid-circuit measurement and reset, with the averaged effect
\[
\rho_k = \ket{\mathbf{0}}\bra{\mathbf{0}}\otimes \mathrm{Tr}_R(\tilde{\rho}_k).
\]
The method is explicitly online, avoids input buffering, and keeps runtime linear in the number of time steps. A proof-of-principle implementation on IBM’s **ibm_basquecountry** Heron r2 device uses \(R=1\) input qubit and \(M=5\) memory qubits, and evaluates Santa Fe prediction and short-term memory tasks [2606.18991].

Learned amplitude encoding has also become a design object in its own right. Quantum Visual Fields construct an energy spectrum from coordinates and latent code,
\[
\boldsymbol{P} = \frac{\exp\left(-\beta \boldsymbol{E}(\boldsymbol{\Theta}, z)\right)}{\mathcal{Z}},
\]
then set real amplitudes by \(\alpha_i=\sqrt{P_i}\) with \(\operatorname{arg}(\alpha_i)=0\). The prepared state is processed by a fully entangled real-Hilbert-space PQC and decoded through local projective measurements
\[
V = \operatorname{Tr}(\hat{\rho} \hat{M}(\boldsymbol{\theta})).
\]
In 2D image representation, the reported gain over QIREN is about \(30\%\) MSE and \(1.6\) dB PSNR, with improved learning of high-frequency details [2508.10900].

## 6. Broader meanings of angle and amplitude encoding

Outside QML, the terms “angle” and “amplitude” often denote encoded physical variables rather than feature maps. Angle-multiplexed metasurfaces provide a clear example. Reflective high-contrast dielectric U-shaped meta-atoms are designed so that the same local unit cell has independent reflection coefficients
\[
r_1 = |r_1|e^{i\phi_1}, \qquad r_2 = |r_2|e^{i\phi_2}
\]
at \(0^\circ\) and \(30^\circ\) incidence. This enables one surface to encode two independent phase masks, such as \(0^\circ \rightarrow\) Caltech logo and \(30^\circ \rightarrow\) LMI logo. The work explicitly distinguishes this angle-selective phase encoding from amplitude encoding: amplitude is present through \(|r_1|\) and \(|r_2|\), but is not the primary encoded degree of freedom [1711.02265].

In atomic interferometry, the encoded quantities can be the angle and amplitude of a collective spin oscillator. The transverse spin components are written as
\[
(F_y,F_z)=\rho(-\sin\psi,\cos\psi),
\]
so that \(\psi\) is the spin angle and \(\rho\) the transverse amplitude. Using dynamically-decoupled QND Faraday probing, simultaneous tracking achieved steady-state sensitivities \(2.9\) dB below the standard quantum limit for angle fluctuations and \(7.0\) dB below Poisson statistics for amplitude fluctuations [1702.08888].

A third contrast arises when amplitude is deliberately excluded from the inference stage. In snapshot radio SLAM, an amplitude-independent method estimates 6-D UE pose and clock bias from angle-delay measurements
\[
\mathbf z_i = [c\tilde\tau_i,\tilde{\bm\phi}_i,\tilde{\bm\theta}_i]
\]
using the unified residual
\[
\mathbf r_i(\mathbf x)\doteq \mathbf M_i\mathbf t-d_i(\mathbf v_i-\mathbf u_i),
\]
thereby avoiding amplitude-based LoS preclassification or path-wise latent variables in the coarse stage [2607.04847].

Taken together, these literatures portray angle and amplitude encoding not as interchangeable defaults but as a recurring design axis. In quantum information processing, the axis is governed by qubit compression versus preparation depth, with strong downstream consequences for fidelity, trainability, and noise tolerance. In adjacent physical and engineering settings, the same terms denote different but structurally related choices about which variables are represented directly, which are inferred indirectly, and which are intentionally omitted.

Source: https://www.emergentmind.com/topics/angle-amplitude-encoding