Equivariant Quantum Clustering (EQC)

• EQC is a framework that integrates symmetry-aware quantum encoding with differential privacy to enable unsupervised clustering of sensitive datasets.
• It employs p4m-equivariant variational quantum circuits combined with kernelized spectral clustering, enhancing parameter efficiency and robustness.
• Empirical evaluations on datasets like NSL-KDD demonstrate that parameter sharing and hybrid quantum-classical encoding significantly improve privacy-utility performance.

Equivariant Quantum Clustering (EQC) denotes a family of clustering methods in which quantum encoders, variational circuits, kernels, or readouts are constrained by symmetry, so that inputs related by a group action are processed consistently. In the most specific current usage, EQC refers to a symmetry-aware, differentially private, parameter-efficient clustering framework that integrates $p4m$-equivariant parameter sharing into a variational quantum circuit and combines it with privacy-preserving kernelized spectral clustering for sensitive tabular data (Haq et al., 9 Jul 2026). The broader concept is rooted in earlier work on equivariant quantum neural networks, equivariant quantum graph circuits, symmetry-preserving embeddings, and measurement-efficient equivariant QCNNs, most of which study representation learning or classification rather than clustering directly (Nguyen et al., 2022, Mernyei et al., 2021, Arnott et al., 2024, Chinzei et al., 2024).

1. Definition, scope, and terminology

In its narrowest sense, EQC is the framework introduced in “Equivariant Quantum Clustering with Differential Privacy: Parameter-Efficient Privacy-Preserving Analysis Across Heterogeneous Sensitive Datasets” (Haq et al., 9 Jul 2026). That framework targets privacy-preserving unsupervised clustering of sensitive datasets and combines three ingredients: a hybrid classical/quantum-style feature encoding, a symmetry-constrained variational quantum circuit with strong parameter sharing, and end-to-end differential privacy accounting across preprocessing, kernel computation, and clustering. Its reported application domains are cybersecurity logs, insider-threat behavior traces, and clinical records, specifically NSL-KDD, CERT Insider Threat v6.2, and a synthetic MIMIC-III clinical dataset (Haq et al., 9 Jul 2026).

In a broader research sense, EQC describes clustering pipelines in which equivariance or invariance is built into the quantum representation-learning stage. This broader meaning does not correspond to a single canonical algorithm in the pre-2026 literature. Rather, it is assembled from several technical strands: exact equivariant quantum channels and QCNNs, permutation-equivariant circuits for graph data, norm-sensitive equivariant embeddings for classical data, learned equivariant maps with VQCs, and split-parallel equivariant QCNNs that improve measurement efficiency (Nguyen et al., 2022, Mernyei et al., 2021, Arnott et al., 2024, Bradshaw et al., 2024, Chinzei et al., 2024).

The acronym is not historically unique. In a distinct line of work on the Traveling Salesman Problem, “EQC” denotes an Equivariant Quantum Circuit rather than Equivariant Quantum Clustering, and the two usages should not be conflated (Teo et al., 13 Nov 2025).

2. Formal symmetry principles

The general quantum notion of equivariance used in this area is representation-theoretic. If $G$ is a group with input and output representations $R^{\mathrm{in}}$ and $R^{\mathrm{out}}$, a quantum map $\phi$ is equivariant when

$$\phi \circ \operatorname{Ad}_{R^{\mathrm{in}}(g)}=\operatorname{Ad}_{R^{\mathrm{out}}(g)}\circ \phi,$$

equivalently,

$$\phi\big(R^{\mathrm{in}}(g)\rho R^{\mathrm{in}}(g)^\dagger\big)=R^{\mathrm{out}}(g)\phi(\rho)R^{\mathrm{out}}(g)^\dagger.$$

An operator $O$ is symmetry-compatible when it lies in the commutant,

$[O,R(g)]=0 \qquad \forall g\in G,$

and equivariant quantum neural networks plus equivariant measurements yield invariant scalar outputs (Nguyen et al., 2022).

A closely related circuit-level formulation appears in equivariant QCNNs. For a parametrized circuit $U(\theta)$ and a group action $G$0, equivariance is imposed by

$G$1

If the observable is also symmetric,

$G$2

then

$G$3

inherits the corresponding label symmetry (Chinzei et al., 2024).

For clustering, the direct supervised target $G$4 is absent, but the same symmetry logic applies to latent representations, similarities, and cluster assignments. A direct reinterpretation, stated explicitly for the QCNN setting, is that if cluster identity should not change under $G$5, then feature maps or pairwise affinities should satisfy the same invariance; equivalently, one may require learned embeddings, similarities, or distances to be invariant or equivariant under the group action (Chinzei et al., 2024). This suggests two distinct design regimes: an invariant regime, in which all points in one orbit should share a cluster identity, and an equivariant latent regime, in which orbit structure is preserved until a final invariant readout.

3. Architectural lineage and precursor frameworks

The most general theoretical foundation is the theory of equivariant quantum neural networks. That framework provides three explicit construction methods for equivariant layers and channels—nullspace, twirling, and Choi-operator methods—and uses them to build group-equivariant QCNNs for finite groups and compact Lie groups, including $G$6-equivariant pooling and convolution. It also characterizes equivariant operators through commutants and isotypic decompositions, making symmetry-compatible latent processing a representation-theoretic design problem rather than an ad hoc regularization heuristic (Nguyen et al., 2022).

A second important precursor is the family of reverse map projection embeddings $G$7. These embeddings were introduced to correct a limitation of amplitude embedding, namely that positive scalar multiples are identified and norm information is lost. Their defining equivariance theorem states that for any unitary representation $G$8 on $G$9,

$R^{\mathrm{in}}$0

where the extra coordinate transforms trivially and encodes norm information. Because the last coordinate depends only on $R^{\mathrm{in}}$1 and $R^{\mathrm{in}}$2, these embeddings preserve directional information while retaining a radial degree of freedom, which is directly relevant when clustering depends on magnitude as well as direction (Arnott et al., 2024).

For graph-structured data, Equivariant Quantum Graph Circuits provide a permutation-equivariant circuit law,

$R^{\mathrm{in}}$3

together with node-state preparation, graph-dependent unitary processing, node-wise measurement, and invariant classical aggregation. On bounded graph domains, EDU-QGCs are universal approximators for graph functions, and they can simulate MPNNs with sum aggregation. This makes them a natural representation-learning backbone for graph clustering or node clustering, even though the paper itself studies graph representation learning rather than clustering (Mernyei et al., 2021).

Another strand learns the equivariant embedding itself. “Learning Equivariant Maps with Variational Quantum Circuits” studies learning an intertwiner between a data-space representation $R^{\mathrm{in}}$4 and a Hilbert-space representation $R^{\mathrm{in}}$5, formalized by

$R^{\mathrm{in}}$6

Its operational message is that symmetry-respecting encoders need not be fully hand-specified; they can be learned approximately with a variational ansatz and an equivariance loss, then followed by invariant post-embedding circuits and measurements (Bradshaw et al., 2024). A plausible implication is that future EQC systems may combine hard architectural symmetry constraints with softly learned symmetry-adapted embeddings.

Finally, resource-efficient equivariant QCNNs contribute an efficiency result that is especially pertinent to clustering. The equivariant split-parallelizing QCNN replaces symmetry-breaking discard pooling with symmetry-preserving branch splitting, achieving order-$R^{\mathrm{in}}$7 improvements in the measurement efficiency of expectation estimation and gradient estimation in favorable regimes, while retaining high trainability and favorable barren-plateau behavior. The paper studies noisy supervised classification, not clustering, but it explicitly argues that the mechanism transfers naturally to settings where many embeddings, similarities, or cluster statistics must be evaluated repeatedly (Chinzei et al., 2024).

4. The differentially private EQC framework

The 2026 EQC framework is defined as a symmetry-aware, differentially private quantum clustering pipeline. Its end-to-end organization is stated as: privacy-preserving preprocessing, hybrid quantum encoding, equivariant quantum circuit transformation, secure quantum kernel computation, and privacy-aware spectral clustering (Haq et al., 9 Jul 2026).

Stage	Mechanism	Concrete specification
Preprocessing	Feature reduction and privatization	NSL-KDD: $R^{\mathrm{in}}$8 features, one-hot encoding, MinMax scaling, autoencoder $R^{\mathrm{in}}$9, Gaussian noise, quantization
Encoding	Hybrid amplitude-angle map	$R^{\mathrm{out}}$0
Circuit	$R^{\mathrm{out}}$1-equivariant VQC	8 qubits on a $R^{\mathrm{out}}$2 grid, depth $R^{\mathrm{out}}$3, gate set $R^{\mathrm{out}}$4
Kernel	Overlap-based similarity with DP release	$R^{\mathrm{out}}$5, then $R^{\mathrm{out}}$6
Clustering	Private spectral clustering	$R^{\mathrm{out}}$7, privacy-preserving eigendecomposition, then DP k-means

For NSL-KDD, the preprocessing pipeline is unusually specific. Raw features are reduced from 41 to 15 dimensions by removing redundant or near-constant attributes, categorical features are one-hot encoded, MinMax scaling is applied, and a nonlinear autoencoder reduces the data to 8 latent dimensions using the architecture

$R^{\mathrm{out}}$8

with ReLU activations, Adam, learning rate $R^{\mathrm{out}}$9, 100 epochs, and batch size 64. Gaussian noise is added as

$\phi$0

and the result is quantized to three decimal places. CERT and synthetic MIMIC-III undergo normalization, imputation, and feature selection, although the paper gives fewer exact details for those datasets (Haq et al., 9 Jul 2026).

The quantum encoding is hybrid. For the first $\phi$1 qubits, the framework uses amplitude encoding,

$\phi$2

and for the remaining $\phi$3 qubits it uses angle encoding,

$\phi$4

The full encoded state is

$\phi$5

The circuit then applies an 8-qubit variational ansatz on a $\phi$6 grid with four alternating layers and hardware-efficient gates $\phi$7 (Haq et al., 9 Jul 2026).

The symmetry mechanism is parameter tying under $\phi$8 orbits. The paper operationalizes equivariance through equality constraints such as

$\phi$9

and similarly for edge parameters,

$$\phi \circ \operatorname{Ad}_{R^{\mathrm{in}}(g)}=\operatorname{Ad}_{R^{\mathrm{out}}(g)}\circ \phi,$$0

The unconstrained circuit has 112 independent parameters; equivariant sharing reduces this to 24, organized into rotational, reflectional, and combined orbit families (Haq et al., 9 Jul 2026).

The downstream clustering stage is kernel spectral clustering. Pairwise similarities are computed from a modified swap-test-style overlap,

$$\phi \circ \operatorname{Ad}_{R^{\mathrm{in}}(g)}=\operatorname{Ad}_{R^{\mathrm{out}}(g)}\circ \phi,$$1

then perturbed as

$$\phi \circ \operatorname{Ad}_{R^{\mathrm{in}}(g)}=\operatorname{Ad}_{R^{\mathrm{out}}(g)}\circ \phi,$$2

with $$\phi \circ \operatorname{Ad}_{R^{\mathrm{in}}(g)}=\operatorname{Ad}_{R^{\mathrm{out}}(g)}\circ \phi,$$3 and sensitivity $$\phi \circ \operatorname{Ad}_{R^{\mathrm{in}}(g)}=\operatorname{Ad}_{R^{\mathrm{out}}(g)}\circ \phi,$$4. The normalized Laplacian is

$$\phi \circ \operatorname{Ad}_{R^{\mathrm{in}}(g)}=\operatorname{Ad}_{R^{\mathrm{out}}(g)}\circ \phi,$$5

the $$\phi \circ \operatorname{Ad}_{R^{\mathrm{in}}(g)}=\operatorname{Ad}_{R^{\mathrm{out}}(g)}\circ \phi,$$6 smallest eigenvectors are extracted with privacy-preserving eigendecomposition, and a differentially private k-means is run on the spectral embedding (Haq et al., 9 Jul 2026).

The objective couples utility and privacy: $$\phi \circ \operatorname{Ad}_{R^{\mathrm{in}}(g)}=\operatorname{Ad}_{R^{\mathrm{out}}(g)}\circ \phi,$$7 The paper notes no explicit sign convention for minimization or maximization, so the printed formula is best read as a joint utility-privacy criterion rather than an unambiguous optimization statement. A second caveat is methodological: the operational task is clustering, but labels are used not only for Hungarian-matched evaluation but also through NMI in the optimization objective. This makes the implementation hybrid or weakly supervised in a strict methodological sense, even though the final output is a clustering (Haq et al., 9 Jul 2026).

5. Empirical performance, privacy, and ablation evidence

The headline result is reported on NSL-KDD at privacy budget $$\phi \circ \operatorname{Ad}_{R^{\mathrm{in}}(g)}=\operatorname{Ad}_{R^{\mathrm{out}}(g)}\circ \phi,$$8, $$\phi \circ \operatorname{Ad}_{R^{\mathrm{in}}(g)}=\operatorname{Ad}_{R^{\mathrm{out}}(g)}\circ \phi,$$9. Under that budget, EQC achieves $$\phi\big(R^{\mathrm{in}}(g)\rho R^{\mathrm{in}}(g)^\dagger\big)=R^{\mathrm{out}}(g)\phi(\rho)R^{\mathrm{out}}(g)^\dagger.$$0 clustering accuracy, $$\phi\big(R^{\mathrm{in}}(g)\rho R^{\mathrm{in}}(g)^\dagger\big)=R^{\mathrm{out}}(g)\phi(\rho)R^{\mathrm{out}}(g)^\dagger.$$1 ARI, $$\phi\big(R^{\mathrm{in}}(g)\rho R^{\mathrm{in}}(g)^\dagger\big)=R^{\mathrm{out}}(g)\phi(\rho)R^{\mathrm{out}}(g)^\dagger.$$2 NMI, $$\phi\big(R^{\mathrm{in}}(g)\rho R^{\mathrm{in}}(g)^\dagger\big)=R^{\mathrm{out}}(g)\phi(\rho)R^{\mathrm{out}}(g)^\dagger.$$3 membership-inference attack success, and $$\phi\big(R^{\mathrm{in}}(g)\rho R^{\mathrm{in}}(g)^\dagger\big)=R^{\mathrm{out}}(g)\phi(\rho)R^{\mathrm{out}}(g)^\dagger.$$4 attribute inference error (Haq et al., 9 Jul 2026). Selected baselines under matched privacy budgets are substantially weaker on the same benchmark: Spectral Clustering reports 57.8 accuracy and 75.8 privacy leakage, VQC reports 63.5 accuracy and 65.4 leakage, DP-K-means reports 43.5 accuracy and 52.1 leakage, and DiffP-Spectral reports 49.3 accuracy and 50.5 leakage (Haq et al., 9 Jul 2026).

The privacy-utility curve is one of the paper’s central empirical claims. On NSL-KDD, EQC reports 68.7 accuracy and 32.5 privacy leakage at $$\phi\big(R^{\mathrm{in}}(g)\rho R^{\mathrm{in}}(g)^\dagger\big)=R^{\mathrm{out}}(g)\phi(\rho)R^{\mathrm{out}}(g)^\dagger.$$5, 75.2 and 35.8 at $$\phi\big(R^{\mathrm{in}}(g)\rho R^{\mathrm{in}}(g)^\dagger\big)=R^{\mathrm{out}}(g)\phi(\rho)R^{\mathrm{out}}(g)^\dagger.$$6, 79.3 and 38.3 at $$\phi\big(R^{\mathrm{in}}(g)\rho R^{\mathrm{in}}(g)^\dagger\big)=R^{\mathrm{out}}(g)\phi(\rho)R^{\mathrm{out}}(g)^\dagger.$$7, and 81.5 and 42.7 at $$\phi\big(R^{\mathrm{in}}(g)\rho R^{\mathrm{in}}(g)^\dagger\big)=R^{\mathrm{out}}(g)\phi(\rho)R^{\mathrm{out}}(g)^\dagger.$$8. The same comparison shows DP-K-means rising only from 32.5 accuracy at $$\phi\big(R^{\mathrm{in}}(g)\rho R^{\mathrm{in}}(g)^\dagger\big)=R^{\mathrm{out}}(g)\phi(\rho)R^{\mathrm{out}}(g)^\dagger.$$9 to 48.2 at $O$0, while also exhibiting higher leakage (Haq et al., 9 Jul 2026).

The paper evaluates three membership-inference attacks—shadow model, threshold, and loss-based—and reports average attack success of 38.3 on NSL-KDD, with individual values 36.5, 39.8, and 38.6. Because balanced binary membership decisions would have 50% random-guessing accuracy, the paper interprets the average result as strong resistance. It further reports model inversion with average reconstruction error 0.568 MSE and privacy protection factor 2.42 relative to k-means, and attribute inference errors of 70.3, 74.2, and 73.0 across demographic, behavioral, and sensitive inference categories, for a 72.5 average (Haq et al., 9 Jul 2026).

Ablation results are unusually direct about mechanism. Removing rotational equivariance drops NSL-KDD performance to 72.5 accuracy and 45.7 leakage; removing reflectional equivariance yields 75.8 and 42.3; removing parameter sharing yields 68.7 and 52.5; removing all equivariance yields 63.5 and 65.4. A capacity-matched comparison among 24-parameter variants—random sharing, permutation equivariance, and EQC $O$1—shows 78.1/39.7, 77.4/40.5, and 79.3/38.3, respectively, with $O$2 for EQC versus random sharing. The paper explicitly concludes that parameter reduction, not $O$3 structure per se, is the primary driver of gains (Haq et al., 9 Jul 2026).

The additional datasets follow the same pattern. Full EQC reports 75.0 accuracy and 35.7 privacy leakage on CERT Insider Threat v6.2, and 70.5 accuracy and 32.8 leakage on synthetic MIMIC-III. Robustness tests under input noise give an average retention of 90.8%, exceeding the reported VQC retention of 76.9%. Hardware-noise studies using the IBM Cairo noise model show substantial degradation: from 78.4 accuracy and 38.3 MIA in the noiseless case to 68.7/48.2 at 100,000 shots, 66.5/50.1 at 50,000 shots, 62.3/54.7 at 10,000 shots, and 53.8/63.2 at 1,000 shots (Haq et al., 9 Jul 2026).

6. Limitations, interpretations, and research directions

The EQC paper is explicit that its gains are not uniquely quantum. Its own interpretation is that improvements arise primarily from parameter reduction through equivariant sharing and from differential privacy noise composition, not from uniquely quantum-mechanical privacy effects (Haq et al., 9 Jul 2026). This is reinforced by the capacity-matched ablation in which random sharing performs similarly to $O$4-sharing. The same paper also notes that $O$5 symmetry is not semantically natural for tabular data, that the experiments are largely simulation-based, and that uniquely quantum privacy effects such as no-cloning and measurement collapse are not truly active in classical simulation (Haq et al., 9 Jul 2026).

Methodological caveats remain substantial. The clustering task is operationally unsupervised, but labels enter the training objective through NMI, so the implementation is not a purely unsupervised clustering algorithm. Dataset realism is also limited: NSL-KDD is acknowledged as imperfect, and both CERT and the MIMIC-III variant are synthetic. Fairness remains unresolved, with NSL-KDD attack-class disparity reported at 18.2 percentage points. Hardware noise is a major obstacle, since both utility and privacy worsen sharply at lower shot counts (Haq et al., 9 Jul 2026).

The broader literature indicates several directions for a more mature EQC field. Exact equivariant channels, invariant commutant measurements, and group-equivariant QCNNs provide a rigorous toolkit for constructing symmetry-respecting latent maps and hierarchical pooling (Nguyen et al., 2022). Reverse map projection embeddings suggest that norm-sensitive equivariant encodings may be preferable when cluster structure depends on magnitude rather than direction alone (Arnott et al., 2024). Equivariant quantum graph circuits indicate that graph clustering and node clustering are especially natural targets because permutation symmetry is built in at the circuit level (Mernyei et al., 2021). Learned equivariant embeddings suggest that the symmetry-adapted encoder itself may be trainable rather than fully prescribed (Bradshaw et al., 2024). Measurement-efficient split-parallel equivariant QCNNs suggest that future EQC architectures could reduce the shot cost of repeated embedding, similarity, or gradient evaluations in measurement-limited unsupervised settings (Chinzei et al., 2024).

Taken together, these results support a narrow and a broad understanding of EQC. Narrowly, EQC currently names a parameter-efficient, differentially private, kernel-based variational clustering framework with strong empirical privacy-utility performance on sensitive tabular benchmarks (Haq et al., 9 Jul 2026). Broadly, it denotes a research program in which clustering is built on symmetry-aware quantum representation learning, with exact equivariance, learned equivariant embeddings, graph permutation symmetry, and measurement efficiency serving as the principal architectural axes (Nguyen et al., 2022, Mernyei et al., 2021, Bradshaw et al., 2024, Chinzei et al., 2024).