Quantum Hyperdimensional Computing

• Quantum Hyperdimensional Computing is a framework that maps high-dimensional hypervectors to quantum states, enabling native implementations of binding, bundling, permutation, and similarity estimation using quantum circuits.
• The approach employs techniques like phase oracles, LCU with OAA, and the Quantum Fourier Transform to recreate classical hyperdimensional operations on quantum hardware for efficient data encoding and compression.
• Recent developments extend QHDC to one-pass training, density-matrix based classification, and qudit-based architectures, offering promising advancements in quantum-enhanced neuromorphic computing.

Searching arXiv for papers on Quantum Hyperdimensional Computing and closely related qudit/high-dimensional quantum computing work. Quantum Hyperdimensional Computing (QHDC) denotes a family of approaches that brings Hyperdimensional Computing (HDC) into the quantum domain by representing hypervectors as quantum states and by realizing the core HDC primitives through native quantum processes. In the foundational formulation, a $D$-dimensional hypervector is mapped to an $N$-qubit state with $D=2^N$, binding is implemented by diagonal phase oracles, bundling by a Linear Combination of Unitaries (LCU) protocol with Oblivious Amplitude Amplification (OAA), permutation by the Quantum Fourier Transform (QFT) and phase shifts, and similarity by interferometric fidelity estimation via the Hadamard Test. Related work extends this program toward one-pass quantum-enhanced classification, hypervector decoding and factorization, and high-dimensional qudit platforms, so the term now covers both a specific quantum-native HDC formalism and a broader research direction at the intersection of HDC, quantum machine learning, and high-dimensional quantum information processing (Cumbo et al., 16 Nov 2025, Xu et al., 21 Jun 2026, Poduval et al., 2024, Giraldo-Carvajal et al., 2021).

1. Conceptual basis and relation to classical HDC

HDC, also described as a brain-inspired model and as a robust computational framework inspired by human cognition, represents information by very long hypervectors and manipulates them with simple algebraic operators. In the formulation used by QHDC, a classical data item is encoded as a random high-dimensional bipolar vector $v\in\{\pm1\}^D$, and the three primitive operations are binding, bundling, and permutation; retrieval and classification rely on vector similarity. HDC also appears in interpretable learning and information retrieval, where complex objects are represented from atomic concepts through high-dimensional vectors and efficient operators (Cumbo et al., 16 Nov 2025, Poduval et al., 2024, Xu et al., 21 Jun 2026).

The central claim of QHDC is that these primitives admit a direct, resource-efficient correspondence with quantum operations. For a bipolar hypervector $v=(v_0,\dots,v_{D-1})\in\{\pm1\}^D$, one chooses $N=\lceil\log_2 D\rceil$ and defines

$$|+\rangle_N=\frac{1}{\sqrt{D}}\sum_{i=0}^{D-1}|i\rangle,\qquad O_v=\sum_{i=0}^{D-1} v_i |i\rangle\langle i|.$$

The encoded quantum state is

$$|\psi_v\rangle = O_v |+\rangle_N = \frac{1}{\sqrt{D}}\sum_{i=0}^{D-1} v_i |i\rangle.$$

This construction interprets the hypervector as the phase profile of an $N$-qubit state, yielding what the foundational paper describes as a truly quantum-native neuromorphic architecture (Cumbo et al., 16 Nov 2025).

A second line of work, QeHDC, preserves the HDC emphasis on efficient training and robust prototypes but replaces bipolar-state preparation by a sinusoidal amplitude encoding and replaces direct prototype summation by density-matrix-based superclass generation. This broadens the meaning of QHDC from a single mapping to a class of quantum-enhanced HDC frameworks that differ in encoding and aggregation strategy while retaining the HDC logic of symbolic composition and similarity-based inference (Xu et al., 21 Jun 2026).

2. Algebraic primitives and their quantum realizations

The foundational QHDC paper presents a one-for-one correspondence between classical HDC operations and quantum circuits (Cumbo et al., 16 Nov 2025).

Classical HDC primitive	Quantum realization	Representative mechanism
Hypervector	Quantum state	Phase oracle on $|+\rangle_N$
Binding	Phase multiplication	Sequential diagonal phase oracles
Bundling	Coherent averaging	LCU + OAA
Permutation	Cyclic shift	QFT + phase shifts + inverse QFT
Similarity	State overlap estimation	Hadamard Test

Binding is the most direct mapping. If $N$0 is the classical Hadamard-product binding of two bipolar hypervectors, then quantumly

$N$1

with $N$2. On each computational basis state $N$3, the phases multiply, reproducing the HDC binding rule exactly. This is the sense in which QHDC treats binding as a native phase-composition primitive rather than as a variationally learned transformation (Cumbo et al., 16 Nov 2025).

Bundling is less trivial because classical HDC forms a normalized prototype from sums such as $N$4. QHDC realizes the corresponding coherent state

$N$5

through LCU with an ancilla register and then amplifies the success amplitude of the desired ancilla branch using OAA. The reflection operators are

$N$6

and the OAA iterate is

$N$7

This construction is central to the claim that HDC bundling can be implemented as a quantum-native averaging process rather than only approximated classically (Cumbo et al., 16 Nov 2025).

Permutation, which in classical HDC often appears as a cyclic shift used to encode order and structure, is implemented by diagonalizing the shift in the Fourier basis:

$N$8

The implementation described uses QFT on $N$9 qubits, single-qubit $D=2^N$0 rotations with angles $D=2^N$1, and inverse QFT, with depth $D=2^N$2 and $D=2^N$3 two-qubit gates (Cumbo et al., 16 Nov 2025).

Similarity, the quantum analogue of cosine similarity in HDC, is measured by the Hadamard Test. For states $D=2^N$4 and $D=2^N$5,

$D=2^N$6

so

$D=2^N$7

This makes retrieval and classification interferometric rather than purely algebraic, but the role in the HDC pipeline remains the same: similarity drives nearest-prototype decisions (Cumbo et al., 16 Nov 2025).

3. One-pass training, quantum binding, and superclass construction

QeHDC proposes a different operationalization of quantum-enhanced HDC, explicitly framed as a one-pass training method. Training proceeds in a single sweep over the data: each sample is encoded into a quantum state, bound to a fixed reference state through a small quantum circuit, accumulated into a class-wise density matrix, and then converted into a class template by eigen-decomposition. There are no gradients and no epochs; the method computes one template state per class after exactly one pass through the training set (Xu et al., 21 Jun 2026).

Its encoder is called Cross-Multiplicative Encoding. For $D=2^N$8, one chooses a random projection $D=2^N$9 with $v\in\{\pm1\}^D$0 and a bias $v\in\{\pm1\}^D$1 drawn from $v\in\{\pm1\}^D$2, then computes

$v\in\{\pm1\}^D$3

followed by

$v\in\{\pm1\}^D$4

The complex vector

$v\in\{\pm1\}^D$5

is normalized into

$v\in\{\pm1\}^D$6

The stated purpose is to project classical data into quantum amplitude states efficiently while using only one weight matrix and ensuring approximately uniform phase angles across the amplitudes (Xu et al., 21 Jun 2026).

QeHDC’s binding operation is not the same as the diagonal phase-oracle binding of the foundational QHDC paper. Instead, it uses a fixed reference amplitude state

$v\in\{\pm1\}^D$7

and a binding unitary $v\in\{\pm1\}^D$8 acting on two $v\in\{\pm1\}^D$9-qubit registers. The circuit has three stages: controlled-$v=(v_0,\dots,v_{D-1})\in\{\pm1\}^D$0 gates from qubits $v=(v_0,\dots,v_{D-1})\in\{\pm1\}^D$1 to $v=(v_0,\dots,v_{D-1})\in\{\pm1\}^D$2 with angle $v=(v_0,\dots,v_{D-1})\in\{\pm1\}^D$3, single-qubit $v=(v_0,\dots,v_{D-1})\in\{\pm1\}^D$4 rotations on the $v=(v_0,\dots,v_{D-1})\in\{\pm1\}^D$5 register, and entangling $v=(v_0,\dots,v_{D-1})\in\{\pm1\}^D$6 gates among the $v=(v_0,\dots,v_{D-1})\in\{\pm1\}^D$7 qubits. The resulting bound state is

$v=(v_0,\dots,v_{D-1})\in\{\pm1\}^D$8

The paper describes this as a compressed entangled state encoding the pair $v=(v_0,\dots,v_{D-1})\in\{\pm1\}^D$9 (Xu et al., 21 Jun 2026).

Superclass generation is based on class density matrices. For class $N=\lceil\log_2 D\rceil$0 with states $N=\lceil\log_2 D\rceil$1,

$N=\lceil\log_2 D\rceil$2

After eigendecomposition,

$N=\lceil\log_2 D\rceil$3

the simplest version keeps the dominant eigenvector

$N=\lceil\log_2 D\rceil$4

Classification of a test sample uses fidelity to the class templates,

$N=\lceil\log_2 D\rceil$5

On real hardware, amplitudes are not read out directly; the procedure instead reconstructs the test density matrix by quantum state tomography and computes $N=\lceil\log_2 D\rceil$6 (Xu et al., 21 Jun 2026).

4. Decoding and factorization as a QHDC subproblem

A distinct but closely related direction concerns decoding rather than classification. “Hyperdimensional Quantum Factorization” introduces HDQF as a quantum algorithm for efficiently decoding hypervectors, defined as the extraction of atomic elements from hypervectors. The problem becomes especially difficult when the target hypervector is the product, or binding, of multiple hypervectors, because factorization is described as prohibitively costly with classical optimization-based methods and specialized recurrent networks, an inherent consequence of the binding operation (Poduval et al., 2024).

HDQF addresses this by exploiting parallels between HDC and quantum computing and by encoding potential factors as a quantum superposition using qubit states and bipolar vector representation. The reported consequence is a quadratic speedup over classical search methods, together with mitigation of Hypervector Factorization capacity issues. Since the available description is limited to the abstract, the technical details of the circuit decomposition, noise model, and complexity proof are not specified here beyond that claim (Poduval et al., 2024).

This suggests that QHDC is not restricted to prototype learning or quantum similarity search. A plausible implication is that the same quantum-native treatment of HDC primitives can support symbolic decomposition tasks in which a composite hypervector must be unbound or factorized into constituent atomic concepts. That role is important for interpretable learning and information retrieval, where decoding is often the operational bottleneck (Poduval et al., 2024).

5. Qudits, simulators, and high-dimensional physical substrates

Although much of the current QHDC discussion is qubit-based, related work on qudits provides a broader high-dimensional quantum context. QuantumSkynet introduces a cloud-hosted C++ simulator for qudit-based quantum algorithms and proposes a unified generalization of high-dimensional quantum gates. A qudit is defined as a $N=\lceil\log_2 D\rceil$7-dimensional system with Hilbert space $N=\lceil\log_2 D\rceil$8 and basis $N=\lceil\log_2 D\rceil$9, with arbitrary pure state

$$|+\rangle_N=\frac{1}{\sqrt{D}}\sum_{i=0}^{D-1}|i\rangle,\qquad O_v=\sum_{i=0}^{D-1} v_i |i\rangle\langle i|.$$0

The paper emphasizes two principal advantages of $$|+\rangle_N=\frac{1}{\sqrt{D}}\sum_{i=0}^{D-1}|i\rangle,\qquad O_v=\sum_{i=0}^{D-1} v_i |i\rangle\langle i|.$$1: information capacity, since a single qudit carries $$|+\rangle_N=\frac{1}{\sqrt{D}}\sum_{i=0}^{D-1}|i\rangle,\qquad O_v=\sum_{i=0}^{D-1} v_i |i\rangle\langle i|.$$2 bits of information, and enhanced error resilience, since high-dimensional Pauli errors can be corrected by generalized stabilizer codes with a larger error threshold (Giraldo-Carvajal et al., 2021).

QuantumSkynet’s gate model includes the generalized Pauli operators

$$|+\rangle_N=\frac{1}{\sqrt{D}}\sum_{i=0}^{D-1}|i\rangle,\qquad O_v=\sum_{i=0}^{D-1} v_i |i\rangle\langle i|.$$3

the Weyl operators, the $$|+\rangle_N=\frac{1}{\sqrt{D}}\sum_{i=0}^{D-1}|i\rangle,\qquad O_v=\sum_{i=0}^{D-1} v_i |i\rangle\langle i|.$$4-dimensional Fourier gate

$$|+\rangle_N=\frac{1}{\sqrt{D}}\sum_{i=0}^{D-1}|i\rangle,\qquad O_v=\sum_{i=0}^{D-1} v_i |i\rangle\langle i|.$$5

and the SUM gate

$$|+\rangle_N=\frac{1}{\sqrt{D}}\sum_{i=0}^{D-1}|i\rangle,\qquad O_v=\sum_{i=0}^{D-1} v_i |i\rangle\langle i|.$$6

The set $$|+\rangle_N=\frac{1}{\sqrt{D}}\sum_{i=0}^{D-1}|i\rangle,\qquad O_v=\sum_{i=0}^{D-1} v_i |i\rangle\langle i|.$$7 is described as universal for qudit quantum computation. QuantumSkynet packages these constructions in an AWS Lambda service behind an API Gateway endpoint and reports noiseless simulations of qudit Deutsch–Jozsa and qudit quantum phase estimation, including 100% success for Deutsch–Jozsa at $$|+\rangle_N=\frac{1}{\sqrt{D}}\sum_{i=0}^{D-1}|i\rangle,\qquad O_v=\sum_{i=0}^{D-1} v_i |i\rangle\langle i|.$$8 and exact phase extraction for QPE at $$|+\rangle_N=\frac{1}{\sqrt{D}}\sum_{i=0}^{D-1}|i\rangle,\qquad O_v=\sum_{i=0}^{D-1} v_i |i\rangle\langle i|.$$9 (Giraldo-Carvajal et al., 2021).

For QHDC, the significance of this qudit work is methodological. The simulator is presented as the first cloud-accessible testbed for prototyping qudit algorithms at arbitrary dimension $$|\psi_v\rangle = O_v |+\rangle_N = \frac{1}{\sqrt{D}}\sum_{i=0}^{D-1} v_i |i\rangle.$$0, and its implications explicitly include exploring new hyperdimensional algorithms and quantum machine learning encodings that leverage richer Hilbert spaces for data embedding. This suggests that QHDC need not be confined to binary amplitude or phase encodings on qubits; qudit registers provide a direct route toward higher-radix hyperdimensional representations (Giraldo-Carvajal et al., 2021).

A hardware-oriented extension of the same idea appears in work on multi-photon, multi-dimensional hyper-entanglement using higher-order-radix photonic qudits. That architecture encodes a radix-$$|\psi_v\rangle = O_v |+\rangle_N = \frac{1}{\sqrt{D}}\sum_{i=0}^{D-1} v_i |i\rangle.$$1 qudit in orbital angular momentum (OAM) modes, so that the joint Hilbert-space dimension of $$|\psi_v\rangle = O_v |+\rangle_N = \frac{1}{\sqrt{D}}\sum_{i=0}^{D-1} v_i |i\rangle.$$2 qudits is $$|\psi_v\rangle = O_v |+\rangle_N = \frac{1}{\sqrt{D}}\sum_{i=0}^{D-1} v_i |i\rangle.$$3. It employs generalized single-qudit gates such as $$|\psi_v\rangle = O_v |+\rangle_N = \frac{1}{\sqrt{D}}\sum_{i=0}^{D-1} v_i |i\rangle.$$4, $$|\psi_v\rangle = O_v |+\rangle_N = \frac{1}{\sqrt{D}}\sum_{i=0}^{D-1} v_i |i\rangle.$$5, the Chrestenson gate

$$|\psi_v\rangle = O_v |+\rangle_N = \frac{1}{\sqrt{D}}\sum_{i=0}^{D-1} v_i |i\rangle.$$6

and two-qudit gates such as $$|\psi_v\rangle = O_v |+\rangle_N = \frac{1}{\sqrt{D}}\sum_{i=0}^{D-1} v_i |i\rangle.$$7 and controlled-$$|\psi_v\rangle = O_v |+\rangle_N = \frac{1}{\sqrt{D}}\sum_{i=0}^{D-1} v_i |i\rangle.$$8, while q-plate unitaries couple spin angular momentum and OAM. The paper also introduces a “quantum capacity” metric,

$$|\psi_v\rangle = O_v |+\rangle_N = \frac{1}{\sqrt{D}}\sum_{i=0}^{D-1} v_i |i\rangle.$$9

intended to compare higher-radix systems with qubit-only architectures. Because hyperdimensional representations are inherently high-cardinality, this photonic-qudit line provides a natural physical backdrop for future QHDC implementations, though that connection is an extrapolation from the reported high-dimensional encoding and gate-set properties (Ashrafi et al., 2020).

6. Empirical status, practical bottlenecks, and open questions

Experimental evidence for QHDC is mixed but substantive. The foundational QHDC paper reports validation on symbolic analogical reasoning and supervised classification across three platforms: classical computation, ideal quantum simulation, and IBM hardware. In the analogical reasoning task “USA : Dollar = Mexico : ?”, a codebook of 9 bipolar hypervectors was encoded in $N$0 using $N$1 qubits, and a noise-free AerSimulator implementation of LCU+OAA with $N$2 and $N$3 rounds reproduced the correct ranking, with “Peso” having the highest fidelity. For supervised classification on MNIST digits $N$4 after downscaling to $N$5 and binarization, the reported QHDC results were F1$N$6 in noise-free simulation at $N$7, F1$N$8 in noisy simulation, and hardware F1$N$9 at circuit depth 510; reducing the dimension to $|+\rangle_N$0 improved hardware F1 to $|+\rangle_N$1 at depth 126. On the same binary MNIST task in noise-free simulation, QHDC at $|+\rangle_N$2 is reported to run in $|+\rangle_N$3 for 5-fold cross-validation versus $|+\rangle_N$4 for VQC and QSVC, corresponding to a $|+\rangle_N$5 speedup, while full quantum bundling for the $|+\rangle_N$6 analogical circuit was infeasible on hardware because transpiled depth exceeded $|+\rangle_N$7 (Cumbo et al., 16 Nov 2025).

QeHDC reports a different empirical profile focused on low-dimensional classification under NISQ constraints. The experiments use ISOLET, MNIST, and UCI HAR, restricted to 16D, 32D, and 64D, and compare against classical HDC baselines including VanillaHD, AdaptHD, OnlineHD, NeuralHD, and QuantHD. The reported binary results for QeHDC-Aer were ISOLET: $|+\rangle_N$8, MNIST: $|+\rangle_N$9, and UCI HAR: $N$00 at 16D/32D/64D. On IBM QPU for a 4-qubit small-batch UCI HAR setting with Train$N$01/Test$N$02, the reported values were Accuracy $N$03 and Fidelity $N$04 for QeHDC-Aer, versus Accuracy $N$05 and Fidelity $N$06 for QeHDC-QPU. The paper states that multi-class results from 3 to 10 classes likewise show that QeHDC retains higher accuracy than most classical HDCs under low-dimensional constraints (Xu et al., 21 Jun 2026).

Framework	Setting	Reported result
QHDC	Analogical reasoning, $N$07, $N$08	Correct ranking; “Peso” highest fidelity
QHDC	MNIST $N$09, $N$10	F1$N$11 sim; $N$12 noisy sim; $N$13 hardware
QHDC	MNIST $N$14, $N$15	Hardware F1$N$16
QeHDC-Aer	UCI HAR, Train$N$17/Test$N$18	Accuracy $N$19; Fidelity $N$20
QeHDC-QPU	UCI HAR, Train$N$21/Test$N$22	Accuracy $N$23; Fidelity $N$24

Several practical misconceptions can therefore be resolved precisely. QHDC is not identical to generic variational quantum machine learning: the foundational formulation explicitly avoids iterative optimization by using direct mappings of HDC primitives, and QeHDC uses one-pass training rather than gradient-based updates. Conversely, QHDC is not yet a uniformly end-to-end quantum workflow on contemporary devices: the foundational classification experiments required hybrid classical-quantum bundling, and QeHDC relies on classical eigen-decomposition of class density matrices and may require tomography overhead, which in the worst case scales as $N$25 measurement bases, although compressed sensing or partial tomography are proposed to reduce the burden (Cumbo et al., 16 Nov 2025, Xu et al., 21 Jun 2026).

The main bottlenecks are also explicit. In the foundational QHDC work, the LCU+OAA bundling protocol is the dominant NISQ limitation, with depth growing linearly in $N$26 and becoming impractical for large prototype sets; performance degrades rapidly when circuit depth exceeds approximately 300 gates on ibm_pittsburgh. In QeHDC, the quantum circuit per sample is lighter—$N$27 controlled-$N$28 gates, $N$29 $N$30 gates, and up to $N$31 entangling $N$32 gates—but superclass generation still requires classical eigendecomposition of a $N$33 matrix per class, and real-device inference may require tomography. At the same time, QeHDC explicitly argues that binding compresses classical hypervectors of dimension $N$34 down onto $N$35 physical qubits, converting $N$36 classical cost into $N$37 quantum-gate cost, and describes this as exponential dimensional compression suited to near-term devices (Cumbo et al., 16 Nov 2025, Xu et al., 21 Jun 2026).

Taken together, the literature presents QHDC as an emerging research area rather than a settled architecture. One branch emphasizes exact algebra-to-circuit correspondence and neuromorphic symbolism; another emphasizes one-pass, density-matrix-based classification under NISQ constraints; a third addresses hypervector factorization and decoding; and adjacent qudit and photonic work points toward richer high-dimensional substrates. The unifying theme is that hyperdimensional representations and quantum state spaces share enough structural affinity that binding, bundling, permutation, retrieval, and even factorization can be reformulated as quantum operations with explicit circuit-level semantics (Cumbo et al., 16 Nov 2025, Xu et al., 21 Jun 2026, Poduval et al., 2024, Giraldo-Carvajal et al., 2021, Ashrafi et al., 2020).