QWHA: Diverse Quantum & Adaptation Methods

• QWHA is an acronym with multiple definitions, including a Hamming-weight-preserving variational ansatz, a quantization-aware Walsh-Hadamard adapter for fine-tuning, and a quantum water health assessment pipeline.
• The Hamming-weight-preserving ansatz reduces effective state-space dimensionality, enhances trainability by mitigating barren-plateau effects, and enables intrinsic error detection.
• Quantization-aware Walsh-Hadamard adaptation leverages fast transform techniques to efficiently refine quantized models, while quantum-walk hash approaches expand design space in cryptographic contexts.

Searching arXiv for exact acronym usage and the provided ids to disambiguate the topic. QWHA is an acronym used in recent arXiv literature for several technically distinct constructs rather than a single standardized concept. In current usage, it denotes a Hamming-weight-preserving variational quantum ansatz for symmetry-constrained VQE, a quantization-aware Walsh-Hadamard adaptation method for parameter-efficient fine-tuning of quantized LLMs, and a quantum water health assessment pipeline based on QSVC and QNN models; in the quantum-walk hash literature, it also appears as a field label surrounding controlled alternate quantum-walk hashing (Yan et al., 2024, Jeon et al., 22 Sep 2025, Khan et al., 2024, Zhou et al., 2021).

Usage	Research object	arXiv id
QWHA	Hamming-weight-preserving ansatz on $\mathcal{H}_w$	(Yan et al., 2024)
QWHA	Quantization-Aware Walsh-Hadamard Adaptation	(Jeon et al., 22 Sep 2025)
QWHA	Quantum Water Health Assessment	(Khan et al., 2024)
QWHA field/design space	Quantum-walk hash setting referenced by QHFM	(Zhou et al., 2021)

1. Terminological scope

The acronym is not used uniformly across the cited works. In one setting, an ansatz $U(\theta)$ is called Hamming-weight preserving if every layer $U_l=e^{i\theta_l H_l}$ has a generator $H_l$ that acts block-diagonally on the computational basis and mixes only basis states of the same Hamming weight, so that the full circuit preserves the $w$-excitation subspace $\mathcal{H}_w$ (Yan et al., 2024). In another, QWHA expands to “Quantization-Aware Walsh-Hadamard Adaptation,” a transform-domain PEFT mechanism inserted in parallel with a frozen quantized weight matrix (Jeon et al., 22 Sep 2025). In a third, the water-quality study explicitly frames its pipeline as a Quantum Water Health Assessment system built around quantum kernels and small-data classification (Khan et al., 2024).

The quantum-walk hash paper uses the term differently again. It states that QHFM “advances the QWHA field” and broadens the “QWHA design space” by combining controlled alternate quantum walks with unequal memory length (Zhou et al., 2021). This suggests that “QWHA” is currently paper-specific terminology whose meaning must be inferred from immediate context, especially when reading titles, abstracts, or code repositories.

2. Hamming-weight-preserving QWHA in variational quantum algorithms

In the variational-quantum usage, an $n$-qubit computational-basis state is $|x\rangle \equiv |x_1\ldots x_n\rangle$ with $x_i\in\{0,1\}$, and its Hamming weight is $w(x)\equiv \sum_{i=1}^n x_i$, the number of ones in the bit string (Yan et al., 2024). The relevant invariant subspace is

$U(\theta)$0

A Hamming-weight-preserving ansatz confines evolution to this subspace, which directly exploits particle-number or spin-conservation symmetry in chemistry and condensed-matter settings.

The fixed-weight restriction has several stated consequences. First, it yields dimensionality reduction, since $U(\theta)$1 when $U(\theta)$2. Second, it improves trainability: gradient scales are reported as $U(\theta)$3 instead of $U(\theta)$4. Third, it provides intrinsic error detection because bit-flip errors change $U(\theta)$5 and can be flagged (Yan et al., 2024). These properties make the ansatz simultaneously symmetry-preserving and resource-aware, which is the central motivation for its use in near-term VQE.

At the two-qubit level, the general Hamming-weight-preserving Hamiltonian on qubits $U(\theta)$6 has nonzero support only on $U(\theta)$7, with coefficients $U(\theta)$8 and $U(\theta)$9. The decomposition introduces

$U_l=e^{i\theta_l H_l}$0

together with basis matrices $U_l=e^{i\theta_l H_l}$1 so that $U_l=e^{i\theta_l H_l}$2 (Yan et al., 2024). This coefficient structure is the basis for the universality classification.

3. Universality, circuit construction, and trainability of the variational ansatz

The subspace-universality theorem states that for an HWP ansatz on $U_l=e^{i\theta_l H_l}$3 qubits restricted to the Hamming-weight-$U_l=e^{i\theta_l H_l}$4 subspace $U_l=e^{i\theta_l H_l}$5 of dimension $U_l=e^{i\theta_l H_l}$6, and assuming full connectivity, the restricted unitary $U_l=e^{i\theta_l H_l}$7 is universal on $U_l=e^{i\theta_l H_l}$8 if and only if one of the following holds: $U_l=e^{i\theta_l H_l}$9 The proof proceeds through the Dynamical Lie Algebra $H_l$0 generated by $H_l$1, with universality equivalent to $H_l$2; the paper reports that examining all 15 vanishing and non-vanishing patterns of $H_l$3 isolates exactly these two universal cases (Yan et al., 2024).

A concrete universal two-qubit gate is the “BS” gate, obtained from a coefficient choice satisfying condition (1) with $H_l$4 and $H_l$5. Its unitary is $H_l$6, and the circuit decomposition uses three CNOTs plus single-qubit $H_l$7-rotations (Yan et al., 2024). For nearest-neighbor connectivity on a linear or circular layout, one layer consists of parallel BS gates on disjoint edges $H_l$8, and full expressivity under nearest-neighbor constraints is obtained by alternating “forward” and “reverse” layers so that interactions occur on both $H_l$9 and $w$0.

The parameter count is $w$1 for $w$2 layers, since each layer introduces $w$3 real parameters. Numerical unitary-approximation results show that exact realization of arbitrary $w$4 is achieved around $w$5, consistent with overparameterization theory in subspaces; expressivity then saturates once the DLA dimension reaches $w$6 (Yan et al., 2024).

Trainability is quantified by the variance estimate

$w$7

Because $w$8 grows polynomially for fixed $w$9, the ansatz avoids the $\mathcal{H}_w$0 barren-plateau regime, and for $\mathcal{H}_w$1 the paper reports $\mathcal{H}_w$2 (Yan et al., 2024). The same work validates the ansatz in fermionic VQE via Jordan-Wigner mappings, where particle-number conservation corresponds to Hamming-weight preservation. Reported energy errors are below $\mathcal{H}_w$3 for H$\mathcal{H}_w$4, LiH, BeH$\mathcal{H}_w$5, and F$\mathcal{H}_w$6, below $\mathcal{H}_w$7 across an H$\mathcal{H}_w$8O bond-length and bond-angle surface, and $\mathcal{H}_w$9 for $n$0, $n$1, and $n$2 Fermi-Hubbard lattices even at $n$3 and with underparameterized depths $n$4 (Yan et al., 2024).

4. Quantization-Aware Walsh-Hadamard Adaptation

In the PEFT usage, QWHA expands to “Quantization-Aware Walsh-Hadamard Adaptation.” The core object is a Walsh-Hadamard Transform adapter in which the frozen quantized weight $n$5 is augmented by a learnable transform-domain correction (Jeon et al., 22 Sep 2025). If $n$6 is the normalized Walsh-Hadamard matrix with orthonormality $n$7, and $n$8 is learnable diagonal or sparse, then the adapter output on activation $n$9 is

$|x\rangle \equiv |x_1\ldots x_n\rangle$0

so that the fine-tuned layer computes

$|x\rangle \equiv |x_1\ldots x_n\rangle$1

Equivalently, the effective weight is $|x\rangle \equiv |x_1\ldots x_n\rangle$2, and back-propagation updates only the entries of $|x\rangle \equiv |x_1\ldots x_n\rangle$3 (Jeon et al., 22 Sep 2025).

The architectural motivation is that the fixed transform has entries $|x\rangle \equiv |x_1\ldots x_n\rangle$4, so its forward and backward passes can be implemented by an $|x\rangle \equiv |x_1\ldots x_n\rangle$5 fast Hadamard transform, replacing dense rank-$|x\rangle \equiv |x_1\ldots x_n\rangle$6 multiplications. The method is inserted in parallel after quantizing the original linear weight $|x\rangle \equiv |x_1\ldots x_n\rangle$7, including attention projections and feed-forward “up” and “down” sublayers in standard transformers (Jeon et al., 22 Sep 2025).

The quantization-aware initialization is posed as sparse approximation of the frozen quantization error

$|x\rangle \equiv |x_1\ldots x_n\rangle$8

The paper describes this as NP-hard and proposes a two-stage approximation. AdaAlloc assigns channel-wise sparsity budgets

$|x\rangle \equiv |x_1\ldots x_n\rangle$9

with temperature $x_i\in\{0,1\}$0, and assigns flooring remainders to the least-served channels, guaranteeing each $x_i\in\{0,1\}$1 and thus full rank in the adapter subspace. For each output channel $x_i\in\{0,1\}$2, the dense spectral coefficients $x_i\in\{0,1\}$3 are computed, the top-$x_i\in\{0,1\}$4 magnitudes are selected, and a small least-squares refinement is solved for the selected values via a single Cholesky solve per channel (Jeon et al., 22 Sep 2025).

The theoretical error analysis is expressed in terms of the layer-wise MSE with zero-mean input covariance $x_i\in\{0,1\}$5. Before adaptation,

$x_i\in\{0,1\}$6

and after inserting the adapter,

$x_i\in\{0,1\}$7

Because $x_i\in\{0,1\}$8 is orthonormal, the residual obeys

$x_i\in\{0,1\}$9

so the AdaAlloc-plus-refinement procedure directly targets the bound on post-adaptation error (Jeon et al., 22 Sep 2025).

5. Efficiency and empirical behavior of the Walsh-Hadamard adapter

The storage and FLOP profile of QWHA is defined relative to a trainable-parameter budget $w(x)\equiv \sum_{i=1}^n x_i$0. QWHA stores $w(x)\equiv \sum_{i=1}^n x_i$1 nonzeros in $w(x)\equiv \sum_{i=1}^n x_i$2, whereas LoRA of rank $w(x)\equiv \sum_{i=1}^n x_i$3 stores $w(x)\equiv \sum_{i=1}^n x_i$4; the paper explicitly notes that one can choose $w(x)\equiv \sum_{i=1}^n x_i$5 so that both methods have the same trainable-parameter count (Jeon et al., 22 Sep 2025). On a minibatch of size $w(x)\equiv \sum_{i=1}^n x_i$6 and sequence length $w(x)\equiv \sum_{i=1}^n x_i$7, the WHT adapter uses two FHTs on $w(x)\equiv \sum_{i=1}^n x_i$8, giving $w(x)\equiv \sum_{i=1}^n x_i$9 cost plus $U(\theta)$00 pointwise multiplies and lookups, while LoRA costs $U(\theta)$01. Conventional 2D-FT adapters such as DCT- and DHT-based schemes incur $U(\theta)$02 twice, once for rows and once for columns, with larger constants (Jeon et al., 22 Sep 2025).

The reported practical consequence is a 1.5–2× speedup over 2D-FT adapters and runtime within 10–20% of LoRA while matching or exceeding LoRA quality. Memory overhead is described as essentially identical to LoRA because both store $U(\theta)$03 floating-point values plus negligible fixed FHT buffers (Jeon et al., 22 Sep 2025).

The experimental evaluation covers 4-, 3-, and 2-bit quantized LLaMA-3.1-8B, LLaMA-3.2-3B, and Mistral-7B-v0.3 on zero-shot CommonsenseQA and GSM8K. For LLaMA-3.2-3B, the reported CSQA accuracies are $U(\theta)$04, $U(\theta)$05, and $U(\theta)$06 for GPTQ$U(\theta)$07 at 4, 3, and 2 bits, $U(\theta)$08, $U(\theta)$09, and $U(\theta)$10 for quant-aware LoRA, and $U(\theta)$11, $U(\theta)$12, and $U(\theta)$13 for QWHA. On GSM8K, the corresponding values are $U(\theta)$14, $U(\theta)$15, and $U(\theta)$16 for GPTQ$U(\theta)$17, $U(\theta)$18, $U(\theta)$19, and $U(\theta)$20 for LoRA, and $U(\theta)$21, $U(\theta)$22, and $U(\theta)$23 for QWHA (Jeon et al., 22 Sep 2025). The paper highlights that at 2 bits QWHA gains a 2–3 pp absolute CSQA improvement over the best LoRA-based PEFT.

Training latency is reported on Alpaca instruction fine-tuning with batch size 4 on A100, averaged over layers: LoRA takes $U(\theta)$24 h, SSH (2D DHT) $U(\theta)$25 h, LoCA (2D DCT) $U(\theta)$26 h, and WHA (1D WHT) $U(\theta)$27 h (Jeon et al., 22 Sep 2025). The combined quantized model plus adapter remains under 8 GB for 8B-scale models, and at 4 bits the end-to-end footprint is approximately 4× smaller than 16-bit float while retaining $U(\theta)$28 of full-precision accuracy on the reported benchmarks (Jeon et al., 22 Sep 2025).

6. Quantum Water Health Assessment

In the water-quality usage, QWHA refers to a quantum machine-learning pipeline for predicting water quality in the U20A region in Durban, South Africa. The study considers a QSVC and a QNN, and reports that the QSVC is easier to implement and yields higher accuracy; for the QSVC, polynomial and RBF kernels have exactly the same performance (Khan et al., 2024). The feature map is written as

$U(\theta)$29

typically via angle encoding with $U(\theta)$30 on qubit $U(\theta)$31, and the quantum kernel is

$U(\theta)$32

The paper also compares the classical kernel forms

$U(\theta)$33

within its evaluation protocol (Khan et al., 2024).

The QSVC optimization follows the standard SVM primal and dual, with a classical quadratic-program solver providing the separator once the kernel matrix is computed. The quantum contribution is restricted to kernel estimation, implemented conceptually by preparing $U(\theta)$34 and $U(\theta)$35 and estimating the overlap through repeated measurements (Khan et al., 2024). The recommended workflow is: collect classical water-quality variables, balance the dataset, choose a quantum feature map, compute the kernel on a QPU or simulator, solve the SVM decision boundary classically, and classify unseen samples by their kernel vector against the training set.

The dataset contains 32 measurement sites in the U20A catchment and features including E.coli count, NO$U(\theta)$36, NO$U(\theta)$37, SO$U(\theta)$38, turbidity, flow rate, sediment, and related variables. The label is “Acceptable” if E.coli $U(\theta)$39 MPN/100 mL and “Not Acceptable” otherwise. The class balance is severe, with $U(\theta)$40 positive and $U(\theta)$41 negative cases, so the study uses random oversampling of the minority class; no further scaling or normalization is described (Khan et al., 2024).

The reported QSVC results after oversampling are: linear kernel accuracy $U(\theta)$42, F1 $U(\theta)$43, precision $U(\theta)$44, recall $U(\theta)$45, AUROC $U(\theta)$46, AUPRC $U(\theta)$47; polynomial kernel accuracy $U(\theta)$48, F1 $U(\theta)$49, precision $U(\theta)$50, recall $U(\theta)$51, AUROC $U(\theta)$52, AUPRC $U(\theta)$53; RBF kernel accuracy $U(\theta)$54, F1 $U(\theta)$55, precision $U(\theta)$56, recall $U(\theta)$57, AUROC $U(\theta)$58, AUPRC $U(\theta)$59 (Khan et al., 2024). By contrast, the QNN persistently encountered the dead neuron problem. Its generic output is

$U(\theta)$60

with MSE loss

$U(\theta)$61

Across optimizers including Adam, gradient descent, RMSProp, and COBYLA, the network loss plateaued near $U(\theta)$62 and accuracy near $U(\theta)$63 after approximately $U(\theta)$64 epochs, with outputs collapsing to a constant value; the paper attributes this to dead-neuron behavior and vanishing or exploding gradients resembling barren-plateau effects (Khan et al., 2024).

7. QWHA as a quantum-walk hash field label

The QHFM paper uses QWHA not as a single algorithm name but as a broader field or design-space label. QHFM is described as a hash function based on controlled alternate quantum walks with memory on cycles, in which the $U(\theta)$65th message bit decides whether the walker uses one-step memory or two-step memory at that time step, and the hash value is computed from the resulting probability distribution (Zhou et al., 2021). The total Hilbert space is

$U(\theta)$66

with a cycle-position register, two direction-memory registers, and a coin qubit. The per-bit evolution is

$U(\theta)$67

and for a message $U(\theta)$68,

$U(\theta)$69

Hash extraction proceeds from the node probabilities $U(\theta)$70. With $U(\theta)$71 and a suitable integer $U(\theta)$72, each node contributes an $U(\theta)$73-bit block

$U(\theta)$74

and the full hash is the concatenation $U(\theta)$75 (Zhou et al., 2021). The paper reports time $U(\theta)$76, space $U(\theta)$77, and near-ideal statistical behavior: avalanche effects around $U(\theta)$78, diffusion and confusion with mean changed-bit ratio $U(\theta)$79 and $U(\theta)$80, flat bit-flip frequency with $U(\theta)$81 and $U(\theta)$82, and collision-resistance behavior characterized by $U(\theta)$83 and $U(\theta)$84 (Zhou et al., 2021).

Within that paper’s terminology, QHFM is stated to “advance the QWHA field” because prior quantum-walk-based hash functions differed only in coin operators, whereas QHFM alternates walks of unequal memory length and therefore adds “memory-length control” to the controlled-alternate-quantum-walk framework (Zhou et al., 2021). This usage reinforces the broader observation that QWHA functions as a contextual acronym whose exact meaning depends on the local research program: symmetry-preserving ansatz design, Walsh-Hadamard PEFT, water-quality assessment, or quantum-walk hash construction.