Quantum Time-series Transformer

• Quantum Time-series Transformer (QTT) is a quantum-enhanced architecture that efficiently models spatio-temporal fMRI data using operations like LCU and QSVT.
• It replaces explicit self-attention with quantum procedures, reducing parameter counts and computational complexity to polylogarithmic levels.
• Empirical results show QTT converges quickly, generalizes well in small-sample scenarios, and yields interpretable neuroimaging biomarkers.

The Quantum Time-series Transformer (QTT) is a quantum-enhanced transformer architecture introduced for modeling spatio-temporal resting-state fMRI data. In the formulation reported in "Resting-state fMRI Analysis using Quantum Time-series Transformer" (Park et al., 31 Aug 2025), it replaces explicit classical self-attention with a quantum pipeline built from Linear Combination of Unitaries (LCU) and Quantum Singular Value Transformation (QSVT), and is presented as operating with polylogarithmic computational complexity while using substantially fewer trainable parameters than classical transformer baselines. The reported empirical setting is neuroimaging, specifically the Adolescent Brain Cognitive Development Study and the UK Biobank, where the model achieved comparable or superior predictive performance relative to classical transformer models, with especially pronounced gains in small-sample scenarios (Park et al., 31 Aug 2025).

1. Concept and problem setting

QTT is motivated by a specific limitation of classical transformer models for time-series analysis: classical self-attention incurs quadratic complexity, and in the fMRI setting this is coupled to large parameter counts and substantial data requirements (Park et al., 31 Aug 2025). The proposed architecture targets these barriers in the context of resting-state functional magnetic resonance imaging, where the task is to model complex spatio-temporal dependencies and predict phenotypes such as biological sex, ADHD diagnosis, and fluid intelligence (Park et al., 31 Aug 2025).

In the cited formulation, QTT is not presented as a minor modification of a standard encoder block. Rather, it adapts the transformer paradigm to the quantum computing domain by recasting sequence embedding, temporal mixing, nonlinearity, and downstream prediction through a sequence of quantum operations followed by classical post-processing (Park et al., 31 Aug 2025). A central conceptual point is that QTT does not compute an explicit $L \times L$ attention matrix. Instead, it uses quantum entanglement and superposition to implicitly encode dependencies across time points, thereby avoiding direct pairwise attention score construction (Park et al., 31 Aug 2025).

A common misconception is to equate any “quantum transformer” with a straightforward quantum analog of $QK^\top$ attention. QTT is explicitly different: the paper states that self-attention is mimicked implicitly through quantum entanglement and the LCU procedure, without explicit pairwise attention score calculation (Park et al., 31 Aug 2025).

2. Architectural formulation

The reported QTT pipeline consists of five stages (Park et al., 31 Aug 2025).

First, Quantum Sequence Embedding linearly maps classical time-series inputs to angular parameters via a weight matrix $W_E$. These angles parameterize gates in a variational quantum circuit (VQC), embedding the data into quantum states (Park et al., 31 Aug 2025).

Second, Quantum Mixing via LCU combines the parametrized unitary operations associated with temporal embeddings. This is intended to create a quantum superposition akin to weighting or attending to different time points (Park et al., 31 Aug 2025). The LCU primitive is described by

$M = \sum_j b_j U_j, \qquad \sum_j |b_j| \leq 1.$

Third, Nonlinearity via QSVT applies polynomial nonlinear transformations to the mixed quantum state (Park et al., 31 Aug 2025). With a degree-$d$ polynomial

$P_{\vec{c}}(x)=c_d x^d+\cdots+c_0,$

the corresponding transformation is

$$P_{\vec{c}}(M)=c_d M^d + c_{d-1} M^{d-1} + \cdots + c_1 M + c_0 I.$$

In the paper’s framing, this plays a role analogous to nonlinearities used in classical attention mechanisms, but implemented through quantum polynomial transformations rather than standard neural activations (Park et al., 31 Aug 2025).

Fourth, Quantum “Attention” via Composition is realized without explicit $QK^\top$ score computation. Dependencies across all time points are encoded implicitly through the composition of superposition, entanglement, and unitary evolution (Park et al., 31 Aug 2025).

Fifth, a Feed-forward Unitary and Measurement stage applies a further unitary transformation $U_{FF}$, after which quantum measurements such as expectation values $\langle H \rangle$ over Hermitian observables are obtained and then passed to a classical layer $QK^\top$0 for final prediction (Park et al., 31 Aug 2025).

Several differences from classical transformers are explicit in the reported design. QTT encodes temporal sequences directly into quantum states, thereby bypassing the need for explicit positional embeddings; it uses quantum operations rather than large-scale matrix multiplication for the core mixing stage; and it realizes nonlinear transformations through QSVT rather than softmax and standard neural activations (Park et al., 31 Aug 2025).

3. Complexity and statistical considerations

The principal computational claim attached to QTT is that, whereas classical self-attention for a sequence of length $QK^\top$1 has $QK^\top$2 complexity, the combination of LCU and QSVT enables a reduction to $QK^\top$3 because neither the attention matrix nor large intermediate results are classically instantiated (Park et al., 31 Aug 2025).

Operation	Classical complexity	Quantum complexity
Computing attention scores	$QK^\top$4	$QK^\top$5
Applying nonlinear transformations	$QK^\top$6	$QK^\top$7
Applying to value matrix $QK^\top$8	$QK^\top$9	$W_E$0

The paper also reports a generalization bound from QML theory: for $W_E$1 parameterized local quantum channels and $W_E$2 training points,

$W_E$3

This is presented as supporting improved generalization and tractability for QML models, particularly in small-sample regimes (Park et al., 31 Aug 2025).

In the empirical comparison, parameter efficiency is a major part of the argument. QTT is reported to use 22K trainable parameters, whereas the classical transformer baselines in the study range from 1.68–11.2 million (Park et al., 31 Aug 2025). A plausible implication is that the claimed small-sample behavior is tied not only to quantum sequence processing but also to this large disparity in parameter count.

4. Empirical evaluation on resting-state fMRI

The reported experiments use two large resting-state fMRI resources: the ABCD Study with approximately 12,000 children aged 9–10, and the UK Biobank with approximately 41,000 adults aged 40–77 (Park et al., 31 Aug 2025). The evaluated phenotypes are biological sex and ADHD diagnosis as binary tasks, and fluid intelligence as a regression task (Park et al., 31 Aug 2025).

Under full-sample evaluation, QTT is reported to achieve the lowest MAE for fluid intelligence on ABCD and the highest AUROC for sex classification in UK Biobank, while being comparable to strong classical transformer baselines such as BolT on other tasks (Park et al., 31 Aug 2025). The reported numbers are:

• ABCD sex classification: Vanilla Transformer $W_E$4, BrainNet Transformer $W_E$5, BolT $W_E$6, QTT $W_E$7 AUROC.
• ABCD ADHD classification: Vanilla Transformer $W_E$8, BrainNet Transformer $W_E$9, BolT $M = \sum_j b_j U_j, \qquad \sum_j |b_j| \leq 1.$0, QTT $M = \sum_j b_j U_j, \qquad \sum_j |b_j| \leq 1.$1 AUROC.
• ABCD fluid intelligence: Vanilla Transformer $M = \sum_j b_j U_j, \qquad \sum_j |b_j| \leq 1.$2, BrainNet Transformer $M = \sum_j b_j U_j, \qquad \sum_j |b_j| \leq 1.$3, BolT $M = \sum_j b_j U_j, \qquad \sum_j |b_j| \leq 1.$4, QTT $M = \sum_j b_j U_j, \qquad \sum_j |b_j| \leq 1.$5 MAE.
• UK Biobank sex classification: Vanilla Transformer $M = \sum_j b_j U_j, \qquad \sum_j |b_j| \leq 1.$6, BrainNet Transformer $M = \sum_j b_j U_j, \qquad \sum_j |b_j| \leq 1.$7, BolT $M = \sum_j b_j U_j, \qquad \sum_j |b_j| \leq 1.$8, QTT $M = \sum_j b_j U_j, \qquad \sum_j |b_j| \leq 1.$9 AUROC.
• UK Biobank fluid intelligence: all listed models report $d$0 MAE, including QTT (Park et al., 31 Aug 2025).

The most distinctive reported results come from the small-sample setting with $d$1, described as “Tiny Data Generalization.” Here QTT is reported to outperform all classical baselines across all tasks and datasets (Park et al., 31 Aug 2025). The reported values are:

• ABCD sex classification: Vanilla Transformer $d$2, BrainNet Transformer $d$3, BolT $d$4, QTT $d$5 AUROC.
• ABCD ADHD classification: Vanilla Transformer $d$6, BrainNet Transformer $d$7, BolT $d$8, QTT $d$9 AUROC.
• ABCD fluid intelligence: Vanilla Transformer $P_{\vec{c}}(x)=c_d x^d+\cdots+c_0,$0, BrainNet Transformer $P_{\vec{c}}(x)=c_d x^d+\cdots+c_0,$1, BolT $P_{\vec{c}}(x)=c_d x^d+\cdots+c_0,$2, QTT $P_{\vec{c}}(x)=c_d x^d+\cdots+c_0,$3 MAE.
• UK Biobank sex classification: Vanilla Transformer $P_{\vec{c}}(x)=c_d x^d+\cdots+c_0,$4, BrainNet Transformer $P_{\vec{c}}(x)=c_d x^d+\cdots+c_0,$5, BolT $P_{\vec{c}}(x)=c_d x^d+\cdots+c_0,$6, QTT $P_{\vec{c}}(x)=c_d x^d+\cdots+c_0,$7 AUROC.
• UK Biobank fluid intelligence: Vanilla Transformer $P_{\vec{c}}(x)=c_d x^d+\cdots+c_0,$8, BrainNet Transformer $P_{\vec{c}}(x)=c_d x^d+\cdots+c_0,$9, BolT $$P_{\vec{c}}(M)=c_d M^d + c_{d-1} M^{d-1} + \cdots + c_1 M + c_0 I.$$0, QTT $$P_{\vec{c}}(M)=c_d M^d + c_{d-1} M^{d-1} + \cdots + c_1 M + c_0 I.$$1 MAE (Park et al., 31 Aug 2025).

The paper further states that QTT converges more quickly and exhibits a smaller generalization error—defined there as the difference between train and validation/test—than the classical models (Park et al., 31 Aug 2025).

5. Interpretability and biomarker analysis

An important part of the reported QTT study is its interpretability pipeline. SHapley Additive exPlanations (SHAP) values are computed using occlusion-based approaches, with beeswarm plots used to highlight the top 20 brain regions contributing to ADHD classification (Park et al., 31 Aug 2025).

The regions identified by QTT are reported to include Prefrontal and Opercular Areas such as Anterior 24, Area 8Ad, Frontal Opercular 2/3, Area 45, and Area 47s; Intra-parietal and Limbic Structures such as Medial/Lateral Intraparietal, Area 25, PreSubiculum, and ParaHippocampal Area 1; and Visual and Multisensory Areas such as Lateral Occipital 2 and TemporoParietoOccipital Junction 2 (Park et al., 31 Aug 2025). The paper states that these regions align with established ADHD neurobiology, involving executive control, attention, reward processing, and sensory integration (Park et al., 31 Aug 2025).

This interpretability layer matters because quantum-enhanced forecasting and classification models are often assumed to be intrinsically opaque. In the reported QTT formulation, interpretability is not omitted as an afterthought; it is integrated into the empirical evaluation, and the identified biomarkers are presented as clinically meaningful (Park et al., 31 Aug 2025).

6. Relation to other quantum time-series transformer models and open issues

QTT belongs to a broader family of quantum or hybrid quantum-classical models for temporal learning, but the insertion point of the quantum computation varies substantially across the literature.

In QASA, reported in "Quantum Adaptive Self-Attention for Quantum Transformer Models" (Chen et al., 5 Apr 2025), the earlier encoder layers remain classical and only the final encoder layer replaces dot-product attention with a parameterized quantum circuit and a residual quantum projection. On a synthetic damped oscillator task, QASA is reported to reach MSE = 0.0085 and MAE = 0.0679, compared with MSE = 0.0122 and MAE = 0.0916 for a complexity-matched classical variant and MSE = 0.5188 and MAE = 0.3946 for a plain transformer (Chen et al., 5 Apr 2025).

In QCAAPatchTF, described in "Integrating Quantum-Classical Attention in Patch Transformers for Enhanced Time Series Forecasting" (Chakraborty et al., 31 Mar 2025), the encoder alternates Quantum Self-Attention and Full Classical Self-Attention inside an advanced patch-based transformer, with the quantum attention stage using superposition, entanglement, and variational quantum eigensolver principles to compute trainable, data-dependent attention scores (Chakraborty et al., 31 Mar 2025).

In QTFT, introduced in "Quantum Temporal Fusion Transformer" (Barik et al., 6 Aug 2025), the quantumization target is broader within the classical TFT pipeline: Gated Residual Networks, Variable Selection Networks, Static Covariate Encoders, and the Interpretable Multi-Head Attention Mechanism are replaced with variational quantum circuits, while LSTM cells can remain classical or be replaced by QLSTM (Barik et al., 6 Aug 2025). On the reported stock-market benchmark, QTFT with QLSTM achieves Train Loss 0.1711 and Test Loss 0.8007, compared with 0.2630 and 0.9856 for classical TFT (Barik et al., 6 Aug 2025).

In iQTransformer, presented in "Quantum Neural Network Architectures for Multivariate Time-Series Forecasting" (Ranilla-Cortina et al., 24 Oct 2025), the model integrates a quantum self-attention mechanism into the iTransformer framework, using inverted tokenization so that variables rather than time steps are treated as tokens. On the reported real-world wind-energy dataset, the long-term MAPE is 0.0849 for iQTransformer versus 0.0874 for classical iTransformer, with approximately 5,295 trainable parameters versus approximately 11,445 (Ranilla-Cortina et al., 24 Oct 2025).

These related systems clarify that “quantum time-series transformer” is not a single architectural recipe. Some models quantumize attention alone, some quantumize feed-forward or gating blocks, and QTT specifically reformulates temporal mixing and nonlinearity through LCU and QSVT rather than through a direct quantum analog of classical attention matrices (Park et al., 31 Aug 2025). This suggests that the field is still exploring which transformer substructures benefit most from quantum enhancement.

The open issues reported in this literature are also specific. For QTT, the authors note challenges such as handling raw high-dimensional data natively in quantum hardware and generalizing to diverse populations (Park et al., 31 Aug 2025). For iQTransformer, the current implementation is simulated rather than executed on real quantum hardware, and hardware noise is identified as a possible performance factor (Ranilla-Cortina et al., 24 Oct 2025). For QTFT, current limitations include circuit depth, number of qubits, and noise (Barik et al., 6 Aug 2025). A plausible implication is that, although the strongest claims concern efficiency, small-sample generalization, and parameter reduction, much of the current evidence remains tied to hybrid or simulation-centric regimes rather than large-scale end-to-end deployment on fault-tolerant quantum hardware.