Complex Linear Combination of Unitaries (CLCU)

• CLCU is a quantum framework that generalizes standard LCUs by using arbitrary complex coefficients to encode both amplitude and phase information.
• It is implemented via block encoding and controlled operations, which facilitate accurate multi-head quantum self-attention in neural networks.
• Empirical results demonstrate improved classification accuracy on benchmarks like MNIST, while highlighting challenges in scalability and error mitigation.

A Complex Linear Combination of Unitaries (CLCU) extends the standard quantum Linear Combination of Unitaries (LCU) framework by allowing coefficients in the superposition to be arbitrary complex scalars, thereby supporting a strictly complex-valued, phase-sensitive quantum mechanism for weighted sum operations. This generalization is critical for fully quantum-native formulations of self-attention, as it captures both amplitude and phase relationships between quantum-embedded tokens, aligning model expressiveness with the inherent structure of quantum Hilbert space. The CLCU framework has recently been formalized and deployed at the core of advanced Quantum Complex-Valued Self-Attention Architectures (Chen et al., 24 Mar 2025).

1. Definition and Mathematical Framework

The CLCU operator generalizes real-weighted LCUs by preparing quantum operations of the form

$A = \frac{1}{\Omega'}\sum_{j=0}^{N-1} \alpha_j U_j$

where $\alpha_j = |\alpha_j| e^{i\theta_j} \in \mathbb{C}$ are complex-valued coefficients, $U_j$ are unitary operators, and $\Omega'$ is a normalization factor dependent on the sum of amplitudes, $\Omega' = \sum_j |\alpha_j|$. The state preparation stage encodes the magnitude and the phase of these weights on an ancillary quantum register by block-encoding: amplitude via controlled-$R_y$ and phase via controlled-$R_z$ gates. Post-selection on the ancilla projects the working register onto the normalized CLCU state.

This construction permits direct quantum realization of weighted sums

$\sum_{j} \alpha_j U_j |\psi\rangle$

for arbitrary complex coefficients, a fundamental extension for precise multi-way quantum information routing and processing, especially for tasks with essential phase interference effects such as quantum self-attention.

2. Role in Quantum Self-Attention and Expressivity

Within the Quantum Complex-Valued Self-Attention Model (QCSAM) (Chen et al., 24 Mar 2025), CLCUs are used to generate multi-head, phase-sensitive quantum attention outputs. The basic self-attention structure proceeds as follows:

• Encode each query, key, and value as quantum states $\{|Q_i\rangle, |K_j\rangle, |V_j\rangle\}$ via a trainable feature map circuit.
• Compute complex-valued attention weights by measuring the full inner product:

$$\alpha_{ij} = \frac{\langle K_j | Q_i \rangle}{\sum_{k=0}^{N-1} \langle K_k | Q_i \rangle} \in \mathbb{C}$$

• Synthesize the weighted sum over value states using a CLCU operator:

$$|\psi_{\text{out}}\rangle = \sum_{j=0}^{N-1} \alpha_{ij} |V_j\rangle$$

• Multi-head extension: sums across heads are realized by an additional CLCU on the set $\{|\psi_{\text{out}}^{(h)}\rangle\}$ with complex weights $\gamma_h$.

The CLCU thus provides strict preservation and manipulation of both amplitude and phase relations, facilitating expressive quantum-native similarity and routing that classical real-valued softmax-based attention, or even real-valued quantum kernels, cannot capture.

Ablation results [(Chen et al., 24 Mar 2025), Table 3, Fig. 7] demonstrate that CLCU-based complex attention weights outperform real-valued overlap techniques, with a measurable accuracy improvement (e.g., +0.10% for 3 qubits and +0.16% for 8 qubits on 3-class MNIST).

3. Quantum Circuit Implementation

The CLCU circuit consists of three core steps [(Chen et al., 24 Mar 2025), Fig. 8]:

1. State Preparation (PREP):

$$U_{\rm PREP} |0\rangle^{\otimes n} = \frac{1}{\Omega} \sum_{j} \sqrt{|\alpha_j|}\, e^{i\theta_j/2}|j\rangle$$

where $|\alpha_j|$ is encoded in the amplitude (controlled-$R_y$) and $\theta_j$ in the phase (controlled-$R_z$).

1. Conditioned Application:
   • Apply $U_j$ on the working register controlled by the $j$-th basis state of the ancilla.
2. Uncompute and Post-select:
   • Apply $U_{\rm PREP}^{T}$ and measure ancilla in $|0\rangle$, post-selecting successful runs.

The success probability scales as $1/\Omega'^2$, which can become a limiting factor on noisy intermediate-scale quantum (NISQ) hardware for large $n$; however, the procedure is asymptotically efficient for moderate $n$ (e.g., $n\leq8$ in current experiments).

4. Training and Differentiability

The complex coefficients $\alpha_j$ are parameterized as

$$\alpha_j(\boldsymbol{\theta}) = \cos(\theta_j)\,e^{i\phi_j}$$

with $\theta_j,\phi_j$ variational parameters. Gradients with respect to these parameters are estimated by the parameter-shift rule, ensuring hardware-efficient, gradient-based optimization: $$\frac{\partial\mathcal L}{\partial\phi_j} = \frac{\mathcal L(\phi_j+\frac{\pi}{2})-\mathcal L(\phi_j-\frac{\pi}{2})}{2}$$ and analogously for $\theta_j$.

The use of CLCUs in training enables end-to-end learning of phase-sensitive attention weights in quantum neural models. Optimization is typically performed with Adam; shot noise and post-selection probability are empirically managed by moderate batch sizes and multiple seeds (Chen et al., 24 Mar 2025).

5. Scalability, Complexity, and Hardware Constraints

The depth of individual CLCU blocks scales linearly with the number of qubits $n$ plus the number of controlled unitary operations. Circuit overhead comprises one $U_{\rm PREP}$, a multi-controlled $U_j$ tree, and a final $U_{\rm PREP}^{T}$ per CLCU, giving overall depth $O(L\,n+H\,\mathrm{depth}(\text{CLCU}))$ per layer, where $L$ is feature map depth and $H$ is the number of attention heads (Chen et al., 24 Mar 2025). Post-selection reduces effective throughput, and ancilla overhead becomes significant at $n\gtrsim8$.

Current experimental deployments of CLCU-QCSAMs on datasets such as MNIST and Fashion-MNIST demonstrate test accuracies of 100% and 99.2% for 4-qubit circuits, and stable improvement with increasing qubits and heads [(Chen et al., 24 Mar 2025), Table 1, Fig. 5–6]. Ancilla and post-selection costs currently limit scaling on NISQ devices, motivating continued work in resource-efficient block-encoding and error mitigation.

6. Connections to Other Quantum and Classical Attention Mechanisms

The CLCU construction is directly contrasted with standard LCUs, which restrict linear combinations to real (non-negative) coefficients and thus cannot represent genuine quantum phase interference in superpositions generated by quantum self-attention. In kernel-based quantum transformers such as SASQuaTCh (Evans et al., 2024), complex-valued SU(2) gates in Fourier space enable operator-valued quantum kernels, but no explicit CLCU mechanism is constructed. The use of CLCUs is unique as it implements weighted sums where both amplitude and phase can be simultaneously optimized across all value branches, yielding richer hypothesis classes.

Hybrid quantum-classical variants (e.g., QCSAM (Smaldone et al., 26 Feb 2025)) encode only inner product similarities as complex-valued quantities, but final aggregation over value states returns to classical matrix multiplication, not a quantum-native complex-weighted sum. Purely real-valued or measurement-based quantum attention (e.g., QSAN, QSANN) lack the ability to route both amplitude and phase exactly (Shi et al., 2022, Li et al., 2022, Shi et al., 2023).

By employing block-encoded CLCUs for multi-way aggregations, QCSAM with CLCU stands as the only model family to date executing fully quantum, complex-weighted self-attention (Chen et al., 24 Mar 2025).

7. Open Challenges and Prospective Developments

Current hardware limits on ancilla, circuit depth, and post-selection success probability present practical barriers to deploying large-scale CLCU-based architectures on NISQ devices. Proposed directions include error-mitigation strategies for CLCUs, dynamic ancilla resource allocation for complex multi-head attention, and optimized block encoding circuits. The demonstrated expressivity and scalability gains suggest that, as hardware matures, CLCUs may become indispensable for quantum-native sequence modeling, generative modeling, and other complex machine learning workloads demanding full quantum mechanical expressivity (Chen et al., 24 Mar 2025).

References:

• Quantum Complex-Valued Self-Attention Model (Chen et al., 24 Mar 2025)
• Learning with SASQuaTCh: a Novel Variational Quantum Transformer Architecture (Evans et al., 2024)
• A Hybrid Transformer Architecture with a Quantized Self-Attention Mechanism (Smaldone et al., 26 Feb 2025)
• QSAN: A Near-term Achievable Quantum Self-Attention Network (Shi et al., 2022)
• Quantum Self-Attention Neural Networks for Text Classification (Li et al., 2022)