SPQCs: Superposed Parameterized Quantum Circuits

• SPQCs are quantum architectures that use an address register to coherently encode multiple parameterized sub-models within a single circuit.
• One formulation employs FFQRAM and repeat-until-success techniques to achieve polynomial activations, while another leverages controlled rotations for compressed ensemble inference.
• SPQCs offer resource efficiency by reducing qubit overhead through logarithmic scaling and enabling multiplexed execution of multiple models.

Searching arXiv for the cited SPQC literature and related parameterized quantum circuit context. Superposed Parameterized Quantum Circuits (SPQCs), also written “Superposed Parameterised Quantum Circuits,” are parameterized quantum-circuit architectures in which multiple parameter sets, sub-models, or ensemble members are encoded coherently within a single circuit through an address register and controlled operations. In the literature, the term has at least two closely related but operationally distinct meanings. One line of work uses flip–flop quantum random access memory (FFQRAM) and repeat-until-success (RUS) subroutines to embed an exponential family of parameterized sub-models and to induce polynomial activations on their amplitudes (Patapovich et al., 10 Jun 2025). Another uses address-conditioned uniformly controlled rotations to compress a quantum ensemble into one circuit for simultaneous multi-model inference inside Quantum DeepONet and QOrthoNN operator-learning pipelines (Matlia et al., 1 May 2026). In both cases, the defining idea is that a single coherent execution replaces a family of separate parameterized quantum circuits.

1. Definition and conceptual scope

The basic SPQC construction introduces an address register prepared in superposition and uses that register to select which parameterized unitary is applied to the data register. In the FFQRAM-based formulation, an SPQC layer prepares

$$\ket{\Phi} = \frac{1}{\sqrt{L}} \sum_{j=0}^{L-1} \ket{j}_a \otimes U(\boldsymbol{\theta}_j) S(\mathbf{x}) \ket{0}^{\otimes n},$$

where $S(\mathbf{x})$ is a data-encoding unitary and $U(\boldsymbol{\theta}_j)$ is the variational block associated with parameter set $\boldsymbol{\theta}_j$. After post-selection on a projector such as $|0\rangle\!\langle 0|^{\otimes n}$ on the data register, the address register stores amplitudes proportional to the predictions $$p_j=\bra{0}^{\otimes n}U(\boldsymbol{\theta}_j)S(\mathbf{x})\ket{0}^{\otimes n}$$ of the $L$ sub-models (Patapovich et al., 10 Jun 2025).

In the ensemble-compression formulation, the same general architecture is specialized to hardware-efficient deployment of quantum ensembles. There the joint state is written as

$$|\psi\rangle = \frac{1}{\sqrt{L}} \sum_{j=1}^{L} |j\rangle_{addr}\otimes U(\boldsymbol{\theta}^{(j)})S(\mathbf{x}^{(j)})|0_D\rangle^{\otimes n},$$

with address qubits indexing ensemble members, controlled data loaders $S(\mathbf{x}^{(j)})$, and controlled model unitaries $U(\boldsymbol{\theta}^{(j)})$. The objective is not nonlinear activation but simultaneous execution of all ensemble members, followed by modified tomography that recovers the independent outputs of all $S(\mathbf{x})$0 models (Matlia et al., 1 May 2026).

These formulations share a common core: superposition over model indices, logarithmic address overhead in the number of indexed models, and controlled routing of parameterized operations. They differ in immediate purpose. The FFQRAM-RUS SPQC is presented as a route to deeper, non-linear quantum machine learning architectures, whereas the Quantum DeepONet/QOrthoNN SPQC is presented as a resource-efficient execution strategy for ensembles.

Formulation	Core mechanism	Principal role
FFQRAM-RUS SPQC	Address-conditioned $S(\mathbf{x})$1, post-selection, RUS amplitude transformations	Multi-layer non-linear QML (Patapovich et al., 10 Jun 2025)
Ensemble SPQC	Address-conditioned $S(\mathbf{x})$2 and $S(\mathbf{x})$3, modified tomography	Compressed ensemble inference in operator learning (Matlia et al., 1 May 2026)

2. Circuit architecture and state preparation

An SPQC typically uses a data register, an address register, and ancillary qubits. In the FFQRAM architecture, the address register has $S(\mathbf{x})$4 qubits and indexes $S(\mathbf{x})$5 parameter sets. Hadamards prepare the uniform superposition

$S(\mathbf{x})$6

and a cascade of multi-controlled rotations writes $S(\mathbf{x})$7 onto the data register conditioned on $S(\mathbf{x})$8. The circuit therefore realizes what the paper explicitly characterizes as “bagging in superposition”: $S(\mathbf{x})$9 appear as amplitudes of the address register in a single quantum run (Patapovich et al., 10 Jun 2025).

In the QOrthoNN-based operator-learning setting, the architectural blocks are RBS-based unitaries that preserve the unary subspace. A single-model QOrthoNN layer uses $U(\boldsymbol{\theta}_j)$0 data qubits and one ancilla qubit; the SPQC version adds $U(\boldsymbol{\theta}_j)$1 address qubits, giving $U(\boldsymbol{\theta}_j)$2 qubits per sub-network for $U(\boldsymbol{\theta}_j)$3 transformations. The paper specifies that the circuit is realized through “a cascade of uniformly controlled rotations conditioned on address qubits initialized in a uniform superposition,” with both the data-loading unitary $U(\boldsymbol{\theta}_j)$4 and the model unitary $U(\boldsymbol{\theta}_j)$5 conditioned on the address state $U(\boldsymbol{\theta}_j)$6 (Matlia et al., 1 May 2026).

A recurrent misconception is that an SPQC is merely a larger conventional PQC. The FFQRAM paper argues directly against this interpretation: a wider or deeper single PQC still uses one parameter vector, whereas an SPQC embeds multiple independent parameter vectors $U(\boldsymbol{\theta}_j)$7 in one coherent construction. Likewise, the operator-learning formulation is not naive parallelism. Its purpose is to replace linear qubit scaling across $U(\boldsymbol{\theta}_j)$8 independent ensemble members by a single circuit with address-conditioned branches.

3. Non-linear activation, layered composition, and training

A distinctive feature of the original SPQC proposal is that it attempts to move beyond the linearity usually associated with unitary models and quantum kernel viewpoints. The mechanism is RUS plus post-selection across coherent copies of a layer. If a single layer produces amplitudes $U(\boldsymbol{\theta}_j)$9 on the address register, then duplicating the structure and post-selecting on identical outcomes yields the quadratic transformation

$\boldsymbol{\theta}_j$0

so that $\boldsymbol{\theta}_j$1. The paper presents this as a polynomial activation acting on intermediate amplitude representations, and notes that $\boldsymbol{\theta}_j$2 coherent copies can realize higher-order polynomial activations $\boldsymbol{\theta}_j$3, with depth and post-selection costs increasing accordingly (Patapovich et al., 10 Jun 2025).

This activation mechanism is central to the paper’s claim that SPQCs can approximate multi-layer, non-linear architectures rather than remaining within the strict kernel interpretation of a single measurement after a unitary. The construction therefore combines three ingredients in one model family: a coherent ensemble of sub-models, amplitude-based intermediate representations on the address register, and post-selected polynomial activation.

Training in that setting follows a standard hybrid loop. For supervised tasks, quantum runs produce classical statistics, the loss is computed classically, and all parameter sets $\boldsymbol{\theta}_j$4 are updated jointly using Adam with learning rate $\boldsymbol{\theta}_j$5. The paper states that gradients are obtained through a black-box gradient-based optimizer at the classical level, rather than deriving an SPQC-specific analytic gradient formalism (Patapovich et al., 10 Jun 2025). This places SPQCs inside the broader variational-circuit framework in which a parameterized model is trained by alternating quantum evaluation and classical optimization. A review of parameterized quantum circuits explicitly notes that quantizing parameters and preparing them in superposition enables backpropagation-like algorithms via phase kickback and tunneling, situating SPQCs within a larger family of parameter-superposition approaches (Benedetti et al., 2019).

The operator-learning SPQC uses a different training regime. There, ensemble members are trained independently and classically as exact orthogonal matrices; trained parameter sets are then transferred into controlled unitaries for inference-time deployment. The SPQC is therefore an execution mechanism rather than a distinct learning objective in that framework (Matlia et al., 1 May 2026).

4. Ensemble compression and operator learning

The most concrete systems-level use of SPQCs to date is ensemble compression in Conformalized Quantum DeepONet Ensembles. The framework combines Quantum DeepONet, QOrthoNN layers, ensemble-based epistemic modeling, and adaptive conformal prediction. A central hardware difficulty is that naive parallel execution scales qubit requirements linearly with the ensemble size. The SPQC solution is to encode all $\boldsymbol{\theta}_j$6 ensemble members into a single unified circuit and to extract the outputs of all members through modified tomography (Matlia et al., 1 May 2026).

The resource tradeoff is explicit. Naive parallelism requires $\boldsymbol{\theta}_j$7 qubits if each model uses an $\boldsymbol{\theta}_j$8-qubit QOrthoNN sub-network. The SPQC requires only $\boldsymbol{\theta}_j$9 qubits per sub-network, at the cost of linear $|0\rangle\!\langle 0|^{\otimes n}$0 circuit-depth scaling. The paper emphasizes that sequential execution on a single circuit would require $|0\rangle\!\langle 0|^{\otimes n}$1 distinct state preparations and measurement cycles, whereas the SPQC performs state preparation, unitary evolution, and measurement exactly once for the entire ensemble (Matlia et al., 1 May 2026).

Within this pipeline, SPQCs serve conformalized uncertainty quantification without changing the ensemble definition itself. The ensemble mean and standard deviation remain

$|0\rangle\!\langle 0|^{\otimes n}$2

and the SPQC merely changes how the underlying quantum sub-networks are executed. The result is a hardware-efficient realization of quantum ensembling rather than a new statistical estimator.

This deployment-oriented interpretation distinguishes the QOrthoNN SPQC from the FFQRAM-RUS SPQC. The former is fundamentally an ensemble multiplexer for a pre-trained collection of models. The latter is a trainable architecture intended to enlarge representational power through amplitude-domain nonlinearity.

5. Resource scaling, readout, and empirical performance

The scalability claims of SPQCs are defined by a qubit–depth tradeoff. In the FFQRAM architecture, embedding $|0\rangle\!\langle 0|^{\otimes n}$3 sub-models requires $|0\rangle\!\langle 0|^{\otimes n}$4 address qubits, so the total qubit count scales approximately as $|0\rangle\!\langle 0|^{\otimes n}$5 plus ancillas. FFQRAM state preparation requires $|0\rangle\!\langle 0|^{\otimes n}$6 controlled gates in general, while the non-Clifford $|0\rangle\!\langle 0|^{\otimes n}$7-depth scales as $|0\rangle\!\langle 0|^{\otimes n}$8 when RUS-based synthesis and ancilla recycling are used. RUS activations introduce a different cost: a degree-$|0\rangle\!\langle 0|^{\otimes n}$9 polynomial requires $$p_j=\bra{0}^{\otimes n}U(\boldsymbol{\theta}_j)S(\mathbf{x})\ket{0}^{\otimes n}$$0 coherent copies and a success probability that scales approximately as $$p_j=\bra{0}^{\otimes n}U(\boldsymbol{\theta}_j)S(\mathbf{x})\ket{0}^{\otimes n}$$1, thereby increasing shot requirements (Patapovich et al., 10 Jun 2025).

The original SPQC paper reports two benchmark tasks. On 1D step-function regression, an SPQC with $$p_j=\bra{0}^{\otimes n}U(\boldsymbol{\theta}_j)S(\mathbf{x})\ket{0}^{\otimes n}$$2 data qubits and $$p_j=\bra{0}^{\otimes n}U(\boldsymbol{\theta}_j)S(\mathbf{x})\ket{0}^{\otimes n}$$3 address qubits achieved MSE $$p_j=\bra{0}^{\otimes n}U(\boldsymbol{\theta}_j)S(\mathbf{x})\ket{0}^{\otimes n}$$4, MAE $$p_j=\bra{0}^{\otimes n}U(\boldsymbol{\theta}_j)S(\mathbf{x})\ket{0}^{\otimes n}$$5, and $$p_j=\bra{0}^{\otimes n}U(\boldsymbol{\theta}_j)S(\mathbf{x})\ket{0}^{\otimes n}$$6 $$p_j=\bra{0}^{\otimes n}U(\boldsymbol{\theta}_j)S(\mathbf{x})\ket{0}^{\otimes n}$$7; the parameter-matched PQC baseline achieved MSE $$p_j=\bra{0}^{\otimes n}U(\boldsymbol{\theta}_j)S(\mathbf{x})\ket{0}^{\otimes n}$$8, MAE $$p_j=\bra{0}^{\otimes n}U(\boldsymbol{\theta}_j)S(\mathbf{x})\ket{0}^{\otimes n}$$9, and $L$0 $L$1. On 2D star-shaped classification, the linear SPQC reached MSE $L$2, MAE $L$3, and accuracy $L$4, while the quadratic SPQC reached MSE $L$5, MAE $L$6, and accuracy $L$7 (Patapovich et al., 10 Jun 2025).

The operator-learning SPQC reports a different empirical profile. For a 4-member ensemble on the synthetic antiderivative operator, the unified SPQC required 7 qubits per sub-network, compared with 20 qubits under naive parallel execution. The transpiled circuit statistics reported for one sub-network were: standard circuit depth 95 with CZ 62, RX 76, RZ 62, SX 21, X 1; SPQC depth 272 with CZ 173, RX 144, RZ 125, SX 44, X 1. The paper notes that this depth remains below the cumulative depth 380 of four independent 95-depth circuits run sequentially, and further states that the unified SPQC “strictly matches the standard ensembling baseline across all predictive and uncertainty metrics,” including relative $L$8 error, coverage, average prediction interval width, and max uncertainty, even under varying depolarizing noise with fixed shots $L$9 (Matlia et al., 1 May 2026).

These results support a narrow but concrete conclusion: SPQCs can improve representational performance in small supervised-learning benchmarks and can compress ensemble inference without degrading predictive accuracy or uncertainty calibration in the specific QOrthoNN/Quantum DeepONet setting that has been reported.

6. Formal analysis and related theoretical frameworks

Although SPQCs are named explicitly only in recent work, several adjacent literatures provide tools for their analysis. A diagrammatic study of parameterized quantum circuits develops a ZX-calculus formalism for linear combinations of parameterized diagrams with explicit complex coefficients, together with pull and product rules for manipulating such sums. That work does not present SPQCs by name, but it gives a direct operator-level representation for superpositions of parameterized circuit blocks and for expectation values containing cross terms between different summands. In that sense, it supplies a rigorous diagrammatic language for SPQCs understood as linear combinations or superpositions of parameterized circuits (Stollenwerk et al., 2022).

Broader PQC theory also clarifies what SPQCs inherit from ordinary variational circuits. The standard decomposition

$$|\psi\rangle = \frac{1}{\sqrt{L}} \sum_{j=1}^{L} |j\rangle_{addr}\otimes U(\boldsymbol{\theta}^{(j)})S(\mathbf{x}^{(j)})|0_D\rangle^{\otimes n},$$0

and the parameterized-gate form

$$|\psi\rangle = \frac{1}{\sqrt{L}} \sum_{j=1}^{L} |j\rangle_{addr}\otimes U(\boldsymbol{\theta}^{(j)})S(\mathbf{x}^{(j)})|0_D\rangle^{\otimes n},$$1

remain the default primitives. The same literature emphasizes both the expressive potential of PQCs and the trainability difficulties associated with barren plateaus in sufficiently random deep circuits (Benedetti et al., 2019). This suggests that SPQCs should be interpreted not as a separate paradigm from PQCs, but as an extension of the PQC model class in which parameter selection, model multiplexing, or architecture selection itself becomes coherent.

A further related direction studies one-parameter PQC families that exhibit non-analytic expectation values and phase transitions in the infinite-volume limit. SPQCs are not explicitly discussed there, but the work analyzes a family $$|\psi\rangle = \frac{1}{\sqrt{L}} \sum_{j=1}^{L} |j\rangle_{addr}\otimes U(\boldsymbol{\theta}^{(j)})S(\mathbf{x}^{(j)})|0_D\rangle^{\otimes n},$$2 whose global parameter enters every layer and for which phase structure, order parameters, and correlation length can be characterized analytically. A plausible implication is that such one-parameter critical families are natural candidates for parameter-superposition constructions, because the control register would then coherently index branches belonging to different phases of the same PQC family (Wang et al., 29 Mar 2026).

7. Limitations, misconceptions, and open problems

Several limitations recur across the literature. The original SPQC proposal assumes FFQRAM, whose realization requires coherent control of multi-qubit address lines and data rotations. It also relies on RUS and post-selection, which reduce retention rates and increase shot complexity as activation degree grows. The numerical studies in that work do not include an explicit noise model, so its performance claims are tied to ideal or near-ideal simulations (Patapovich et al., 10 Jun 2025). By contrast, the Quantum DeepONet/QOrthoNN study does include depolarizing-noise experiments, but its SPQC validation is still confined to small circuits, simplified noise models, and an ensemble size fixed at $$|\psi\rangle = \frac{1}{\sqrt{L}} \sum_{j=1}^{L} |j\rangle_{addr}\otimes U(\boldsymbol{\theta}^{(j)})S(\mathbf{x}^{(j)})|0_D\rangle^{\otimes n},$$3 (Matlia et al., 1 May 2026).

A second misconception is that “superposition” automatically implies unrestricted quantum interference among models at the level of classical outputs. The literature does not support such a blanket statement. In the operator-learning setting, the practical emphasis is on multiplexed execution and tomography-based extraction of independent model outputs, not on interference as a learning primitive. In the FFQRAM-RUS setting, the crucial effect is amplitude transformation plus post-selection, not merely the presence of a coherent sum over branches.

Open problems follow directly from these constraints. The literature repeatedly points to scaling SPQCs to larger ensembles and deeper circuits, reducing multi-controlled-gate overhead, improving post-selection efficiency, and incorporating realistic hardware noise. The operator-learning paper also leaves open how SPQC noise accumulation affects conformal coverage, while the original SPQC paper highlights future directions such as fixed-point amplitude amplification, partial post-selection, alternative memory models, and deeper hierarchies (Patapovich et al., 10 Jun 2025). Taken together, these works indicate that SPQCs are best understood as a family of resource-conscious superposition mechanisms whose practical value depends on whether address-controlled execution, readout, and noise management can be made to scale beyond the small regimes demonstrated so far.