Variational quantum circuits are parameterized sequences of quantum gates tuned by a classical optimizer. They train far better when they are built to respect the symmetries of the problem. One example is a circuit that only ever produces states with exactly $k$ qubits set to one, which is the natural encoding of "choose exactly $k$ items." Restricting the circuit this way shrinks the space it must search from exponential to polynomial and avoids the flat, gradient-free landscapes that plague unconstrained circuits. Separately, classical multi-task learning has spent years on the problem of several objectives pulling one shared model in conflicting directions. A decade of "gradient surgery" optimizers resolves that tug-of-war, but each requires one gradient per task. That cost has pushed the field toward cheap alternatives that never look at per-task gradients.
This note argues that in symmetry-reduced quantum circuits, both halves of that classical story change. First, the cost argument disappears. When all tasks are different measured quantities on one shared quantum state, every per-task gradient comes from the same circuit runs as the summed gradient, and the parameter count is small. Second, we hypothesize that most of the gradient conflict one observes is not intrinsic to the tasks. It is manufactured when the symmetric circuit cannot reach every direction inside its allowed subspace. The proposal tests this with exact subspace simulations on constrained multi-objective selection problems. If it holds, the practical rule is to fix the circuit's expressiveness first, after which plain weighted-sum training is enough.
Core Claim
The mechanism is a projection effect. Let $g_t$ be the gradient of task $t$'s energy with respect to arbitrary motions of the state inside the symmetry subspace, which is what a fully expressive circuit could follow. Let $P$ be the projector onto the directions a particular circuit can actually move the state at its current parameters. The circuit only sees $P g_t$. Two vectors can have a positive inner product while their projections onto a smaller subspace have a negative one. We therefore define the excess conflict between tasks $i$ and $j$ as
$$\Delta_{ij} = \cos(P g_i, P g_j) - \cos(g_i, g_j),$$
which is the conflict created by the circuit rather than by the objectives. It is computed in the natural quantum-state geometry (the quantum Fisher metric) rather than in raw parameter coordinates, so that reparameterizing the circuit does not change it.
Rethinking the symmetry-preserving circuits for constrained variational quantum algorithms supplies the method. It shows that common $k$-preserving gates (the reconfigurable beam splitter and the XY mixer) fail to be subspace-controllable on some qubit connectivities. Subspace-controllable means the circuit's generators span every motion in the $\binom{n}{k}$-dimensional subspace, a property measured by the dimension of their dynamical Lie algebra, the space of motions those generators can produce. The same paper shows that an engineered beam-splitter gate reaches the full dimension $d_k^2$. That result gives us a clean knob: circuits with matched parameter counts that differ only in controllability.
Our prediction is threefold. In controllable circuits that are overparameterized relative to the subspace, $\Delta_{ij}$ tends to zero. In that regime simple scalarization, meaning a fixed weighted sum of task losses, matches conflict-averse optimizers. In non-controllable or over-pruned circuits, $\Delta_{ij}$ is systematically negative, and there conflict-averse methods help.
This transposes the classical finding of Scalarization for Multi-Task and Multi-Domain Learning at Scale that added capacity buffers task interference. The transposition replaces "more parameters" with "full reach inside the symmetric subspace," which costs no extra qubits.
Research Program
The setting is multi-objective, constrained selection: one circuit prepares one state, and $T$ objectives are energies of different cost operators on that state. Because all cost operators here are diagonal in the computational basis, a single set of measured bitstrings yields every task's energy. Under the parameter-shift rule, the standard way to estimate quantum gradients by running the circuit at shifted parameter values, the same $2P$ shifted circuits (for $P$ parameters) yield every task's gradient. Holding $T$ gradients of size $P$ in memory is negligible.
This removes the $O(T)$ backward-pass and $O(T \times N)$ memory bottleneck that motivates FAMO: Fast Adaptive Multitask Optimization. FAMO avoids per-task gradients by tracking changes in each task's log-loss. Once the bottleneck is gone, the only honest currency left is measurement shots.
We are careful about what the symmetry buys. Shot-sharing here comes from the shared state and diagonal observables, not from symmetry alone. Symmetry buys trainability, a small $P$, and the controllability knob. For non-diagonal chemistry observables, seniority-zero restriction (Gámez et al.) plays the analogous role of collapsing all observables into a few shared measurement settings. We leave that extension as a secondary check.
The work proceeds in three parts.
Measuring intrinsic versus excess conflict. We build Hamming-weight-preserving circuits on $n = 10$ to $16$ qubits with $k = 3$ to $5$. The subspace has at most a few thousand dimensions, so $g_t$, $P$, and $\Delta_{ij}$ can be computed exactly along training trajectories. Three gate families are compared: beam-splitter (controllable), reconfigurable beam-splitter, and XY on ring and line connectivities. Each family is also run at depths below and above the overparameterization threshold. That threshold is predicted by the effective-dimension analysis of Symmetric Pruning in Quantum Neural Networks, whose tangent-kernel theory says the critical parameter count scales with effective subspace dimension rather than $2^n$.
Comparing aggregators under an equal shot budget. We compare the following on how many shots each needs to reach a target solution quality:
- unitary scalarization;
- weight-swept scalarization;
- FAMO;
- the minimum-norm multiple gradient descent algorithm;
- Conflict-Averse Gradient Descent for Multi-task learning, which maximizes worst-case task improvement near the average gradient;
- Nash-MTL from Multi-Task Learning as a Bargaining Game.
Scaling $T$. We vary the number of objectives from 2 to 12 to see whether excess conflict grows with $T$ in non-controllable circuits. That growth would be the quantum version of the classical "many tasks, one compressed model" failure.
The two testbeds are chosen so that $T$ can be scaled naturally. The first is multi-objective MaxCut with $T$ edge-weightings of one graph, in the spirit of the hypervolume evaluation in Variational Quantum Multi-Objective Optimization. That paper's method (QMOO) adds one phase operator per objective and tunes parameters with a derivative-free evolution strategy (CMA-ES). We use gradients instead and add a partition-size constraint via Hamming weight.
The second is multi-target feature selection. It extends the $k$-feature selection task of the Rethinking paper to choosing $k$ molecular descriptors jointly informative for up to 12 property targets of QM-9, the multi-property molecular dataset also used in the FAMO evaluation. Each target contributes one relevance-minus-redundancy cost operator.
Metrics are the hypervolume of sampled bitstrings against a fixed reference point, the average per-task relative drop versus single-task optima, constraint-violation rate, and shots to 95% of the best achieved hypervolume. Everything is simulated exactly first, then with finite shots down to 128 per setting, the budget at which QMOO reports working Pareto fronts.
Baselines
Naive reference: unitary scalarization. This minimizes the plain sum of task energies with the same symmetric circuit and optimizer. In Defense of the Unitary Scalarization for Deep Multi-Task Learning found that, with basic regularization, this matches specialized optimizers across vision and reinforcement-learning benchmarks. It is therefore the default a skeptic expects to win. Under our hypothesis it should indeed win in the controllable regime and lose in the non-controllable one. A regime-dependent reversal, rather than a uniform margin, is what would count as support.
Strongest scalable competitor: tuned scalarization. Weights are chosen by population-based search in the style of Scalarization for Multi-Task and Multi-Domain Learning at Scale. On DomainNet that paper reports 48.01% average accuracy, against 46.78% for uniform weights and 42–43% for PCGrad and CAGrad. These are classical numbers that frame the prior, not targets on our testbeds. We give this search the same total shot budget as every other method, including the shots spent on weight search.
Strongest gradient-based competitors: CAGrad, Nash-MTL, and FAMO. They are run with identical circuits, learning-rate grids, and shot budgets. On NYU-v2, CAGrad reports a 0.20% average per-task drop against 3.97% for PCGrad. We found no reported result on Lacuna for any multi-task gradient optimizer on symmetry-constrained quantum circuits.
Derivative-free reference: QMOO. This is QMOO's CMA-ES optimization of hypervolume, plus NSGA-II as a classical reference. QMOO reports parity with NSGA-II and IBEA at depth 1 on 3- and 5-objective benchmarks. No hypervolume value is given in the file, so we reproduce both ourselves on our instances.
Beating the baselines. A conflict-aware method beats scalarization if it reaches the scalarization hypervolume with at least 30% fewer shots, consistently across seeds. It must do so in the regime where $\Delta_{ij}$ is negative, and not do so in the regime where $\Delta_{ij} \approx 0$.
Key ablation. The ingredient that matters is controllability. We hold parameter count, depth, qubit connectivity, and two-qubit gate count fixed and swap only beam-splitter gates for reconfigurable beam-splitter or XY gates. A second ablation replaces the symmetric circuit with a hardware-efficient circuit plus a soft penalty on Hamming weight.
Why not just…?
Why not just use weighted-sum scalarization and skip the conflict analysis? This is the real threat, and our own hypothesis says it is correct whenever the circuit is controllable and wide enough. The contribution is not a new optimizer that beats it everywhere. It is a diagnostic ($\Delta_{ij}$) and a regime map saying when scalarization suffices and when it does not. If scalarization matches CAGrad even in non-controllable circuits with strongly negative $\Delta_{ij}$, the hypothesis is refuted, and we report that.
Why not just use FAMO, since it never needs per-task gradients? FAMO's advantage is avoiding $T$ backward passes. In our setting those passes cost nothing extra, because all tasks are read from the same shots. What remains is whether FAMO's log-loss differences, which are ratios of noisy shot estimates, stay informative at 128 shots. The finite-shot sweep answers this directly: FAMO is run at exactly the budget of the gradient-based methods.
Why not just add a penalty term and use an unconstrained circuit? The Rethinking paper reports that hardware-efficient circuits with soft constraints failed to reach chemical accuracy on LiH and H₂O, while Hamming-weight-preserving circuits reached errors below $10^{-10}$. It also reports that penalty-based feature selection violated the $k$ constraint without careful tuning. Our penalty ablation checks whether this holds with multiple objectives, where the penalty becomes one more conflicting task.
Why not just scale up an unconstrained circuit until interference vanishes? Unconstrained circuits lose gradient signal exponentially with qubit count. Exploiting symmetry in variational quantum machine learning shows polynomial rather than exponential gradient decay once circuits are made symmetry-equivariant. The capacity that buffers conflict must therefore be capacity inside the subspace, which is exactly what our depth sweep varies.
Decisive Checks
The first check is whether $\Delta_{ij}$, computed exactly, is near zero for overparameterized beam-splitter circuits and significantly negative for reconfigurable beam-splitter and XY circuits on rings at matched depth. The second is whether the shot-efficiency gap between CAGrad or Nash-MTL and scalarization tracks $\Delta_{ij}$ across all circuit variants and both testbeds. A rank correlation near zero would falsify the mechanism. The third is whether this gap closes as depth crosses the effective-dimension threshold of the Symmetric Pruning analysis. That would give the quantum version of the classical finding that capacity buffers interference.
A positive result would tell practitioners to diagnose controllability before reaching for multi-task optimizers. A negative result would show that multi-task conflict in quantum circuits is intrinsic to the objectives and deserves dedicated optimizers even in small symmetric subspaces.
References
- Rethinking the symmetry-preserving circuits for constrained variational quantum algorithms
- Scalarization for Multi-Task and Multi-Domain Learning at Scale
- FAMO: Fast Adaptive Multitask Optimization
- Simulating quantum chemistry in the seniority-zero space on qubit-based quantum computers
- Symmetric Pruning in Quantum Neural Networks
- Conflict-Averse Gradient Descent for Multi-task learning
- Multi-Task Learning as a Bargaining Game
- Variational Quantum Multi-Objective Optimization
- In Defense of the Unitary Scalarization for Deep Multi-Task Learning
- Exploiting Symmetry in Variational Quantum Machine Learning