Main result: every ℚ-structured circuit has its dynamical Lie algebra inside 𝔰𝔭(2ⁿ⁻¹) or 𝔰𝔬(2ⁿ), and its loss variance is exactly 1/(2ⁿ − (−1)ⁿ), the Haar value up to ±1; the quaternionic restriction alone therefore does not remove the barren plateau.
Variational quantum computing schemes train a loss function by sending an initial state through a parametrized quantum circuit, and measuring the expectation value of some operator. Despite their promise, the trainability of these algorithms is hindered by barren plateaus (BPs) induced by the expressiveness of the circuit, the entanglement of the input data, the locality of the observable, or the presence of noise. Up to this point, these sources of BPs have been regarded as independent. In this work, we present a general Lie algebraic theory that provides an exact expression for the variance of the loss function of sufficiently deep parametrized quantum circuits, even in the presence of certain noise models. Our results allow us to understand under one framework all aforementioned sources of BPs. This theoretical leap resolves a standing conjecture about a connection between loss concentration and the dimension of the Lie algebra of the circuit’s generators. The barren plateau problem represents one of the major bottlenecks for parametrized quantum circuits algorithms. Here, the authors study the known sources of BP using the lens of Lie algebraic theory, finding an expression of the variance of the loss function depending on the dynamical Lie algebra of the circuit.
Variational quantum algorithms (VQAs) are a leading approach to near-term quantum computation, but their utility is limited by barren plateaus and other pathologies in their loss landscapes. Existing landscape theories based on dynamical Lie algebras, Jordan-algebraic Wishart systems, approximate t-designs, and Haar-random circuits are foundational, but they often neglect the finite-depth geometry of realistic ansätze and are therefore ill-suited to the shallow-depth regime, where VQAs are poor approximators of 2-designs and trainability is most feasible. This work introduces Krylov algebras, algebraic structures induced by the Krylov span of a finite generator set acting on one or more seed vectors, as a framework for VQA landscape theory. We show that VQA reachable manifolds can be approximated in a numerically robust, geometrically faithful fashion by Krylov-Lie algebras and groups, and that these structures induce canonical invariant measures for computing expectation values and variances under general sampling measures. In particular, we derive weighted non-Haar variance formulas that recover the usual Lie-algebraic Haar formulas as a special case while isolating non-Haar effects into explicit correction terms. We also show that the common heuristic that sufficiently deep circuit ensembles must converge to Haar fails in general without additional hypotheses, identify concrete obstructions to naive Haar convergence, and recover convergence under natural necessary and sufficient ergodic conditions. Lastly, our formulas further imply that non-Haar contributions to landscape statistics may mitigate barren plateaus by reweighting the visible sectors of the loss landscape, suggesting that VQAs may be more trainable than recent literature has posited.
We give an analytical expression for the dynamical Lie algebra corresponding to the QAOA-MaxCut problem on complete graphs, and show that the variance of the associated loss function scales linearly in the number of qubits. This solves an open problem from [ASYZ26] and confirms that such systems do not exhibit barren plateaus. The proof is based on projecting the dynamical Lie algebra generators onto subspaces given by the Schur-Weyl duality between irreducible representations of the unitary and symmetric groups.
Barren plateaus have emerged as a pivotal challenge for variational quantum computing. Our understanding of this phenomenon underwent a transformative shift with the recent introduction of a Lie algebraic theory capable of explaining most sources of barren plateaus. However, this theory requires either initial states or observables that lie in the circuit's Lie algebra. Focusing on parametrized matchgate circuits, in this work we are able to go beyond this assumption and provide an exact formula for the loss function variance that is valid for arbitrary input states and measurements. Our results reveal that new phenomena emerge when the Lie algebra constraint is relaxed. For instance, we find that the variance does not necessarily vanish inversely with the Lie algebra's dimension. Instead, this measure of expressiveness is replaced by a generalized expressiveness quantity: The dimension of the Lie group modules. By characterizing the operators in these modules as products of Majorana operators, we can introduce a precise notion of generalized globality and show that measuring generalized-global operators leads to barren plateaus. Our work also provides operational meaning to the generalized entanglement as we connect it with known fermionic entanglement measures, and show that it satisfies a monogamy relation. Finally, while parameterized matchgate circuits are not efficiently simulable in general, our results suggest that the structure allowing for trainability may also lead to classical simulability.
Barren plateaus hinder training of parameterized quantum circuits by making loss gradients exponentially small in the number of qubits. For angle-encoded product states, we show how the input distribution affects output variation through algebraic input purity, i.e. the input's overlap with the circuit's dynamical Lie algebra. With the specified readouts and Haar or exact group 2-design sampling, matchgate circuits on n qubits retain output variance of order 1/n for every pure product input. The off-diagonal family instead has zero output on computational-basis inputs at any depth and for every parameter choice. Independent uniform angles yield mean variance of the same order as matchgates. A count of diagonal Pauli strings identifies these zero-output endpoints. For any fixed dataset of nonzero inputs with at least 2n coordinates, we construct a classical preprocessing certificate. A shared random rotation is accepted only when the dataset's mean purity passes a computable threshold. This certifies mean output variance of order 1/n for the off-diagonal family before circuit execution, with a constant expected number of trials. Binary and ternary weighted encodings also recover the independent-angle mean purity from one uniform scalar input. Numerical experiments examine training behavior and extensions to larger algebras. The guarantee concerns output variance averaged over inputs, while successful learning also depends on the task and information preserved by the encoding.
Abstract We present a Lie-geometric framework for analyzing the trainability of quantum neural networks (QNNs) through the algebraic structure of their generator spaces. We prove that QNNs whose dynamics are restricted to polynomial-dimensional Lie subalgebras of su ( 2 n ) admit a polynomial lower bound on gradient variance, thereby avoiding barren plateaus by construction. In contrast, architectures exploring the full Lie algebra su ( 2 n ) exhibit exponentially vanishing gradient variance, O ( 2 − n ) , due to the exponential growth of the associated unitary orbit dimensions. Our results interpret barren plateaus as a geometric concentration phenomenon arising from high-dimensional Lie-group actions, rather than as an intrinsic limitation of parameterized quantum models. We further illustrate the mechanism using a finite-size collective-spin benchmark and a minimal two-qubit sanity check, both of which are consistent with the theoretical predictions.