Dynamical Lie Algebra Bounds

• Dynamical Lie algebras are the smallest Lie subalgebras generated by control Hamiltonians that determine the reachable operator space in quantum systems.
• Explicit DLA bounds reveal polynomial scaling in structured models and exponential growth for generic interacting systems, which informs quantum algorithm design.
• These bounds directly influence system controllability, trainability, and classical simulability, impacting quantum control strategies and variational optimization.

A dynamical Lie algebra (DLA) is the smallest Lie subalgebra of the system's operator algebra (e.g., $\mathfrak{su}(2^n)$ for $n$-qubit systems) generated by a prescribed set of Hamiltonians or control generators through repeated commutators and real linear combinations. DLA bounds refer to explicit inequalities or classifications governing the possible dimension, structure, or scaling of this algebra under various physical, algorithmic, or symmetry constraints. These bounds are central in quantum control, variational algorithm expressivity, complexity theory, and the mathematical classification of accessible operator spaces.

1. Formal Definition and General Properties

Given a set of generators $\mathcal{G} = \{ H_1, \dots, H_m \}$ (usually Hermitian operators or Pauli strings), the associated DLA is

$$\mathfrak{g} = \mathrm{Lie}\left\{ i H_1, \ldots, i H_m \right\} \subseteq \mathfrak{su}(2^n)$$

where $\mathrm{Lie}\{\cdots\}$ denotes closure under commutators $[A, B]=AB-BA$ and real linear combinations. The structure, dimension, and decomposition of $\mathfrak{g}$ encapsulate the accessible unitaries and dynamical trajectories generated by the control set, determining the reachable manifold $\exp(\mathfrak{g})$ and the ultimate power of quantum algorithms or simulation protocols employing these gates.

Any DLA satisfies the universal upper bound $\dim \mathfrak{g} \leq 4^n - 1$, with the maximum realized for the full unitary algebra $\mathfrak{su}(2^n)$. The structure of $n$0 may include direct sums of simple algebras (e.g., $n$1, $n$2, $n$3), Abelian ideals, or more intricate configurations depending on the system, symmetry, and generator connectivity (Cuypers, 9 Mar 2026).

2. Explicit Dimension and Structure Bounds: Key Cases

Dimension bounds for DLAs vary sharply depending on physical symmetry, generator connectivity, and circuit architecture:

Model class	DLA dimension $n$4	Reference
Non-entangling, product	$n$5 (on-site $n$6)	(Bärligea et al., 17 Apr 2026)
Free-fermion/matchgate	$n$7 ($n$8)	(Bärligea et al., 17 Apr 2026, Wiersema et al., 2023)
Permutation-inv.	$n$9	(Bärligea et al., 17 Apr 2026, Allcock et al., 2024)
Hamming-weight $\mathcal{G} = \{ H_1, \dots, H_m \}$0	$\mathcal{G} = \{ H_1, \dots, H_m \}$1	(Bärligea et al., 17 Apr 2026)
Ising chain ($\mathcal{G} = \{ H_1, \dots, H_m \}$2)	$\mathcal{G} = \{ H_1, \dots, H_m \}$3 (Abelian)	(Wiersema et al., 2023)
MaxCut QAOA on $\mathcal{G} = \{ H_1, \dots, H_m \}$4	$\mathcal{G} = \{ H_1, \dots, H_m \}$5	(Allcock et al., 2024, Tan, 2 Dec 2025)
Random all-to-all (SK)	$\mathcal{G} = \{ H_1, \dots, H_m \}$6 (full $\mathcal{G} = \{ H_1, \dots, H_m \}$7)	(Tan, 2 Dec 2025, Cuypers, 9 Mar 2026)

This demarcation underpins several major findings:

• Polynomial dimension ($\mathcal{G} = \{ H_1, \dots, H_m \}$8, $\mathcal{G} = \{ H_1, \dots, H_m \}$9) arises only in highly structured circuits: free-fermionic, symmetry-adapted, or product models.
• Generic interacting or random 2-local models almost always saturate the exponential bound, with $$\mathfrak{g} = \mathrm{Lie}\left\{ i H_1, \ldots, i H_m \right\} \subseteq \mathfrak{su}(2^n)$$0, implying maximal controllability but exponentially rising algorithmic complexity (Kökcü et al., 2024).

Special classifications exist for translation-invariant chains (Wiersema et al., 2023) and arbitrary 2-local undirected graphs (Kökcü et al., 2024), showing that polynomial-size DLAs are exclusive to product, Abelian, or free-fermionic (quadratic) models; all other topologies yield exponential growth.

3. DLA Bounds in Quantum Algorithms and Expressivity

The DLA dimension provides a rigorous metric for the expressivity of variational ansätze and quantum walks. In the context of QAOA and general hybrid algorithms:

• QAOA on complete graphs ($$\mathfrak{g} = \mathrm{Lie}\left\{ i H_1, \ldots, i H_m \right\} \subseteq \mathfrak{su}(2^n)$$1): DLA dimension is $$\mathfrak{g} = \mathrm{Lie}\left\{ i H_1, \ldots, i H_m \right\} \subseteq \mathfrak{su}(2^n)$$2—provably below the Hilbert space exponential, but rapidly growing (Allcock et al., 2024, Tan, 2 Dec 2025).
• Hybrid Quantum Walks (HQW): The DLA expands beyond QAOA through the inclusion of coin-controlled directions and, crucially, the Jordan–Lie algebra structure (Chen et al., 28 Apr 2026). Here:
  • QAOA: $$\mathfrak{g} = \mathrm{Lie}\left\{ i H_1, \ldots, i H_m \right\} \subseteq \mathfrak{su}(2^n)$$3
  • HQW: $$\mathfrak{g} = \mathrm{Lie}\left\{ i H_1, \ldots, i H_m \right\} \subseteq \mathfrak{su}(2^n)$$4, where $$\mathfrak{g} = \mathrm{Lie}\left\{ i H_1, \ldots, i H_m \right\} \subseteq \mathfrak{su}(2^n)$$5 includes both commutators and Jordan products, yielding

  $$\mathfrak{g} = \mathrm{Lie}\left\{ i H_1, \ldots, i H_m \right\} \subseteq \mathfrak{su}(2^n)$$6

  and (strictly) enhanced expressivity. Transverse directions associated with negativity in the normalized Jordan product $$\mathfrak{g} = \mathrm{Lie}\left\{ i H_1, \ldots, i H_m \right\} \subseteq \mathfrak{su}(2^n)$$7 are generically inaccessible to QAOA, but accessible to HQW (Chen et al., 28 Apr 2026):

  $$\mathfrak{g} = \mathrm{Lie}\left\{ i H_1, \ldots, i H_m \right\} \subseteq \mathfrak{su}(2^n)$$8

  When $$\mathfrak{g} = \mathrm{Lie}\left\{ i H_1, \ldots, i H_m \right\} \subseteq \mathfrak{su}(2^n)$$9, this strict bound enables HQW to escape barren or flat directions intrinsic to QAOA.

4. DLA Bounds, Barren Plateaus, and Trainability

The dimension of the DLA directly governs variational quantum circuit trainability via gradient concentration:

• Variance Scaling: For LASA-type ansätze (observable and state in the DLA), the variance of the cost function derivative scales inversely with the DLA dimension:

$\mathrm{Lie}\{\cdots\}$0

as shown in (Fontana et al., 2023).

• For full $\mathrm{Lie}\{\cdots\}$1, the variance decays exponentially, leading to barren plateaus and intractable optimization. Polynomial-size DLAs (e.g., MaxCut on cycles, matchgate circuits) avoid such plateaus (Allcock et al., 2024, Tan, 2 Dec 2025).
• The DLA also determines generalized module-dimension scaling for matchgate circuits, where plateau onset is dictated by the dimension of Majorana-component modules, not the DLA per se (Diaz et al., 2023).
• Efficiency bounds in ancilla-feedback optimization exhibit sharp transitions: for polynomial DLA, algorithmic efficiency $\mathrm{Lie}\{\cdots\}$2 increases with system size; for exponential DLA, $\mathrm{Lie}\{\cdots\}$3 collapses at $\mathrm{Lie}\{\cdots\}$4 qubits, coincident with vanishing gradient variance and the onset of plateaus (Tan, 2 Dec 2025).

5. DLA Bounds under Symmetry Constraints

Symmetry imposition severely constrains the possible dimension of the DLA:

• Abelian symmetries restrict the DLA to direct sums of 1-dimensional $\mathrm{Lie}\{\cdots\}$5, yielding only toric manifolds of unitaries, which is a measure-zero subset within the full non-Abelian case (Silva et al., 22 Mar 2026).
• For permutation-invariant circuits, Schur–Weyl duality restricts $\mathrm{Lie}\{\cdots\}$6, supporting polynomially-large—but not exponential—DLAs (Bärligea et al., 17 Apr 2026).
• For circuits acting within specified Hamming-weight subspaces, the dimension is squared in the excitation number: $\mathrm{Lie}\{\cdots\}$7 (Bärligea et al., 17 Apr 2026).
• Graph-theoretic constraints: On connected, undirected graphs that are not simple chains, all nontrivial 2-local interacting DLAs are exponential in system size (Kökcü et al., 2024).

6. Classification of Pauli-Generated DLAs

The DLA generated by a set of Pauli strings admits an explicit classification via quadratic spaces over $\mathrm{Lie}\{\cdots\}$8 (Cuypers, 9 Mar 2026). Each Pauli string is mapped to a vector in $\mathrm{Lie}\{\cdots\}$9, a space of dimension $[A, B]=AB-BA$0, equipped with a quadratic form $[A, B]=AB-BA$1 and symplectic bilinear form $[A, B]=AB-BA$2. Key results:

• Classification Theorem: Any Pauli-generated DLA, given a connected frustration graph, is (up to direct summands) a sum of simple Lie algebras of types $[A, B]=AB-BA$3, or $[A, B]=AB-BA$4.
• Dimension bounds: All such DLAs satisfy

$[A, B]=AB-BA$5

with equality if and only if the generators span the full quadratic space.

• Recognition of the DLA type (including dimension and structure) for any Pauli-set is algorithmically feasible in $[A, B]=AB-BA$6 time.

7. Generalization, Simulability, and Operator Complexity

DLA dimension bounds not only quantify expressivity but also determine model generalization and classical simulability:

• Generalization Bound: The Rademacher complexity and associated PAC-style generalization gap for quantum neural networks is $[A, B]=AB-BA$7, where $[A, B]=AB-BA$8 is the number of training examples (2504.09771).
• Lie-Algebraic Simulation: If $[A, B]=AB-BA$9, classical simulation of circuit dynamics and expectation values via adjoint representation is tractable (Bärligea et al., 17 Apr 2026).
• Krylov Complexity Speed Limits: In time-dependent systems whose evolution is governed by a DLA closed on a simple algebra (e.g., $\mathfrak{g}$0), the growth rate of Krylov complexity $\mathfrak{g}$1 obeys

$\mathfrak{g}$2

with equality conditioned on commutativity of time-slices; this applies irrespective of time-dependence (Grabarits et al., 6 May 2026).

• DLA classification and Pauli algebra bounds (Cuypers, 9 Mar 2026)
• DLA in QAOA, Max-Cut, polynomial and exponential scaling (Allcock et al., 2024, Tan, 2 Dec 2025)
• DLA and barren plateau theory (Fontana et al., 2023, Diaz et al., 2023)
• Quantum walk, coin-extended DLA bounds (Chen et al., 28 Apr 2026)
• Graph-theoretic DLA structure (Kökcü et al., 2024)
• Translation-invariant chains (Wiersema et al., 2023)
• Generalization and machine learning theory (2504.09771)
• Classical simulatability via Lie algebra dimension (Bärligea et al., 17 Apr 2026)
• Symmetry-restricted and non-Abelian DLA bounds (Silva et al., 22 Mar 2026)
• Krylov operator growth and DLA-based speed limits (Grabarits et al., 6 May 2026)