Beyond the Expressivity-Trainability Paradox: A Dynamical Lie Algebra Perspective on Navigating Barren Plateaus in Quantum Machine Learning

Abstract: As Quantum Machine Learning (QML) transitions toward practical implementation, the field faces a critical architectural bottleneck that challenges the fundamental assumptions of classical statistical learning theory. In classical deep learning, increasing model capacity typically risks overfitting. However, this study advances a counter-intuitive paradigm: unstructured contemporary QML architectures suffer from a profound state of quantum underfitting, driven by the "expressivity-trainability paradox." We demonstrate that the vast Hilbert space capacity of Parameterized Quantum Circuits (PQCs)-traditionally chased as the source of quantum advantage is the direct mathematical cause of Barren Plateaus (BPs), where gradient landscapes become exponentially flat. By synthesizing recent breakthroughs in Dynamical Lie Algebras (DLAs) and Geometric QML, we establish a comprehensive framework linking the algebraic dimension of circuit generators to their optimization dynamics. Furthermore, we empirically validate this framework on a non-linear binary classification task, illuminating a uniquely quantum manifestation of the bias-variance tradeoff: while unstructured architectures achieve near-perfect training accuracy via unscalable parameterization (quantum overfitting), embedding group-theoretic geometric priors acts as a structural regularizer. By restricting the DLA growth to a polynomial regime, our symmetry-preserving approach sacrifices raw memorization capacity to guarantee scalable, gradient-rich training landscapes, offering a robust roadmap for "Trainability-by-Design" in scalable quantum neural networks.

• The paper establishes dynamical Lie algebra dimension as a design-level predictor of trainability, linking exponential growth in hardware-efficient ansätze to gradient variance scaling as O(1/4^n).
• The framework compares hardware-efficient ansätze with symmetry-preserving circuits, showing that polynomially growing Lie algebras maintain usable gradients while restricting reachable dynamics and expressivity.
• Numerical results show hardware-efficient circuits achieving nearly 100% training accuracy through apparent memorization on Make Moons, while a compact symmetry-preserving ansatz reaches about 83.5% with better scalability, though noise and generalization remain unresolved.

The expressivity-trainability paradox

This paper addresses a structural obstacle to scalable Quantum Machine Learning (QML): the "expressivity-trainability paradox." In classical statistical learning, increasing model capacity risks overfitting but preserves trainability; the author argues that Parameterized Quantum Circuits (PQCs) invert this logic. Highly expressive, unstructured circuits—typically Hardware-Efficient Ansatzes (HEAs)—can memorize small datasets via brute-force parameterization ("quantum overfitting"), yet once their depth suffices to approximate a unitary 2-design, the variance of cost-function gradients decays exponentially as $\mathcal{O}(1/4^n)$ with qubit number $n$ (McClean et al., 2018), producing Barren Plateaus (BPs). At scale, the model cannot fit training data at all—a regime the paper terms "quantum underfitting." The proposed remedy is architectural rather than algorithmic: restrict the Dynamical Lie Algebra (DLA) of the circuit's generators to polynomial growth by embedding group-theoretic symmetries, following the Geometric QML program (Sarti et al., 2024) and the Lie-algebraic theory of barren plateaus (Siegel et al., 2023).

Theoretical framework

The formal object of study is a PQC $U(\bm{\theta}) = \prod_l e^{-i\theta_l H_l} W_l$ minimizing a cost $$C(\bm{\theta}) = \langle \psi_0 | U^\dagger O U |\psi_0\rangle$$. The DLA $\mathfrak{g} = \text{Lie}(\{H_l\})$ is the real span of all nested commutators of the generators, and its dimension bounds the reachable unitary manifold $e^{\mathfrak{g}}$. The central diagnostic claim is a dichotomy: if $\dim(\mathfrak{g}) \sim \mathcal{O}(4^n)$, the circuit is controllable on $SU(2^n)$ and gradient signals dilute into an exponentially large operator space, triggering BPs; if $$\dim(\mathfrak{g}) \sim \mathcal{O}(\text{poly}(n))$$, gradient variance decays only polynomially and the model remains trainable. This reframes the paradox as an explicitly algebraic trade-off.

Two architectures instantiate the dichotomy:

• HEA: single-qubit rotations $RX, RY, RZ$ on every qubit plus global entanglers. The Lie closure fills $n$0, giving $n$1.
• Symmetry-Preserving Ansatz (SPA): generators restricted to operators commuting with a symmetry $n$2 (e.g., $n$3 terms and exchange interactions $n$4), so that $n$5 lies within the commutant of $n$6 and grows polynomially.

The paper also notes a practical corollary: exact commutator-closure computation costs $n$7 with $n$8, approaching $n$9 for HEAs—so symmetry confinement is what makes DLA analysis tractable at all beyond very small systems.

Numerical results

Three experiments compare HEA and SPA for $U(\bm{\theta}) = \prod_l e^{-i\theta_l H_l} W_l$0 using PennyLane state-vector simulation.

Gradient variance. For the HEA, $U(\bm{\theta}) = \prod_l e^{-i\theta_l H_l} W_l$1 follows the exponential $U(\bm{\theta}) = \prod_l e^{-i\theta_l H_l} W_l$2 scaling once depth permits 2-design behavior, becoming unusable for $U(\bm{\theta}) = \prod_l e^{-i\theta_l H_l} W_l$3. The SPA exhibits only polynomial decay, keeping gradients above optimizer precision limits. This empirically confirms that geometric constraints immunize against BPs—and, importantly, the effect is attributed directly to the algebraic restriction rather than to reduced parameter count alone.

DLA scaling. Explicit commutator closure confirms $U(\bm{\theta}) = \prod_l e^{-i\theta_l H_l} W_l$4 versus polynomial growth for the SPA, mirroring the gradient results and supporting the claimed causal link between algebraic dimension and concentration of measure.

Classification task. On Make Moons binary classification at $U(\bm{\theta}) = \prod_l e^{-i\theta_l H_l} W_l$5, the HEA (30 layers, >700 parameters) reaches nearly 100% training accuracy—identified as memorization rather than generalization—while the SPA (5 layers, <30 parameters) plateaus at approximately 83.5% accuracy. The paper's strong claim here is that the SPA's lower accuracy is not a defect but the necessary price of architectural stability: because its DLA is polynomially bounded, trainability is guaranteed at any system size, whereas the HEA's success at $U(\bm{\theta}) = \prod_l e^{-i\theta_l H_l} W_l$6 would not survive scaling to $U(\bm{\theta}) = \prod_l e^{-i\theta_l H_l} W_l$7.

Limitations and open questions

Several caveats bear directly on the strength of these conclusions. First, the empirical validation rests on a single toy dataset (Make Moons) at a single system size ($U(\bm{\theta}) = \prod_l e^{-i\theta_l H_l} W_l$8); no test-set generalization numbers are reported, so the characterization of HEA behavior as "overfitting" rather than genuine fitting is asserted from architecture and scale arguments rather than measured out-of-sample performance. Second, the SPA's 83.5% ceiling demonstrates that symmetry constraints can impose an irreducible expressivity deficit; the framework does not provide a method for selecting which symmetry yields sufficient expressivity for a given task while retaining polynomial DLA growth. Third, the DLA dimension is argued to be a necessary condition for avoiding exponential gradient vanishing, but the paper does not establish it as sufficient—other BP sources such as hardware noise (Alexos et al., 2021) and global cost functions (Jiang et al., 2020) operate independently of ansatz structure. Fourth, whether geometrically restricted DLAs confer resilience against noise-induced barren plateaus is explicitly left open, as is extension to non-Abelian continuous symmetries (e.g., $U(\bm{\theta}) = \prod_l e^{-i\theta_l H_l} W_l$9) and discrete topological symmetries relevant to quantum chemistry and condensed-matter datasets. Finally, the exponential cost of exact DLA computation means the diagnostic itself cannot certify large unstructured circuits, limiting its use as a practical screening tool at scale.

Conclusion

The paper consolidates recent Lie-algebraic theory into an actionable design principle: treat $$C(\bm{\theta}) = \langle \psi_0 | U^\dagger O U |\psi_0\rangle$$0 as the primary predictor of trainability, and engineer symmetry-preserving architectures whose DLAs grow polynomially. Its empirical contribution—correlating exponential DLA growth with $$C(\bm{\theta}) = \langle \psi_0 | U^\dagger O U |\psi_0\rangle$$1 gradient decay and contrasting it with polynomial-bounded, trainable SPA dynamics on a concrete classification task—is consistent with prior theory but adds a compact demonstration of the quantum bias-variance trade-off. The evidence remains limited to small-scale noiseless simulations and toy data, so the framework's value as a "Trainability-by-Design" roadmap will depend on resolving the open questions of symmetry selection, sufficiency of the DLA criterion, and robustness under realistic noise.