Effective Quantum Geometric Tensor (QGT)

• Effective QGT is a gauge-invariant metric that encodes the local geometry of quantum states, revealing accessible directions in the Hilbert space.
• Its spectrum, obtained via eigenvalue analysis, serves as a practical diagnostic for evaluating variational model efficiency and detecting over-parameterization.
• The efficiency indicator ρ (d_r/d_q) directly correlates with infidelity, offering a quantitative probe to optimize and benchmark neural and tensor network states.

The effective quantum geometric tensor (QGT) is a fundamental object encoding the local geometry of quantum states parameterized by variational or physical parameters. It arises in contexts ranging from variational ansätze in quantum many-body theory to information geometry and quantum machine learning. The rank and spectrum of the QGT quantify, in a precise geometric sense, the extent to which a variational family of quantum states can explore the accessible Hilbert space directions. This notion is central to understanding and diagnosing the expressivity, efficiency, and representational limitations of parameterized quantum models, particularly neural quantum states (NQS).

1. Mathematical Structure of the Quantum Geometric Tensor

Consider a variational wavefunction $\vert\psi(\theta)\rangle$ depending smoothly on complex parameters $\theta=(\theta_1, ..., \theta_N)$. The quantum geometric tensor is defined via the Fubini–Study metric on the projective Hilbert space: $$G_{ij} = \operatorname{Re} \left[ \langle \partial_i\psi | \partial_j\psi \rangle - \langle \partial_i\psi | \psi \rangle \langle \psi | \partial_j\psi \rangle \right]\,,$$ where $$|\partial_i\psi\rangle = \partial/\partial\theta_i |\psi(\theta)\rangle$$. The first term measures the susceptibility of the wavefunction to parameter variations in both directions $i$ and $j$, while the subtraction projects out the component along the global phase, ensuring $G$ is positive semi-definite and measures distinguishable directions only. This tensor provides a rigorous, gauge-invariant metric on the landscape of variational states (Dash et al., 2024).

2. Practical Computation and Interpretation of the QGT Spectrum

In practice, the spectrum of $G$ is obtained by evaluating its elements, typically through Monte Carlo sampling or, for small systems, by exact enumeration over the Hilbert space sector under study. Upon optimizing a variational ansatz—e.g., minimizing the infidelity

$$I(\theta) = 1 - \frac{|\langle \psi_\mathrm{exact} | \psi(\theta)\rangle|^2}{\langle\psi(\theta)|\psi(\theta)\rangle}$$

using natural gradient descent (where the QGT acts as a metric preconditioner)—the converged $G$ is regularized and diagonalized to yield its eigenvalues $\theta=(\theta_1, ..., \theta_N)$0 (Dash et al., 2024). Directions with $\theta=(\theta_1, ..., \theta_N)$1 represent accessible, linearly independent local tangent directions in Hilbert space utilized by the ansatz, while null eigenvalues ($\theta=(\theta_1, ..., \theta_N)$2) indicate redundant or inaccessible directions.

Numerically, the rank is defined via a cutoff $\theta=(\theta_1, ..., \theta_N)$3 (e.g., $\theta=(\theta_1, ..., \theta_N)$4), so the effective rank is $\theta=(\theta_1, ..., \theta_N)$5.

3. Effective Quantum Dimension and the Efficiency Indicator

For a given subspace (e.g., the $\theta=(\theta_1, ..., \theta_N)$6 sector of a spin model with global symmetries), the number of independent basis states sets an upper bound on the geometric directions accessible. The maximal attainable QGT rank is termed the "effective quantum dimension" $\theta=(\theta_1, ..., \theta_N)$7: $\theta=(\theta_1, ..., \theta_N)$8 for the case with dimension $\theta=(\theta_1, ..., \theta_N)$9 and spin-flip symmetry (Dash et al., 2024).

The key efficiency indicator is the ratio

$$G_{ij} = \operatorname{Re} \left[ \langle \partial_i\psi | \partial_j\psi \rangle - \langle \partial_i\psi | \psi \rangle \langle \psi | \partial_j\psi \rangle \right]\,,$$0

which quantifies the fraction of the accessible Hilbert space actually used by the converged ansatz. $$G_{ij} = \operatorname{Re} \left[ \langle \partial_i\psi | \partial_j\psi \rangle - \langle \partial_i\psi | \psi \rangle \langle \psi | \partial_j\psi \rangle \right]\,,$$1 signals full utilization (i.e., 'universal' expressibility given the variational family and the subspace), while $$G_{ij} = \operatorname{Re} \left[ \langle \partial_i\psi | \partial_j\psi \rangle - \langle \partial_i\psi | \psi \rangle \langle \psi | \partial_j\psi \rangle \right]\,,$$2 indicates under-utilization, either due to architectural bottlenecks or premature optimization plateaus.

4. Relationship to Representational Accuracy and Phase Dependence

Empirical studies show that the variational infidelity $$G_{ij} = \operatorname{Re} \left[ \langle \partial_i\psi | \partial_j\psi \rangle - \langle \partial_i\psi | \psi \rangle \langle \psi | \partial_j\psi \rangle \right]\,,$$3 decays almost exponentially with increasing $$G_{ij} = \operatorname{Re} \left[ \langle \partial_i\psi | \partial_j\psi \rangle - \langle \partial_i\psi | \psi \rangle \langle \psi | \partial_j\psi \rangle \right]\,,$$4: $$G_{ij} = \operatorname{Re} \left[ \langle \partial_i\psi | \partial_j\psi \rangle - \langle \partial_i\psi | \psi \rangle \langle \psi | \partial_j\psi \rangle \right]\,,$$5 up to a saturation point, beyond which further increases in network width or parameter count yield minimal improvements (Dash et al., 2024). Crucially, different quantum phases (e.g., AKLT, Heisenberg, critical) display distinct slopes $$G_{ij} = \operatorname{Re} \left[ \langle \partial_i\psi | \partial_j\psi \rangle - \langle \partial_i\psi | \psi \rangle \langle \psi | \partial_j\psi \rangle \right]\,,$$6 in this relation, reflecting the intrinsic difficulty of efficiently parameterizing ground states in those regions. This provides a nuanced, quantitative probe of the landscape complexity across quantum phases.

As the network width $$G_{ij} = \operatorname{Re} \left[ \langle \partial_i\psi | \partial_j\psi \rangle - \langle \partial_i\psi | \psi \rangle \langle \psi | \partial_j\psi \rangle \right]\,,$$7 (with $$G_{ij} = \operatorname{Re} \left[ \langle \partial_i\psi | \partial_j\psi \rangle - \langle \partial_i\psi | \psi \rangle \langle \psi | \partial_j\psi \rangle \right]\,,$$8 the number of hidden units/parameters and $$G_{ij} = \operatorname{Re} \left[ \langle \partial_i\psi | \partial_j\psi \rangle - \langle \partial_i\psi | \psi \rangle \langle \psi | \partial_j\psi \rangle \right]\,,$$9 the number of sites) grows, both $$|\partial_i\psi\rangle = \partial/\partial\theta_i |\psi(\theta)\rangle$$0 and $$|\partial_i\psi\rangle = \partial/\partial\theta_i |\psi(\theta)\rangle$$1 increase, and the QGT spectrum 'fills up' the available directions until all are exploited or the ansatz becomes over-parameterized.

5. Practical Implications: Over-Parameterization, Expressivity, and Diagnostics

The QGT rank $$|\partial_i\psi\rangle = \partial/\partial\theta_i |\psi(\theta)\rangle$$2 directly measures the local dimension of the variational manifold explored during optimization:

• Saturation of $$|\partial_i\psi\rangle = \partial/\partial\theta_i |\psi(\theta)\rangle$$3 as a function of parameter count indicates a regime where adding more parameters leads to redundancy without an actual increase in variational expressivity: the ansatz is 'over-parameterized'.
• Low $$|\partial_i\psi\rangle = \partial/\partial\theta_i |\psi(\theta)\rangle$$4 at low infidelity is rare; more often, efficient approximation of the target state is achieved only when $$|\partial_i\psi\rangle = \partial/\partial\theta_i |\psi(\theta)\rangle$$5 approaches unity, revealing the practical limitations to so-called universal representability of generic neural network states.
• The efficiency indicator $$|\partial_i\psi\rangle = \partial/\partial\theta_i |\psi(\theta)\rangle$$6 is a concise diagnostic of parameter utilization: a high $$|\partial_i\psi\rangle = \partial/\partial\theta_i |\psi(\theta)\rangle$$7 at low infidelity flags a genuinely expressive and efficiently optimized ansatz.

This framework allows diagnostically distinguishing between algorithmic bottlenecks (insufficient optimization, barren plateaus, etc.) and fundamental expressivity constraints of the ansatz architecture. Convergence to low infidelity with low $$|\partial_i\psi\rangle = \partial/\partial\theta_i |\psi(\theta)\rangle$$8 is typically not observed for ground states of strongly correlated phases.

6. Generalization Beyond Neural Quantum States

Although the empirical focus is on neural quantum states (e.g., shallow Boltzmann machines), the effective QGT and its spectrum can be similarly computed for any variational family:

• Deep neural networks,
• Tensor networks,
• Variational quantum circuits (NISQ-era parameterized circuits), by Monte Carlo or other tensor-tracing techniques, and analyzed as a function of model size or architectural depth.

This geometric spectral analysis offers a universal method to interrogate the representational 'reach' of arbitrary parameterized quantum models as they scale. By tracking the growth and saturation of $$|\partial_i\psi\rangle = \partial/\partial\theta_i |\psi(\theta)\rangle$$9 and $i$0 across model families, architectures, and training regimes, one can identify algorithmic bottlenecks, detect over-parameterization plateaus, and quantitatively benchmark ansatz efficiency (Dash et al., 2024).

7. Summary Table: Effective QGT Key Quantities

Quantity	Definition / Role	Typical Use
$i$1	Re $i$2	QGT matrix; local metric on variational manifold
$i$3	Eigenvalues of $i$4	Quantifies accessible directions
$i$5	Number of $i$6	Geometric rank ("effective dimension used")
$i$7	Maximal possible rank ("effective quantum dimension"); $i$8	Normalizes $i$9 for Hilbert space size
$j$0	$j$1	Efficiency: fraction of Hilbert space used
$j$2	Infidelity w.r.t. exact target state	Accuracy of variational approximation

Increasing $j$3 is highly correlated with decreasing $j$4, until saturation. The spectral analysis of the effective quantum geometric tensor thus provides a foundational, quantitative tool for assessing, optimizing, and comparing variational ansätze in quantum science, especially as parameterizations become high-dimensional and nontraditional (Dash et al., 2024).