Exploring the Effect of Basis Rotation on NQS Performance

This paper demonstrates that local basis rotations in Neural Quantum States, while preserving the optimization landscape, geometrically displace the target ground state in parameter space to induce trainability failures and incorrect wavefunction structures even when the state is theoretically representable.

The Gist

Imagine you are trying to teach a robot to recognize a specific shape, like a perfect circle. You give the robot a set of instructions (a "neural network") and a goal: find the shape that matches the circle best.

This paper is about what happens when you rotate that circle before showing it to the robot. You haven't changed the circle itself—it's still a perfect circle with the same size and properties. But you've turned it sideways.

The researchers found that even though the circle hasn't changed, the robot's ability to learn it changes dramatically depending on how it's rotated. Sometimes the robot learns it instantly; other times, it gets stuck in a corner and learns a "fake" circle that looks almost right but is actually wrong.

Here is a breakdown of their findings using simple analogies:

1. The Setup: The Robot and the Rotated Circle

The scientists used a famous physics model called the Ising model (think of it as a row of tiny magnets that can point up or down). They wanted to find the "ground state," which is the most stable, lowest-energy arrangement of these magnets.

• The Trick: They applied a "local basis rotation." Imagine taking every single magnet in the row and twisting it slightly by the same amount.
• The Result: The physics of the system didn't change. The magnets still interact the same way, and the energy levels are identical. However, the description of the perfect solution changed. It's like taking a map of a city and rotating the paper 45 degrees. The city is the same, but the coordinates you need to type into your GPS are now completely different.

2. The Problem: The "Saddle Point" Trap

The researchers discovered that this rotation moves the "perfect solution" to a different spot in the robot's "learning landscape."

• The Landscape Analogy: Imagine the robot is trying to roll a ball down a hill to find the lowest point (the best solution).
  • Normal Rotation: Sometimes, the rotation moves the target to a smooth, gentle slope. The ball rolls down easily.
  • Bad Rotation: Other times, the rotation moves the target to a spot that looks like a saddle (like the seat of a horse). It's a high point in one direction and a low point in another.
  • The Trap: When the robot tries to roll down, it gets stuck on the saddle. It thinks it has reached the bottom because the ground feels flat around it, but it hasn't actually reached the true lowest point.

3. The Deceptive "Low Energy"

This is the most surprising part of the paper. When the robot gets stuck on that saddle point:

• It calculates the energy and says, "Hey, this is very low! I'm doing a great job!"
• But if you check the actual structure of the solution (the wavefunction), it's wrong. The robot has found a "fake" solution that is a messy mix of two different states, rather than the single, pure state it was supposed to find.

The Analogy: Imagine you are trying to mix a perfect cup of coffee. You accidentally mix in some tea. If you only measure the temperature (energy), the cup might be the perfect temperature. But if you taste it (check the structure), it's a terrible, muddy mess. The robot was fooled by the temperature reading.

4. Why "Shallow" Robots Struggle

The paper tested "shallow" neural networks (simple, small robots).

• These simple robots are very sensitive to where the target is placed.
• When the target is rotated into a "bad" spot (near a saddle point), the simple robot gets lost.
• Even if you make the robot slightly bigger (add more neurons), it still struggles to escape these traps without taking an impossibly long time.

5. The Solution: Checking the Map, Not Just the Altitude

The researchers showed that if you only look at the "energy" (altitude), you might think the robot is successful. But if you also check the "fidelity" (how closely the shape matches the target) and the "coherence" (how organized the internal parts are), you can see the robot is actually stuck.

The Big Takeaway

The paper concludes that how you describe a problem matters just as much as the problem itself.

Even if the physics of a quantum system is perfect and unchanged, the way a computer "sees" it (the basis) can make the problem easy or impossible to solve. The computer isn't failing because the problem is too hard; it's failing because the "map" it's using has been rotated into a confusing shape that leads it into a trap.

In short: You can't just look at the final score (energy) to see if a quantum computer is working. You have to look at the path it took to get there, because a rotated map can trick even the smartest algorithms into thinking they've won when they've actually lost.

Technical Summary

Technical Summary: Exploring the effect of basis rotation on the performance of Neural Quantum States

Problem Statement
Neural Quantum States (NQS) have emerged as a powerful variational framework for representing quantum many-body wavefunctions, capable of handling volume-law entanglement and leveraging modern machine learning hardware. However, their performance is known to depend sensitively on the choice of basis used to represent the quantum state. While basis dependence is an intrinsic feature of neural-network representations, the specific mechanisms by which a change of basis impacts variational optimization remain poorly understood. A critical gap exists in distinguishing whether optimization failures arise from the representational limitations of the ansatz (i.e., the network cannot express the state) or from optimization-induced trainability effects (i.e., the geometry of the loss landscape prevents convergence). Previous works have investigated basis dependence and optimization geometry separately, but a controlled separation between changes in the physical properties of the target state and purely optimization-induced effects has been lacking.

Methodology
The authors introduce a controlled, analytically transparent framework to isolate optimization-induced contributions to basis dependence. The study employs the one-dimensional transverse-field Ising model with periodic boundary conditions, considering both ferromagnetic ($J = -1$) and antiferromagnetic ($J = +1$) couplings on chains with an odd number of spins.

1. Basis Rotation Framework: Instead of modifying the Hamiltonian's spectrum, the authors apply a site-resolved local basis rotation $U_y(\phi) = \bigotimes_i e^{i\phi \sigma^y_i}$ to the exact ground state $|\psi_0\rangle$. This generates a rotated target state $|\psi_\phi\rangle = U_y(\phi)|\psi_0\rangle$. Crucially, this transformation preserves the physical energy spectrum and entanglement structure while relocating the target state within the Hilbert space relative to the variational parametrization.
2. Variational Ansatz: The study utilizes shallow neural architectures: Restricted Boltzmann Machines (RBM) and small fully connected feedforward neural networks (FCNN). These are trained using Stochastic Reconfiguration (Quantum Natural Gradient), which incorporates the geometry of the variational manifold via the quantum geometric tensor.
3. Optimization Metrics: To disentangle optimization effects from representational limits, the authors compare two objective functions:
   • Variational Energy ($E$): The standard cost function used when the target state is unknown.
   • Infidelity ($I$): Used when the exact target state is available, providing a direct measure of wavefunction overlap.
4. Geometric Analysis: The authors employ information-geometric measures, specifically the Fubini–Study distance, to quantify the geometric displacement between the initialization state and the rotated target. They also utilize Uniform Manifold Approximation and Projection (UMAP) to visualize optimization trajectories in parameter space.

Key Results
The study reveals that local basis rotations, despite leaving the physical spectrum unchanged, significantly alter the trainability of NQS by modifying the geometric relationship between the initialization and the target state.

• Geometric Displacement and Optimization Trapping: The optimization landscape is not invariant under basis rotation. Rotating the target state increases the information-geometric distance from typical initializations (e.g., equal-weight superpositions). This displacement steers optimization trajectories toward saddle points and high-curvature regions of the variational landscape.
• Divergence Between Energy and Wavefunction Accuracy: In the ferromagnetic regime ($J=-1$), where low-energy eigenstates are nearly degenerate, shallow NQS architectures often converge to states with very low relative energy errors but high infidelity. The optimization gets trapped in a superposition of the two lowest-energy eigenstates rather than the true ground state. This "mismatch" indicates that low energy error is not a sufficient metric for wavefunction quality in these geometric regimes.
• Sensitivity to Low-Energy Structure: The antiferromagnetic regime ($J=+1$), characterized by topological frustration and a gapless low-energy band, shows even more severe optimization failures for most rotation angles, with the shallow ansatz failing to converge to the target state within practical iteration limits.
• Role of Quantum Coherence: The authors demonstrate that rotated ground states possess different global quantum resources, such as quantum coherence (measured by the Shannon entropy of wavefunction coefficients), compared to unrotated states. Optimized states often fail to reproduce the correct coherence statistics, even when energy errors are small.
• Universality of Geometric Obstructions: The geometric mechanisms causing optimization slowdowns (saddle points, poor conditioning) are observed across different variational algorithms, including Lanczos, DMRG, and RBM, when initialized from the same state and subjected to the same rotated target.

Significance and Claims
The paper claims to identify a geometric mechanism contributing to basis dependence in NQS, distinct from limitations in representational expressivity. The primary contribution is the demonstration that:

1. Optimization is Geometry-Dependent: Trainability is strongly influenced by the geometric relation between the target state and the variational manifold, not just the expressivity of the ansatz.
2. Basis Choice as a Control Parameter: Local basis rotations serve as a controlled probe to disentangle physical properties from optimization artifacts. They show that a "good" basis is one that places the target state in a region of the parameter space with favorable curvature and distance from the initialization.
3. Limitations of Energy Metrics: In regimes with near-degenerate spectra, minimizing energy alone can lead to convergence on incorrect wavefunction structures (e.g., superpositions of degenerate states), highlighting the need for metrics that capture wavefunction structure, such as infidelity or coherence.
4. Landscape-Aware Design: The results motivate the development of "landscape-aware" variational designs. The authors suggest that future strategies for improving variational optimization should consider the geometric displacement of the target state, potentially utilizing more expressive architectures (deeper networks) or adaptive regularization to navigate high-curvature regions induced by basis rotations.

The work concludes that while shallow NQS architectures perform reliably near the computational basis, their performance degrades significantly when required to represent rotated states with enhanced quantum resources, underscoring the critical role of optimization geometry in variational quantum algorithms.