Generalized Nonlinear Imaginary-Time Evolution

Beyond Energy Minimization

Researchers have developed a new way to prepare specific quantum states more efficiently. By generalizing a standard method called imaginary-time evolution, they can now optimize for complex goals. These include finding excited states or maximizing sensing sensitivity. This new method outperforms standard gradient descent.

In variational quantum algorithms, the goal is often to find a specific quantum state. This state must minimize a cost function, typically the energy of a Hamiltonian (an operator representing the total energy of a system). For years, the industry standard has relied on Imaginary-Time Evolution (ITE). This method acts like a physical process. It naturally decays a system toward its lowest energy state.

While powerful, standard ITE is mathematically locked into one specific type of objective. It only works for the expectation value of a Hamiltonian. But modern quantum tasks are moving past simple ground-state preparation. We now need to minimize variance to find excited states. We also need to maximize Fisher information for high-precision sensing. The central question is whether the geometric elegance of ITE can be decoupled from energy minimization. Can it serve these broader, more complex optimization landscapes?

The limits of linear evolution

Until now, the community has treated ITE as a specialized tool rather than a general optimizer. Standard ITE is effectively a gradient flow of the energy expectation value. This flow occurs with respect to the Fubini–Study metric (a geometric way to measure distance between quantum states).

As seen in, standard ITE induces a predictable descent toward the ground-state subspace.

This works perfectly if your goal is the energy expectation value. However, if your cost function is something else, the math breaks. Examples include the variance of the energy or a penalty term for maintaining orthonormal states. You cannot simply plug a non-linear cost function into the existing ITE framework. Doing so would not yield a valid quantum evolution.

The field has largely compensated for this by using standard gradient descent (GD). GD is much more flexible. However, it lacks the geometric awareness of the quantum state space. This often leads to slower convergence or suboptimal regions.

Generalizing the generator

The authors of this paper investigate a "Nonlinear Imaginary-Time Evolution" (NITE). This method allows the generator to depend on the state itself. The move is mathematically heavy. They invoke the Riesz representation theorem to guarantee the existence of a unique Hermitian operator, $H(\phi)$. This operator ensures the evolution follows the steepest descent direction for any differentiable cost function.

To test this, the researchers targeted three distinct subroutines. First, they tackled variance minimization to find excited states. Second, they looked at maximizing Quantum Fisher Information (QFI). This is essential for quantum metrology (using quantum systems to measure physical constants). Finally, they attempted simultaneous preparation of the ground and first excited states. They used penalty terms to enforce orthogonality (ensuring states remain perpendicular).

For near-term devices, they implemented this via a variational approach. They proved that NITE is formally equivalent to a generalized version of Quantum Natural Gradient Descent (QNGD).

Faster convergence and higher precision

The results suggest that NITE is significantly more robust than the status quo. In the variance minimization task, NITE converges much faster than standard gradient descent [Figure 2a]. When the researchers used full-space parameterization to remove ansatz constraints (the mathematical templates used to represent states), NITE maintained a linear convergence rate. In contrast, standard GD stalled at a precision floor [Figure S1 in supplemental material].

In the QFI maximization task, the difference is even sharper. As shown in [Figure 2b], NITE tracks closely toward the theoretical upper bound. Meanwhile, gradient descent exhibits sluggish, sublinear behavior.

The most striking result is in the multi-state preparation task [Figure 2c, d]. Standard GD failed to reliably identify both the ground and first excited states. NITE successfully drove both states to their respective exact eigenenergies. The authors note that NITE reaches these targets faster. It also demonstrates much lower variance across different random initializations. This suggests it is less sensitive to the starting position in the optimization landscape.

Mapping the new optimization path

If NITE generalizes the geometric advantages of ITE, the implications are significant. First, it provides a unified framework for "multi-state" physics. Instead of running separate optimizations, one can evolve coupled nonlinear equations. This maps out a low-energy subspace simultaneously.

Second, it offers a way to upgrade existing quantum machine learning workflows. Many ML loss functions are non-linear combinations of expectations. NITE could replace standard optimizers. This would provide "geometry-aware" training that respects the underlying quantum manifold.

The paper does not address how NITE scales with noise. In a real NISQ (Noisy Intermediate-Scale Quantum) environment, the state-dependent generator might pose challenges. Computing the state-dependent $H(\phi)$ likely requires repeated measurements. This could lead to increased noise exposure during the optimization process. Testing this NITE-based QNGD update on a noisy simulator is a necessary next step.

Figures from the paper