Solving “Barren Plateaus” Problem With Quantum Algorithms
Quantum Algorithm Breakthrough: Researchers Prevent "Barren Plateaus"
Quantum algorithms have the potential to alter materials research and medicine development, but the “barren plateau” problem often prevents their adoption. Variational quantum algorithm (VQA) optimisation problems with vanishing gradients make training these algorithms harder as system size increases.
Recent advancements in quantum circuit and tensor network mathematics are revealing the obvious solutions to avoid these computational dead ends, making quantum optimisation more reliable and scalable.
The Barren Plateaus Challenge
Barren plateaus occur when VQAs' average energy gradient amplitude decreases exponentially with system scale. Brickwall quantum circuits and quantum neural networks exhibit this phenomenon when circuit depth grows polynomially with system size. A “random walk in a flat region of the energy landscape” is often caused by statistical errors in measurements that make tiny gradients hard to detect. They provide a key quantum computer optimisation challenge. Such circumstances cause the algorithm to lose direction, halting optimisation.
Tensor Network States: Trainability Architectures
Communications in Mathematical Physics shows that variational optimisation problems for MERA, TTNS, and MPS are inherently barren plateau-free. Tensor Network States (TNS) are vital for variational quantum eigensolvers and are used in classical quantum many-body simulations.
The paper shows that energy optimisation of isometric TNS does not undergo exponential gradient amplitude deterioration for prolonged Hamiltonians with finite-range interactions. Even with randomly initialised TNS, gradient variance scales well, ensuring trainability analytically.
Important reasons why TNS avoids desolate plateaus:
Analytical assessments are simplified by articulating optimisation difficulties using Riemannian gradients, which are components of the tangent space of the unitary groups parametrising the TNS. Energy gradient variance scaling: For large bond dimensions, MPS energy gradient variance assumes a system-size independent value, especially when nearest-neighbor interactions are involved. This makes random MPS initialisation efficient. TTNS and MERA have an average energy-gradient amplitude of (bη)^τ for tensors in layer τ. This is layer-dependent decay for hierarchical TNS. The second-largest eigenvalue amplitude of a "doubled layer transition channel" is η, whereas b denotes the branching ratio, or number of sites mapped to one renormalised site in MERA/TTNS. To ensure optimisation, the number of layers T should not exceed logarithmic system size (log_b L) and bη should be less than 1 for bond dimensions greater than 1, preventing barren plateaus. Also See VQC: Variational Quantum Circuits & BVQC Protects Quantum IP
These results suggest initialisation methods include optimising tensors in initial layers before going on to deeper ones or starting with lower bond dimensions and increasing them.
Structure-agnostic trainability of quantum walk optimisation algorithm
Another major accomplishment is research on the Quantum Walk Optimisation Algorithm (QWOA), a continuous-time quantum walk variant of the Quantum Approximate Optimisation Algorithm (QAOA). QWOA focusses on structurally constrained combinatorial optimisation.
QWOA avoids barren plateaus for NP optimisation problems with polynomially bounded cost functions (NPO-PB), a large class of problems. This is significant because QWOA was intended for NPO-PB difficulties to implement effectively.
Some elements assist QWOA avoid barren plateaus:
QWOA's Dynamic Lie Algebra (DLA) input size is polynomially constrained for NPO-PB concerns. DLA describes the circuit's generators' unitary operating space. Variance-DLA Relationship: Using theoretical frameworks that relate DLA to loss function variance, this polynomially constrained DLA dimension shows that QWOA's loss function variance decays polynomially with input size. This eliminates barren plateaus' exponential decrease. Simple Structure: The statistical distribution of costs (the spectrum of the problem Hamiltonian) determines QWOA's performance, not its structure, since its expected value is invariant under basis state permutations. Its “agnostic character” makes DLA analysis easier by avoiding complex Pauli decompositions and group-theoretic approaches used by other algorithms. Also see Quantum Circuit Complexity Reveals Hidden Phases.
QWOA may require to overparameterize (have more layers in its circuit than the DLA dimension) to find optimal or approximate solutions for problems that are not in the “easy” complexity classes (BPPO and BP-APX), according to the article. QWOA's performance is limited by a quadratic speed-up similar to Grover's.
Key Role of Dynamic Lie Algebra
Both studies emphasise the importance of the Dynamic Lie Algebra (DLA) in recognising and understanding barren plateaus. Ansatz expressivity is measured by its dimension, which indicates how well it can explore a limited subspace or the unitary group. It is noteworthy that the DLA dimension is inversely related to loss function variance.
New frameworks, such as the adjoint representation and the Heisenberg image, provide a formal mathematical foundation for coupling the DLA dimension to gradient variance scaling in “Lie algebra supported ansätze” (LASAs). This allows exact gradient variance calculations, which assist explain observed events and predict barren plateau episodes.
This accomplishment in variational quantum algorithm creation and implementation is significant. By defining trainable architectural traits and problem classes, researchers are closer to constructing scalable quantum computers that can solve difficult real-world problems.