Qiskit Algorithms V2 - FAdam VQE Custom Implementation

This repository contains a custom implementation for integrating the Fisher Adam (FAdam) optimizer into the Variational Quantum Eigensolver (VQE) algorithm, adapted to Qiskit's latest Primitives API, EstimatorV2.

1. Project Overview

With the significant update to Qiskit's Primitives API to EstimatorV2 in Qiskit 1.0+, some functionalities within the existing qiskit-algorithms library (specifically VQE and EstimatorGradient) encountered compatibility issues with the new API. This project addresses these API discrepancies by providing custom implementations and integrations for the following components:

• Custom VQE Algorithm (VariationalQuantumEigensolverV2): The VQE's energy and gradient evaluation logic has been re-implemented to align with the new EstimatorV2's run() signature and result handling (result[0].data.evs[0]).
• [cite_start]Fisher Adam (FAdam) Optimizer (FAdam, FAdamOptimizer): The FAdam algorithm, as presented in the relevant research paper[cite: 1], is implemented using NumPy and wrapped to conform to Qiskit's Optimizer interface (compatible with scipy.optimize.OptimizeResult).
• Finite Difference-based Gradient Calculator (FiniteDiffEstimatorGradientV2): This component is implemented as an independent callable class, not inheriting from qiskit-algorithms's BaseEstimatorGradient. It directly conforms to the EstimatorV2's run() signature (pubs=...) for performing gradient calculations using the finite difference method.

This implementation aims to leverage the advanced features of Qiskit's EstimatorV2 while providing a workaround for current version compatibility issues, enabling the application of advanced optimizers like FAdam to quantum algorithms.

2. Introduction to Fisher Adam (FAdam)

Fisher Adam (FAdam) is an enhanced optimization algorithm building upon the Adam optimizer's mathematical foundation, re-examined through the lens of Information Geometry and Riemannian Geometry[cite: 1]. [cite_start]According to the paper, Adam approximates Natural Gradient Descent (NGD) by utilizing a diagonal empirical Fisher Information Matrix (FIM).

FAdam proposes several corrections to the original Adam algorithm to address its inherent flaws:

• Enhanced momentum calculations
• Adjusted bias corrections
• Adaptive epsilon
• Gradient clipping
• Refined weight decay term based on a new theoretical framework

These modifications enable FAdam to demonstrate superior performance across diverse domains such as Large Language Models (LLMs), Automatic Speech Recognition (ASR), and Vector Quantized Variational Autoencoders (VQ-VAE), achieving state-of-the-art results in ASR.

3. Project Structure

This repository is structured with the following key files:

• fadam_optimizer.py: Contains the definition of the FAdam class (for qiskit VQE).
• vqe_v2.py: Contains the definition of the VariationalQuantumEigensolverV2 class (the custom VQE).
• finite_estimator_gradient_v2.py: Contains the definition of the FiniteDiffEstimatorGradientV2 class (the custom gradient calculator).
• FAdam_V1.ipynb: Example using fadam optimizer with EstimatorV1, qiskit_algorithms VQE
• FAdam_V2.ipynb: Example using fadam optimizer with EstimatorV2, custom VQE for EstimatorV2

4. Setup and Usage

4.1. Conda Environment Setup (Recommended)

It is highly recommended to set up a Conda virtual environment to ensure proper version compatibility among Qiskit and other required libraries.

# 1. Create a new Conda virtual environment (Python 3.10 or 3.11 recommended)
conda create -n qiskit_fadam_env python=3.10
conda activate qiskit_fadam_env

# 2. Install Qiskit metapackage and necessary libraries
#    This command installs all core Qiskit libraries (qiskit-terra, qiskit-algorithms, qiskit-aer, etc.)
#    at compatible, up-to-date versions.
#    Include `pytorch-cuda=X.Y` for GPU-enabled PyTorch (e.g., 11.8 or 12.1 for CUDA version).
conda install qiskit pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia -c conda-forge

# 3. Install other required packages
conda install numpy scipy

5. Example to use fadam

check FAdam_V1.ipynb, FAdam_V2.ipynb

References

[1] Hwang, D. (2024). FAdam: Adam is a natural gradient optimizer using diagonal empirical Fisher information. arXiv.org. https://doi.org/10.48550/arXiv.2405.12807