Qadence: a differentiable interface for digital-analog programs

TL;DR

Qadence presents a differentiable, high-level interface for digital-analog quantum programming that unifies block-based circuit construction, symbolic parameters, and flexible qubit registers with differentiable backends for both simulators and hardware. It enables end-to-end DAQC workflows by providing Hamiltonian translation facilities, a Hamiltonian-construction toolkit, and a Transform framework for mapping target evolutions to build Hamiltonians, all integrated within a PyTorch-friendly QuantumModel. The paper demonstrates two core applications: differentiable quantum machine learning for solving differential equations and a QUBO solver via an analog QAOA-like approach on Rydberg-atom hardware, highlighting the compatibility with AD, adjoint, and GPSR differentiation modes. Overall, Qadence offers a scalable, modular platform that could become a standard tool for developing variational DAQC algorithms and deploying them on near-term hardware.

Abstract

Digital-analog quantum computing (DAQC) is an alternative paradigm for universal quantum computation combining digital single-qubit gates with global analog operations acting on a register of interacting qubits. Currently, no available open-source software is tailored to express, differentiate, and execute programs within the DAQC paradigm. In this work, we address this shortfall by presenting Qadence, a high-level programming interface for building complex digital-analog quantum programs developed at Pasqal. Thanks to its flexible interface, native differentiability, and focus on real-device execution, Qadence aims at advancing research on variational quantum algorithms built for native DAQC platforms such as Rydberg atom arrays.

Figures (5)

• Figure 1: Qadence high-level architecture diagram together with main components and their relations. Programs are primary focused on using the object, which ties a composed with a , useful for machine learning (ML) workloads execution with fully integrated differentiation engines. Here, Qadence builds on top of PyTorch pytorch for automatic differentiation of statevector simulators, while also providing a native integration of generalized parameter-shift rules gpsr2021 for general differentiability of quantum variational programs.
• Figure 2: Qadence uses a flexible block system focused on operations modularity, heavily inspired by the Yao package and Quipper yaoquipper. By default, shows the respective block tree in the console output, as exemplified on the left-hand side for a single QFT block. Blocks can also be tagged with and the circuit visualized with the function available in the module.
• Figure 3: Registers in Qadence encode a NetworkX graph to represent the topology, where each qubit coordinates are node properties. Typically, register constructors will define an edge for each pair of neighboring qubits accessible with the method. While these edges can be used to represent interactions in abstract Hamiltonians, they are not necessarily representative of the interaction topology in real qubit systems. Instead, all node pairs and respective distances can be accessed with the and methods. These are convenient tools for the creation of arbitrary $\mathcal{H}_{\textrm{int}}$, which sum over the complete graph pairs with an interaction strength that decays with the distance (exemplified in Sample \ref{['lst:arb-ham']}).
• Figure 4: A) Solution of the simple non-linear ODE presented in the text using DQC. The solution is computed within the $(-1,1)$ domain for $x$ and the results are compared to the known closed-form solution after gradient-based training with a random uniform sample of 20 collocation points. B) Solution of 2D Laplace equation presented in the text using DQC. The solution is computed within the $[0,1]$ domains for both $x$ and $y$ and the results are compared to the known closed-form solution after gradient-based training with a random uniform sample of 100 collocation points. For both examples, 1000 epochs of training with the Adam optimizer and a learning rate of 0.01 was used. The DQC comprised a hardware-efficient ansatz with 4 qubits and depth = 3, a Chebyshev feature map for data encoding, and a transverse-field Ising Hamiltonian as a cost function.
• Figure 5: QUBO solutions before optimization (left panel) and after optimization (right panel). The red bars are the expected solutions of the problem.