Skip to content

Meeting 2026 06 08

Thomas Naughton edited this page Jul 2, 2026 · 5 revisions

Meeting - June 8, 2026

Upcoming meetings

Agenda

  • Architecture Seminar Series: IBM - QeMCMC (Kate Marshall)

Attendance

Person Institution
Amir Shehata Oak Ridge National Laboratory (ORNL)
Kate Marshall IBM Quantum / University of Oxford
Ross Duncan Quantinuum
In-Saeng Suh Oak Ridge National Laboratory (ORNL)
Masoud Mohseni Hewlett Packard Enterprise (HPE)
Tom Beck Oak Ridge National Laboratory (ORNL)
Brandon Neth Hewlett Packard Enterprise (HPE)
Laura Schulz Argonne National Laboratory (ANL)
Ryan Hoffman Quantinuum
Peter Groszkowski Oak Ridge National Laboratory (ORNL)
Claire Marvinney Oak Ridge National Laboratory (ORNL)
Ryan Landfield Oak Ridge National Laboratory (ORNL)
Daniel Chaves Claudino Oak Ridge National Laboratory (ORNL)
Jack Lange Oak Ridge National Laboratory (ORNL)
Marcelo Amaral IBM
Yoonho Park IBM
Alex Chernoguzov Quantinuum
John Children Quantinuum
Neal Erickson Quantinuum
Gaurav Gyawali Hewlett Packard Enterprise (HPE)
Christian Heiter Hewlett Packard Enterprise (HPE)
Humberto Muñoz Bauza Hewlett Packard Enterprise (HPE)
Andy Stone Hewlett Packard Enterprise (HPE)
Namit Anand Hewlett Packard Enterprise (HPE)
Smriti Bajaj Dell Technologies
Burns Healy Dell Technologies
Stefano Mensa NVIDIA
Doug Jacobsen NVIDIA
Josh Moles IonQ
Sakshi Chhabra Indian Institute of Science (IISc)
Yogesh Simmhan Indian Institute of Science (IISc)
Abdur Rahman Hatim Indian Institute of Science (IISc)
Matthieu Moreau Pasqal
Aleksander Wennersteen Pasqal
Lukas Burgholzer Munich Quantum Valley (MQV)
Arman Agha-Alikhani Munich Quantum Valley (MQV)
Jorge Alfonso Echavarria Gutiérrez Technical University of Munich (TUM)
Doriana Medic University of Turin (UniTO)
Jonathan Skone NERSC / Lawrence Berkeley National Laboratory
Fernanda Foertter University of Arizona

Notes

Meeting Objective

The session focused on presenting recent research on a hybrid quantum-classical optimization algorithm based on Quantum Enhanced Markov Chain Monte Carlo (QMCMC) and its application to solving large-scale combinatorial optimization problems. The discussion explored implications for quantum computing, HPC integration, optimization benchmarking, and future scalability.


Presentation Summary

1. Introduction and Research Context

Kate Marshall introduced herself as a Research Scientist at IBM Quantum and a PhD researcher at the University of Oxford. She explained that the work presented was the result of a collaboration among:

  • IBM Quantum
  • Hartree Centre (UK Science and Technology Facilities Council)
  • Hartree National Centre for Digital Innovation
  • Open Quantum Institute

The work is described in a recent preprint titled:

Quantum Enhanced Markov Chain Monte Carlo for Combinatorial Optimization

The project investigates whether Quantum Enhanced Markov Chain Monte Carlo, originally developed as a sampling algorithm, can be adapted into a practical optimization framework for difficult combinatorial problems.


2. Background: Quantum Enhanced Markov Chain Monte Carlo (QMCMC)

Kate reviewed the original QMCMC approach introduced in prior IBM research.

Classical MCMC

Classical Markov Chain Monte Carlo is commonly used to approximate difficult probability distributions such as low-temperature Gibbs/Boltzmann distributions without explicitly calculating computationally expensive partition functions.

The process consists of:

  1. Starting from an initial configuration.
  2. Proposing transitions to new configurations.
  3. Accepting or rejecting transitions according to an acceptance criterion (typically Metropolis-Hastings).
  4. Building a Markov chain that converges toward the target distribution.

The method satisfies detailed balance conditions that guarantee convergence to the desired distribution.

Quantum Enhancement

In QMCMC:

  • The proposal distribution is generated by a quantum circuit.
  • Acceptance and rejection decisions remain classical.
  • The workflow becomes hybrid quantum-classical.

The original QMCMC work demonstrated:

  • Up to quartic speedup in convergence to target distributions.
  • Strong resilience to noise.
  • Successful demonstrations on small quantum devices.

The original application focused on sampling Boltzmann distributions of spin glasses at low temperatures.


3. Extending QMCMC to Optimization

The central question addressed by the research was:

Can QMCMC be adapted from a sampling framework into an optimization framework?

Instead of converging toward a probability distribution, the objective becomes finding a global optimum of a combinatorial optimization problem.

The team leveraged concepts commonly used in classical optimization, including:

  • Simulated Annealing
  • Parallel Tempering
  • Warm Starting

to construct a hybrid quantum optimization algorithm.


4. Parallel Tempering Framework

A major component of the algorithm is Parallel Tempering.

Concept

Multiple replicas of the optimization process run simultaneously at different temperatures.

Characteristics:

  • High-temperature replicas explore aggressively.
  • Low-temperature replicas focus on exploiting promising regions.
  • Periodic exchanges of information occur between neighboring replicas.

Benefits

This approach helps:

  • Escape local minima.
  • Cross large energy barriers.
  • Improve robustness in rugged optimization landscapes.

Higher-temperature chains can discover promising regions and pass that information to colder chains, which refine solutions more effectively.


5. Relationship to QAOA

Kate discussed how the QMCMC circuits resemble the Quantum Approximate Optimization Algorithm (QAOA).

Key Observations

  • A first-order Trotter decomposition transforms the QMCMC Hamiltonian into circuits resembling QAOA.

  • Unlike traditional QAOA:

    • Parameters remain static.
    • The method does not rely on iterative variational optimization.

Advantages

Because of this similarity, the algorithm benefits from existing QAOA infrastructure:

  • Compilation tools
  • Mapping and routing methods
  • Parameter optimization frameworks
  • Utility-scale execution workflows

available within IBM's ecosystem and elsewhere.


6. Warm-Start Strategy

The algorithm incorporates warm-start techniques commonly used in classical optimization.

Approach

Rather than beginning from random states:

  • Good solutions discovered during earlier iterations are reused.
  • Accepted configurations initialize subsequent circuit executions.

Expected Benefits

  • Faster convergence
  • Better search efficiency
  • Improved utilization of previously discovered information

This warm-start mechanism becomes an important component of the iterative QMCMC process.


7. Overall Algorithm Workflow

The resulting optimization algorithm combines:

Quantum Components

  • QAOA-like proposal circuits
  • Quantum sampling engine

Classical Components

  • Metropolis-Hastings acceptance step
  • Parallel tempering exchanges
  • Warm-start initialization
  • Noise mitigation and post-processing

The workflow follows:

  1. Generate candidate solutions via quantum circuits.
  2. Collect many samples from hardware.
  3. Apply classical filtering/noise mitigation.
  4. Evaluate solutions with Metropolis-Hastings.
  5. Exchange accepted solutions among temperature replicas.
  6. Warm-start the next iteration.

This cycle repeats until convergence.


8. Benchmark Problem: Maximum Independent Set (MIS)

The team evaluated the algorithm using Maximum Independent Set problems.

Problem Definition

Find the largest subset of graph vertices such that no two selected vertices share an edge.

Reasons for choosing MIS:

  • NP-hard combinatorial optimization problem.
  • More challenging than standard QUBO formulations.
  • Well-established benchmarking resources available.
  • Existing IBM tooling supports mapping MIS instances onto heavy-hex hardware.

9. Experimental Setup

Hardware

Experiments were conducted on IBM Heron quantum processors.

Largest Problem

  • 117-node graph
  • 117 physical qubits utilized

Benchmark Library

Problem instances were drawn from the Quantum Optimization Benchmarking Library.

The chosen graphs intentionally did not map naturally onto heavy-hex hardware, ensuring the problems remained non-trivial.


10. Noise Mitigation Strategy

A major challenge involved sampling-based quantum workloads.

Traditional expectation-value error mitigation techniques are less applicable.

Instead, the team adopted a strategy based on previous IBM work:

Method

  • Collect many samples per iteration.
  • Retain only low-energy candidate solutions.
  • Discard poor-quality samples likely affected by noise.
  • Resample from the filtered distribution.

Scale

  • Approximately 10,000 shots per iteration.
  • Roughly 3 seconds per experiment iteration on hardware.

This simple strategy proved highly effective.


11. Results

Convergence

The algorithm successfully reached optimal solutions for all tested MIS instances.

Solutions were verified against:

  • IBM CPLEX classical solver

which confirmed optimality.


Scaling Analysis

Five graph instances were studied, ranging from:

  • 17 qubits
  • Up to 117 qubits

Comparison 1: Classical Proposal vs Quantum Proposal

Classical baseline:

  • MCMC with classical bit-flip proposals

Quantum variant:

  • QMCMC proposal generation using quantum circuits

Observations:

  • Classical proposal strategy outperformed one-shot quantum simulations.
  • However, multi-shot hardware executions showed improved performance.

Comparison 2: Hardware vs Simulation

Interestingly:

  • Real quantum hardware outperformed tensor-network simulations at larger scales.
  • Around 60+ qubits, simulation truncation errors became more limiting than hardware noise.

This result was viewed as particularly encouraging.


Device Generational Improvements

Performance comparisons across IBM Heron generations showed:

  • Heron R3 significantly outperformed Heron R1 and R2.

This demonstrated tangible hardware progress and improved optimization performance.


Discussion and Q&A

Question: Fairness of Classical Baselines

Masoud Mohseni raised concerns regarding benchmarking methodology.

Key points:

  • Standard MCMC may not be the most appropriate classical counterpart.
  • Modern variational probabilistic classical methods can mimic some QAOA behavior.
  • More sophisticated non-local classical moves may provide stronger baselines.
  • Claims of quantum advantage require increasingly rigorous comparisons.

Response

Kate agreed.

She clarified:

  • No quantum advantage claims were being made.
  • Alternative classical approaches were explored.
  • The work should be viewed as an early exploratory study.
  • Future large-scale experiments should include stronger benchmark methods.

Question: Did Solutions Reach Classical Optima?

Ross Duncan asked whether optimal solutions were achieved.

Response

Kate confirmed:

  • Every tested instance reached the classical optimum.
  • CPLEX was used for validation.
  • Hardware size, not solution quality, was the limiting factor.

Question: HPC and Parallelism

Amir Shehata requested clarification regarding:

  • Quantum parallelism
  • Classical parallelism
  • Data exchange mechanisms

Response

Kate explained:

Quantum Parallelism

Multiple temperature replicas ideally execute simultaneously on:

  • Separate quantum devices
  • Or partitions of larger quantum systems

Due to hardware limitations, experiments were run sequentially rather than fully in parallel.

Classical Parallelism

Classical post-processing includes:

  • Shot filtering
  • Noise mitigation
  • Replica exchange
  • Acceptance/rejection evaluation

These stages can naturally be parallelized.

Data Exchange

Neighboring temperature chains exchange accepted candidate solutions.

This enables hotter chains to help colder chains escape local minima.


Question: HPC Requirements

Amir asked whether HPC resources were required.

Response

Kate noted:

Current experiments did not require runtime HPC resources because:

  • Only 10,000 shots were used.
  • Post-processing workloads remained manageable.

However:

  • Larger-scale experiments
  • Millions of shots
  • Thousands of nodes

would likely require tightly integrated HPC and quantum systems.


Question: Tight Coupling Requirements

Ross Duncan questioned whether quantum-classical coupling truly needed to be tight.

Response

Kate explained:

While experiments could theoretically pause and resume:

  • Queue delays would dramatically increase execution times.
  • Hardware calibration drift could impact assumptions used for shot estimation.
  • Long-running iterative workflows benefit substantially from low-latency quantum-classical interaction.

Thus, tighter integration becomes increasingly valuable at scale.


Question: Queue Times and Runtime Sessions

Amir asked about queue delays.

Response

Kate reported:

  • Typical queue wait: under 24 hours.
  • Often under 6 hours.
  • IBM Runtime Sessions were used.
  • Once a session began, the experiment retained access throughout execution.

This avoided repeated queue delays between iterations.


Question: Fault-Tolerant Quantum Computing

Amir asked how fault-tolerant hardware would affect the workflow.

Response

Kate stated that fault tolerance would likely:

  • Reduce required shot counts.
  • Eliminate much of the noise mitigation overhead.
  • Improve solution quality.
  • Potentially reduce total iteration counts.

The overall algorithm structure would remain largely unchanged.


Key Takeaways

  1. QMCMC can be adapted from sampling to combinatorial optimization.

  2. The proposed algorithm combines:

    • QMCMC
    • Parallel Tempering
    • Warm-Start QAOA-inspired circuits.
  3. Maximum Independent Set problems up to 117 qubits were solved on IBM hardware.

  4. All benchmark instances reached verified classical optima.

  5. Hardware performance exceeded tensor-network simulations at larger scales.

  6. No quantum advantage claims were made; results are exploratory.

  7. HPC resources were primarily used for:

    • Simulation
    • Parameter tuning
    • Benchmarking rather than runtime execution.
  8. Future work will focus on:

    • Larger graph instances
    • Multi-device execution
    • Stronger classical baselines
    • HPC-quantum integration at scale.

Files

Clone this wiki locally