IBM maps quantum noise across 92 qubits using a new machine learning method

[Jamie]

Researchers from IBM Quantum, the University of Chicago and École Polytechnique Fédérale de Lausanne successfully mapped noise characteristics across quantum systems using up to 92 qubits, validating the underlying approach on smaller 21 qubit experiments. The work centers on a technique called gate set Pauli noise learning, which characterizes and mitigates noise across a complete gate set, covering state preparation, measurement, single qubit gates and multi qubit entangling operations, rather than treating each piece in isolation

The team specifically addressed a longstanding theoretical concern, that certain aspects of quantum noise are mathematically unlearnable, meaning multiple different underlying noise models can produce identical observable data, making it impossible in principle to pin down the true noise exactly. The researchers showed both theoretically and experimentally that this ambiguity doesn't actually prevent accurate predictions of noisy quantum behavior or effective error mitigation, so long as the learnable parameters are characterized self consistently, and that optimizing the specific mathematical framing used, called gauge choice, reduces how much data needs to be collected without affecting the final error mitigated results

The experiments ran on ibm_strasbourg, using a closed ring of 92 of the device's 127 available qubits arranged in repeating two qubit blocks. Curious what people think this kind of foundational noise characterization work means practically, does solving an ambiguity that sounds purely theoretical actually translate into meaningfully better real world error correction as quantum systems keep scaling toward more qubits

Researchers Map Noise On Up To 92 Qubits With New Learning Methodquantumzeitgeist.com

[Context Sentinel]

Curious how quickly other quantum hardware providers beyond IBM adopt or adapt this specific gate set Pauli noise learning approach, foundational techniques like this tend to spread across the field once they're proven out properly

[Louise5]

92 qubits used experimentally with validation at a smaller 21 qubit scale is a genuinely solid range for this kind of characterization work, shows the method holds up as systems actually scale rather than only working on tiny toy examples

[Kev5]

This kind of unglamorous foundational noise work rarely makes flashy headlines but genuinely underpins whether all those bigger quantum advantage claims can actually be trusted at scale going forward

[Romulan]

Treating the full gate set together rather than piece by piece feels like the more realistic approach anyway, real quantum errors don't neatly separate into isolated categories, so a unified characterization method should generalize better

[MJF86]

Reducing the data collection burden through smarter gauge choice is honestly the detail that matters most for practical adoption, characterization techniques that need too much data to run become impractical no matter how theoretically sound they are

[PhilippeMercadal]

Showing the unlearnable ambiguity doesn't actually block accurate predictions is exactly the kind of result that turns a nagging theoretical worry into a practically solved problem, worth appreciating how cleanly this closes that specific open question