Research on diagnosing the errors of quantum sensors and quantum computers.
These are research examples. The sensing examples simulate observations from published instrument parameters, and eight verifications reproduce the conclusions of published experiments. The computing examples run on simulated evidence. The project site, quantum.deepcausality.com, documents each check, its proof status, and its boundaries.
Gravimeter error diagnosis
Problem
An atom gravimeter reads 5 µGal low, and seven known effects could each explain the offset, from the Coriolis force to the laser wavefront. More passive data fits all seven equally well, and isolating one effect can take hours of setup.
Approach
DeepCausality Quantum prices each instrument change in instrument time and plans the cheapest set that separates every pair of explanations. It runs the set as a campaign, re-planning as the tide moves, until one explanation remains.
Result
Planning runs in simulation, so the instrument spends its time only on the changes that separate the explanations. In the simulated campaign, those changes take at most 83 minutes of instrument time, whichever of the seven effects is responsible, and under three minutes when it is the quadratic Zeeman shift.
Limits
The example simulates observations from a published instrument; its setup times and tide stand in for a lab’s own values. The library weighs only the listed causes: if the true cause is missing from the list, a listed cause takes the blame.
A gravity gradiometer compares two atom clouds that share one laser. Light leaking between the clouds or a rotating platform biases the gradient, yet the passive data cannot tell either from the shared laser, which leaves the gradient correct.
Approach
The same planner selects the instrument changes that separate the four explanations and runs the cheapest first.
Result
One 60 s test exposes a leak: with one cloud lit and the other dark, the dark cloud’s signal rises from 0.020 to 0.079. A 6 % leak biases the gradient by −356 E. The full plan, including a deliberate turn of the platform, takes about 722 s.
Limits
The example simulates its observations; the 6 % leak and the 2 % dark background stand in for a lab’s values.
A planning method earns trust only if it reaches the answers that real instruments gave.
Approach
Eight verifications feed the published numbers of atom-interferometry and crosstalk papers into the same planner and test whether it reaches each paper’s conclusion.
Result
All 53 checks pass, and each verification reaches the paper’s own answer: for example, the wavefront behind a temperature-dependent shift (Karcher et al.) and detection clipping behind a wrong-sign Coriolis result (Farah et al.). Several also expose inconsistencies in a paper’s own numbers.
Limits
The verifications read published values and data regenerated from published models, not raw instrument records.
Two qubits show correlated errors. Qubit 1 may drive qubit 2, qubit 2 may drive qubit 1, or a shared bath may drive both. The three structures are Markov equivalent, so passive data cannot separate them.
Approach
The Quantum Causal Language screens each candidate as a quantum causal model, solves a minimum-cost cover for the experiments that separate every pair by at least 5 bits, and adjudicates against the observation.
Result
The plan holds qubit 1 excited, then qubit 2, at a total cost of 2 against 200 for process tomography, with the tightest pair 99.5 bits apart. On 1,024 simulated shots the observation, 0.377 ± 0.015, leaves one structure standing: qubit 1 drives qubit 2.
Limits
Costs and predicted read-outs are inputs, so the 2-to-200 ratio is the example’s, not a laboratory measurement. The observation is drawn from the surviving candidate’s own prediction.
A state-vector check of a code is bounded by simulator size; this library’s simulator stops at 24 qubits.
Approach
Exact checks over 𝔽₂ derive the code’s parameters from its lattice and decide check weights, logical-gate invariance on homology classes, and Clifford action, with exact rational phases and no quantum state.
Result
On Kitaev’s toric code on a 4 × 4 torus, the pipeline derives [[32, 2]] and accepts all four checks. At a weight bound of 3, the weight check rejects at the first Z check, which shows that it can fail.
Limits
No code distance or threshold is computed, one lattice size is shown, and only the weight check is driven to rejection in this run.
Switching codes reaches gates that one code cannot perform transversally, and noise inside the switching gadget adds logical error.
Approach
A composition law bounds the error of stacked abstractions from the residual of each link, and each run reports the bound beside the measured residual at f32, f64, and 106-bit precision.
Result
When two logical qubits move from a [[4,2,2]] code to an [[8,2,2]] toric code with 1/10 depolarizing noise on one logical wire, the law bounds the residual at 3.695 and the run measures 0.4619 at all three precisions. Without noise, the switch is exact.
Limits
Residuals use the Frobenius norm on Choi operators, not the diamond norm, so the bound is loose by a factor of eight. One query is checked, and Lean proves only the exact case.
The source code and verification records for every example are public.
The Center undertakes consulting, contract research, and training for
industry partners.