Channel-Shaped Bridges for Recovering Quantum States from Decohered Estimates

When a quantum device loses coherence, an experimenter is left with a mixed density matrix. A density matrix is the object that describes a quantum state, including its classical uncertainty. Decoherence does not scramble this matrix uniformly. Dephasing erases the off-diagonal "coherence" terms and leaves populations intact, while amplitude damping pulls the state toward the ground state along a known direction. Generative models for quantum states ignore this structure. The only density-matrix diffusion model in our file, the structure-preserving diffusion model of Quantum State Generation with Structure-Preserving Diffusion Model, starts every sample from pure Gaussian noise. That is the same waste image restoration abandoned when it moved to data-to-data bridges: methods that transport a degraded input directly to a clean output instead of first dissolving it into noise.

We propose a bridge that starts from the decohered state itself and runs in the unconstrained "mirror" coordinates where every point maps back to a physically valid state. It adds noise only along the Pauli directions that the physical channel actually damped. The idea in two sentences: treat decoherence like a known image degradation and run a short stochastic bridge from the decohered state back toward the clean one. Do this in log-density coordinates so that every sample is a valid quantum state, and inject randomness only where the channel destroyed information, which mirrors how frequency-controlled samplers inject only high-frequency noise in images.

Core Claim

A diffusion bridge (a learned stochastic process that interpolates between paired endpoints) stays exactly as tractable after the mirror map. The mirror map takes a density matrix $\rho$ to $\theta = \log \rho$, a real vector of Hermitian-matrix coordinates. Its inverse, $\rho = e^\theta/\mathrm{Tr}\, e^\theta$, always returns a matrix that is positive semi-definite, Hermitian and of unit trace. Because $\theta$ lives in flat space, the closed-form Gaussian transition kernels used by I2SB: Image-to-Image Schrödinger Bridge and by the mean-reverting process of Image Restoration with Mean-Reverting Stochastic Differential Equations apply without modification. This resolves the geometric mismatch the file flags between Euclidean bridges and the cone of valid states. It needs no Lie-group machinery such as TDM.

The substantive claim sits on top of that. The bridge noise should be shaped by the channel. Write $\theta$ in the Pauli basis and let $\lambda_P \in [0,1]$ be how much the channel shrinks Pauli component $P$. These are the diagonal entries of the Pauli transfer matrix for Pauli channels, and its singular values otherwise. We set the bridge noise covariance to

$$\Sigma = \mathrm{diag}\big(\sigma_0^2\,(1-\lambda_P)^{\gamma} + \epsilon\big)_P ,$$

where $\sigma_0$ sets the overall scale, $\gamma$ controls how sharply noise concentrates on damped directions, and $\epsilon$ is a small floor. Components the channel preserved get almost no noise, so the sampler does not re-estimate them. This is the quantum analogue of the observation in NFCDS that low-frequency noise damages structure the observation already fixes. It also replaces that paper's hand-tuned spectral cutoff with a covariance derived from the physics.

We predict three things. With noise shaped this way, the bridge reaches a given fidelity in 2–10 function evaluations. It produces calibrated posterior samples rather than a single averaged state. Its advantage over simple inversion appears exactly where the channel is strong enough to be non-invertible in practice.

Research Program

The data come from the synthetic 4-qubit and 6-qubit state families of the structure-preserving diffusion paper: product, pairwise-entangled and fully entangled states. Each clean state is pushed through a Lindblad channel, meaning a standard master-equation model of decoherence. We use dephasing, amplitude damping, depolarizing, and a mixed channel whose relaxation and dephasing times follow the calibrated IBM Kingston snapshot used in QMaxCal. The decohered state is then estimated from a finite number of shots through Pauli or classical-shadow measurements, so the input carries both physical decoherence and statistical noise.

Clean pure targets sit on the boundary of the valid-state cone, which the mirror map sends to infinity. We therefore regularize targets to $(1-\delta)\rho + \delta I/d$ and report how results depend on $\delta$. This is a real cost of the design, and we state it openly.

Training follows the I2SB recipe of simulation-free denoising score matching on paired endpoints, carried out in $\theta$-coordinates with the shaped covariance $\Sigma$. We also run a mean-reverting variant anchored at $\log \rho_{\text{decohered}}$ for comparison. The network is a transformer over Pauli-coefficient tokens, and it is conditioned on the channel parameters.

The analysis part asks when a bridge can beat a point estimate at all. Strong dephasing makes many pure states consistent with one decohered matrix, so the true posterior is multimodal. The best single fidelity estimate is then a mixed average of those states. The value of sampling lies in the calibrated spread of the samples, and a single estimate cannot supply it. For 2–3 qubits we compute exact posteriors by Markov chain Monte Carlo against the known channel and prior. This gives a ground-truth reference for how well the samples match.

Baselines

The naive reference is to do nothing beyond standard maximum-likelihood tomography of the decohered state. This means reporting the physically projected estimate as the answer. It shows how much information recovery is actually available.

The second baseline is channel inversion followed by projection back onto valid states. We apply the inverse Pauli transfer matrix to the estimate and clip negative eigenvalues. It uses the same channel knowledge our method uses, so it is the fairest physics baseline.

The third is one-shot regression. The same transformer, with the same data and parameter count, is trained to output $\theta$ directly with a fidelity or trace-distance loss. Because it decodes through the matrix softmax, it is also guaranteed valid.

The fourth is the strongest existing generative method, the structure-preserving diffusion model of Quantum State Generation with Structure-Preserving Diffusion Model. It is extended to conditional restoration with classifier-free guidance on the decohered state, keeps its pure-Gaussian start, and is run at 10 to 100 steps. That paper reports 100% validity for its own samples. It also reports that Euclidean diffusion fails the positive semi-definite check in nearly all 6-qubit samples. Its distributional metrics are described only as "orders of magnitude better," and no reported result found on Lacuna for fidelity of density-matrix restoration from decohered inputs by any method.

The fifth is a Euclidean I2SB applied directly to matrix entries. We include it to confirm that the mirror coordinates are necessary.

The key ablation replaces the shaped covariance $\Sigma$ with isotropic noise of matched total variance, keeping everything else fixed. All methods get identical training pairs, identical GPU budgets, and the same number of hyperparameter-search trials.

Beating the baselines means two things. At a fixed budget of 10 or fewer evaluations, the shaped bridge must match or exceed the guided diffusion model run at 100 steps on mean fidelity. It must also beat regression and inversion on posterior calibration in the strong-channel regime, measured by sliced Wasserstein distance to the exact posterior and by coverage of observable intervals.

Why not just…?

Why not just invert the known channel and project? For mild, invertible Pauli channels with many shots, this may well win, and we expect it to. The inverse transfer matrix amplifies shot noise by $1/\lambda_P$, so it degrades exactly as components approach full erasure. Our experiments sweep channel strength and shot count to locate the crossover. Our claim is restricted to the regime beyond it. If no crossover appears, the proposal fails.

Why not just train a one-shot regressor with a valid-state output? This is the most serious threat to the fidelity metric, because a point estimate of the posterior mean is close to fidelity-optimal. We concede that the regressor may tie on mean fidelity. The test that separates the two is posterior calibration against exact Markov chain Monte Carlo posteriors on small systems, together with downstream decisions that depend on uncertainty, such as whether a state's entanglement exceeds a threshold. If the regressor paired with a simple variance head matches the calibration, the bridge is not worth its cost, and we will report that.

Why not just condition the existing structure-preserving diffusion model on the decohered state? This is the natural extension of prior work. It re-synthesizes the populations and surviving coherences from noise. The image evidence from DiffIR and I2SB says this costs steps and fidelity, since I2SB holds near-peak quality at 2–10 steps where a noise-started conditional model collapses. Whether that transfers to density matrices is what the step-versus-fidelity curves will show.

Why not a neural tomography ansatz fit to shadows? Neural-Shadow Quantum State Tomography fits a pure-state transformer by minimizing infidelity against classical shadows. It fixes the phase failure of cross-entropy neural tomography, which stagnates near infidelity 1.0 on its QCD benchmark. It still represents a pure state fit per instance, with neither mixed outputs nor an amortized posterior. We include it as a per-instance reference on the pure-target tasks.

Proposed Experiments and Decisive Checks

On the 4- and 6-qubit families under each channel, we report the following: mean fidelity and trace distance to the regularized clean state; validity rate; number of function evaluations; sliced Wasserstein and energy distances between the sample set and the clean ensemble; and, on 2–3 qubits, the distance to exact posteriors plus the coverage of 90% intervals on Pauli expectations and entanglement negativity.

Three results decide the proposal. First, the isotropic-noise ablation must lose to the shaped bridge at low step counts. If it does not, the channel-shaping claim, which is the central new ingredient, is false. Second, a strength sweep must show the crossover where the bridge overtakes channel inversion. Third, the bridge must beat the variance-headed regressor on posterior calibration while tying or winning on fidelity. A Kingston-calibrated 6-qubit run, where the noise is non-uniform with a single noisy site, tests whether shaping still helps when the damped directions are spatially localized rather than uniform.