The Coherence Cost of Telling Quantum Channels Apart

Suppose you want to learn which of two unknown quantum processes, or channels, is acting on your system, and every query costs you something. A natural guess is that the scarce ingredient is entanglement or quantum memory. For many pairs of channels, though, the ingredient that actually does the work is coherence: superposition relative to a fixed, cheap-to-prepare basis. Two phase rotations look identical to any probe that is diagonal in that basis. Coherent probes are expensive to make, but a probe is not necessarily used up after one query. A phase rotation leaves its coherence intact, while a dephasing channel erodes it. This note asks how much coherence an experimenter must pay, per channel query, to tell channels apart at a given error, and whether reusing the leftover probe brings that per-query cost below the obvious "one fresh coherent probe per query."

The second ingredient comes from the theory of implementing channels. Using and reusing coherence to realize quantum processes showed two things: the coherence needed to build a channel equals the coherence that channel can generate, and leftover coherence in a helper register can be recycled, so the long-run "amortized" cost per use falls below the one-shot cost. That paper never asked what recycling buys in a discrimination task. Channel discrimination theory, for its part, counts queries and memory but treats coherence as free. We put the two together. The tester may use only operations that cannot create coherence on their own, plus a metered supply of maximally coherent qubits ("cosbits"). We then measure the error achievable with $n$ queries and $m$ cosbits.

Core Claim

Define $P_e(n,m)$ as the optimal error for discriminating channels $\mathcal{N}_0$ and $\mathcal{N}_1$ using $n$ queries. The tester is any sequence of maximally incoherent operations (MIO, maps that send every basis-diagonal state to a basis-diagonal state) interleaved with the queries, starting from $m$ cosbits. The quantity of interest is the coherence rate

$$r(E) = \liminf_{n\to\infty} \frac{m_n}{n} \quad \text{such that} \quad -\tfrac{1}{n}\log P_e(n,m_n) \ge E,$$

which is the number of cosbits per query needed to sustain an error exponent $E$.

We make three claims. The first is that $r(E)$ separates channel families sharply. For coherence-preserving families such as phase rotations, recycling drives $r$ to zero, so one probe can serve many queries. For coherence-eroding families such as dephasing channels of differing strength, $r$ is strictly positive under recycling and strictly below the fresh-probe value of one cosbit per query. The second is a conjecture: for recycled strategies, $r$ is set by the coherence each query destroys in the operating probe, measured as log-robustness of coherence lost per round, and not by the coherence each query needs. This would be a discrimination analogue of the García Díaz amortization result.

The third claim is where accessible coherence enters. Inequivalent multipartite coherence classes and two operational coherence monotones defines accessible coherence as how many target states a given state can be turned into under free operations. Our claim is that the reusable value of a leftover probe is governed by this reachability, not by how much coherence the probe contains. The same paper shows that coherence conversions are often incomparable and fall into infinitely many classes. That is direct evidence that a single scalar amount, such as robustness, can mispredict which leftover states can be reshaped into the next good probe.

Research Program

The computational core is a semidefinite program, the convex optimization class used throughout this literature, over the Choi operator of an $n$-slot tester. The Choi operator is the matrix encoding of the whole multi-round strategy. Each inter-query operation carries the linear MIO constraint, and the input is fixed to $m$ cosbits. MIO keeps the problem tractable in the same way it did for Using and reusing coherence to realize quantum processes. We expect to solve it exactly for qubit channels up to about four queries and for qutrits up to about three. Larger $n$ will use structured ansätze: one recycled probe topped up with fresh coherence each round, then measured with an incoherent measurement at the end. Finite-$n$ exponents are fitted from these curves.

We pair each optimum with a lower-bound argument. The argument tracks the log-robustness of the probe through the rounds, using the monotonicity of channel robustness $C_R$ under free operations. The test of the second claim is whether this bound becomes tight on the numerical frontier.

Because MIO allows some coherence creation under post-selection, we repeat the qubit cases under strictly incoherent operations (SIO). SIO operations preserve incoherence Kraus operator by Kraus operator, so they are closer to what a laboratory can do. The SIO computations use direct parametrization of small Kraus decompositions. These give inner bounds only, and we will report them as such. The comparison speaks directly to the MIO-versus-physical question that García Díaz et al. left open.

For the third claim we extend accessible coherence from pure to mixed qubit states numerically. For qubits, Luo et al. already have closed forms, and reachable sets are low-dimensional, so the mixed-state extension is feasible. We then ask whether accessible coherence predicts recycling value better than robustness does.

Finally, the flagpole-state result says that in dimension three or more, any pure coherent state can implement coherent unitaries. That result suggests a dimension divide in $r$. We predict that qutrit testers recycle strictly better than qubit testers on matched families. This is a falsifiable consequence of the framework, not an assumption.

Baselines

The naive reference point is the fully incoherent tester with $m=0$. For phase-rotation families it can do no better than guessing the more likely hypothesis. For dephasing families it may do somewhat better, and we compute it exactly with the same program by setting $m=0$.

The first real competitor is the fresh-probe tester. It uses one cosbit per query and a new probe each time, which corresponds to the product strategies that Ultimate Limits of Quantum Channel Discrimination shows fall short of the optimal exponent when coherence is unrestricted. It runs in the same program with the same $n$, the same incoherent measurement class, and the leftover probe forced to be discarded. That configuration is also our key ablation: it differs from the recycled tester only in whether the leftover probe is kept.

The ceiling is the unconstrained optimum. This means dropping the MIO constraint in the same program, so testers can use arbitrary operations and unlimited coherence; it is the regime whose limits Fang, Gour, and Wang characterize.

For the Hamiltonian-recognition family, the strongest reported number comes from Optimal Hamiltonian recognition of unknown quantum dynamics. With $k$ queries and prior $p$, they report average binary success probability $(k+\max\{p,1-p\})/(k+1)$, so the error decays as $O(1/k)$, and they confirm this on a 13-qubit processor. Our coherence-constrained tester counts as matching it if it reaches this curve with $m$ growing sublinearly in $k$, and as beating it only in the specific sense of using fewer cosbits for equal success. Nothing unconstrained can exceed this curve.

On the higher-dimensional families, Unambiguous discrimination of general quantum operations saturates the theoretical success bound for up to six displacement operators. It reports process fidelities of roughly 95% to 99% for block-Pauli noise at error rate 0.5 in four dimensions. For any coherence-limited version of these tasks, we found no reported result on Lacuna.

Two single-shot resource predictors serve as baselines for the third claim. One is channel robustness $C_R$ from García Díaz et al. The other is projective robustness $\Omega$ from Projective robustness for quantum channels and measurements and their operational significance, which equals the one-shot advantage in a combined discrimination-and-exclusion task. Both are scored on how well they rank channel families by their measured $r$.

Why not just…?

Why not just use the memory hierarchy of Ohst et al.? Characterising memory in quantum channel discrimination via constrained separability problems restricts the dimension and type of the memory carried between queries. It does not ask how coherent that memory is. A large memory that stays diagonal in the reference basis is effectively classical memory, so coherence is a separate axis. The real risk is that, for every family we test, the coherence-budget frontier simply reproduces the memory-dimension frontier. We will compute both on the same channel families. If they coincide everywhere, the coherence rate adds nothing and we will say so.

Why not just predict performance with $C_R$ or $\Omega$? These are the natural one-number summaries, and $\Omega$ is exactly operational for a single query. They assign a channel a fixed value, though, and say nothing about how a probe degrades across rounds. Phase rotations and weak dephasing can have similar single-shot values but very different $r$ once recycling is allowed. The rank-correlation comparison settles whether this gap is real.

Why not just give each query a fresh probe? That is the simplest protocol, and for unconstrained exponents Fang, Gour, and Wang already show product inputs are suboptimal. The fresh-probe ablation answers the narrower question we care about: how many cosbits recycling saves at equal error. If the savings vanish for noisy families, the recycling claim fails there.

Why bring in accessible coherence rather than just robustness of the leftover probe? Robustness is cheaper to compute and may be enough. We treat accessible coherence as a hypothesis that has to win this comparison. If robustness predicts recycling value equally well, accessible coherence will be reported as unnecessary for this task.

Proposed Experiments

We test four families, all at both equal and skewed priors. The first is a pair of qubit phase rotations by $\pm\theta$. The second is a pair of qubit dephasing channels with different strengths. The third is the X-versus-Z Hamiltonian-recognition task with unknown angle. The fourth is the block-dephasing channels in dimensions three and four, plus a small displacement set from Cai et al. truncated to the Fock levels used there.

For each family we produce two outputs: the full $P_e(n,m)$ surface, and $r(E)$ for fresh-probe and recycled strategies under MIO, with SIO inner bounds for qubits.

Four outcomes are decisive. The first is whether recycled $r$ is zero for phase families and strictly between zero and one for dephasing families. The second is whether the log-robustness-loss lower bound is tight. The third is whether accessible coherence of the leftover probe beats robustness at predicting recycling value, measured as rank correlation across sampled operating points. The fourth is whether qutrits show the predicted recycling advantage over qubits.

If the dephasing-family savings survive under SIO, a two-round recycled protocol is a natural follow-up on a cavity-transmon platform like that of Cai et al., which already runs generalized measurements in dimension six. Hardware decay there produces nonzero error rates, so any such demonstration would be reported against their measured fidelities rather than ideal bounds.