The Signal
Mele and Bittel match channel tomography’s dimension scaling
A Nature Physics paper gives a channel-tomography protocol whose query count scales with input dimension, output dimension and Kraus rank. Its upper bound still has an accuracy dependence that the paper says is not fully matched by lower bounds.

A dimension-scaled query bound
A protocol for learning unknown quantum channels uses a number of channel queries that scales with the channel’s degrees of freedom, according to a paper by Antonio Anna Mele and Lennart Bittel published in Nature Physics on 12 November 2026. The work establishes matching upper and lower bounds on the dependence of query complexity on input dimension, output dimension and Kraus rank, for a fixed accuracy in diamond distance.
The result is theoretical: it gives an algorithm and proves query-complexity bounds, rather than reporting a test on quantum hardware. Channel tomography seeks a classical description of an unknown process; diamond distance measures how distinguishable two processes can be, including when the input is entangled with an auxiliary system.
Queries and the remaining gap
For a channel with input dimension din, output dimension dout and Kraus rank at most k, the paper’s upper bound is O(dindoutk/ε²) uses to reach diamond-distance error ε with a chosen constant success probability. The corresponding lower bound is Ω(dindoutk); when the allowed Kraus rank exceeds the minimum, the authors also prove a separate Ω(1/ε²) lower-bound contribution.
For general channels on a d-dimensional system, earlier schemes cited in the paper required about d6 queries, while the number of independent parameters is Θ(d4) and earlier lower bounds scaled as about Ω(d4) at constant accuracy. The new bound brings the dimensional scaling to Θ(dindoutk) at fixed accuracy, but it does not settle the full accuracy dependence: the paper says its upper and lower bounds do not yet match in ε.
From channel uses to a description
The procedure is non-adaptive at the channel-use stage: it makes all calls in one parallel block. It feeds half of a maximally entangled state into each channel use to prepare copies of the channel’s Choi state, then applies a purification step and performs pure-state tomography on the resulting states.
Mele and Bittel show that this route can yield a diamond-distance guarantee without a dimension-dependent loss in converting the estimated Choi state back into a channel description. The output is a classical description whose diamond distance from the unknown channel is at most the target error.
Related tomography tasks
The framework also covers other learning problems as special cases. For rank-one channels, the paper gives a bound for learning isometries; it also gives an O(d2/ε2) query bound for learning binary quantum measurements in operator norm.
The authors note that their channel-tomography bounds match in their dependence on input dimension, output dimension and Kraus rank, while finding the sharp accuracy exponent for different allowed Kraus ranks remains open.
Source
Related stories
- Quantum channel paper sets new average-case testing bounds
- Stony Brook Proves Quantum Algorithms Need Exponential Queries for Welded-Tree Paths
- Preprint expects no exponential DQI speedup with local decoders
- Study reports higher simulated Max-Cut ratios with cat-qubit QAOA
- DLR researchers report shallower circuits for quantum data loading