A new arXiv preprint reports that a conserved phase charge, referenced to the pump, constrains both the functions a squeezed-light reservoir can read and the cost of measuring them. The result comes from a hardware-faithful digital twin; no physical device was built.
The rule narrows the machine’s reach
Here, the charge works as a selection rule for readout order. For the specified designed mask ensemble, order-D readout exactly spans the D-assembleable character class—the set of patterns available at that order—and stays inside charge sectors with |q|≤D. At a fixed order, some higher-charge patterns remain outside reach.
The paper’s universality claim is correspondingly qualified. It says the union across polynomial orders is dense in the space of fading-memory functionals, meaning outputs whose dependence on past inputs weakens over time. But at every fixed setting, a charge D+1 character remains a positive distance from the order-D reachable class.
Nonlinearity moves before the detector
The proposed approach places the nonlinear step in the optical dynamics. A χ(2) transducer turns a second moment into a harmonic first-moment signal before detection. Under the theory’s assumptions, the paper says this degree-one readout has polynomial shot cost in inverse accuracy, rather than the super-exponential cost of direct moment readout.
The gains appear in simulation
On the synthetic NARMA2 and NARMA10 benchmarks, optimized QRC had lower normalized mean squared error (NMSE) than the ESN ensemble at both reported finite budgets. For NARMA2, the QRC-versus-ESN values were 0.0110 versus 0.0473 at one budget and 0.0024 versus 0.0147 at the other, margins of 4.3× and 6.1×. On NARMA10, they were 0.168 versus 0.282 and 0.095 versus 0.201, margins of 1.7× and 2.1×. The comparisons were optimized separately by task and budget.
On a four-family RadioML scope—BPSK, QAM16, PAM4 and GFSK—the study used 12,000 snapshots, split 60/20/20. At matched search scale, the phase-encoded machine reached 69.1 ± 0.6% accuracy, compared with 62.0 ± 3.4% for an architecture-matched ESN bank based on 200 draws.
A matched control points in the same direction. With hardware and photon number held identical in the model, phase encoding scored 74.4% on RadioML, versus 56.7% for displacement encoding—a 17.7-point deficit for the control. The reported per-seed deficits were 17.6, 16.9 and 18.5 points.
What the preprint does not settle
A separate control makes the interpretation more cautious. At optimized operating points, a classical-light twin agreed with the squeezed-light machine within 0.4% in a noiseless check; the reported parity ratios were 0.996 and 1.001. That is simulated noiseless parity, not evidence of a finite-budget resource advantage.
The same simulated register showed quantum witnesses without a simple pairwise picture: 47–52 of its 60 bins were sub-vacuum, logarithmic negativity reached 0.59, and no bin pair was entangled. These are simulated covariance witnesses, not physical measurements.
The document is an arXiv version-1 preprint dated 20 Aug 2026, and its exact reachability result applies to a specified designed mask ensemble; whether one frozen mask is sufficient remains open. The universality statement concerns the union over polynomial orders, not a single fixed setting, where the positive charge-sector gap remains.
In short, the preprint offers a mathematical design rule and a simulated performance case, not a hardware demonstration. Its central claim is the combination of charge-limited reach with degree-one optical readout; whether those predictions survive a real device remains unanswered. Numerical results, source data and analysis code are reported as deposited at Zenodo, with the code released under the MIT license.
Paper data and sources
Original title: A charge selection rule fixes what a squeezed-light reservoir computer can compute and afford
Authors: Daniel Soh
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text