On the Optimality and Generalization of Quantum Oracle Sketching

[Zha+26] introduce quantum oracle sketching (QOS), a streaming algorithm that constructs
an approximate phase oracle of a Boolean function f : [N] → {0, 1} from M =
Θ(NQ2/ε) classical samples and then executes any quantum query algorithm of complexity
Q on the sketched oracle. Combined with an interferometric variant of classical shadow
tomography, QOS yields exponential separations in machine size between quantum and classical
learners for linear systems, classification, and dimensionality reduction. The argument
is information–theoretic and unconditional, depending only on quantum mechanics.
This work offers a rigorous critique and constructive extension. We (i) sharpen the diamond–
norm guarantee using non–commutative matrix Bernstein, recovering the linear N/M
bias scaling without the spurious logarithmic factor and with explicit constants; (ii) recast
the optimal Q2 scaling as a quantum Cram´er–Rao bound on the symmetric–logarithmic–
derivative (SLD) Fisher information of the random oracle channel; (iii) generalize QOS
to non–uniform priors by showing that sample complexity is governed by the effective dimension
Neff = ∥p∥2
1/2, allowing genuine improvements when data are heavy–tailed; (iv) replace
the hard spectral–gap assumption in PCA by a polynomial–filter analysis that pays only
quadratically in the soft gap; (v) building on the parallel manuscript of [Sep26], extend
QOS to non–linear kernel learning via random Fourier features, deriving the explicit sample
complexity for shift–invariant kernels via an independent route and connecting it to the information–
geometric framework of (ii) and the effective–dimension framework of (iii); (vi) prove
that QOS is structurally compatible with R´enyi differential privacy, achieving a per–sample
privacy budget that scales as ε2
DP = Θ(1/(NM)); (vii) analyse a hybrid scheme that pre—
composes QOS with a Johnson–Lindenstrauss embedding, isolating the regimes in which the
quantum advantage survives classical compression; (viii) building further on [Sep26], give a
matrix Bernstein proof of the QOS noise threshold pg ≲ ε2/(NQ2) with explicit constants,
and pose the matching tight–threshold question as an open problem alongside the recent
noisy–random–circuit obstruction [Aha+23]. We close with a candid critique of the resource
accounting underlying the “60–logical–qubit” demonstration and a list of open problems in
soft–gap, non–Hermitian, and continuous–domain QOS.