Quantum Oracle Sketching

• Quantum oracle sketching is a quantum algorithmic framework that creates compressed representations of large datasets using sequential, randomized quantum operations.
• It employs incremental quantum rotations to simulate phase oracles, achieving exponential memory compression and enhanced sample efficiency.
• Empirical applications in sentiment analysis and single-cell RNA-seq validate its theoretical guarantees and practical quantum advantage over classical methods.

Quantum oracle sketching is a quantum algorithmic framework for succinctly accessing and processing massive classical datasets entirely through random classical samples, enabling exponential space advantages and quantum speedup in machine learning and classical data analysis tasks. The technique constructs compressed quantum representations—“sketches”—of classical oracles or data streams via sequential, randomized quantum operations, circumventing the classical data loading bottleneck and outperforming any classical algorithm in memory-constrained or streaming scenarios, particularly when both prediction accuracy and sample efficiency are required (Zhao et al., 8 Apr 2026).

1. Problem Formulation and Oracle Model

The principal objective of quantum oracle sketching (QOS) is the estimation of a Boolean function $f(o) \in \{0,1\}$ defined on an unknown oracle $o : [N] \to \{0,1\}$, where $o \in \{0,1\}^N$. Direct access to the oracle via coherent quantum queries is restricted; instead, the only available interaction is through passive, noisy classical samples.

Sampling occurs via a noisy encoding function $g: \{0,1\}^b \times \{0,1\}^b \to \{0,1\}$ with a small discrepancy $\mathrm{disc}(g) \le 2^{-\eta b}$. Letting $G = (g, \ldots, g)$ act in parallel on $N$ coordinates, data is streamed as $$z_t = (x_t, Y^{(0)}_{x_t} \text{ or } Y^{(1)}_{x_t}, \alpha_t)$$, where $x_t$ is uniformly random, $\alpha_t \in \{0,1\}$ (toggling every $o : [N] \to \{0,1\}$0 steps), and $o : [N] \to \{0,1\}$1 are uniformly random preimages of $o : [N] \to \{0,1\}$2 under $o : [N] \to \{0,1\}$3. Both classical and quantum learners process the sample stream, with a classical algorithm constrained by memory $o : [N] \to \{0,1\}$4 and sample complexity $o : [N] \to \{0,1\}$5.

2. Quantum Oracle Sketching Algorithm

QOS replaces the classical storage of $o : [N] \to \{0,1\}$6 with an in-place quantum simulation of the phase oracle

$o : [N] \to \{0,1\}$7

built up via a sequence of incremental quantum rotations informed by streaming classical samples. The target unitary evolution is defined as

$o : [N] \to \{0,1\}$8

with $o : [N] \to \{0,1\}$9. For each sample pair $o \in \{0,1\}^N$0 with $o \in \{0,1\}^N$1, the quantum algorithm applies the channel

$o \in \{0,1\}^N$2

to an $o \in \{0,1\}^N$3-qubit register. After $o \in \{0,1\}^N$4 steps,

$o \in \{0,1\}^N$5

the average diamond-norm distance to $o \in \{0,1\}^N$6 is upper bounded by

$o \in \{0,1\}^N$7

provided

$o \in \{0,1\}^N$8

where $o \in \{0,1\}^N$9. The following table summarizes major theoretical performance bounds from (Zhao et al., 8 Apr 2026):

Setting	Quantum (QOS)	Classical
Space	$g: \{0,1\}^b \times \{0,1\}^b \to \{0,1\}$0	$g: \{0,1\}^b \times \{0,1\}^b \to \{0,1\}$1
Required Samples (M)	$g: \{0,1\}^b \times \{0,1\}^b \to \{0,1\}$2	$g: \{0,1\}^b \times \{0,1\}^b \to \{0,1\}$3 if $g: \{0,1\}^b \times \{0,1\}^b \to \{0,1\}$4
Time per Sample	$g: \{0,1\}^b \times \{0,1\}^b \to \{0,1\}$5	Problem dependent

3. Theoretical Guarantees and Lower Bounds

Three principal results characterize the efficiency of QOS:

Theorem 1 (Oracle Sketching, IID): To approximate $g: \{0,1\}^b \times \{0,1\}^b \to \{0,1\}$6 within diamond-norm error $g: \{0,1\}^b \times \{0,1\}^b \to \{0,1\}$7, $g: \{0,1\}^b \times \{0,1\}^b \to \{0,1\}$8 samples suffice.

Theorem 2 (Sequential Queries): Any quantum query algorithm making $g: \{0,1\}^b \times \{0,1\}^b \to \{0,1\}$9 queries to $\mathrm{disc}(g) \le 2^{-\eta b}$0 and its adjoint can be simulated via QOS using $\mathrm{disc}(g) \le 2^{-\eta b}$1 classical samples and the same quantum space complexity.

Theorem 3 (Sample-Space Lower Bound): If the classical randomized query complexity of evaluating $\mathrm{disc}(g) \le 2^{-\eta b}$2 is $\mathrm{disc}(g) \le 2^{-\eta b}$3, any classical learner with space $\mathrm{disc}(g) \le 2^{-\eta b}$4 and samples $\mathrm{disc}(g) \le 2^{-\eta b}$5 must satisfy $\mathrm{disc}(g) \le 2^{-\eta b}$6. Especially when $\mathrm{disc}(g) \le 2^{-\eta b}$7, classical learners using $\mathrm{disc}(g) \le 2^{-\eta b}$8 samples must store $\mathrm{disc}(g) \le 2^{-\eta b}$9 bits; those with $G = (g, \ldots, g)$0 require $G = (g, \ldots, g)$1 samples.

Corollary (Exponential Classical Hardness): For tasks with $G = (g, \ldots, g)$2, QOS achieves exponential memory compression and sample-efficiency unattainable by any sub-exponentially sized classical machine.

4. Quantum versus Classical Complexity

In QOS, both the space and time complexities are polylogarithmic in database size:

• Quantum: Space cost is $G = (g, \ldots, g)$3 qubits; sample complexity $G = (g, \ldots, g)$4; each sample processed in $G = (g, \ldots, g)$5 time.
• Classical: For full data access, QRAM or explicit storage requires $G = (g, \ldots, g)$6 space; streaming algorithms for $G = (g, \ldots, g)$7 data require at least $G = (g, \ldots, g)$8 space; the sample-space product must satisfy $G = (g, \ldots, g)$9.

These contrasts persist even under assumptions favorable to classical algorithms (unlimited time or $N$0), establishing unconditional quantum advantage at the intersection of memory, sample, and computational efficiency.

5. Interferometric Classical Shadows and Readout

After constructing the quantum sketch state $N$1, QOS relies on a measurement technique dubbed “interferometric classical shadows” to extract classical information from compact quantum representations.

The process involves preparing the state,

$N$2

and measuring the observable $N$3, which gives unbiased estimates of $N$4. Iterating this with randomized Clifford shadows allows prediction from $N$5 for all sparse $N$6, enabling classical model reconstruction of size $N$7.

6. Empirical Applications

QOS demonstrates practical quantum advantage in large-scale real-world tasks:

• Sentiment Analysis (IMDb): Given an $N$8 TF–IDF matrix (with $N$9, $$z_t = (x_t, Y^{(0)}_{x_t} \text{ or } Y^{(1)}_{x_t}, \alpha_t)$$0), the LS-SVM classifier $$z_t = (x_t, Y^{(0)}_{x_t} \text{ or } Y^{(1)}_{x_t}, \alpha_t)$$1, $$z_t = (x_t, Y^{(0)}_{x_t} \text{ or } Y^{(1)}_{x_t}, \alpha_t)$$2, is learned by block-encoding the matrix on $$z_t = (x_t, Y^{(0)}_{x_t} \text{ or } Y^{(1)}_{x_t}, \alpha_t)$$3 qubits using $$z_t = (x_t, Y^{(0)}_{x_t} \text{ or } Y^{(1)}_{x_t}, \alpha_t)$$4 samples and quantum singular value transformation (QSVT). Interferometric shadow readout enables efficient classical prediction. QOS achieves a $$z_t = (x_t, Y^{(0)}_{x_t} \text{ or } Y^{(1)}_{x_t}, \alpha_t)$$5 fold reduction in memory, requiring only $$z_t = (x_t, Y^{(0)}_{x_t} \text{ or } Y^{(1)}_{x_t}, \alpha_t)$$6 qubits and $$z_t = (x_t, Y^{(0)}_{x_t} \text{ or } Y^{(1)}_{x_t}, \alpha_t)$$7 samples.
• Single-cell RNA-seq (PBMC): For $$z_t = (x_t, Y^{(0)}_{x_t} \text{ or } Y^{(1)}_{x_t}, \alpha_t)$$8 cells across $$z_t = (x_t, Y^{(0)}_{x_t} \text{ or } Y^{(1)}_{x_t}, \alpha_t)$$9 genes, QOS block-encodes the data with $x_t$0 samples. Quantum PCA is performed with QSVT in $x_t$1 qubits; classical readout of the principal component via shadows is achieved with $x_t$2 qubits and $x_t$3 samples, far below classical streaming requirements.

These examples substantiate that QOS enables learning and inference on massive datasets using resources inaccessible to classical machines of comparable size.

7. Key Circuits and Operational Summary

QOS operates through a sequence of resource-efficient quantum circuits for incremental phase sketching. The unitary applied at each step is

$x_t$4

with the target oracle transformation

$x_t$5

The following summarizes the core protocol:

$x_t$6 After construction, interferometric measurements yield succinct classical predictors.

A schematic quantum circuit for a single query involves entangling operations, incremental phase exponentials indexed by the sampled data, and final interferometric measurement.

---

Quantum oracle sketching thus establishes machine learning and classical data processing as a naturally quantum-advantageous domain, delineating tasks where quantum devices achieve exponential compression and efficient prediction from massive data streams, provably unattainable by classically feasible algorithms under constrained space (Zhao et al., 8 Apr 2026).