Unitary Single-Sampling (USS)

• Unitary Single-Sampling (USS) is a non-iterative paradigm that exploits inherent unitary or probabilistic structures to achieve immediate, exact selection without the need for iterative resampling.
• It underpins diverse applications including streaming data sampling, harmonic analysis, determinantal point process sampling, quantum circuit design, and active learning in medical imaging.
• USS methods deliver stable, unbiased sampling with minimal computational overhead, eliminating iterative reweighting and Markov chain equilibration while ensuring theoretical guarantees.

Unitary Single-Sampling (USS) refers to a conceptual and algorithmic paradigm where an entire selection, generation, or reconstruction procedure can be executed in a single, non-iterative pass or step, exploiting underlying unitary or probabilistic structure. USS arises in streaming sampling, group-based harmonic analysis, determinantal point processes, tensor network simulation, quantum circuit design, and active learning. The term encompasses concrete algorithms and frameworks where each candidate or data element is sampled or labeled immediately, with no revisiting, reweighting, or Markovian equilibration.

1. Immediate-Decision Streaming Sampling

In the context of classical stream sampling, USS denotes a one-pass, immediate-decision algorithm for selecting random samples from a linear or grouped data stream, supporting both equal and unequal probability inclusion (Panahbehagh et al., 2021). The protocol is as follows:

• Each incoming unit $j$ has a specified inclusion probability $\pi_j$.
• The algorithm maintains a running sum $F$ of inclusion probabilities and a count $n$ of selections so far.
• For each $j$, it draws $u \sim \mathrm{Unif}(0,1)$, updates $F$, and computes auxiliary scalars:
  • $$a = \lfloor F \rfloor - \lfloor F_{\mathrm{old}}\rfloor - 1$$
  • $b = \lfloor F \rfloor - n$
  • $$m = F_{\mathrm{old}} - \lfloor F_{\mathrm{old}} \rfloor$$
• The inclusion threshold $\pi_j$0 is computed explicitly:

$\pi_j$1

• $\pi_j$2 is included in the sample $\pi_j$3 if $\pi_j$4; $\pi_j$5 is incremented accordingly.

Correctness is established inductively—marginal inclusion probabilities $\pi_j$6 and, in fixed-size sampling $\pi_j$7, $\pi_j$8 almost surely. No data beyond $\pi_j$9’s arrival is needed for decision, and the time complexity is $F$0 per unit with $F$1 state. The same scheme adapts seamlessly to weighted sampling ($F$2), and batched arrivals are handled by iterating USS per unit.

2. Unitary Structure in Harmonic Analysis and Sampling Theory

In functional and harmonic analysis, USS describes sampling schemes derived from unitary group actions on Hilbert spaces, particularly for continuous frames (García, 2020). Let $F$3 be a second-countable locally compact abelian (LCA) group and $F$4 a separable Hilbert space. A family $F$5 forms a continuous unitary frame if $F$6 for a unitary representation $F$7.

• Sampling occurs on a discrete subgroup (lattice) $F$8.
• The analysis operator yields samples $F$9, where $n$0 is restricted to the closed span of $n$1.
• These samples are represented as a discrete convolution $n$2, with $n$3.
• Exact reconstruction is possible iff the convolution operator by $n$4 is invertible; equivalently, its group Fourier transform $n$5 is bounded away from zero.

The Unitary Single-Sampling Theorem establishes that USS provides stable and exact reconstruction over $n$6 under these invertibility conditions. This framework subsumes Shannon sampling, Gabor sampling, and diverse group-invariant signal processing paradigms.

3. Determinantal Point Process Sampling for Unitary Ensembles

USS in random matrix theory is associated with determinantal point process (DPP) sampling algorithms for unitary invariant ensembles (UIE) (Olver et al., 2014). The eigenvalue distribution of an $n$7 UIE is given by a DPP with kernel $n$8 constructed from orthogonal polynomials with respect to potential $n$9.

The USS procedure here comprises:

• Construction of feature vectors $j$0 from the orthogonal polynomial basis.
• Recursive single-sample deletion via deflation of the feature space; each eigenvalue $j$1 is sampled from the conditional marginal density $j$2.
• Basis vectors are orthogonalized and deflated after each sample, guaranteeing that each $j$3 is sampled exactly once and all samples are mutually independent given the kernel.

The algorithm delivers machine-precision accuracy in single-sample eigenvalue draws, facilitating Monte Carlo analysis of UIE statistics not amenable to direct analytic evaluation.

4. Unitary Single-Sampling for Unitary Tensor Networks

USS enables perfect sampling (i.e., direct generation of samples according to the true probability distribution in the wavefunction) in unitary tensor networks—such as uMPS, uTTN, and MERA (Ferris et al., 2012). The key is the unitarity of the network tensors:

• The amplitude $j$4 for a configuration $j$5 on the "causal cone" $j$6 can be factorized as a product of conditional single-site probabilities: $j$7.
• Each conditional is obtained from the diagonal of the one-site reduced density matrix $j$8.
• Sampling $j$9 component-by-component yields statistically exact, autocorrelation-free draws at a computational cost $u \sim \mathrm{Unif}(0,1)$0 per sample.
• Partial sampling—a basis-dependent contraction over a subset of indices—can further reduce sampling variance without bias.

By exploiting the cancelation properties of unitarity, USS circumvents the equilibration and autocorrelation penalties inherent to Markov chain Monte Carlo in many-body simulation.

5. USS in Quantum Diffusion Models

In quantum generative modeling, USS refers to the construction of a single-step—or "consistency"—unitary variational quantum circuit $u \sim \mathrm{Unif}(0,1)$1 parameterizing the full denoising transformation between noisy and target quantum states (Kölle et al., 2024). Given a chain of reversible, step-wise unitaries mapping $u \sim \mathrm{Unif}(0,1)$2, the USS architecture collapses these into a single, parameterized quantum circuit:

$u \sim \mathrm{Unif}(0,1)$3

The circuit consists of $u \sim \mathrm{Unif}(0,1)$4 layers of trainable single-qubit rotations and CNOT entanglement, acting directly on the data qubits (optionally with ancilla for guided models). Training is performed using mean-absolute-error loss over the full state vector, leveraging the state vector simulator. This dramatically accelerates inference compared to sequential quantum diffusion, and is conceptually analogous to recent classical consistency models.

6. USS for Active Learning in Medical Image Segmentation

In active learning, Uncertainty Slice Sampling (USS) designates a pool-based algorithm for optimizing labeling effort in 3D medical image segmentation (Chlebus et al., 2021). The process is as follows:

• Predictive entropy is estimated at the voxel level via Monte Carlo dropout.
• The mean entropy for each 2D slice, $u \sim \mathrm{Unif}(0,1)$5, is computed by averaging over all voxels in the slice.
• Candidate slices are identified by searching for local maxima in the slice-uncertainty sequence, enforcing a minimal slice distance between peaks.
• The $u \sim \mathrm{Unif}(0,1)$6 most uncertain slices across all volumes are selected for annotation.

This approach allows for fine-grained, high-efficacy annotation targeting and yields robust segmentation models using only a fraction of the available annotation budget, with better or comparable robustness compared to full-volume annotation. In experiments, only 4% of all possible slices annotated with USS sufficed to match or exceed the performance of models trained on the entire dataset.

7. Contextual and Unifying Perspective

Across its applications, Unitary Single-Sampling encapsulates the exploitation of algebraic or probabilistic structure—unitarity, determinantal kernels, or precise conditional factorization—to enable exact, unbiased sampling or reconstruction in a single pass. Core advantages include:

• Elimination of iterative resampling, reweighting, or Markov chain equilibration.
• Minimal memory and computational overhead—often $u \sim \mathrm{Unif}(0,1)$7 or $u \sim \mathrm{Unif}(0,1)$8 state, with per-sample cost matching or improving upon traditional alternatives.
• Immediate and interpretable error/variance properties, and often provable correctness guarantees.
• Applicability in settings as diverse as streaming data, group-invariant signal processing, quantum Monte Carlo, and uncertainty-driven data selection.

The breadth of USS demonstrates the unifying power of single-pass, structure-exploiting algorithms in probabilistic computation, harmonic analysis, and quantum simulation (Panahbehagh et al., 2021, García, 2020, Olver et al., 2014, Ferris et al., 2012, Chlebus et al., 2021, Kölle et al., 2024).