Sample Amplification Lower Bounds

• The paper establishes that amplification lower bounds are determined by statistical risk, information-theoretic measures, and the geometry of the sample space.
• It employs techniques like Le Cam’s lemma, Hellinger tensorization, and quantum query analysis to delineate the limits of simulating extra samples.
• The results indicate that in high-dimensional and noisy computational settings, amplification is fundamentally restricted, impacting streaming, communication, and quantum error mitigation protocols.

Sample amplification lower bounds characterize the fundamental limitations of algorithms and protocols purporting to "amplify" the information obtained from a finite number of samples. These limitations are critical in settings across theoretical computer science, statistics, information theory, and quantum computing. The sample amplification paradigm asks: given $n$ samples from some unknown distribution or source, or $n$ uses of a noisy computational component, by what mechanisms and to what extent can one generate outputs statistically indistinguishable from genuine samples, or otherwise simulate additional access, and what are the ultimate lower bounds constraining such amplification? Lower bound results reveal how these limitations manifest, often tied to core principles of statistical risk, information-theoretic distinguishability, adversarial communication, and the geometry of the underlying sample space.

1. Formal Definitions and General Lower Bound Paradigms

The classical sample amplification problem considers a domain $\mathcal{X}$, a family of distributions $\mathcal{P}$, and an $(n, n+m, \epsilon)$ sample amplifier $T: \mathcal{X}^n \to \mathcal{X}^{n+m}$ such that, for all $P \in \mathcal{P}$,

$$\left\| P^{\otimes n} \circ T^{-1} - P^{\otimes(n+m)} \right\|_{\text{TV}} \leq \epsilon,$$

i.e., the joint law of amplified outputs is close (in total variation) to $n+m$ i.i.d. samples from $P$ (Axelrod et al., 2022). The minimax error $n$0 and amplification complexity $n$1 formalize the sharpness of this indistinguishability.

Lower bounds typically leverage risk-based inequalities (such as Le Cam’s lemma), information-theoretic measures (total variation, Hellinger, $n$2-divergence), and adaptivity considerations. In quantum contexts, analogous definitions use the number of independent prepared states (samples) or oracle queries, distinguishing between copy complexity and query complexity (Wang et al., 2023).

2. Statistical and Information-Theoretic Lower Bounds

For regular parametric families, tight lower bounds are available. For $n$3, the minimax indistinguishability is exactly the total variation between the $n$4-sample and $n$5-sample mean distributions: $n$6 For $n$7, this stays bounded away from 0, so no amplification of more than $n$8 extra samples is possible (Axelrod et al., 2022). For general $n$9-dimensional continuous exponential families,

$\mathcal{X}$0

asymptotically ruling out amplification beyond $\mathcal{X}$1 samples.

For product models, where $\mathcal{X}$2, uniform lower bounds show that, under mild regularity, amplification by more than $\mathcal{X}$3 is uniformly impossible. The extremal case is governed by the local change in Bayes risk: $\mathcal{X}$4 where $\mathcal{X}$5 is the Bayes risk with $\mathcal{X}$6 samples for an appropriate loss function (Axelrod et al., 2022). Le Cam’s two-point method, bowl-shaped loss analysis, Hellinger tensorization, and combinatorial interpolation are key techniques.

This statistical framework demonstrates that, in high-dimensional problems, sample amplification is fundamentally limited by the geometry and risk structure of the family, unlike general (total variation) closeness, which admits powerful amplification in trivial or degenerate scenarios.

3. Sample Amplification in Data Streams and Communication

In the streaming and communication complexity regime, amplification lower bounds are articulated via reductions from canonical hard problems such as the Universal Relation (UR). For the one-way strict-turnstile $\mathcal{X}$7-sampling problem (where $\mathcal{X}$8) with failure probability $\mathcal{X}$9, any solution must use

$\mathcal{P}$0

bits of space (Nelson et al., 2017, Kapralov et al., 2017). The core encoding/decoding argument translates an optimal public-coin UR protocol into a streaming sampler, using adaptivity lemmas to handle correlated queries and injection of random noise to control mutual information leakage. Amplifying success probability from a constant to $\mathcal{P}$1 requires an additional $\mathcal{P}$2 bits, as each "boosting" run costs an $\mathcal{P}$3 contribution.

Both encoding-based arguments and reductions from Augmented Indexing establish that these bounds are sharp for all $\mathcal{P}$4 above $\mathcal{P}$5. No significant improvement is possible with current techniques.

4. Lower Bounds in Quantum and Noisy Computation

Information-theoretic lifting results in quantum property testing and lower bounds for noisy classical computation provide structural results on sample amplification in these non-classical contexts. The quantum sample-to-query lifting theorem establishes a generic lower bound

$\mathcal{P}$6

where $\mathcal{P}$7 is the minimum required number of state copies (samples) and $\mathcal{P}$8 is the minimum number of block-encoding oracle queries (Wang et al., 2023). Specializations yield:

• Quantum state discrimination: $\mathcal{P}$9 for infidelity $(n, n+m, \epsilon)$0.
• Gibbs state preparation: $(n, n+m, \epsilon)$1.
• Entanglement entropy testing: $(n, n+m, \epsilon)$2 for entropy gap $(n, n+m, \epsilon)$3.
• Matrix spectrum testing: various $(n, n+m, \epsilon)$4, $(n, n+m, \epsilon)$5, and $(n, n+m, \epsilon)$6 bounds.

These results match known upper bounds (up to polylogarithmic factors), confirming optimality. The mechanism leverages density matrix exponentiation, Quantum Singular-Value Transformation (QSVT), and risk-to-query conversions.

In noisy computation, especially in distributed sensor networks (with $(n, n+m, \epsilon)$7 nodes and $(n, n+m, \epsilon)$8-bit noise), protocols with fewer than $(n, n+m, \epsilon)$9 transmissions cannot reliably compute functions (parity, majority) whose value depends symmetrically on every input bit (Dutta et al., 2015). The reduction from randomized decision trees to sampling algorithms encapsulates the amplification limitation: noisy protocols with insufficient depth can be simulated by sampling a vanishing subset of the inputs, and robust functions cannot be computed reliably from such sketches.

5. Lower Bounds for Quantum Error Mitigation and Adversarial Noise Models

Universal bounds on the sample complexity of quantum error mitigation demonstrate that, under broad conditions (including nonlinear, adaptive, or entangled postprocessing), the number of runs required grows at least exponentially with the circuit depth: $T: \mathcal{X}^n \to \mathcal{X}^{n+m}$0 for $T: \mathcal{X}^n \to \mathcal{X}^{n+m}$1 circuit layers with per-layer depolarizing noise $T: \mathcal{X}^n \to \mathcal{X}^{n+m}$2 (Takagi et al., 2022). The generic information-theoretic argument reduces error mitigation to state discrimination. Entropic contraction (e.g., of quantum relative entropy) is the limiting factor: every layer shrinks the informative part of the output, requiring exponentially more samples to maintain distinguishability.

This renders constant-accuracy mitigation fundamentally infeasible for deep circuits without full error correction. Only in shallow-depth, low-noise, or highly structured settings (e.g., probabilistic error cancellation saturating the bound under local dephasing) are the limits achievable.

6. Privacy Amplification and the Tightness of Sample Amplification Bounds

Privacy amplification by shuffling in differential privacy is another regime where lower bounds on the amplification effect are tight. Using the general clone paradigm, lower bounds on privacy amplification translate to exact formulas for the divergence between outputs with and without amplification: $T: \mathcal{X}^n \to \mathcal{X}^{n+m}$3 where $T: \mathcal{X}^n \to \mathcal{X}^{n+m}$4 is a lower-bound random variable derived from the local randomizer (Su et al., 10 Apr 2025). The privacy blanket decomposition is shown to be optimal for all decomposition-based analyses, and computational techniques (FFT convolution) allow empirical bounds to saturate the lower bound to within numerical precision.

No general upper bound on privacy amplification can improve the divergence below this threshold; thus, the lower bounds are both theoretically and practically unimprovable within the established framework.

7. Implications and Open Problems

The landscape of sample amplification lower bounds presents broad unification across statistical learning, streaming complexity, quantum algorithms, and privacy. Amplification is generally limited by the available information and the inherent risk geometry, with exponential or quadratic reductions often provably tight. Specific regimes—high dimensions, deep noise, or adversarial adaptivity—pose hard barriers.

Open problems include sharpening the lower bounds in models outside the product and exponential-family settings, extending communication lower bounds to multi-pass and distributed streaming models, and fully characterizing amplification tightness in non-unital or non-classical noise channels. Identifying models or tasks where amplification can beat the universal lower bounds remains of foundational interest.