Theoretical Limits of Synthetic Data

• Synthetic data is algorithmically generated data that simulates real-world samples, crucial for enhancing data availability and preserving privacy in sensitive domains.
• Key theoretical limits include trade-offs in privacy amplification, statistical efficiency, and generalization, which necessitate techniques like secret random seed usage and bias correction.
• Practical guidelines emphasize optimal mixing of real and synthetic data, covariance matching in high dimensions, and proper noise tuning to minimize distribution gaps and ensure reliable model ranking.

Synthetic data refers to data instances generated algorithmically rather than collected from actual measurement or user behavior. Its use spans data augmentation, privacy-preserving data release, and bolstering scarce domain-specific datasets for tasks such as LLM post-training. While synthetic data offers flexibility and, under certain generative protocols, privacy guarantees, its theoretical limits are governed by trade-offs among statistical efficiency, privacy, generalization, and fidelity. This article surveys the latest theoretical developments quantifying these limits from the perspectives of privacy amplification, statistical optimality, generalization, information-theoretic bounds, and selection/verification in high-dimensional regimes.

1. Privacy Amplification and Differential Privacy Boundaries

The privacy guarantees of synthetic data are inherited from the generative model's underlying mechanisms, typically formulated through Differential Privacy (DP). For linear regression employing output perturbation, the standard estimator satisfies $(\varepsilon,0)$-DP, with $\varepsilon = \Delta/\sigma_\theta$ where $\Delta$ represents sensitivity and $\sigma_\theta$ is the noise scale.

A sharp dichotomy arises depending on seed management:

• No Amplification under Adversarial Seed Control: If the data recipient can choose the generation seed arbitrarily, releasing even a single synthetic sample exposes as much information as releasing the model parameters—a formal negative result. Specifically, for every single direction $z\in\mathbb{R}^d$, there exist adjacent datasets $\mathcal{D},\mathcal{D}'$ such that the privacy guarantee for the released quantity $\mathcal{M}(\mathcal{D})z$ is exactly $(\varepsilon,0)$-DP, matching that of the private model itself. This holds because an adversary can select $z$ to align perfectly with the model sensitivity direction, nullifying any amplification (Pierquin et al., 5 Jun 2025).
• Amplification through Hidden Random Seeds: If the generative seeds are drawn independently at random and kept secret, synthetic data releases lead to genuine privacy amplification. Constructively, releasing $m$ such points yields a new $(\varepsilon',\delta)$-DP guarantee with the scaling

$$\varepsilon' \approx \frac{m}{2d}\varepsilon + \sqrt{\frac{2m}{d}\log(1/\delta)} + o(1/d)$$

for large $d$ and $m\ll d$. This result is established via total-variation arguments and Rényi divergence calculations for Gaussian mixtures. When the number of synthetic samples is small relative to the data dimension and seeds remain secret, privacy leakage is thus suppressed by a factor proportional to $m/d$ (Pierquin et al., 5 Jun 2025).

• Extensions: These results depend crucially on the output appearing (approximately) as a high-dimensional Gaussian post-privatization and seed secrecy. In more complex generative settings, analogous high-dimensional central limit behaviors are expected to provide similar amplification, provided the model output is Gaussian-like and seeds are not exposed.

A design principle immediately follows: restrict synthetic data release to small batches from hidden, random seeds, and never allow public access to the generative process or adaptive querying.

2. Statistical Efficiency and Distributional Consistency

Naive synthetic data generation—i.i.d. sampling from a fitted $f_{\hat\theta(X)}$—is fundamentally limited in both estimator efficiency and distributional fidelity:

• Efficiency Degradation: For parametric models with an efficient real-data estimator $\hat\theta(X)$, the variance of $\hat\theta$ estimated from synthetic samples doubles in the limit, as

$$\sqrt{n}\bigl(\hat\theta(Z)-\theta_0\bigr) \to N(0,2I^{-1}(\theta_0))$$

compared to $N(0,I^{-1}(\theta_0))$ for the genuine estimator (Awan et al., 2020).

• Distributional Non-Convergence: The joint law of $n$ i.i.d. real versus naive synthetic samples does not converge in total variation, no matter the sample size. There remains a strictly positive gap, evidencing that the two can always be statistically distinguished.
• One-Step Correction for Optimality: The one-step synthetic data algorithm remedies these pathologies. By performing a single Newton-type bias-correction step post-bootstrap, it achieves match in first-order estimator law and vanishing distributional KL-divergence to the real data:

$\hat\theta(Y) = \hat\theta(X) + o_p(n^{-1/2})$

with the same Cramér–Rao-optimal efficiency, even under DP when a DP-efficient estimator is used (Awan et al., 2020). No theoretically stronger result is attainable within this general class—simpler approaches are demonstrably suboptimal.

• Lower Bound: This is the "statistical possibility frontier": it is impossible to beat the Cramér–Rao bound or make the total variation gap of naive bootstrapping vanish without this bias correction.

3. Generalization, Ranking, and Learning-Theoretic Utility

The impact of synthetic data on model utility is governed by generalization error (test performance loss relative to real data) and model ranking reliability. Theoretical characterizations reveal nontrivial trade-offs:

• Generalization Difference Bound: In nonparametric regression,

$$\Delta_{gen}(\widetilde{f},\widehat{f}) \leq |R(\widehat{f})-R(f^*_{\mathcal{F}})| + |\widetilde{R}(\widetilde{f})-\widetilde{R}(\widetilde{f}^*_{\mathcal{F}})| + 4D_{\mathcal{H}}(P_X,P_{\widetilde{X}}) + \dots$$

where $D_{\mathcal{H}}$ is an IPM over quadratic functions quantifying feature distribution mismatch. With model class well-specified and correct synthetic regression function $\widehat\mu\approx\mu$, synthetic feature distributions can be substantially mismatched—provided functional error is small, generalization optimality is preserved (Xu et al., 2023).

• Model-Ranking Consistency: Synthetic data can reliably rank models (i.e., preserve performance gaps) even with moderate feature distribution mismatch, as long as the difference in minimal real-data risks exceeds a bound determined by the synthetic generator's fidelity and the approximation error (Xu et al., 2023). Thus, model comparison by synthetic data is more robust than global error metrics.

This demonstrates that perfect feature matching is not necessary; synthetic generators should instead focus on capturing relevant target function structure and minimizing the approximation gap.

4. High-Dimensional Phenomena and Selection Criteria

Recent high-dimensional analyses clarify which synthetic data properties matter most for generalization in over/under-parameterized regimes:

• Covariance over Mean: When augmenting real data with synthetic in linear regression, excess risk for minimum-norm interpolation depends on the alignment of covariances between real ($\Sigma_t$) and synthetic ($\Sigma_s$) data, not on mean shifts:

$R_X(\hat\beta;\beta) \to \frac{\sigma^2}{n}\,\Tr\big[(\alpha_1 M^\top M + \alpha_2 I_p)^{-1}\big]$

with $M = \Sigma_s^{1/2}\Sigma_t^{-1/2}$, $\alpha_1,\alpha_2$ determined by relative sample sizes. 'Mean shift' contributions vanish asymptotically under mixed-training (Rezaei et al., 9 Oct 2025).

• Covariance Matching is Optimal: Among all selection criteria for synthetic data (within fixed norm constraints), matching the covariance to that of the target data minimizes generalization risk. Greedy forward selection guided by minimizing the Frobenius distance between synthetic and real covariance achieves this optimum. Empirical findings confirm that this principle generalizes across architectures and generative models (Rezaei et al., 9 Oct 2025).
• Phase Transitions in Label Noise: In high-dimensional binary classification, the critical label noise threshold for synthetic data to confer positive performance is $\epsilon^* = (1+\rho/\phi)^{-1}$, where $\phi$ and $\rho$ are pruning accuracies for correct and incorrect labels, respectively; transitions are smooth for finite sample-to-feature ratios, but sharp in infinite-sample limits (Firdoussi et al., 2024).

5. Regularization, Data Mixing, and Domain Adaptation

Blending synthetic and real data introduces an explicit bias-variance trade-off characterized by algorithmic stability:

• Generalization Error Bound: For a convex loss and a mixture of $n$ real and $m$ synthetic samples with mixture weight $\alpha\approx m/(n+m)$, the expected test error scales as

$$E_{\rm test} \lesssim \alpha W_2^2(p_x,p'_x) + (1-\alpha)\left(\frac{\text{const}}{N\alpha}\right)^{1/(d^*+1)}$$

where $W_2$ is the Wasserstein-2 distance between real $(p_x)$ and synthetic $(p'_x)$ feature distributions, and $d^*$ is the packing dimension of $p'_x$ (Shidani et al., 9 Oct 2025).

• Optimal Data Ratio and U-shaped Error: The test error curve as a function of synthetic fraction $\alpha$ is U-shaped, minimized at

$$\alpha^* \asymp \left(\frac{1}{nW_2^{2(d^*+1)}}\right)^{1/(d^*+2)}$$

signifying that excessive synthetic augmentation backfires due to distributional discrepancy. A regime with moderate $W_2$ permits synthetic data to dominate when $n$ is small, but as $W_2$ grows, $\alpha^*$ drops, favoring real data.

• Domain Adaptation: When synthetic data is generated to approximate a target domain, the benefit depends on relative $W_2$ distances: use larger $\alpha$ (more synthetic) if $W_2(p_x^*,p'_x)<W_2(p_x^*,p_x)$, and vice versa. The bound extends to out-of-domain adaptation settings (Shidani et al., 9 Oct 2025).

6. Information-Theoretic and Mutual Information Limits in Generative Modeling

In post-training for LLMs, the effectiveness of synthetic data is governed by mutual information flow from the pre-trained generator and the information bottleneck imposed on the fine-tuned model:

• Generalization Bound: The error of the synthetic-data-fine-tuned model is upper-bounded by the sum of total variation divergences ("task divergence", "generation divergence") and a contraction-attenuated square-root term involving the information gain $\Delta I$ from the generative model:

$$\mathbb{E}[\text{Err}(\pi^{S_{\text{gen}}})] \leq C \left(D_{\text{TV}}(\mathcal{D}, \mathcal{D}_M) + D_{\text{TV}}(\mathcal{D}_M, \mathcal{D}_{\text{gen}})\right) + \exp\left(-\frac{L}{2}\log\frac{1}{\eta}\right) \sqrt{\frac{2\sigma^2(-\Delta I + B_{\text{syn}} + H(e_M) + \delta_{\epsilon,p})}{n}}$$

where $\Delta I$ is the model's synthesized information gain, $B_{\mathrm{syn}}$ is a compression bottleneck term, $H(e_M)$ codes the entropy of the model factor, and $\delta_{\epsilon,p}$ reflects curation/prompt inefficiencies (Gan et al., 2024).

• Design Levers: Maximizing $\Delta I$ (generator exposes new target modes), minimizing $B_{\mathrm{syn}}$ (proper regularization), reducing $H(e_M)$ (model-task alignment), and increasing synthetic data quantity $n$, jointly optimize generalization. If synthetic data diverges too far from the downstream task distribution or the generative model is not tuned to cover the proper modes, these bounds collapse.

7. Synthesis: Foundational Limits and Practical Takeaways

Recent work has solidified several theoretical impossibility and possibility results for synthetic data:

• Impossible: Under adversarial seed choice, for naive bootstrapping, or without bias correction, synthetic data cannot achieve vanishing statistical distance to the true distribution nor optimal statistical efficiency. Privacy cannot be amplified in the worst case.
• Possible: With (i) bias correction (one-step method), (ii) carefully tuned mixing strategies leveraging high-fidelity synthetic generators, (iii) covariance-matching in high dimensions, and (iv) privacy-aware generation with hidden seeds, one can approach the optimal trade-off curve in generalization, privacy, and statistical efficiency—subject to explicit bounds as summarized above.

These results collectively delineate the structural and quantitative frontiers governing synthetic data's theoretical limits and establish precise protocols for practitioners seeking to exploit synthetic augmentation without breaching foundational statistical and privacy constraints (Awan et al., 2020, Pierquin et al., 5 Jun 2025, Xu et al., 2023, Gan et al., 2024, Firdoussi et al., 2024, Rezaei et al., 9 Oct 2025, Shidani et al., 9 Oct 2025).