Conditional Probability Curvature

• Conditional probability curvature is a geometric measure that captures the local structure of probability distributions in response to conditional events, revealing intrinsic statistical dependencies.
• It integrates Riemannian geometry with statistical inference to evaluate Gaussian approximations and characterize high-dimensional correlation structures.
• This concept is applied in machine learning to distinguish between human-authored and machine-generated texts using robust, curvature-based statistical tests.

Conditional probability curvature quantifies the local geometric structure of a probability distribution or model with respect to conditional events or sequential decisions. In modern applications, this notion arises both in the Riemannian geometry of statistical manifolds—where curvature reveals fundamental properties of irreducible correlation—and in applied machine learning, particularly for the task of distinguishing machine-generated from human-authored text. Conditional probability curvature formalizes how a candidate event or sequence sits at a local extremum of the model's probability surface, with curvature values signaling both probabilistic dependence and suitability for Gaussian approximation. The concept integrates fundamental differential geometric invariants with practical, robust statistical tests in high-dimensional language modeling.

1. Statistical Manifolds and the Geometric Framework

A statistical manifold $\mathcal{M}$ is defined by a parametrized family of conditional probability distributions

$$dp(x\mid \theta) = \rho(x\mid \theta)\,dx, \quad x\in\mathcal{R}_x\subset\mathbb{R}^n,\; \theta\in\mathcal{R}_\theta\subset\mathbb{R}^m.$$

Equipped with fluctuation geometry, $\mathcal{M}$ becomes a Riemannian manifold whose structure captures the intrinsic constraints and dependence encoded in the distributions. The Riemannian metric is given by

$ds^2 = g_{ij}(x\mid\theta)\,dx^i\,dx^j,$

where the metric tensor $g_{ij}(x\mid\theta)$ satisfies covariant equations involving the log-density and its derivatives, or equivalently, the information potential $\mathcal{S}(x\mid\theta) = \log\omega(x\mid\theta)$, with the invariant weight $\omega$ including a term $\sqrt{|2\pi g^{ij}|}$ (Velazquez, 2013).

2. Curvature Tensors and Their Statistical Significance

The Levi–Civita (metric) connection $\Gamma^k_{ij}$ and the associated Riemann curvature tensor $R^l{}_{ijk}$ are defined from the metric in the standard manner: $$\Gamma^k_{ij} = \frac12\,g^{k\ell}(\partial_i g_{\ell j} + \partial_j g_{\ell i} - \partial_\ell g_{ij}),$$

$$R^l{}_{ijk} = \partial_i \Gamma^l_{jk} - \partial_j \Gamma^l_{ik} + \Gamma^l_{im}\Gamma^m_{jk} - \Gamma^l_{jm}\Gamma^m_{ik}.$$

The curvature tensor $R_{ijkl} = g_{lm} R^m{}_{ijk}$ and scalar curvature $R(x\mid\theta)$ encode the deviation from local flatness. These geometric objects provide a means to detect structural features of the underlying distributions: the sign and magnitude of curvature correspond to the presence of irreducible statistical correlations and to the suitability of Gaussian approximations for fluctuations (Velazquez, 2013).

3. Curvature as Indicator of Irreducible Statistical Correlation

A fundamental result states that the statistical manifold $\mathcal{M}$ is flat ($R^l{}_{ijk}\equiv 0$) if and only if the joint distribution $dp(x\mid\theta)$ can be transformed—under some coordinate change $x\mapsto \check{x}$—into a product of independent marginals,

$$dp(\check{x}\mid\theta) = \prod_{i=1}^n dp^{(i)}(\check{x}^i \mid \theta).$$

Nonzero curvature ($R_{ijkl}\neq 0$) is therefore a direct geometric certificate of irreducible correlations: no coordinate system exists in which the components become independent. Conversely, flatness implies reducibility of statistical dependence (Velazquez, 2013). This geometric criterion functions as a general test for the presence of non-factorizable dependencies in high-dimensional distributions.

4. Conditional Probability Curvature in Language Modeling

The concept of conditional probability curvature has been operationalized in machine learning for detecting machine-generated text (Bao et al., 2023). For a candidate sequence $x = (x_1,\ldots,x_n)$, and a reference LLM $p_\theta$, define the local curvature as

$$\hat{d}(x) = \frac{\log p_\theta(x) - \tilde{\mu}}{\tilde{\sigma}},$$

where $\tilde{\mu}$ and $\tilde{\sigma}^2$ are the sample mean and variance, respectively, of log-probabilities of alternative passages $\tilde{x}^{(i)}$ sampled from a surrogate model $q_\phi(\cdot|x)$. Here,

$$p_\theta(\tilde{x}|x) = \prod_{j=1}^n p_\theta(\tilde{x}_j | x_{<j}).$$

A high positive curvature $\hat{d}(x)$ indicates that the original passage $x$ lies near a sharp local maximum of the probability surface, a signature characteristic of sequences produced by sampling from the target model. In human-authored text, values of $\hat{d}(x)$ tend to be closer to zero, as variations sampled around the human passage often have similar or higher likelihood (Bao et al., 2023).

5. Algorithmic Realization: Fast-DetectGPT

The Fast-DetectGPT algorithm leverages conditional probability curvature for efficient zero-shot detection. Its workflow is summarized as follows:

1. Sampling: Generate $N$ perturbed variants $\tilde{x}^{(i)} \sim q_\phi(\cdot|x)$.
2. Scoring: For each variant, compute $\log p_\theta(\tilde{x}^{(i)}|x)$, the conditional likelihood under the scoring model.
3. Aggregation: Compute the empirical mean $\tilde{\mu}$ and variance $\tilde{\sigma}^2$.
4. Curvature Calculation: Evaluate $\hat{d}(x)$ as above. A threshold $\epsilon$ is selected to classify passages.

If $q_\phi = p_\theta$, the expressions for $\tilde{\mu}$ and $\tilde{\sigma}^2$ can be computed analytically, further reducing computational overhead, and eliminating sampling noise (Bao et al., 2023). Fast-DetectGPT achieves speedups of up to $340\times$ compared to DetectGPT, with relative AUROC improvements of approximately $75\%$ when tested across diverse LLMs and datasets.

Method	5-Model AUROC	Speedup
DetectGPT	0.9554	$1\times$
Fast-DetectGPT	0.9887	$340\times$

6. Theoretical and Practical Connections

Conditional probability curvature encompasses both foundational statistical and practical algorithmic implications:

• Gaussian Approximation: The curvature scalar $R(x\mid\theta)$ of the statistical manifold sets a criterion for the validity of Gaussian approximations. Gaussian behavior prevails when $nR(\bar{x}) \ll 1$, where $n$ is manifold dimension and $\bar{x}$ the most likely point. Deviations from Gaussianity are controlled by curvature-induced corrections, entering at second order in geodesic distance from the mean (Velazquez, 2013).
• Distributional Watermarking: As a statistical signal dependent only on distributional geometry, conditional probability curvature functions as a distributional watermark, orthogonal to explicit watermarking approaches. It can be combined with cryptographic watermarks to augment detection robustness (Bao et al., 2023).
• Invariant Fluctuation Theorems: Curvature-based results underlie the derivation of invariant fluctuation theorems, with expectation values involving “generalized restituting forces” $\eta_i = \partial_i \mathcal{S}$ obeying exact identities when averaged over $dp(x\mid\theta)$, such as $\langle \eta^2(x) \rangle = n k$ (Velazquez, 2013).

7. Representative Examples and Empirical Validation

Curvature-based analysis admits analytic computation and geometric interpretation in concrete distributional families:

• Gaussian Distributions: Both one-dimensional and multivariate normal laws produce flat ($R\equiv 0$) statistical manifolds, reflecting complete reducibility and lack of population-level correlation.
• Nontrivial Correlated Models: Examples such as the 2D density $$dp = (\text{const})\, r/\sqrt{r^2+\theta^2}\, e^{-r^2/2}\, dr\, d\varphi$$ yield nonzero scalar curvature $$R(r,\varphi) = \frac{6\,\theta^2}{(\theta^2 + r^2)^2}$$, which vanishes in the limit $\theta \to \infty$, recovering Gaussianity.
• Empirical Detection: Fast-DetectGPT achieves near-perfect detection of machine-generated content from open-source and API LLMs (e.g., AUROC $0.9887$ on five open models, $0.9338$ on ChatGPT/GPT-4), with monotonic accuracy gains as passage length increases. False positive rates under operating thresholds are demonstrably low; e.g., $87\%$ recall at $1\%$ false alarm on ChatGPT (Bao et al., 2023).

Conditional probability curvature thus occupies a central role at the intersection of statistical geometry and practical AI detection, serving as both a theoretical marker of irreducible dependence and a robust, efficiently computable feature for model-based analysis in modern machine learning.