TL;DR
Can you perform meaningful statistical inference with only a single data point? This paper proves that you can, providing mathematical frameworks to construct valid confidence intervals for both the mean ($\mu$) and standard deviation ($\sigma$) from just one observation. The authors extend these "n=1" results from normal distributions to more complex cases, including heavy-tailed Cauchy distributions and high-dimensional multivariate settings where the number of observations is less than the number of dimensions.
What Problem Does This Paper Address?
In introductory statistics, we are taught that you need at least two observations to estimate variance—after all, you can't calculate a "deviation" from a mean if the data point is the mean. This creates a challenge in fields dealing with rare events, extremely expensive experiments, or "high-dimensional, low-sample size" data (like genomics), where we might have very few observations relative to the complexity of the system.
This paper addresses the theoretical gap of constructing valid confidence intervals—ranges that are guaranteed to contain the true parameter with a specific probability (e.g., 95%)—when only one observation ($X$) is available. While previous work focused on the mean of a normal distribution, this paper expands the toolkit to include the scale (variance), non-normal distributions, and multivariate data.
Methods
The authors employ several sophisticated statistical techniques to bypass the "n=1" limitation:
- Scale Invariance: For normal distributions, they show that the coverage probability of certain intervals depends only on the ratio $\lambda = \mu/\sigma$. By analyzing the "worst-case" $\lambda$, they can find a constant $c$ that guarantees the interval works regardless of the true parameters.
- Monotone Likelihood Ratio (MLR): To bound the variance ($\sigma^2$), the authors use the MLR property of non-central chi-square distributions. This allows them to prove that the probability of the true variance being smaller than a scaled version of the squared observation ($X^2/c^2$) is at least $1-\alpha$.
- Paley-Zygmund Inequality: For distributions with "compact support" (meaning the data is guaranteed to fall within a specific range $[a, b]$), they use this inequality to provide a lower bound on the probability that the observation is "large enough" to provide information about the variance.
- Bonferroni Correction: They suggest using this standard method to combine individual intervals for $\mu$ and $\sigma$ into a joint "confidence box" for both parameters simultaneously.
Datasets
Not discussed in the paper. As a theoretical statistics paper, the work relies on mathematical proofs and derivations rather than empirical testing on specific datasets.
Key Insights and Ideas
- The Power of Symmetry: A core "aha moment" is that if a distribution is symmetric and unimodal (has one peak), the single observation $X$ carries enough information about the "spread" of the distribution to bound $\sigma$.
- Cauchy is Not a Dealbreaker: Even for the Cauchy distribution—a "pathological" case in statistics where the mean and variance don't technically exist—the authors show you can still construct a valid confidence interval for the location parameter as long as your desired confidence level is greater than 50%.
- Multivariate Utility: While $n=1$ for a single number might seem like a curiosity, the multivariate extension is practically useful. In modern data science, we often have a sample size $m$ that is much smaller than the dimension $p$ (e.g., 10 patients but 20,000 genes). The paper provides a way to bound the mean vector $||\mu||$ even when the covariance matrix cannot be estimated.
Key Results
- Variance Bounds: For a normal distribution, the interval $\sigma^2 \leq X^2 / [\Phi^{-1}(\frac{1+\alpha}{2})]^2$ is a valid $1-\alpha$ confidence interval.
- Optimal Intervals: The paper references and builds upon "randomized" intervals that are strictly better (shorter on average) than classical approaches.
- Multivariate Constant: For a sample of size $m$ from a $p$-dimensional normal distribution, the authors provide a specific formula for a constant $c$ (based on $\alpha, p,$ and $m$) that ensures the true mean vector lies within a certain distance of the observed average.
Limitations and Future Work
- Interval Width: The authors candidly describe the univariate results as an "intriguing curiosity." Because you have so little data, the resulting confidence intervals are necessarily very wide to maintain their validity.
- Knowledge of Bounds: The results for distributions with compact support require the researcher to know the absolute bounds ($a$ and $b$) of the distribution beforehand.
- Simultaneous Optimality: While they show how to combine intervals for $\mu$ and $\sigma$, finding the mathematically optimal (minimax) simultaneous set is described as "extremely difficult" and remains an open problem.
- Practicality: The multivariate results are more likely to see real-world application in "large $p$, small $n$" problems than the univariate $n=1$ cases.
Summary updated