Robust Univariate Mean Estimator (RUME)

• RUME is a robust estimator framework that minimizes sensitivity to outliers and heavy-tailed data using techniques like median-of-means, trimmed means, and M-estimators.
• It guarantees strong statistical properties, including high breakdown points, sub-Gaussian deviation bounds, and minimax optimality under minimal moment assumptions.
• Recent advances in RUME provide efficient, self-tuned algorithms for reliable mean estimation in applications such as heavy-tailed regression and robust statistical learning.

A Robust Univariate Mean Estimator (RUME) refers to any estimator of the mean for univariate data that delivers strong robustness guarantees—specifically, diminished sensitivity to outliers and heavy tails, high breakdown point, and, when possible, optimal non-asymptotic deviation bounds. The RUME framework encompasses a wide range of techniques, including median-of-means, trimmed means, quantile-based methods, M-estimators, block-wise filter approaches, and more recent semiparametric generalizations. RUMEs have become the central tools for robust location estimation under minimal moment assumptions, offering provable sub-Gaussian performance even under heavy-tailed and contaminated data scenarios.

1. Core Definitions and Methodologies

The canonical RUME is the median-of-means (MoM) estimator. For i.i.d. data $X_1,\ldots,X_n$ with $\mathbb{E}X_i=\mu$, split the data into $k$ disjoint blocks ($B_1,\ldots,B_k$), compute each block mean $m_j = \frac{1}{|B_j|}\sum_{i\in B_j} X_i$, and return $$\widehat\mu_{\rm MoM} = \mathrm{median}(m_1,\ldots,m_k)$$. Choosing $k \asymp \log(1/\delta)$ provides, under finite variance, the sub-Gaussian deviation: $$|\widehat\mu_{\rm MoM} - \mu| \leq C \sigma \sqrt{\frac{\log(1/\delta)}{n}}$$ with probability at least $1-\delta$, where $\sigma^2 = \mathrm{Var}(X_i)$ (Lugosi et al., 2019, Lerasle et al., 2011, Joly et al., 2015).

Advanced RUMEs include:

• Trimmed mean: sorts the data, discards the smallest and largest $\mathbb{E}X_i=\mu$0 observations, and averages the center (Oliveira et al., 2023).
• Self-normalized block estimators: use block means, compute empirical variances per block, and aggregate with weights inversely proportional to variance ($\mathbb{E}X_i=\mu$1) (Minsker et al., 2020).
• (Pseudo-)Huber M-estimators: solve $\mathbb{E}X_i=\mu$2 with adaptive tuning of $\mathbb{E}X_i=\mu$3 (Sun, 2021, Minsker, 2018).
• Sample-reuse interval estimators: deliver interval bounds for $\mathbb{E}X_i=\mu$4 under partial knowledge of distributional uncertainty (0810.4727).

Recent innovations classify RUMEs semiparametrically using the orderliness of quantile combinations, motivating new block/binomial means and $\mathbb{E}X_i=\mu$5-means (e.g., median Hodges-Lehmann) with superior bias properties (Tuobang, 2024). Other RUMEs leverage empirical characteristic functions (Bahmani, 2020), Bayesian aggregation (Orenstein, 2019), or iterative filter-based procedures (Hopkins et al., 2020).

2. Statistical Guarantees and Robustness Properties

RUMEs are characterized by:

• High breakdown point: Many attain breakdown $\mathbb{E}X_i=\mu$6 or close; MoM tolerates up to a $\mathbb{E}X_i=\mu$7 fraction of contaminated blocks (Lerasle et al., 2011, Oliveira et al., 2023). Trimmed/winsorized means and block/binomial RUMEs are tunable to the contamination level, trading bias and breakdown.
• Sub-Gaussian deviation: Under only finite variance, RUMEs achieve

$\mathbb{E}X_i=\mu$8

for suitable $\mathbb{E}X_i=\mu$9 and $k$0 (Lugosi et al., 2019, Minsker, 2018).

• Minimax optimality: For Huber's $k$1-contamination,

$k$2

is minimax-optimal under finite variance (Prasad et al., 2019). For only finite $k$3th moments, the optimal error is $k$4 (Oliveira et al., 2023).

• Redescending influence: Truncation or weighting guarantees that the effect of extreme outliers is suppressed (Jonsson et al., 2016, Minsker et al., 2020).
• Asymptotic efficiency: Recent advances yield RUMEs asymptotically achieving the Cramér–Rao lower bound for finite-variance families, e.g., self-normalized estimators (Minsker et al., 2020, Sun, 2021).

3. Representative Algorithms and Computational Complexity

A comparison of primary RUME classes and their computational profiles is as follows:

RUME Type	Operations	Complexity	Tuning
Median-of-means (MoM)	Partition, medians	$k$5	Number of blocks
Trimmed mean	Sort, discard extremes	$k$6	Trimming level $k$7
Self-normalized block mean	Block stats, weighted avg	$k$8	$k$9
Huber/Catoni M-estimator	Solve fixed-point eqn	$B_1,\ldots,B_k$0	Scale param
Filter/QUE-based (regret)	Iterative weights	$B_1,\ldots,B_k$1	None
Bayesian MoM (BMM)	Dirichlet resample	$B_1,\ldots,B_k$2	$B_1,\ldots,B_k$3
Truncated quadratic (RMQ)	Sliding window	$B_1,\ldots,B_k$4	Cutoff $B_1,\ldots,B_k$5
Sample-reuse bounds	Monte Carlo w/ reuse	$B_1,\ldots,B_k$6	None

Median-of-means and block-weighted estimators are fully linear-time, only requiring a partition and aggregation (Lerasle et al., 2011, Minsker et al., 2020). Truncated quadratic RMQ (robust mean quadratic) is $B_1,\ldots,B_k$7 due to sorting (Jonsson et al., 2016). Pseudo-Huber self-tuned estimators require convex optimization but are practical and need no variance input (Sun, 2021).

4. Theoretical Insights: Breakdown, Bias, and Influence

Modern RUME designs leverage statistical theory to balance resistance to contamination with efficiency:

• Breakdown point: MoM achieves 50%. Standard trimmed means and winsorized means achieve breakdown $B_1,\ldots,B_k$8 (with $B_1,\ldots,B_k$9 set to the projected contamination level) (Oliveira et al., 2023, Tuobang, 2024). Median Hodges-Lehmann achieves $m_j = \frac{1}{|B_j|}\sum_{i\in B_j} X_i$0 (Tuobang, 2024).
• Bias/Variance trade-offs: For lightly contaminated data, asymptotic bias is negligible for MoM and blockwise RUMEs. Median-of-means and blockwise estimators can incur a constant factor loss in efficiency compared to the mean (e.g., $m_j = \frac{1}{|B_j|}\sum_{i\in B_j} X_i$1 for MoM), but new constructions close this gap (Minsker et al., 2020, Sun, 2021).
• Redescending and adaption: Influence functions for RUMEs are either bounded and redescending (e.g., truncated quadratic, Huber-type) or null beyond a region (MoM, trimmed mean) (Jonsson et al., 2016, Oliveira et al., 2023). In weighted blockwise RUMEs, block means with inflated variance due to outliers get downweighted (Minsker et al., 2020).

5. Extensions, Generalizations, and Recent Developments

• Unified RUME frameworks: Filter-based meta-problems and quantum-entropy (QUE) scores unify robust and heavy-tailed mean estimation for both one- and high-dimensional data, obtaining both computational efficiency and minimax optimality (Hopkins et al., 2020).
• Semiparametric classification: Distributional orderliness via quantile averages ($m_j = \frac{1}{|B_j|}\sum_{i\in B_j} X_i$2), as in (Tuobang, 2024), enables a taxonomy of robust means—trimmed, Winsorized, block/binomial, and $m_j = \frac{1}{|B_j|}\sum_{i\in B_j} X_i$3-means, with explicit bias and breakdown formulas.
• Adaptive/self-tuned M-estimators: Pseudo-Huber estimators with self-tuned scale parameters remove the need for variance estimation while achieving both sub-Gaussian finite-sample rates and asymptotic normality (Sun, 2021).
• Empirical characteristic function estimators: These obtain tight (nearly sub-Gaussian) deviation, even for unbounded mean, and can be refined to achieve exact shift-equivariant optimality (Bahmani, 2020).

6. Comparative Evaluation and Application Domains

Compared to classical mean, sample median, and traditional M-estimators, RUMEs:

• Require only finite variance for sub-Gaussian deviation (mean requires exponential concentration/outlier-free data) (Lerasle et al., 2011, Minsker, 2018).
• Are fundamentally superior in presence of outliers or heavy tails. The mean fails completely under even small contamination; MoM, trimmed, and blockwise estimators remain accurate up to the breakdown threshold (Lerasle et al., 2011, Oliveira et al., 2023).
• Are widely applied in statistical learning with adversarial noise, heavy-tailed regression, robust risk minimization, bandit problems, image processing (RMQ-RUME for edge-preserving smoothing (Jonsson et al., 2016)).

Practical recommendations:

• Use MoM or trimmed mean when only variance or contamination bound is known, as they are tuning-free and provide worst-case guarantees.
• Employ adaptive/self-tuned methods when efficiency and absence of variance information are critical (Sun, 2021).
• For distributional uncertainty models, use sample-reuse RUME for interval bounding (0810.4727).
• For superior bias under right-skewed or nonparametric families, recent semiparametric block/binomial or $m_j = \frac{1}{|B_j|}\sum_{i\in B_j} X_i$4-means are recommended (Tuobang, 2024).

7. Summary Table of Principal RUME Types

Estimator	Breakdown	Optimality	Key Reference
Median-of-means (MoM)	$m_j = \frac{1}{|B_j|}\sum_{i\in B_j} X_i$5	Sub-Gaussian, robust	(Lerasle et al., 2011, Lugosi et al., 2019)
Trimmed mean	$m_j = \frac{1}{|B_j|}\sum_{i\in B_j} X_i$6	Minimax optimality	(Oliveira et al., 2023)
Self-normalized block mean	$m_j = \frac{1}{|B_j|}\sum_{i\in B_j} X_i$7	Asymptotically efficient	(Minsker et al., 2020)
Pseudo-Huber/adaptive $m_j = \frac{1}{|B_j|}\sum_{i\in B_j} X_i$8-est.	---	Asymptotically efficient	(Sun, 2021)
Robust mean (characteristic fn.)	---	Nearly sub-Gaussian	(Bahmani, 2020)
Filter/QUE meta-algorithms	tunable	Minimax/efficient	(Hopkins et al., 2020)
Median Hodges-Lehmann	$m_j = \frac{1}{|B_j|}\sum_{i\in B_j} X_i$9	Lowest max bias	(Tuobang, 2024)
Truncated quadratic (RMQ)	$$\widehat\mu_{\rm MoM} = \mathrm{median}(m_1,\ldots,m_k)$$0	Redescending, metric use	(Jonsson et al., 2016)

• "Efficient Robust Mean Value Calculation of 1D Features" (Jonsson et al., 2016)
• "Uniform bounds for robust mean estimators" (Minsker, 2018)
• "Trimmed sample means for robust uniform mean estimation and regression" (Oliveira et al., 2023)
• "Robust mean estimation and regression under heavy-tailed distributions—a survey" (Lugosi et al., 2019)
• "Robust empirical mean Estimators" (Lerasle et al., 2011)
• "Robust and efficient mean estimation: an approach based on the properties of self-normalized sums" (Minsker et al., 2020)
• "A Unified Approach to Robust Mean Estimation" (Prasad et al., 2019)
• "Do we need to estimate the variance in robust mean estimation?" (Sun, 2021)
• "Nearly Optimal Robust Mean Estimation via Empirical Characteristic Function" (Bahmani, 2020)
• "Robust and Heavy-Tailed Mean Estimation Made Simple, via Regret Minimization" (Hopkins et al., 2020)
• "Robust Mean Estimation with the Bayesian Median of Means" (Orenstein, 2019)
• "Robust Estimation of Mean Values" (0810.4727)
• "Robust estimations from distribution structures: I. Mean" (Tuobang, 2024)
• "Robust estimation of U-statistics" (Joly et al., 2015)