Root Mean Squared Error (RMSE)

• Root Mean Squared Error (RMSE) is a metric that measures average prediction error by computing the square root of the mean of squared deviations, maintaining dimensional consistency.
• It is widely applied in forecast verification, Bayesian estimation, and dynamical system reduction, leveraging its quadratic penalization to highlight large errors.
• Recent research integrates RMSE into skill score frameworks and model reduction techniques, optimizing model selection and enhancing computational efficiency in high-dimensional systems.

Root Mean Squared Error (RMSE) is a fundamental measure of average squared deviation between estimated and true values, extensively employed across disciplines including statistical learning, dynamical system modeling, and forecast verification. RMSE possesses several mathematically advantageous properties: it is directly interpretable in the units of the quantity of interest, penalizes large errors quadratically, and provides analytic tractability for both continuous and discrete settings. The metric serves roles as both a raw error quantifier and a building block for more sophisticated evaluation frameworks, including skill scores and model reduction objectives. Modern research addresses not only the raw computation of RMSE but also its integration into model selection, consistency analysis between optimization and verification directives, and structure-preserving reduction techniques for large-scale systems.

1. Mathematical Definition and Core Properties

The RMSE between a forecast vector $f = (f_1, ..., f_n)$ and an observation vector $o = (o_1, ..., o_n)$ is defined as

$$\mathrm{RMSE}(f,o) = \sqrt{\frac{1}{n} \sum_{i=1}^n (f_i - o_i)^2}$$

for discrete samples, and

$$\mathrm{RMSE} = \sqrt{ \frac{1}{T} \int_0^T [y(t)]^2 dt }$$

for a scalar time-series $y(t)$ over a continuous interval $[0, T]$ (Reiter et al., 2024).

RMSE is dimensionally homogeneous with the predictand, and its strict convexity with respect to the estimator makes it particularly suited to analytic minimization properties. Quadratic penalization renders the metric sensitive to outliers, enhancing its discriminative power in signal and system assessment.

2. RMSE in Probabilistic Estimation and Classification

In Bayesian MMSE error estimation contexts such as linear discriminant analysis (LDA) for the multivariate Gaussian model, RMSE quantifies the expected squared deviation between an estimator $\hat{\varepsilon}$ and the true error $\varepsilon$: $$\mathrm{RMSE}[\hat{\varepsilon}] = \sqrt{ \mathbb{E}\left[ (\hat{\varepsilon} - \varepsilon)^2 \right] }$$ where the expectation is taken over sampling randomness and, if appropriate, prior distributional uncertainty (Zollanvari et al., 2013).

Key analyses involve explicit closed-form characterizations of the moments $E[\hat{\varepsilon}]$, $o = (o_1, ..., o_n)$0, and $o = (o_1, ..., o_n)$1, enabling analytic calculation of RMSE under both asymptotic and finite-sample regimes. These results deliver practical computation of RMSE for evaluating classifier performance even in small-sample, high-dimensional scenarios where model-free approaches fail. The framework allows derivation of sample size requirements to guarantee RMSE below user-specified thresholds, a critical consideration in experimental design.

3. RMSE in Model Reduction of Dynamical Systems

In linear time-invariant (LTI) system modeling, RMSE is recast as a quadratic output functional. For a system $o = (o_1, ..., o_n)$2 with state $o = (o_1, ..., o_n)$3 and output $o = (o_1, ..., o_n)$4, the root-mean-squared value is

$o = (o_1, ..., o_n)$5

This quadratic structure is exploited to enable direct preservation of the RMSE measure under projection-based model reduction. Both Galerkin and Petrov-Galerkin methods utilize carefully constructed projection matrices $o = (o_1, ..., o_n)$6 to produce reduced-order surrogates $o = (o_1, ..., o_n)$7 that interpolate the quadratic transfer function

$o = (o_1, ..., o_n)$8

at designated frequency points. When these Hermite interpolation conditions are met, the reduced model accurately matches the original RMSE over the frequency range of interest, enabling dramatic computational savings for high-dimensional problems (e.g., a $o = (o_1, ..., o_n)$9 system reduced to order $$\mathrm{RMSE}(f,o) = \sqrt{\frac{1}{n} \sum_{i=1}^n (f_i - o_i)^2}$$0 achieves order-of-magnitude speedup with errors $$\mathrm{RMSE}(f,o) = \sqrt{\frac{1}{n} \sum_{i=1}^n (f_i - o_i)^2}$$1) (Reiter et al., 2024).

4. RMSE-Based Skill Scores and Consistency

Skill scores contextualize RMSE by comparing a candidate forecasting method to a reference or naive forecast. The canonical RMSE-based skill score is

$$\mathrm{RMSE}(f,o) = \sqrt{\frac{1}{n} \sum_{i=1}^n (f_i - o_i)^2}$$2

where $$\mathrm{RMSE}(f,o) = \sqrt{\frac{1}{n} \sum_{i=1}^n (f_i - o_i)^2}$$3 is a reference forecast (Mayer et al., 2024).

A crucial consideration is consistency—the requirement that the forecast optimization criterion match the verification directive. When forecasts are optimized for mean absolute error (MAE) but scored by RMSE, inconsistency arises: RMSE penalizes squared errors, whereas MAE penalizes linearly, leading to bias in skill score interpretation and unfair penalization or reward, especially prevalent in fields (e.g., deterministic solar forecasting) where practice diverges from theoretical consistency requirements.

5. Potential RMSE Skill Score: Addressing Inconsistency

Mayer and Yang (2024) introduce the "potential RMSE skill score," which evaluates what skill a given forecast could achieve if it were linearly recalibrated to minimize MSE. This potential skill score depends only on the Pearson correlation coefficient $$\mathrm{RMSE}(f,o) = \sqrt{\frac{1}{n} \sum_{i=1}^n (f_i - o_i)^2}$$4 between the (possibly non-MSE-optimized) forecast and the observations, and the lag-$$\mathrm{RMSE}(f,o) = \sqrt{\frac{1}{n} \sum_{i=1}^n (f_i - o_i)^2}$$5 autocorrelation $$\mathrm{RMSE}(f,o) = \sqrt{\frac{1}{n} \sum_{i=1}^n (f_i - o_i)^2}$$6 for the reference (CLIPER) forecast: $$\mathrm{RMSE}(f,o) = \sqrt{\frac{1}{n} \sum_{i=1}^n (f_i - o_i)^2}$$7 No explicit recalibration is required; all necessary information is contained in the relevant cross- and auto-correlation statistics. This approach corrects for metric inconsistency: highly correlated, non-MSE-optimized forecasts (e.g., MAE minimizers) are not unfairly penalized and can be validly compared alongside MSE-optimized solutions (Mayer et al., 2024).

Experimental results demonstrate that the potential skill score eliminates artificial performance gaps between MAE- and MSE-optimized forecasts when correlation structure is similar, supporting its use as a cross-directive, consistency-preserving verification tool.

6. Computational Efficiency in RMSE Evaluation

For large-scale dynamical models, evaluating RMSE directly from raw outputs rapidly becomes computationally prohibitive, particularly when the output set is high-dimensional or time-discretized at fine scales. By leveraging the quadratic output formulation, model order reduction techniques construct low-dimensional surrogates that preserve the RMSE measure to high accuracy. Methods enforcing Hermite interpolation of the quadratic transfer function reliably achieve relative errors of $$\mathrm{RMSE}(f,o) = \sqrt{\frac{1}{n} \sum_{i=1}^n (f_i - o_i)^2}$$8 or lower with substantial reductions in solve time and memory footprint (Reiter et al., 2024). A plausible implication is that RMSE-driven model reduction is feasible for real-time, high-dimensional simulation workflows.

7. Guidelines for RMSE Use and Sample Size Determination

RMSE's closed-form behaviors under Bayesian and frequentist sampling regimes allow derivation of practical guidelines. In MMSE error estimation for LDA, analytic formulas for RMSE enable direct computation of the minimum sample size $$\mathrm{RMSE}(f,o) = \sqrt{\frac{1}{n} \sum_{i=1}^n (f_i - o_i)^2}$$9 required as a function of feature dimension $$\mathrm{RMSE} = \sqrt{ \frac{1}{T} \int_0^T [y(t)]^2 dt }$$0, desired error threshold $$\mathrm{RMSE} = \sqrt{ \frac{1}{T} \int_0^T [y(t)]^2 dt }$$1, and prior informativeness $$\mathrm{RMSE} = \sqrt{ \frac{1}{T} \int_0^T [y(t)]^2 dt }$$2. This capability is particularly valuable for small-sample, high-dimensional classifier design (Zollanvari et al., 2013).

When verifying forecasts with non-MSE-optimized directives, it is advisable to compute both the actual and potential RMSE skill scores. This dual reporting allows for robust, consistent evaluation across heterogeneous forecast generation strategies, minimizing the risk of misleading cross-model or cross-method comparisons (Mayer et al., 2024).

---

In conclusion, Root Mean Squared Error remains a central metric across statistical estimation, model reduction, and forecast verification. Contemporary developments in structure-preserving model reduction, skill score theory, and analytic error estimation collectively extend RMSE's utility while ensuring both computational tractability and methodological consistency in cutting-edge research applications.