Correlated Uncertainties & Covariance Evaluation
In basic uncertainty analysis, metrologists frequently assume that all input quantities entering a measurement model are completely mutually independent. Under this simplifying assumption, the combined standard uncertainty is evaluated via simple root-sum-of-squares (RSS).
However, in high-precision calibration and scientific metrology, input quantities are frequently correlated. Whether due to the use of a shared calibration standard across multiple measurement points, simultaneous exposure to ambient temperature fluctuations, or common instrument drift, neglecting correlation violates ISO/IEC Guide 98-3 (GUM) Clause 5.2 and can lead to dangerous underestimations—or massive overestimations—of measurement uncertainty.
Mathematical Formulation of Covariance & Correlation
When two input quantities $X_i$ and $X_j$ are interdependent, their mutual dispersion is quantified by their covariance, denoted as $u(x_i, x_j)$ or $\text{cov}(x_i, x_j)$.
For simultaneously observed pairs of independent repeated measurements $(x_{i,k}, x_{j,k})$ over $n$ observations (Type A evaluation), the sample covariance is calculated as:
The degree of linear interdependence is normalized into the dimensionless correlation coefficient $r(x_i, x_j)$:
Where $u(x_i)$ and $u(x_j)$ are the standard uncertainties of the respective estimates. When $r = 0$, the quantities are uncorrelated; when $r = \pm 1$, there is perfect linear correlation.
The General Law of Propagation of Uncertainty (GUM Clause 5.2)
For any measurement model $Y = f(X_1, X_2, \dots, X_N)$ whose inputs may be correlated, the full first-order Taylor series expansion yields the complete combined variance:
Or, equivalently expressed in terms of standard uncertainties and correlation coefficients:
Where $c_i = \frac{\partial f}{\partial x_i}$ is the partial derivative (sensitivity coefficient) evaluated at the input estimates. The double summation represents the covariance cross-terms.
Matrix Formulation: Modern Metrological Computation
For multi-dimensional or complex systems, expressing the law of propagation in matrix notation provides mathematical elegance and prevents bookkeeping errors. The combined variance is evaluated as:
Where $\mathbf{J}$ is the row vector of partial derivatives (the Jacobian matrix):
And $\mathbf{V}_x$ is the symmetric $N \times N$ Variance-Covariance Matrix:
Notice that the diagonal entries $V_{ii}$ are the variances $u^2(x_i)$, while the off-diagonal entries $V_{ij} = V_{ji}$ are the covariances.
Physical Causes of Correlation in Calibration Laboratories
Correlation rarely arises by chance; it is usually the deterministic consequence of shared physical conditions or measurement procedures:
1. Shared Reference Standards
When calibrating two different test points, or when calibrating two artifacts against the identical reference calibrator (e.g. Fluke 5730A), the systematic calibration uncertainty of the reference calibrator affects both measurements identically. The errors are positively correlated ($r \approx +1$).
2. Common Environmental Drivers
In dimensional calibration, ambient temperature fluctuation ($\Delta T$) affects both the reference gage block and the workpiece through thermal expansion: $\Delta L_s = L_s \alpha_s \Delta T$ and $\Delta L_w = L_w \alpha_w \Delta T$. Because $\Delta T$ is identical for both, their length uncertainties are highly correlated.
3. Calibration Curve Parameters
When an instrument is fitted to a linear calibration curve $y = \beta_0 + \beta_1 x$, the estimated slope $\hat{\beta}_1$ and intercept $\hat{\beta}_0$ obtained via ordinary least squares are inherently correlated: $\text{cov}(\hat{\beta}_0, \hat{\beta}_1) = -\bar{x} \cdot \text{var}(\hat{\beta}_1)$. Treating them as independent produces invalid uncertainty bands.
4. Comparator Instrument Drift
In substitution weighing (Borda or substitution design on a mass comparator), any linear gain drift or thermal zero drift in the comparator affects both the reference weight reading and the unknown test weight reading during the sequence.
Positive vs. Negative Correlation: Case Studies
The sign of the covariance cross-term ($2 c_i c_j u(x_i, x_j)$) depends on both the sign of the correlation coefficient $r$ and the signs of the sensitivity coefficients $c_i$ and $c_j$:
| Measurement Scenario | Model Function | Sensitivity Coeffs | Correlation Impact on Combined Uncertainty |
|---|---|---|---|
| Difference / Substitution (e.g., Comparator mass, differential voltage) | $Y = X_1 - X_2$ | $c_1 = +1$ $c_2 = -1$ | $u_c^2(y) = u^2(x_1) + u^2(x_2) - 2 u(x_1) u(x_2) r$ Beneficial common-mode rejection: If $r > 0$, the covariance term is subtracted! When $u(x_1) = u(x_2)$ and $r = +1$, $u_c(y) = 0$. The shared systematic error cancels completely. |
| Sum / Series Combination (e.g., Wringing gage blocks in stack, resistor strings) | $Y = X_1 + X_2$ | $c_1 = +1$ $c_2 = +1$ | $u_c^2(y) = u^2(x_1) + u^2(x_2) + 2 u(x_1) u(x_2) r$ Compounding error: If $r > 0$, the uncertainty is amplified. If two gage blocks calibrated with the same standard ($r \approx 1$) are wrung together, $u_c(y) = u(x_1) + u(x_2)$, which is far greater than the independent RSS estimate $\sqrt{u^2(x_1) + u^2(x_2)}$. |
Evaluating Type B Covariance without Paired Data
When repeated paired observations are not available (Type B evaluation), GUM Clause F.1.2 provides an analytical method by decomposing input quantities into independent components and a shared common influence:
Where $Z$ is a shared underlying systematic parameter (such as room temperature $T$ or calibrator voltage drift) with uncertainty $u(z)$, and $g_i$ is the functional dependence. Using partial differentiation:
This analytical decomposition allows metrologists to rigorously evaluate covariance terms directly from physical principles and manufacturer specifications without needing empirical bivariate Monte Carlo trials.