Measurement Models & Sensitivity Coefficients
In metrology and precision calibration, we rarely measure the target quantity (the measurand) in direct isolation. Instead, we measure several related physical parameters and compute the result through a mathematical equation. That governing equation is the measurement model, and the levers that scale every input uncertainty into the output are the sensitivity coefficients.
Level 1: The Intuition: Levers, Knobs, and Recipes
Before diving into multi-variable calculus, consider three physical analogies that explain why sensitivity coefficients exist and what they actually represent:
1. The Baking Recipe
Imagine baking artisan sourdough bread. The quality depends on flour, water, and salt. If you accidentally add an extra 10 grams of flour to a 500 g batch (a 2% error), the bread tastes virtually identical. But if you add an extra 10 grams of salt to a 10 g recipe (a 100% error), the bread is completely ruined.
The output recipe is far more sensitive to salt than to flour. In metrology, the salt has a huge sensitivity coefficient; flour has a tiny one.
2. Gear Ratios & Control Knobs
Think of an analog machine with two tuning dials coupled by internal gears to an output needle. Turning Knob A by 1 millimeter moves the needle 10 millimeters (a 10:1 gear ratio). Turning Knob B by 1 millimeter moves the needle only 0.1 millimeters (a 1:10 ratio).
If your hand shakes by $\pm 0.5\text{ mm}$ on both dials, Knob A creates $\pm 5.0\text{ mm}$ of needle flutter, while Knob B creates only $\pm 0.05\text{ mm}$. The gear ratio is the sensitivity coefficient.
3. Speed, Distance & Time
Average velocity is $v = d / t$. Suppose you drive 60 miles in 1 hour ($60\text{ mph}$). If your odometer reading has an uncertainty of $\pm 1\text{ mile}$, your computed speed varies by $\pm 1\text{ mph}$.
However, if your wristwatch has an uncertainty of just 1 minute ($1/60\text{ hr}$), that tiny 60-second error also shifts your computed speed by approximately $\pm 1\text{ mph}$. The sensitivity to elapsed time is 60 times steeper per hour than distance.
Level 2: The Functional Relationship $Y = f(X_1, X_2, \dots, X_N)$
According to ISO/IEC Guide 98-3 (GUM Clause 4.1), in almost all measurement scenarios, the measurand $Y$ is not observed directly, but is determined from $N$ other input quantities $X_1, X_2, \dots, X_N$ through a mathematical functional relationship:
In practice, the true values of the input quantities $X_i$ are unknown. We estimate them using measured values, mean observations, calibration certificate corrections, or reference handbook values, denoted as lowercase estimates $x_1, x_2, \dots, x_N$. The output estimate $y$ is then:
Linear Measurement Models
A model is linear when the output is a weighted sum or difference of the input quantities:
Key characteristics: The sensitivity coefficients $c_i$ are constant everywhere across the entire measurement range. Input probability distributions pass through without warping, and first-order uncertainty propagation is mathematically exact.
Non-Linear Measurement Models
A model is non-linear whenever it involves products, quotients, powers, square roots, trigonometric functions, or exponentials:
Key characteristics: The sensitivity coefficients change depending on the operating point $(x_1, \dots, x_N)$. We must calculate the local tangent slope at the specific operating point using partial differentiation.
Level 3: Rigorous Calculus: Partial Derivatives as Sensitivity Coefficients
To understand how an uncertainty in an input $X_i$ propagates into the output $Y$, we expand the non-linear function $f(X_1, \dots, X_N)$ in a multi-variable Taylor series about the input estimates $(x_1, x_2, \dots, x_N)$:
In standard GUM uncertainty evaluation, when non-linearities are mild and input uncertainties are small relative to the input values, we truncate the Taylor series after the first-order (linear) terms. The first partial derivatives evaluated at the input estimates are defined as the sensitivity coefficients, denoted $c_i$:
Physical Meaning & Dimensional Analysis of $c_i$
The sensitivity coefficient $c_i$ is a physical conversion rate:
- Units: The dimension of $c_i$ is always $\frac{\text{units of } Y}{\text{units of } X_i}$. For instance, if $Y$ is power in Watts and $V$ is voltage in Volts, $c_V$ has units of $\text{W/V}$.
- Uncertainty Contribution: Multiplying $c_i$ by the standard uncertainty $u(x_i)$ produces the component uncertainty contribution $u_i(y)$:$u_i(y) \equiv |c_i| \cdot u(x_i)$Notice that $u_i(y)$ now shares the exact same unit as the measurand $Y$, allowing direct comparison across completely different physical input quantities (e.g. comparing temperature in kelvins vs resistance in ohms).
- Combined Standard Uncertainty: For mutually independent (uncorrelated) inputs, the combined variance is the sum of the squared individual contributions:$u_c^2(y) = \sum_{i=1}^N c_i^2 \, u^2(x_i) = \sum_{i=1}^N [u_i(y)]^2$Taking the square root gives the combined standard uncertainty:$u_c(y) = \sqrt{\sum_{i=1}^N \left( \frac{\partial f}{\partial x_i} \, u(x_i) \right)^2}$
Level 4: Multiplicative Power-Product Models & Relative Sensitivity
Many of the most important measurement models across electrical, mechanical, and thermodynamic metrology consist purely of products, quotients, and power laws:
Where $K$ is a constant and each exponent $p_i$ can be positive, negative, integer, or fractional (e.g., $-1$ for inverse terms, $1/2$ for square roots).
The Logarithmic Shortcut: Exponents as Relative Sensitivity Coefficients
Taking the natural logarithm of both sides converts multiplication into simple addition:
Differentiating implicitly with respect to each variable yields the fractional change:
This proves a remarkable metrological theorem: for any power-product model, the relative (fractional) sensitivity coefficient of each input is simply its exponent $p_i$!
The combined relative standard uncertainty $u_{c,r}(y) = \frac{u_c(y)}{|y|}$ can therefore be evaluated directly from the relative uncertainties $u_r(x_i) = \frac{u(x_i)}{|x_i|}$ without computing any absolute derivatives:
Or in percentage terms: $u_{\%}^2(y) = \sum_{i=1}^N p_i^2 \cdot u_{\%}^2(x_i)$.
| Physical Law | Model Equation | Input & Exponent | Relative Sensitivity Impact |
|---|---|---|---|
| Kinetic Energy | $E_k = \frac{1}{2} m v^2$ | $p_m = 1, \; p_v = 2$ | Velocity uncertainty is magnified by a factor of 2 in relative terms (and 4x in variance contribution). |
| Simple Pendulum | $T = 2\pi \sqrt{L / g} = 2\pi L^{0.5} g^{-0.5}$ | $p_L = 0.5, \; p_g = -0.5$ | Square root attenuates relative uncertainty: a 1% error in length causes only a 0.5% error in period. |
| Electrical Power | $P = V^2 / R = V^2 R^{-1}$ | $p_V = 2, \; p_R = -1$ | A 1% uncertainty in voltage contributes 2% to power; a 1% uncertainty in resistance contributes 1%. |
| Ideal Gas Law | $P = n R_u T / V = n R_u T V^{-1}$ | $p_T = +1, \; p_V = -1$ | Equal relative weighting: temperature and volume both propagate at a 1:1 relative scale. |
Level 5: Numerical Finite-Difference (When Calculus Is Impractical)
In modern metrology, measurement models are frequently embedded inside software algorithms, complex look-up calibration tables, iterative root finders, finite element models, or polynomials containing dozens of terms. Calculating symbolic partial derivatives $\frac{\partial f}{\partial X_i}$ by hand can be error-prone or analytically impossible.
Under GUM Clause 5.1.3 Note 2, sensitivity coefficients may be determined numerically using the two-point central finite-difference method:
Where $h_i$ is a small numerical perturbation step size applied to input $x_i$ while all other inputs are held strictly constant at their estimated values.
Why Central Difference Beats Forward Difference
The forward difference formula $\frac{f(x_i + h_i) - f(x_i)}{h_i}$ carries a truncation error of order $O(h_i)$. In contrast, the two-point central difference cancels out all even-order Taylor series error terms, yielding a truncation error of order $O(h_i^2)$:
This delivers orders of magnitude higher numerical precision for the identical computational cost of just two function evaluations.
The Metrologist's Rule of Thumb for $h_i$
How should you choose the step size $h_i$?
- GUM Clause 5.1.3 Recommendation: Set the perturbation equal to the standard uncertainty itself: $h_i = u(x_i)$!
- When $h_i = u(x_i)$, the uncertainty contribution simplifies directly to:$u_i(y) \approx \frac{1}{2} \left| f(x_i + u(x_i)) - f(x_i - u(x_i)) \right|$
- Numerical Software: For automated solvers, setting $h_i = \max\left(10^{-6} |x_i|, 10^{-8}\right)$ balances roundoff error against Taylor truncation error.
Level 6: Step-by-Step Practical Worked Examples
Worked Example 1: Electrical Power Dissipation ($P = V^2 / R$)
A precision current shunt is calibrated by applying a steady DC voltage $V$ and measuring resistance $R$. The measurand is the electrical power dissipated:
Estimate: $v = 120.00\text{ V}$
Standard Uncertainty: $u(v) = 0.24\text{ V}$
Relative Uncertainty: $u_r(v) = \frac{0.24}{120.00} = 0.0020 = 0.20\%$
Estimate: $r = 25.000\ \Omega$
Standard Uncertainty: $u(r) = 0.050\ \Omega$
Relative Uncertainty: $u_r(r) = \frac{0.050}{25.000} = 0.0020 = 0.20\%$
Step 1: Compute the nominal measurand estimate $p$:
Step 2: Analytical Partial Differentiation (Absolute Method):
Sensitivity to Voltage:
$c_V = \frac{\partial P}{\partial V} = \frac{2V}{R} = \frac{2(120.00)}{25.000} = +9.600\text{ W/V}$Uncertainty contribution from $V$:
$u_V(P) = |c_V| \cdot u(v) = 9.600 \times 0.24\text{ V} = 2.304\text{ W}$Sensitivity to Resistance:
$c_R = \frac{\partial P}{\partial R} = -\frac{V^2}{R^2} = -\frac{14400}{625.000} = -23.040\text{ W/}\Omega$Uncertainty contribution from $R$:
$u_R(P) = |c_R| \cdot u(r) = |-23.040| \times 0.050\ \Omega = 1.152\text{ W}$Step 3: Verification via Relative Sensitivity (Logarithmic Method):
Because $P = V^2 R^{-1}$, the exponents are $p_V = 2$ and $p_R = -1$:
Step 4: Verification via Central Finite-Difference:
Perturb voltage by $u(v) = 0.24\text{ V}$:
$P(v + 0.24, r) = \frac{120.24^2}{25} = 578.3055\text{ W}$
$P(v - 0.24, r) = \frac{119.76^2}{25} = 573.6975\text{ W}$
$u_V(P) \approx \frac{578.3055 - 573.6975}{2} = 2.3040\text{ W}$ (Matches analytical!)
Perturb resistance by $u(r) = 0.050\ \Omega$:
$P(v, r + 0.050) = \frac{14400}{25.050} = 574.8503\text{ W}$
$P(v, r - 0.050) = \frac{14400}{24.950} = 577.1543\text{ W}$
$u_R(P) \approx \frac{|574.8503 - 577.1543|}{2} = 1.1520\text{ W}$ (Matches analytical!)
Uncertainty Budget Summary Table:
| Quantity $X_i$ | Estimate $x_i$ | Std Uncertainty $u(x_i)$ | Sensitivity $c_i$ | Contribution $u_i(y)$ | Variance Share |
|---|---|---|---|---|---|
| Voltage $V$ | 120.00 V | 0.24 V | +9.600 W/V | 2.304 W | 80.0% |
| Resistance $R$ | 25.000 Ω | 0.050 Ω | -23.040 W/Ω | 1.152 W | 20.0% |
| Power $P$ | 576.00 W | Combined $u_c(P)$ | 2.576 W | 100.0% | |
Key Insight: Even though voltage and resistance both had the same 0.20% relative uncertainty, voltage is responsible for 80% of the combined variance because its exponent is 2 ($2^2 = 4$ times the variance weighting)!
Worked Example 2: Density Determination of a Cylinder ($\rho = \frac{4m}{\pi d^2 h}$)
A metrology laboratory calibrates the density $\rho$ of a solid cylindrical stainless steel mass standard by determining mass $m$, cylinder diameter $d$, and cylinder height $h$:
Estimate: $1.000000\text{ kg}$
Std Unc: $15\text{ mg} = 1.5\times 10^{-5}\text{ kg}$
Relative: $15\text{ ppm} = 0.0015\%$
Estimate: $0.050000\text{ m}$ (50 mm)
Std Unc: $2.0\ \mu\text{m} = 2.0\times 10^{-6}\text{ m}$
Relative: $40\text{ ppm} = 0.0040\%$
Estimate: $0.063662\text{ m}$ (63.662 mm)
Std Unc: $3.0\ \mu\text{m} = 3.0\times 10^{-6}\text{ m}$
Relative: $47.1\text{ ppm} = 0.00471\%$
Step 1: Nominal density calculation:
Step 2: Analytical Sensitivity Coefficients:
Step 3: Combined Uncertainty Evaluation:
Density Uncertainty Budget & Pareto Variance Contribution:
| Input Parameter | Exponent $p_i$ | Relative Unc | Contribution $u_i(\rho)$ | Variance Share $u_i^2 / u_c^2$ | Pareto Impact |
|---|---|---|---|---|---|
| Diameter ($d$) | -2 | 40.0 ppm | 0.640 kg/m³ | 72.4% | Dominant error driver |
| Height ($h$) | -1 | 47.1 ppm | 0.377 kg/m³ | 25.1% | Secondary contributor |
| Mass ($m$) | +1 | 15.0 ppm | 0.120 kg/m³ | 2.5% | Negligible |
| Combined Density ($\rho = 8000.00\text{ kg/m}^3$) | 0.752 kg/m³ | 100.0% | $U_{95\%} = 1.50\text{ kg/m}^3$ ($k=2$) | ||
Metrology Engineering Takeaway: Diameter measurement uncertainty contributes $72.4\%$ of the total variance! If a calibration laboratory wants to halve the uncertainty of their density calibration, investing in a higher-accuracy laser micrometer or CMM for diameter yields nearly 30 times the return on investment compared to purchasing a more expensive mass comparator.
Common Pitfalls & When First-Order Models Fail
1. Forgetting the Negative Signs
Because standard uncertainties are squared in independent RSS combinations ($c_i^2 u^2(x_i)$), metrologists sometimes disregard negative signs in $c_i$. This is catastrophic when input quantities are correlated! In correlated models, the cross-terms are $2 c_i c_j u(x_i, x_j)$. If one coefficient is positive and one is negative, positive correlation reduces uncertainty (common-mode rejection). Ignoring the sign produces completely invalid uncertainty budgets.
2. Unit Inconsistencies
A classic error is mixing units when computing $c_i = \partial f / \partial X_i$. If length is in millimeters in the formula, but $u(x_i)$ is quoted in micrometers or meters, your calculated uncertainty will be off by a factor of 1,000 or 1,000,000. Always verify that $c_i \times u(x_i)$ resolves to the identical unit as the measurand $Y$.
3. Significant Non-Linearity & Curvature
When the second-order Taylor term is non-negligible compared to the first-order term:
The first-order linear approximation begins to underestimate uncertainty and introduce an uncorrected asymmetry bias into the mean estimate $y$. In such cases, one must use GUM Clause 5.1.2 higher-order terms or deploy Monte Carlo simulation (JCGM 101).
4. Numerical Finite-Difference Step Trap
When using central difference $\frac{f(x_i + h) - f(x_i - h)}{2h}$, choosing an excessively tiny step size (e.g. $h = 10^{-16}$) triggers severe floating-point roundoff catastrophe due to catastrophic cancellation of near-identical numbers. Conversely, choosing $h$ too large samples non-linear curvature instead of the tangent. Stick to $h_i = u(x_i)$ or $h_i = 10^{-5} x_i$.
Continue Your Uncertainty Exploration
Correlated Uncertainty
What happens when your input quantities depend on each other? Learn how covariance matrices and cross-terms alter uncertainty propagation.
CMC Uncertainty Budgets
How sensitivity coefficients integrate directly into ISO/IEC 17025 Calibration and Measurement Capability (CMC) budgets.
Monte Carlo (JCGM 101)
When analytical derivatives and linear Taylor approximations fail, use numerical probability distribution propagation.