Propagation of Uncertainty

A titration delivers 24.68 mL of titrant. Each drop is measured with a buret good to ±0.02 mL. A balance weighs a sample to ±0.0001 g. Neither number alone tells you how trustworthy the final answer actually is. This applies to the molarity, the percent purity, or the pH. Propagation of uncertainty is the machinery that carries error bars through arithmetic. It lets a result like "0.1042 M" honestly become "0.1042 ± 0.0006 M." Otherwise it stays a bare number pretending to be exact. This matters because every quantity reported in analytical chemistry is a function of several measured inputs. Those inputs never combine their errors by simple addition. A common student instinct — "just add up all the uncertainties" — badly overstates the total error. This happens because independent random errors partially cancel. The correct approach is built on the variance of a sum of random variables. It gives smaller and more realistic uncertainties. It also explains why some steps in a calculation inflate or shrink error in surprising ways. Squaring, or taking a logarithm for pH, are two such steps. The big picture is that uncertainty propagation is just calculus applied to statistics. It uses a first-order Taylor expansion of a multivariable function. This is combined with the rule that variances of independent errors add. From this one idea fall four practical rules. They cover sums/differences, products/quotients, powers, and logarithms/exponentials. Together they cover essentially every computation in Standard Deviation and Variance-based error analysis. These rules are the direct ancestors of how uncertainty is reported for calibration curves. The same holds for titration endpoints and instrumental signals throughout quantitative analysis. This article derives the general propagation formula from a Taylor series. It specializes the formula into the four rules chemists use daily. It then works through three realistic examples. These are a gravimetric calculation, a buret-and-balance titration, and a pH/logarithm problem. Each example includes full significant-figure bookkeeping. It closes with the classic pitfalls. These include correlated errors, systematic vs. random error, and why you never propagate through significant-figure rounding alone. It also points toward how these same variance rules underlie Reporting Analytical Results. They also underlie the uncertainty budgets used in Uncertainty in Calibration Results.

Key takeaways

Uncertainty propagation converts uncertainties in measured inputs into an uncertainty in a calculated result, using variance, not simple addition.
For addition/subtraction, absolute standard deviations combine in quadrature: $s_y = \sqrt{s_a^2+s_b^2+\cdots}$ .
For multiplication/division, relative standard deviations combine in quadrature: $(s_y/y)^2 = (s_a/a)^2+(s_b/b)^2+\cdots$ .
For powers $y=x^n$ , relative uncertainty scales by $|n|$ : $s_y/y = |n|\,(s_x/x)$ .
For logarithms and exponentials (as in pH, pOH, absorbance), the rules change form: $s_{\log x} = 0.434\,(s_x/x)$ and $s_{10^x} = 2.303\,x\,s_x$ .
The general recipe is a first-order Taylor expansion: variances of independent errors add when weighted by squared partial derivatives.

Key idea: Errors add as variances (squares), not as raw magnitudes — because independent random fluctuations partially cancel, and only the mean square deviation is additive.

The core insight

If $y = f(a,b,c,\ldots)$ with independent random errors in $a,b,c,\ldots$ , then to first order

s_y^2 = \left(\frac{\partial f}{\partial a}\right)^2 s_a^2 + \left(\frac{\partial f}{\partial b}\right)^2 s_b^2 + \left(\frac{\partial f}{\partial c}\right)^2 s_c^2 + \cdots

Every propagation rule in analytical chemistry — sums, products, powers, logs — is this one equation evaluated for a specific $f$ .

Where this lives today

Propagation of uncertainty is the quantitative backbone of every analytical result that gets reported with an error bar, and it underlies the international "measurement uncertainty" framework (the GUM) used by accredited testing laboratories worldwide. In pharmaceutical quality control, propagating uncertainty through an HPLC assay — from injection volume, to peak area, to calibration-curve slope, to final potency — determines whether a drug batch's stated purity (say, 99.2 ± 0.3%) can be certified to meet a regulatory specification. In environmental and clinical chemistry, propagating uncertainty through serial dilutions and instrument calibration curves is what allows a lab to report a blood-lead concentration or a water contaminant level with a defensible confidence interval rather than a bare number. Calibration itself is a direct application: the slope and intercept of a linear calibration curve, each with its own standard error from linear regression, must be propagated through the working equation $x = (y - b)/m$ to get the uncertainty in an unknown's concentration — exactly the union of the sum and quotient rules developed here, and the direct subject of Uncertainty in Calibration Results. In electrochemistry and potentiometry, converting a measured cell potential or pH into a concentration always requires the exponential/antilog rule, because Nernstian responses are logarithmic in concentration; the same 2.303 factor from Example 3 reappears whenever an ion-selective electrode reading is converted into a concentration for a certificate of analysis. In materials science and geochronology, propagating uncertainty through ratios of isotopic abundances (for radiometric dating) and through power-law scaling relationships (for particle-size or surface-area determinations) both rely on the same power and quotient rules. And whenever a final analytical number is reported — as required for Reporting Analytical Results — the propagated standard deviation, not an arbitrary sig-fig guess, is what determines how many digits are meaningful and how a confidence interval should be constructed around the mean.

:::quiz[A pH meter reads 4.50 ± 0.01. What is the approximate relative uncertainty in [H⁺] = 10⁻ᵖᴴ?]

1%, equal to the absolute pH uncertainty
0.23%, using 1/2.303 instead of 2.303
2.3%, from 2.303 × 0.01 ✓
10%, assuming the whole pH value scales linearly

:::faq[Why does a small pH uncertainty translate into a large uncertainty in [H⁺]?] Because pH is a logarithm of concentration, and the propagation rule for exponentiating an uncertain exponent multiplies the absolute uncertainty by $\ln 10 \approx 2.303$ to get a relative uncertainty in the antilog. A seemingly tiny pH uncertainty of ±0.01 becomes a relative uncertainty of $2.303\times0.01 \approx 2.3\%$ in $[\mathrm{H^+}]$ — much larger in percentage terms than the pH digit itself suggests. This is the mathematical reason glass-electrode pH measurements, though precise in pH units, translate into comparatively coarse concentration precision, and it is essential in any calculation feeding a pH-derived concentration into an equilibrium constant.

Ask the tutor: How would this propagation change if the buret reading and the sample mass came from instruments with a shared, non-random calibration bias?

The physical picture

Every measured quantity — a buret reading, a balance mass, an absorbance — is really a random variable: repeated trials would scatter around a mean value with some standard deviation. When you combine several such measurements arithmetically to compute a result (a concentration, a rate constant, a pH), the scatter in the inputs jointly produces scatter in the output. The central question of propagation of uncertainty is: given the standard deviations of the inputs, what is the standard deviation of the output? The naive answer — that uncertainties simply add — is wrong for a specific and instructive reason. If a buret delivery has a small random error that happens to push the reading high on one measurement, an independent random error in a second reading is just as likely to push low. These fluctuations partially cancel on average, so the typical combined error is smaller than the sum of the two individual errors. What survives this cancellation, on average, is not the sum of the errors but the sum of their squares — the variances. This is why standard deviations of sums combine "in quadrature" (as a square-root-of-sum-of-squares), the same Pythagorean-style combination that appears throughout statistics whenever independent random variables are added. The mathematical engine behind every specific rule is a first-order Taylor expansion of the function $f$ around the mean values of its arguments. Small deviations in the inputs translate into a small deviation in the output through the local slope of $f$ — the partial derivative. Squaring and averaging that linear approximation, and using the fact that cross terms between independent errors average to zero, yields the general variance-propagation formula shown in the key idea above. Everything else in this article is that formula, evaluated for particular choices of $f$ : addition, multiplication, powers, and logarithms.

Figure

Random errors in inputs a and b combine in quadrature to give the error ellipse for y = a + b

Physically, each rule below answers the same question for a different arithmetic operation: how does the shape of $f$ near the operating point stretch or compress the input uncertainty? Addition has slope 1 in each variable, so absolute uncertainties add in quadrature unchanged. Multiplication has a slope proportional to the other variable, which is exactly what converts the rule into one about relative uncertainties. Powers and logarithms have derivatives that scale nonlinearly, which is why they get their own rules rather than falling out of the multiplication rule directly.

Figure

How the four propagation rules map to the shape of f near the operating point

The core ideas

s_y^2 = \sum_i \left(\frac{\partial f}{\partial x_i}\right)^2 s_{x_i}^2 \qquad \text{(general propagation law, independent errors)}

where: $y=f(x_1,x_2,\ldots)$ is the calculated result, $s_{x_i}$ is the standard deviation of input $x_i$ , and $\partial f/\partial x_i$ is evaluated at the mean values of all inputs. Reading: the variance of the output is a weighted sum of the input variances, with weights given by how sensitively $f$ responds to each variable.

\textbf{Sum/difference: } y = a \pm b \pm c \;\Rightarrow\; s_y = \sqrt{s_a^2+s_b^2+s_c^2}

where: $s_a, s_b, s_c$ are the absolute standard deviations of the terms (same units as $y$ ). Reading: absolute errors combine in quadrature regardless of whether the operation is addition or subtraction — subtraction does not cancel error.

\textbf{Product/quotient: } y = \frac{ab}{c} \;\Rightarrow\; \left(\frac{s_y}{y}\right)^2 = \left(\frac{s_a}{a}\right)^2+\left(\frac{s_b}{b}\right)^2+\left(\frac{s_c}{c}\right)^2

where: $s_a/a$ , etc., are relative standard deviations (dimensionless fractions or percent). Reading: relative errors combine in quadrature; the operation (multiply or divide) doesn't matter, only that the variables enter as products/ratios.

\textbf{Power: } y = x^n \;\Rightarrow\; \frac{s_y}{y} = |n|\,\frac{s_x}{x}

where: $n$ is any real exponent (including fractional, as in a square root, $n=1/2$ ). Reading: raising to a power multiplies the relative uncertainty by the magnitude of the exponent — squaring a number roughly doubles its percent uncertainty.

\textbf{Natural log: } y=\ln x \;\Rightarrow\; s_y = \frac{s_x}{x}\qquad \textbf{Base-10 log: } y=\log x \;\Rightarrow\; s_y = 0.434\,\frac{s_x}{x}

where: $0.434 = 1/\ln 10$ converts the natural-log derivative into base-10 form; this is the rule behind pH from $[\mathrm{H}^+]$ . Reading: a logarithm converts a relative input uncertainty into an absolute output uncertainty — pH uncertainty is nearly independent of the concentration's own relative error once you're near neutral pH.

\textbf{Exponential (antilog): } y = 10^{x} \;\Rightarrow\; \frac{s_y}{y} = 2.303\,s_x \qquad y=e^{x}\;\Rightarrow\;\frac{s_y}{y}=s_x

where: $2.303 = \ln 10$ ; $x$ here is an absolute quantity such as pH, and $s_x$ its absolute standard deviation. Reading: exponentiating turns a small absolute uncertainty in the exponent into a relative uncertainty in the result — this is why one uncertain digit in pH corresponds to a large percent uncertainty in $[\mathrm{H}^+]$ .

Derivation AGeneral propagation formula from a Taylor expansion

Goal: show that $s_y^2 = \sum_i (\partial f/\partial x_i)^2 s_{x_i}^2$ follows directly from linearizing $f$ near the mean values of independent random inputs.

Step 1 — Set up the random variables. Let $x_1,x_2,\ldots,x_n$ be independent measured quantities, each with mean $\bar{x}_i$ and small random deviation $\delta_i = x_i - \bar{x}_i$ , so that $\mathrm{Var}(\delta_i) = s_{x_i}^2$ . Let $y=f(x_1,\ldots,x_n)$ , and write $\bar{y}=f(\bar{x}_1,\ldots,\bar{x}_n)$ .

Step 2 — Linearize. Because the deviations $\delta_i$ are small compared to the scale over which $f$ curves appreciably, expand $f$ in a first-order Taylor series about the means:

y - \bar{y} \approx \sum_{i=1}^n \left(\frac{\partial f}{\partial x_i}\right)_{\bar{x}} \delta_i.

Higher-order terms (involving $\delta_i^2$ or cross-products of second derivatives) are dropped, which is valid as long as the relative uncertainties are modest (a few percent or less) — a condition satisfied by nearly all analytical-chemistry measurements.

Step 3 — Square and average. The variance of $y$ is the expectation of $(y-\bar{y})^2$ :

s_y^2 = E\left[\left(\sum_i \left(\frac{\partial f}{\partial x_i}\right)\delta_i\right)^2\right] = \sum_i \left(\frac{\partial f}{\partial x_i}\right)^2 E[\delta_i^2] + \sum_{i\neq j}\left(\frac{\partial f}{\partial x_i}\right)\left(\frac{\partial f}{\partial x_j}\right)E[\delta_i\delta_j].

Step 4 — Use independence to kill the cross terms. Because the inputs are independent random variables, $E[\delta_i \delta_j]=0$ for $i\neq j$ (the covariance vanishes). Only the diagonal terms survive:

s_y^2 = \sum_{i=1}^n \left(\frac{\partial f}{\partial x_i}\right)^2 s_{x_i}^2.

This is the general propagation law. Every specific rule below is this formula with a particular $f$ substituted in and the derivatives worked out explicitly, which is what makes the sum/product/power/log rules internally consistent rather than a list of unrelated recipes.

Go deeper · Derivation B · 1 minSpecializing the general law: sums vs. products

Derivation BSpecializing the general law: sums vs. products

Goal: derive the addition-in-quadrature rule for sums and the relative-quadrature rule for products directly from the Step 4 result, showing precisely why the two rules look so different.

Step 1 — Sum/difference case. Let $y = a \pm b$ (extension to more terms and to subtraction is identical, since only the square of the derivative matters). The partial derivatives are

\frac{\partial y}{\partial a} = 1, \qquad \frac{\partial y}{\partial b} = \pm 1.

Substituting into the general law,

s_y^2 = (1)^2 s_a^2 + (\pm 1)^2 s_b^2 = s_a^2 + s_b^2,

so $s_y = \sqrt{s_a^2+s_b^2}$ . Note the sign of $b$ 's coefficient drops out upon squaring — this is exactly why subtraction does not reduce uncertainty, a common misconception addressed below.

Step 2 — Product case. Let $y = ab$ . The partial derivatives are

\frac{\partial y}{\partial a} = b, \qquad \frac{\partial y}{\partial b} = a.

Substituting,

s_y^2 = b^2 s_a^2 + a^2 s_b^2.

Step 3 — Convert to relative form. Divide both sides by $y^2 = a^2b^2$ :

\frac{s_y^2}{y^2} = \frac{b^2 s_a^2}{a^2b^2} + \frac{a^2 s_b^2}{a^2b^2} = \left(\frac{s_a}{a}\right)^2 + \left(\frac{s_b}{b}\right)^2.

This is exactly the product rule stated above. The division is what makes the multiplicative-sensitivity factors ( $b$ and $a$ ) cancel neatly, leaving a rule purely in terms of relative uncertainties — a structural difference from the additive case, where absolute uncertainties are the natural currency because $f$ 's derivatives are dimensionless constants (±1) rather than the other variable's magnitude.

Step 4 — Quotient case, for completeness. Let $y=a/b$ . Then $\partial y/\partial a = 1/b$ and $\partial y/\partial b = -a/b^2$ , giving

s_y^2 = \frac{s_a^2}{b^2} + \frac{a^2 s_b^2}{b^4}.

Dividing by $y^2 = a^2/b^2$ :

\frac{s_y^2}{y^2} = \frac{s_a^2}{a^2} + \frac{s_b^2}{b^2},

identical in form to the product rule — confirming that division behaves exactly like multiplication under quadrature combination of relative uncertainties.

Worked examples

Example 1 (foundational) — gravimetric chloride determination, sums and products together. A student determines the mass of AgCl precipitate by difference: crucible + precipitate weighs $24.1863 \pm 0.0002$ g, and the empty crucible weighs $23.7550 \pm 0.0002$ g (each balance reading has the same standard deviation, from repeated calibration checks). Step 1 — subtraction. The precipitate mass is

m = 24.1863 - 23.7550 = 0.4313\ \text{g}.

Applying the sum/difference rule (subtraction still adds variances):

s_m = \sqrt{(0.0002)^2 + (0.0002)^2} = \sqrt{2\times 4\times 10^{-8}} = 0.000283\ \text{g} \approx 0.0003\ \text{g}.

So $m = 0.4313 \pm 0.0003$ g — note the absolute uncertainty barely grew past a single weighing's, even though two independent weighings were involved, but it did not shrink to zero despite being a "well-controlled" difference. Step 2 — convert to percent chloride. Suppose the original sample mass was $0.5217 \pm 0.0003$ g (also from a by-difference weighing with the same balance uncertainty per weighing propagated identically). The mass fraction of chloride in AgCl is a fixed constant, $\mathrm{Cl}/\mathrm{AgCl} = 35.453/143.32 = 0.24738$ (treated as exact for this problem). Percent chloride is

\%\mathrm{Cl} = 100 \times 0.24738 \times \frac{m}{m_{\text{sample}}} = 100\times 0.24738 \times \frac{0.4313}{0.5217}.

Numerically, $0.4313/0.5217 = 0.82672$ , so

\%\mathrm{Cl} = 100 \times 0.24738 \times 0.82672 = 20.454\%.

Step 3 — propagate through the quotient and the constant multiplier. Since $0.24738$ and $100$ are treated as exact (no uncertainty), only the ratio $m/m_{\text{sample}}$ needs the product/quotient rule:

\left(\frac{s_r}{r}\right)^2 = \left(\frac{0.0003}{0.4313}\right)^2 + \left(\frac{0.0003}{0.5217}\right)^2 = (6.957\times10^{-4})^2 + (5.751\times10^{-4})^2.

\left(\frac{s_r}{r}\right)^2 = 4.840\times10^{-7} + 3.307\times10^{-7} = 8.147\times10^{-7} \;\Rightarrow\; \frac{s_r}{r} = 9.026\times10^{-4}.

Multiplying by exact constants leaves the relative uncertainty unchanged, so

s_{\%\mathrm{Cl}} = 20.454 \times 9.026\times10^{-4} = 0.01846\% \approx 0.02\%.

Result: %Cl = 20.45 ± 0.02%. The final answer is reported to a precision consistent with its propagated uncertainty (two decimal places, matching the ±0.02% figure), illustrating how propagation — not an arbitrary sig-fig rule — determines how many digits are meaningful.

Example 2 (applied) — molarity of a titrant from buret volume and standard mass. A 0.1 M NaOH solution is standardized against potassium hydrogen phthalate (KHP, molar mass $M = 204.22$ g/mol, taken as exact for this problem). A mass of KHP, $m = 0.6134 \pm 0.0002$ g, is weighed by difference (two weighings, each ±0.0001 g — the ±0.0002 g already reflects propagation through that subtraction). Titration to the phenolphthalein endpoint requires a buret volume $V = 27.42 \pm 0.03$ mL, where the ±0.03 mL reflects the combined uncertainty of the initial and final buret readings (each read to ±0.02 mL, propagated as a difference: $\sqrt{0.02^2+0.02^2}=0.028\approx0.03$ mL). Step 1 — compute the molarity.

C_{\text{NaOH}} = \frac{m/M}{V} = \frac{0.6134/204.22}{0.02742\ \text{L}} = \frac{0.0030036\ \text{mol}}{0.02742\ \text{L}} = 0.10955\ \text{mol/L}.

Step 2 — relative uncertainty in mass.

\frac{s_m}{m} = \frac{0.0002}{0.6134} = 3.260\times10^{-4}.

Step 3 — relative uncertainty in volume.

\frac{s_V}{V} = \frac{0.03}{27.42} = 1.094\times10^{-3}.

Step 4 — combine by the product/quotient rule (molar mass is exact, so it contributes no term):

\left(\frac{s_C}{C}\right)^2 = (3.260\times10^{-4})^2 + (1.094\times10^{-3})^2 = 1.063\times10^{-7} + 1.197\times10^{-6} = 1.303\times10^{-6}.

\frac{s_C}{C} = 1.141\times10^{-3}.

Step 5 — convert to absolute uncertainty.

s_C = 0.10955 \times 1.141\times10^{-3} = 1.250\times10^{-4}\ \text{mol/L} \approx 0.0001\ \text{mol/L}.

Result: $C_{\text{NaOH}} = 0.1096 \pm 0.0001$ M (rounded to four significant figures, matching the propagated uncertainty at the fourth decimal). This example demonstrates the everyday chemist's shortcut: because the buret's relative uncertainty (0.109%) dominates the mass's relative uncertainty (0.033%), the volume measurement controls the overall precision — a diagnostic insight propagation reveals that "add up the uncertainties" would obscure.

Example 3 (exam/competition-level) — pH, hydronium concentration, and Ka with mixed operations. A weak acid HA has concentration $C = 0.0850 \pm 0.0004$ M (a volumetric dilution with relative uncertainty $0.47\%$ ). Its measured pH is $2.68 \pm 0.02$ (typical glass-electrode reproducibility). Find $[\mathrm{H}^+]$ with its uncertainty, and then $K_a$ with its uncertainty, assuming $[\mathrm{H^+}] = [\mathrm{A^-}]$ and $C - [\mathrm{H^+}] \approx$ equilibrium $[\mathrm{HA}]$ . Step 1 — convert pH to $[\mathrm{H^+}]$ using the exponential rule. Since $[\mathrm{H^+}] = 10^{-\text{pH}}$ , treat $x = -\text{pH} = -2.68$ with $s_x = 0.02$ (the sign doesn't matter once squared). By the antilog rule,

[\mathrm{H^+}] = 10^{-2.68} = 2.089\times10^{-3}\ \text{M}.

\frac{s_{[\mathrm{H^+}]}}{[\mathrm{H^+}]} = 2.303 \times s_{\text{pH}} = 2.303\times 0.02 = 0.04606\quad(4.61\%).

So $s_{[\mathrm{H^+}]} = 2.089\times10^{-3}\times 0.04606 = 9.62\times10^{-5}$ M, giving $[\mathrm{H^+}] = (2.09 \pm 0.10)\times10^{-3}$ M. Notice a modest absolute pH uncertainty (±0.02, less than 1% of the pH value itself) balloons into a 4.6% relative uncertainty in concentration — the hallmark of exponentiation amplifying error. Step 2 — compute $[\mathrm{HA}]$ at equilibrium via subtraction.

[\mathrm{HA}] = C - [\mathrm{H^+}] = 0.0850 - 0.002089 = 0.082911\ \text{M}.

Propagate as a difference (absolute uncertainties in quadrature):

s_{[\mathrm{HA}]} = \sqrt{(0.0004)^2 + (9.62\times10^{-5})^2} = \sqrt{1.600\times10^{-7}+9.26\times10^{-9}} = \sqrt{1.693\times10^{-7}} = 4.114\times10^{-4}\ \text{M}.

So $[\mathrm{HA}] = 0.08291 \pm 0.00041$ M — the subtraction is dominated by $C$ 's own uncertainty since $[\mathrm{H^+}]$ is small, but $[\mathrm{H^+}]$ 's comparatively large relative error still contributes a small absolute term. Step 3 — compute $K_a$ .

K_a = \frac{[\mathrm{H^+}][\mathrm{A^-}]}{[\mathrm{HA}]} = \frac{[\mathrm{H^+}]^2}{[\mathrm{HA}]} = \frac{(2.089\times10^{-3})^2}{0.082911} = \frac{4.364\times10^{-6}}{0.082911} = 5.264\times10^{-5}.

Step 4 — propagate through the square (power rule) and the quotient rule together. Because $[\mathrm{H^+}]$ appears squared, its relative uncertainty in the numerator doubles first:

\frac{s_{[\mathrm{H^+}]^2}}{[\mathrm{H^+}]^2} = 2 \times \frac{s_{[\mathrm{H^+}]}}{[\mathrm{H^+}]} = 2\times 0.04606 = 0.09212\quad(9.21\%).

The relative uncertainty in $[\mathrm{HA}]$ is

\frac{s_{[\mathrm{HA}]}}{[\mathrm{HA}]} = \frac{4.114\times10^{-4}}{0.082911} = 4.962\times10^{-3}\quad(0.496\%).

Combine numerator and denominator relative uncertainties in quadrature (product/quotient rule):

\left(\frac{s_{K_a}}{K_a}\right)^2 = (0.09212)^2 + (0.004962)^2 = 8.486\times10^{-3} + 2.462\times10^{-5} = 8.510\times10^{-3}.

\frac{s_{K_a}}{K_a} = 0.09225\quad(9.23\%).

s_{K_a} = 5.264\times10^{-5} \times 0.0923 = 4.86\times10^{-6}.

Result: $K_a = (5.26 \pm 0.49)\times10^{-5}$ . The dominant contributor is squaring $[\mathrm{H^+}]$ (contributing 9.21% out of a total 9.23%) — an instructive limiting-case check: if the pH electrode's precision improved to $s_{\text{pH}} = 0$ , the $K_a$ uncertainty would collapse essentially to the 0.5% coming from the concentration $C$ alone, confirming that the exponent/log steps, not the volumetric dilution, dominate the error budget here.

Limiting cases, scope & common traps

Situation	What happens to $s_y$	Why
Adding $n$ equal, independent uncertainties $s$	$s_y = s\sqrt{n}$ , not $ns$	Quadrature grows slower than linear sum
Subtracting two nearly equal large numbers	Absolute $s_y$ unchanged, but relative $s_y/y$ can become huge	$y$ itself shrinks toward zero while $s_y$ does not
One input's relative uncertainty ≫ others	$s_y/y \approx$ that one term	Quadrature sum is dominated by its largest term
Squaring ( $n=2$ ) a measured quantity	Relative uncertainty doubles	Power rule: $s_y/y = \|n\|\,s_x/x$
Uncertainty in exponent (e.g., pH) small in absolute terms	Still produces large relative uncertainty in $[\mathrm{H^+}]$	Factor of $2.303$ amplifies through antilog
$n \to 1$ in the power rule	$s_y/y = s_x/x$	Recovers the trivial identity $y=x$
Correlated (non-independent) errors, e.g. same balance zero-point drift affecting both mass readings	Simple quadrature underestimates or overestimates $s_y$	Cross terms $E[\delta_i\delta_j]\neq 0$ must be included via covariance

Common trapMisconception: subtracting two measurements cancels their error

It is tempting to think that because $24.1863 - 23.7550$ "removes" the crucible's contribution, the uncertainty should shrink. It does not: the sign in $\partial y/\partial b = -1$ disappears upon squaring in the general propagation law, so $s_y=\sqrt{s_a^2+s_b^2}$ regardless of whether the operation is $a+b$ or $a-b$ . What does shrink is $y$ itself when subtracting nearly equal numbers — so the relative uncertainty $s_y/y$ can balloon even though the absolute uncertainty stays modest. This is the statistical reason "difference of two large, nearly equal numbers" is flagged as a dangerous calculation in any analytical method.

The chemist behind it

The variance-addition rule that underlies all of this traces to the broader mathematical tradition of error analysis founded by Carl Friedrich Gauss, who developed the theory of combining independent measurement errors while working on astronomical and geodetic observations in the early nineteenth century. Gauss showed that when many small independent errors combine, their resultant follows what is now called the normal (Gaussian) distribution, and that the natural "size" of a combined error is not the sum of the individual errors but the square root of the sum of their squares — precisely the quadrature rule used throughout this article. This insight, originally aimed at astronomical position-fixing and surveying, became the statistical backbone of every physical science that combines multiple imperfect measurements, including quantitative chemical analysis a century later when Gaussian statistics were formally absorbed into analytical method validation.

Origins · 1 minCarl Friedrich Gauss and how this was discovered

Carl Friedrich Gauss

1777–1855 · Göttingen, celestial mechanics and geodesy

Gauss derived the method of least squares and the normal distribution while reconciling redundant, noisy astronomical observations of the orbit of the asteroid Ceres. His treatment of how independent errors combine — variances add, not magnitudes — became the mathematical foundation later adopted wholesale into the statistical treatment of chemical measurement uncertainty, long after his own death, as analytical chemistry matured into a quantitative discipline in the twentieth century.

History

- 1801 — Gauss predicts the recovery position of the asteroid Ceres using least-squares reconciliation of noisy observations.

- 1809 — Gauss publishes Theoria Motus, formalizing the normal distribution and the combination of independent observational errors.

- 1900s — R.A. Fisher and colleagues extend error-combination theory into general statistical inference used across the sciences.

- 1960s — Propagation-of-error formulas become standard material in analytical chemistry textbooks (Skoog, Harris) as instrumentation and calibration curves demand rigorous uncertainty budgets.

- 1990s–present — International guidelines (the ISO "Guide to the Expression of Uncertainty in Measurement," GUM) formalize propagation of uncertainty as the universal standard for metrology across chemistry, physics, and engineering.

Why do we add variances instead of just adding the uncertainties?

Because uncertainties represent random, independent fluctuations, and independent random fluctuations do not always point the same direction. If one measurement's error happens to be high, an unrelated measurement's error is equally likely to be low, so the two partially cancel rather than reinforcing. The Taylor-expansion derivation shows precisely why: when you square the sum of independent deviations, the cross terms (the "interference" between two different errors) average to zero over many repetitions, leaving only the sum of the squared deviations — the variances. Taking the square root of that sum recovers a standard deviation with the same units as the original measurements, giving the familiar quadrature rule.

Does the sum/difference rule change if I subtract instead of add?

No. The rule $s_y=\sqrt{s_a^2+s_b^2}$ is identical for $y=a+b$ and $y=a-b$ , because the partial derivative with respect to the subtracted term is $-1$ , and $(-1)^2=1$ just like $(+1)^2=1$ . The sign only matters for computing the mean value $y$ , not for computing its uncertainty. This is precisely why subtracting two large, nearly equal numbers is dangerous: the absolute uncertainty stays the same modest size, but because the result itself becomes small, the relative uncertainty in that result can become very large.

How is the product rule different in form from the sum rule?

The sum rule combines absolute standard deviations (same units as the answer); the product rule combines relative standard deviations (dimensionless fractions or percentages). This difference comes directly from the partial derivatives: for a sum, $\partial y/\partial a=1$ , a dimensionless constant, so absolute uncertainties propagate directly. For a product $y=ab$ , $\partial y/\partial a = b$ , which has units and scales with the size of the other variable — dividing through by $y=ab$ converts this into a clean statement about relative sizes, which is why chemists always convert to percent or fractional uncertainty before combining terms in a multiplication or division.

Why does squaring a quantity double its relative uncertainty rather than quadrupling it?

Intuitively you might expect squaring to square the error too, but the power rule $s_y/y = |n|\,(s_x/x)$ comes from differentiating $y=x^n$ : the derivative $dy/dx = n x^{n-1}$ , and after dividing by $y=x^n$ to get a relative form, the exponent survives only as a linear multiplicative factor $n$ , not as $n^2$ . This is a first-order (small-uncertainty) approximation; it's why the rule is only reliable when the relative uncertainty in $x$ is modest (a few percent), consistent with the same first-order Taylor-expansion assumption used to derive the general propagation law itself.

What happens if the input errors are not independent — for example, the same balance is used for both weighings?

The simple quadrature formulas assume the covariance between different errors is zero. If the same systematic bias (say, an uncalibrated balance zero point) affects two measurements identically, the errors are correlated, and the general propagation law must include cross terms, $2\sum_{i<j}(\partial f/\partial x_i)(\partial f/\partial x_j)\,\mathrm{Cov}(x_i,x_j)$ . Ignoring correlation when it exists can make you overestimate or underestimate the true uncertainty in $y$ , depending on the sign of the covariance and the operation involved (subtraction of two positively correlated quantities, for instance, actually reduces uncertainty because the shared bias cancels).

Is propagation of uncertainty the same thing as significant-figure rounding?

No — significant-figure conventions are a rough, qualitative mnemonic that assumes uncertainties are comparable in relative size and roughly one unit in the last reported digit. Propagation of uncertainty is the exact, quantitative procedure using the actual numerical standard deviations of each input. In practice, you compute the propagated $s_y$ using the rules above, and only then decide how many digits to report in $y$ — typically rounding $y$ so its last reported digit matches the first significant digit of $s_y$ . Sig-fig rules can get you into the right ballpark quickly, but they can also mislead when uncertainties are markedly unequal, as seen when one term dominates a quadrature sum.