Absolute and Relative Errors
CSIR NET Chemical Sciences: absolute and relative error, accuracy vs precision, systematic and random errors, significant figures and propagation of errors.
- 19 min
On this page · 14 sections
Prerequisites
- Mean and Standard Deviation: the SD is the measure of random error used throughout this page.
- Percentages and percentage change from Percentages, Ratios, Averages and Mixtures. A relative error is a percentage of the measured value.
- Logarithms and the Beer–Lambert law, used in the pH and absorbance examples. See Spectroscopic Methods of Analysis.
The Hook: The Same 0.1 mg on Two Weighings
Error, Absolute and Relative
If \( x \) is a measured value and \( x_t \) the true (or accepted) value:
$$\boxed{E_a = x - x_t} \qquad \boxed{E_r = \frac{x - x_t}{x_t}} \qquad \text{percentage error} = E_r \times 100\%$$- The absolute error \( E_a \) has the units of the measurement and a sign: positive when the result is high, negative when low.
- The relative error \( E_r \) is dimensionless; it is also quoted in parts per thousand (ppt, × 1000) in analytical work.
Error vs uncertainty
When the true value is unknown, which is most of the time, we cannot calculate the error. We state an uncertainty instead: a range expected to contain the true value. The same absolute/relative split applies:
- Absolute uncertainty: \( \pm 0.02 \) mL on a burette reading.
- Relative uncertainty: \( 0.02/24.32 = 0.08\% \) of that reading.
The uncertainty may be an instrument tolerance (the “maximum error” of a pipette or balance) or a standard deviation from replicates. The propagation rules below work for both, with one difference noted where it matters.
Accuracy and Precision
| Accuracy | Precision | |
|---|---|---|
| Meaning | closeness of a result to the true value | closeness of replicate results to each other |
| Measured by | absolute or relative error (mean − true) | SD, RSD, range |
| Spoilt by | systematic (determinate) errors | random (indeterminate) errors |
| Needs a true value? | yes (reference material, standard) | no |
Precision is necessary for accuracy but not sufficient: a set of results can agree closely with each other and all be wrong by the same amount.
Types of Error
graph TD
E[Errors in a measurement] --> S["Systematic (determinate)
same sign, has a cause,
affects accuracy"]
E --> R["Random (indeterminate)
both signs, Gaussian,
affects precision"]
E --> G["Gross
blunders, outliers"]
S --> S1[Instrumental
uncalibrated pipette, drift]
S --> S2[Method
indicator end point ≠ equivalence,
coprecipitation, side reaction]
S --> S3[Personal
parallax, colour judgement]
S --> S4["By size: constant
or proportional"]Systematic errors
A systematic error has a definite cause and, in principle, a definite value and sign. Repeating the measurement does not reduce it.
| Source | Example |
|---|---|
| Instrumental | a 25 mL pipette that actually delivers 24.95 mL; a balance out of calibration; a drifting reference electrode |
| Method | an indicator that changes colour after the equivalence point; slight solubility of a precipitate; an interfering ion |
| Personal | consistently reading the meniscus from above; a bias in judging a colour change |
Constant vs proportional. A constant error has the same absolute size whatever the sample size (for example, the volume of titrant needed to change the indicator colour, or a fixed solubility loss on washing). Its relative effect shrinks as the sample grows: a 0.5 mg loss is 1% of a 50 mg precipitate but 0.1% of a 500 mg one. A proportional error grows with sample size (an interfering impurity in the sample, an error in a standard’s concentration), so its relative size stays constant.
Detecting and removing systematic error:
- Analyse a certified reference material and compare with the certified value (a t-test, see the previous page).
- Use an independent method on the same sample.
- Run a blank (all reagents, no analyte) to correct for reagent contributions and indicator error.
- Vary the sample size: a constant error shows up as a result that changes with the mass taken.
- Calibrate glassware and instruments.
Random errors
Random errors come from many small, uncontrollable fluctuations: reading a scale, thermal noise in a detector, small temperature changes. They are equally likely to be positive or negative, follow a Gaussian distribution, and are reduced by averaging (as \( 1/\sqrt{n} \)). They are expressed as an SD.
Gross errors
Blunders: a transcription error, a spilled sample, the wrong pipette. They produce outliers, which are handled by a test such as the Q-test or by repeating the work.
Significant Figures
The significant figures of a number are all the certain digits plus the first uncertain one. Writing 24.32 mL claims the 2 in the hundredths place is uncertain.
| Number | Significant figures | Note |
|---|---|---|
| 0.00450 | 3 | leading zeros are not significant; the trailing zero is |
| 1.040 | 4 | zeros between and after digits after a decimal point count |
| \( 2.50 \times 10^{3} \) | 3 | scientific notation makes this unambiguous |
| 2500 | ambiguous | write \( 2.5 \times 10^3 \) or \( 2.500 \times 10^3 \) |
Rules for results:
- Addition and subtraction: keep the decimal places of the term with the fewest decimal places. \( 12.11 + 0.3 = 12.4 \).
- Multiplication and division: keep roughly the number of significant figures of the factor with the fewest. \( 4.3 \times 2.112 = 9.1 \).
- Logarithms: the number of digits after the decimal point in the log (the mantissa) equals the significant figures of the number. \( \log(2.0 \times 10^{-3}) = -2.70 \) (two figures in 2.0, two decimals in −2.70).
- Antilogarithms: the reverse. pH 4.37 has two decimals, so \( [\text{H}^+] = 10^{-4.37} = 4.3 \times 10^{-5} \) M, two significant figures.
- Carry extra digits through a calculation and round only the final answer.
- Exact numbers (stoichiometric coefficients, the 2 in \( 2\pi \), counted objects) do not limit the figures.
Propagation of Errors
A result \( y \) is usually computed from several measured quantities \( a, b, c \), each with its own uncertainty. How big is the uncertainty in \( y \)?
There are two conventions:
- Maximum (limiting) error: assume every error has its worst-case sign, so absolute or relative errors simply add. Used for instrument tolerances and in most “maximum percentage error” questions.
- Standard deviation (random) error: independent random errors partly cancel, so they add in quadrature (square, add, square-root). Used when the uncertainties are SDs.
Quadrature always gives a smaller number than straight addition. Read the question: “maximum error” means add; “standard deviation of the result” means quadrature.
The rules
| Calculation | Maximum error | Standard deviation |
|---|---|---|
| \( y = a + b - c \) | \( E_y = E_a + E_b + E_c \) (absolute) | \( s_y = \sqrt{s_a^2 + s_b^2 + s_c^2} \) (absolute) |
| \( y = ab/c \) | \( \frac{E_y}{y} = \frac{E_a}{a} + \frac{E_b}{b} + \frac{E_c}{c} \) | \( \frac{s_y}{y} = \sqrt{\left(\frac{s_a}{a}\right)^2 + \left(\frac{s_b}{b}\right)^2 + \left(\frac{s_c}{c}\right)^2} \) |
| \( y = k a \) (\( k \) exact) | \( E_y = k E_a \) | \( s_y = k s_a \) |
| \( y = a^n \) | \( \frac{E_y}{y} = n\frac{E_a}{a} \) | \( \frac{s_y}{y} = n\frac{s_a}{a} \) |
| \( y = \log_{10} a \) | \( E_y = 0.434\,\frac{E_a}{a} \) | \( s_y = 0.434\,\frac{s_a}{a} \) |
| \( y = \ln a \) | \( E_y = \frac{E_a}{a} \) | \( s_y = \frac{s_a}{a} \) |
| \( y = 10^{a} \) | \( \frac{E_y}{y} = 2.303\,E_a \) | \( \frac{s_y}{y} = 2.303\,s_a \) |
| \( y = e^{a} \) | \( \frac{E_y}{y} = E_a \) | \( \frac{s_y}{y} = s_a \) |
The one-line summary: sums and differences carry absolute errors; products and quotients carry relative errors; a power multiplies the relative error by the exponent; a log turns a relative error into an absolute one.
Note that \( y = a^2 \) is not the same as \( y = a \times b \) with \( b = a \): the two factors are the same measurement, so their errors are fully correlated and the relative error doubles (\( 2 s_a/a \)), rather than growing by \( \sqrt{2} \).
The general formula
All the rows above come from one rule. For \( y = f(a, b, c, \dots) \) with independent errors,
$$\boxed{s_y^2 = \left(\frac{\partial y}{\partial a}\right)^2 s_a^2 + \left(\frac{\partial y}{\partial b}\right)^2 s_b^2 + \left(\frac{\partial y}{\partial c}\right)^2 s_c^2 + \cdots}$$For the maximum error, drop the squares and add the absolute values of each term. If the errors in \( a \) and \( b \) are not independent, a covariance term \( 2\,\frac{\partial y}{\partial a}\frac{\partial y}{\partial b}\,\text{cov}(a, b) \) is added; see Linear Regression, Covariance and Correlation.
Worked examples
Which measurement to improve
Because random errors add in quadrature, the largest relative uncertainty dominates. If one term is 0.08% and another 0.01%, the result is \( \sqrt{0.0064 + 0.0001} = 0.0806\% \): the small term is irrelevant. Spend the effort on the biggest contributor.
A related classic: in spectrophotometry, an uncertainty in transmittance produces the smallest relative error in concentration near \( A = 0.434 \) (for a constant absolute uncertainty in \( T \)). The derivation uses the log rule above; the result is discussed in Spectroscopic Methods of Analysis.
Memory Tricks
Don’t Confuse
| This | vs | That |
|---|---|---|
| Absolute error (units, signed) | relative error (dimensionless, %, ppt) | |
| Accuracy: mean vs true value | precision: results vs each other | |
| Systematic error: one sign, not reduced by repeats | random error: both signs, reduced by averaging | |
| Constant error: relative size falls with bigger sample | proportional error: relative size constant | |
| Maximum error: \( E_a + E_b \) | standard error: \( \sqrt{s_a^2 + s_b^2} \) | |
| Error of \( a^2 \): \( 2 s_a/a \) | error of \( a \times b \) (independent): \( \sqrt{(s_a/a)^2 + (s_b/b)^2} \) | |
| Error: needs the true value | uncertainty: an estimated range when the truth is unknown |
Common Mistakes
Practice
All questions below are original practice questions written for this page. They are not from any CSIR NET paper.
Level 1: Foundation
- Distinguish absolute error from relative error in one line each.
- The radius of a spherical particle is measured with a 2% error. What is the percentage error in its cross-sectional area?
- State the rule for the uncertainty in \( y = \log_{10} x \).
Answers
- Absolute error = measured − true value, in the units of the measurement. Relative error = absolute error ÷ true value, dimensionless (often in % or ppt).
- 4%. Area \( \propto r^2 \), so the relative error doubles.
- \( s_y = 0.434\,s_x/x \): the relative error in \( x \), times \( 1/\ln 10 \), is the absolute error in \( y \).
Level 2: Part B Style
Practice Q1. A reference sample contains 50.0 mg of analyte. An analyst finds 49.6 mg. The relative error is:
(a) −0.4 mg (b) −0.8% (c) +0.8% (d) −0.4%
Answer: (b) −0.8%
Absolute error \( = 49.6 - 50.0 = -0.4 \) mg. Relative error \( = -0.4/50.0 = -0.008 = -0.8\% \).
(a) is the absolute error, not the relative one; (c) has the wrong sign; (d) treats the absolute error as a percentage.
Practice Q2. The density of a cubic crystal is found from its mass and edge length. The edge is measured with a maximum error of 1% and the mass with 0.5%. The maximum percentage error in the density is:
(a) 1.5% (b) 2.5% (c) 3.5% (d) 4.0%
Answer: (c) 3.5%
\( \rho = m/a^3 \). Maximum relative error \( = 0.5 + 3(1) = 3.5\% \).
(a) forgets the cube.
Practice Q3. A pH meter reading is uncertain by ±0.01 pH unit. The corresponding relative uncertainty in [H⁺] is about:
(a) 0.01% (b) 0.43% (c) 1% (d) 2.3%
Answer: (d) 2.3%
\( [\text{H}^+] = 10^{-\text{pH}} \), so \( s_{[\text{H}^+]}/[\text{H}^+] = 2.303 \times 0.01 = 0.023 \), about 2.3%.
(b) uses 0.434 (the rule for the reverse direction).
Practice Q4. Which of the following is a systematic (determinate) error?
(a) Fluctuation in the last digit of a balance reading caused by air currents (b) Use of an indicator whose colour change consistently occurs after the equivalence point (c) Small random variation in reading a burette meniscus (d) Electronic noise in a detector signal
Answer: (b)
An indicator that always changes late gives a titre that is always high: a definite cause with a definite sign, a method error. Removing it needs an indicator blank or a better indicator, not more replicates.
(a), (c) and (d) are random errors: both signs, reduced by averaging.
Level 3: Part C Style
Practice Q5. A first-order reaction has \( E_a = 80 \) kJ mol⁻¹. Its rate constant is measured at 310 K, but the bath temperature is uncertain by ±0.5 K. Take \( R = 8.314 \) J K⁻¹ mol⁻¹. The resulting relative uncertainty in \( k \), and in the half-life, are closest to:
(a) 5.0% in \( k \); 5.0% in \( t_{1/2} \) (b) 5.0% in \( k \); 2.5% in \( t_{1/2} \) (c) 10% in \( k \); 10% in \( t_{1/2} \) (d) 0.16% in \( k \); 0.16% in \( t_{1/2} \)
Answer: (a)
\( \frac{dk}{k} = \frac{E_a}{RT^2}\,dT = \frac{80\,000}{8.314 \times 310^2} \times 0.5 = 0.1001 \times 0.5 = 0.050 \), i.e. 5.0%.
\( t_{1/2} = 0.693/k \). The constant 0.693 is exact, and a quotient carries the relative error unchanged, so \( t_{1/2} \) is also uncertain by 5.0% (in the opposite direction).
(c) forgets the 0.5 K (uses 1 K). (d) is just \( 0.5/310 \), the relative error in \( T \), ignoring the exponential.
Practice Q6. The molar mass of a volatile liquid is found from \( M = \frac{mRT}{pV} \) (Dumas method). The relative standard uncertainties are: mass 0.40%, temperature 1 K at 300 K, pressure 0.50%, volume 0.20% (\( R \) exact). Consider:
- The relative uncertainty in \( T \) is 0.33%.
- The relative standard uncertainty in \( M \) is about 0.75%.
- The maximum relative error in \( M \) is about 1.43%.
- Halving the volume uncertainty would reduce the standard uncertainty in \( M \) more than halving the pressure uncertainty would.
Which statements are correct?
(a) 1 and 2 only (b) 1, 2 and 3 only (c) 2, 3 and 4 only (d) 1, 2, 3 and 4
Answer: (b) 1, 2 and 3 only
- \( 1/300 = 0.33\% \). Correct.
- \( \sqrt{0.40^2 + 0.33^2 + 0.50^2 + 0.20^2} = \sqrt{0.160 + 0.111 + 0.250 + 0.040} = \sqrt{0.561} = 0.75\% \). Correct.
- \( 0.40 + 0.33 + 0.50 + 0.20 = 1.43\% \). Correct.
- Halving the volume term: \( \sqrt{0.160 + 0.111 + 0.250 + 0.010} = 0.73\% \). Halving the pressure term: \( \sqrt{0.160 + 0.111 + 0.0625 + 0.040} = 0.61\% \). Improving the pressure, the largest contributor, helps more. Incorrect.
Practice Q7. A certified reference solution contains 0.2500 M Cu²⁺. Two methods give (five replicates each):
- Method X: mean 0.2501 M, \( s = 0.0030 \) M
- Method Y: mean 0.2560 M, \( s = 0.0004 \) M
Consider:
- Method Y is more precise than method X.
- Method X is more accurate than method Y.
- The relative error of the mean of method Y is +2.4%.
- Averaging more replicates with method Y would bring its mean close to 0.2500 M.
Which statements are correct?
(a) 1 and 2 only (b) 1, 2 and 3 only (c) 2 and 4 only (d) 1, 2, 3 and 4
Answer: (b) 1, 2 and 3 only
- Y’s \( s \) (0.0004 M) is far smaller than X’s (0.0030 M). Correct.
- X’s mean is 0.0001 M off (+0.04%); Y’s is 0.0060 M off. Correct.
- \( (0.2560 - 0.2500)/0.2500 = 0.024 = +2.4\% \). Correct.
- Y’s error is systematic. Averaging reduces random error only; the mean of Y stays near 0.2560 M. Incorrect.
Quick Revision Table
| Idea | Formula or fact |
|---|---|
| Absolute error | \( E_a = x - x_t \) (units, sign) |
| Relative error | \( E_r = E_a/x_t \); × 100 for %, × 1000 for ppt |
| Accuracy / precision | error of mean / SD of replicates |
| Systematic error | one sign, cause can be found, not cut by repeats |
| Random error | both signs, Gaussian, falls as \( 1/\sqrt{n} \) |
| Constant vs proportional | constant: relative effect falls with sample size |
| Sum, difference | absolute errors: add (max) or \( \sqrt{\sum s^2} \) |
| Product, quotient | relative errors: add (max) or \( \sqrt{\sum (s/x)^2} \) |
| Power \( a^n \) | relative error × \( n \) |
| \( \log_{10} a \) | \( s_y = 0.434\,s_a/a \) |
| \( 10^a \) | \( s_y/y = 2.303\,s_a \) |
| Arrhenius sensitivity | \( dk/k = (E_a/RT^2)\,dT \) |
| Sig figs in logs | decimals in the log = sig figs in the number |
Key Takeaways
What’s Next
- Linear Regression, Covariance and Correlation: fitting straight lines to data, the uncertainty in slope and intercept, and what the correlation coefficient does and does not tell you.
- Mean and Standard Deviation: revisit confidence intervals and the Q-test.
- Radio-analytical Techniques: the same quadrature rule applied to net count rates.
- Data Analysis: Quick Revision.