All lessons

Module 7: Electronics and experimental methods · Lesson 7.2 · about 55 min

Measurement, errors and fitting data

How NET Part C tests data analysis: random and systematic errors, resolution, standard error, propagation through non-linear and correlated formulas, Gaussian confidence levels, weighted means, least-squares lines with the errors and covariance of slope and intercept, linearising plots, chi-square and goodness of fit, and when (not) to reject a point.

Key points, meanings and formulas on one page: choose "Save as PDF" in the print box.

Module 7 · about 55 minutes. By the end you will be able to separate random from systematic errors, quote a result with the right number of figures, propagate uncertainties through any formula (including correlated inputs and stationary points), combine measurements with weights, fit a straight line by least squares and give the errors and covariance of its slope and intercept, choose the right plot to linearise a law, and judge a fit with chi-square. The basic distributions (binomial, Poisson, Gaussian) are in the GATE lesson on probability and statistics; here we use them.

A measurement is a number with a spread. Everything in data analysis follows from two facts: variances of independent quantities add, and the best fit is the one that makes the data most probable, which for Gaussian errors means the smallest chi-square.

1. Two kinds of error, and resolution

Random errors scatter repeated readings both ways; averaging N readings shrinks them as 1/N. Systematic errors push every reading the same way (zero error, calibration, parallax from one side, thermal expansion of a scale); averaging never removes them. Precision is about the spread, accuracy about the offset: a set of readings can be precise and still wrong.

For N readings xi:

x‾=1N∑xi,s=∑(xi−x‾)2N−1,σx‾=sN

The N−1 appears because x‾ was fitted to the same data, using up one degree of freedom. The standard deviation s describes one reading; the standard error σx‾ describes the mean.

Resolution. A scale or display that rounds to steps of Δ leaves the true value anywhere in a band of width Δ: a uniform distribution with standard deviation Δ/12 (0.29 of a step). It adds in quadrature with the statistical error and, when every reading rounds to the same value, it does not average away.

Reporting. Round the uncertainty to one or two significant figures, then round the value to the same decimal place: g=9.82±0.04 m s⁻², never 9.8164±0.0372.

standard deviation s
the spread of single readings; it does not shrink with more data
standard error σ/√N
the spread you would see in the mean if you repeated the whole experiment
Δ/√12
the standard uncertainty from rounding to steps of Δ; the half-step Δ/2 is the worst case, not the standard deviation
degrees of freedom
number of data points minus number of parameters fitted to them

2. Propagation of uncertainties

For f(x,y,…) with small, independent errors, expand to first order: δf=fxδx+fyδy+…, and since variances of independent terms add,

σf2=(∂f∂x)2σx2+(∂f∂y)2σy2+2∂f∂x∂f∂ycov(x,y)

The last term vanishes for independent inputs. Two cases cover most problems:

  • sums and differences: absolute errors add in quadrature, σx−y2=σx2+σy2. Subtracting nearly equal numbers keeps the absolute error but destroys the relative one.
  • products of powers f=xayb/zc: relative errors add in quadrature, each times its power, (σf/f)2=(aσx/x)2+(bσy/y)2+(cσz/z)2.

Worked example. A pendulum of length 1.000±0.002 m makes 50 swings in 100.0±0.2 s. Find g=4π2L/T2 and its error.

  1. T=2.000±0.004 s (the timing error is shared by 50 swings).
  2. g=4π2×1.000/4.000=9.870 m s⁻².
  3. σg/g=(0.002)2+(2×0.002)2=4.5×10−3, so g=9.87±0.04 m s⁻². The period counts twice because it is squared.

When a variable appears twice, differentiate the whole formula; do not combine “the error of the numerator” with “the error of the denominator”. For f=x/(x+y) with x=2.0±0.1, y=3.0±0.1: fx=y/(x+y)2=0.12, fy=−x/(x+y)2=−0.08, so σf=0.0122+0.0082=0.014.

Correlated errors. If two lengths are read with the same tape whose scale is off by a factor (1+ϵ), the ratio L1/L2 does not feel ϵ at all. Common calibration errors cancel in ratios and add linearly in sums; the covariance term above is how the formula knows.

When first order fails. At a stationary point (f′=0) or for large relative errors, keep the second-order term: ⟨f⟩≈f(μ)+12f′′(μ)σ2. For sin⁡θ near 90°, linear propagation predicts zero spread; the true spread is σ2/2 (in radians). The figure lets you see the breakdown.

Twenty thousand Gaussian values of x pushed through f(x). The histogram is the truth; the curve is what first-order propagation predicts. Increase the spread, or pick cos(x − 1) where the slope is zero, and watch the two part company.

3. Combining measurements: the weighted mean

Results xi±σi of the same quantity combine with weights wi=1/σi2:

x‾w=∑wixi∑wi,σx‾w=(∑wi)−1/2

Worked example. Three labs report 4.2±0.2, 4.6±0.4, 4.0±0.1. Weights 25, 6.25 and 100 (sum 131.25) give x‾w=533.75/131.25=4.07 with error 131.25−1/2=0.09. The plain average, 4.27, lets the worst measurement count as much as the best. Before averaging, check that the results agree within their errors; if they do not, something systematic is going on.

4. Fitting a straight line by least squares

For data (xi,yi±σi) with exact x, the most probable line y=a+bx under Gaussian errors minimises

χ2=∑i(yi−a−bxiσi)2

Setting ∂χ2/∂a=∂χ2/∂b=0 with equal errors gives, with Δ=N∑x2−(∑x)2:

b=N∑xy−∑x∑yΔ,a=y‾−bx‾,σb=σNΔ,σa=σ∑x2Δ

Three facts save time in the exam: the line passes through the centroid (x‾,y‾); the residuals add to zero; and cov(a,b)=−x‾σb2, negative when the data sit at positive x (a steeper slope pivots the line about the centroid and lowers the intercept). The slope error does not depend on the y values at all, only on σ and the spread of the x values: σb=σ/∑(xi−x‾)2. Spread your x values widely.

If the errors are unknown, estimate them from the scatter: s2=∑ri2/(N−2) (two parameters fitted). For a line forced through the origin, y=kx, the result is k=∑xy/∑x2 with σk=σ/∑x2.

Worked example. Fit (1,2.1), (2,3.9), (3,6.2), (4,7.8).

  1. ∑x=10, ∑y=20.0, ∑xy=59.7, ∑x2=30, Δ=4×30−100=20.
  2. b=(4×59.7−10×20.0)/20=1.94; a=5.0−1.94×2.5=0.15.
  3. Residuals: 0.01,−0.13,0.23,−0.11 (they sum to zero). s2=0.082/2=0.041, s=0.20, so σb=0.204/20=0.09.

Linearising. Fit a straight line to the right variables: ln⁡y against x for y=Ae−kx (slope −k); ln⁡y against ln⁡x for y=Cxn (slope n); T2 against L for a pendulum; ln⁡ρ against 1/T for activated processes. Transforming changes the error bars: for counts N, σln⁡N=1/N, so points with few counts deserve less weight on a log plot.

5. Judging the fit: chi-square

If the model is right and the errors are honest, each term of χ2 is about 1, and the minimum follows a chi-square distribution with ν=N−p degrees of freedom: mean ν, standard deviation 2ν. So

χmin⁡2ν≈1±2/ν

Much larger means the errors are underestimated or the model is wrong; much smaller (say 0.2 for ν=20) means the errors are overestimated or the points are not independent. Always look at the residuals: random signs mean the model is fine; a smile or a wave means a missing term, whatever χ2 says.

Ten points from a known line with Gaussian scatter, fitted by least squares. Change the real scatter and the error bar you claim: χ²/ν follows their ratio squared. Switch to a slightly curved law and the residuals (lower panel) form a smile that no choice of error bar can hide.

Confidence regions. For one parameter, the 1σ (68.3%) interval is where χ2 rises by 1 from its minimum. For two parameters taken together, the 68.3% region is bounded by Δχ2=2.30.

Rejecting data. Chauvenet’s criterion drops a point if fewer than half a point that far out is expected among N: for N=10 the threshold is |z|>1.96, for N=100 about 2.8. Use it sparingly and say so; never reject a point just because it spoils a fit.

Counting time. To measure a net rate Rg−Rb in a fixed total time, split the time as tg/tb=Rg/Rb. With 64 and 16 counts/s, two thirds of the time goes to the source run.

“A χ² close to zero means a perfect experiment.” It means the error bars are too big, or the data were smoothed. An honest fit scatters about its line by about one error bar.

“Random errors and systematic errors add the same way.” Independent random errors add in quadrature; a systematic error common to several readings is fully correlated and can cancel (in a ratio) or add linearly (in a sum). Treat it through the covariance, not by habit.

  • Standard error of the mean s/N; resolution adds Δ/12 in quadrature; report the value to the decimal place of a one- or two-figure error.
  • Propagation: σf2=∑(∂f/∂xi)2σi2+2∑fifjcovij; powers multiply relative errors; differentiate the whole formula when a variable appears twice; go to second order at stationary points.
  • Weighted mean with w=1/σ2; its error (∑w)−1/2.
  • Least squares: b=(N∑xy−∑x∑y)/Δ, line through (x‾,y‾), σb=σN/Δ, cov(a,b)=−x‾σb2, s2=∑r2/(N−2).
  • χ2/ν≈1±2/ν; Δχ2=1 for one parameter, 2.30 for two jointly; residual patterns reveal a wrong model.

Practise: exam questions on this lesson

NET Part C style, from definitions to multi-step problems, with full solutions. More on the CSIR NET practice page. The Uncertainty propagation calculator and the Curve fitter check your hand calculations.

Next: lesson 7.3 turns physical quantities into the voltages you have just learned to analyse.

Check yourself

1A straight line is fitted to 12 points with known errors and gives χ² = 30. What is the reduced chi-square?
2Two lengths are measured with the same tape, which has an unknown 0.5% scale error. What does that error contribute to the ratio of the lengths?
3A digital meter rounds to whole millivolts. What standard uncertainty does the rounding alone contribute?

Sources

  • P. R. Bevington and D. K. Robinson, Data Reduction and Error Analysis for the Physical Sciences, 3rd ed. (2003), chapters 3, 4, 6, 8 and 11
  • J. R. Taylor, An Introduction to Error Analysis, 2nd ed. (1997), chapters 3, 7, 8, 9 and 12
  • I. G. Hughes and T. P. A. Hase, Measurements and their Uncertainties (2010), chapters 4 to 8
  • JCGM 100:2008, Evaluation of measurement data: Guide to the expression of uncertainty in measurement (GUM), open access from BIPM
  • CSIR-UGC NET Physical Sciences syllabus, Part C unit on experimental techniques and data analysis

Your ticks stay in this browser. Save them as a file on the course home.

Updated 2026-11-22. Not affiliated with CSIR, UGC or the National Testing Agency. Check the official CSIR NET website for the current pattern. All questions are original.