Skip to content

Lesson 41 · Data Analysis and Projects

Chapter 8 review: data analysis

A recap of fitting, uncertainty estimation and Fourier analysis in one picture and one table, plus a ten-question quiz mixing the whole chapter.

Simulations and experiments both end with numbers, and this chapter taught you how to turn them into defensible results. You fit models by least squares, sized the uncertainties three different ways, and moved between time and frequency with the FFT. This page puts everything on one sheet.

Four tiles: fit, uncertainty, spectrum and report
Chapter 8 in one picture: fit, size the uncertainty, look at the frequencies, report.

Recap table

Topic Key formula or rule Check
Statistical error σ_mean = σ/√N 100 times more data gives 10 times smaller error
Systematic error bias that does not average away vary the time step, box size, calibration
Chi-square χ² = Σ((y − f)/σ)², dof = N − p χ²/dof ≈ 1, residuals structureless
Line fit σ_m² = S/Δ, σ_b² = S_xx/Δ agrees with np.polyfit
1σ interval where χ² rises by 1 matches covariance and bootstrap
Resolution Δf = 1/T longer record, finer lines
Nyquist f_max = f_s/2 above it, frequencies alias
Leakage non-integer cycles smear the peak use a Hann window

Check your understanding

Quick check · 10 questions

of 10 answered You scored / 10 · perfect!

  1. 1. A mean of 400 measurements has statistical uncertainty 0.05. How large is it for 1600 measurements?

    Four times more data halves the uncertainty: 0.05/2 = 0.025.
  2. 2. You fit 15 points with a 3-parameter model and get χ² = 30. What is χ²/dof, and what does it suggest?

    dof = 12 and 30/12 = 2.5, well above 1.
  3. 3. Which method gives parameter uncertainty by refitting resampled data?

    The bootstrap resamples with replacement and uses the spread of refit results.
  4. 4. A record lasts 5 s. What is the smallest frequency separation you can resolve?

    Δf = 1/T = 0.2 Hz.
  5. 5. A 130 Hz tone sampled at 200 Hz appears at...

    Nyquist is 100 Hz; 130 Hz folds to 200 − 130 = 70 Hz.
  6. 6. Which is a systematic error in a simulation?

    Too large a step biases every result in the same way, and a longer run does not remove it.
  7. 7. Why divide the Hann-windowed spectrum by the window sum?

    The window reduces the signal amplitude; dividing by its mean (sum/N) compensates.
  8. 8. Residuals of a fit show a clear upward trend. The best action is to...

    A trend means the model misses a systematic effect.
  9. 9. Which statement about zero padding before an FFT is correct?

    Resolution is set by the record length only.
  10. 10. A 'result' is reported as '12.3'. What is missing?

    A number is not a result without its unit and uncertainty.

Summary

  • Statistical errors fall like 1/√N; systematic errors need different checks, not more data.
  • Judge fits by χ²/dof and residuals; get uncertainties from covariance, Δχ² = 1 or the bootstrap.
  • The FFT resolves Δf = 1/T up to f_s/2; avoid aliasing and leakage.
  • Report value, uncertainty, unit and method.

Next: Project: the chaotic double pendulum