Experimental Uncertainties: A Practical Guide What you should already know well What you need to know, and use, in this lab More details available in handout ‘Introduction to Experimental Error’ (2 nd Year Web page). In what follows I will use the word uncertainty instead of error, although in the literature the both are used: –Everybody uses the term “error-bar” in graphs.

Why are Uncertainties Important? Uncertainties absolutely central to the scientific method. Uncertainty on a measurement at least as important as measurement itself! Example 1: “The observed frequency of the emission line was 8956 GHz. The expectation from quantum mechanics was 8900 GHz” Nobel Prize?

Why are Uncertainties Important? Example 2: “The observed frequency of the emission line was 8956 ± 10 GHz. The expectation from quantum mechanics was 8900 GHz” Example 3: “The observed frequency of the emission line was 8956 ± 10 GHz. The expectation from quantum mechanics was 8900 ± 50 GHz”

Types of Uncertainty Statistical Uncertainties ( aka random error ): –Quantify random uncertainties in measurements between repeated experiments –Mean of measurements from large number of experiments gives correct value for measured quantity –Measurements often approximately gaussian- distributed Systematic Uncertainties ( aka syst error, bias ): –Quantify systematic shift in measurements away from ‘true’ value –Mean of measurements is also shifted  ‘bias’

Examples Statistical Uncertainties: –Measurements gaussian- distributed –No system. uncertainty (bias) –Quantify uncertainty in measurement with standard deviation (see later) –In case of gaussian-distributed measurements std. dev. =  in formula –Probability interpretation (gaussian case only): 68% of measurements will lie within ± 1  of mean. True Value

Examples True Value Statistical + Systematic Uncertainties: –Measurements still gaussian- distributed –Measurements biased –Still quantify statistical uncertainty in measurement with standard deviation –Probability interpretation (gaussian case only): 68% of measurements will lie within ± 1  of mean. –Need to quantify systematic uncertainty separately  tricky!

Systematic Uncertainties How to quantify uncertainty? What is the ‘true’ systematic uncertainty in any given measurement? –If we knew that we could correct for it (by addition / subtraction) What is the probability distribution of the systematic uncertainty? –Often assume gaussian distributed and quantify with  syst. –Best practice: propagate and quote separately True Value

Calculating Statistical Uncertainty Mean and standard deviation of set of independent measurements (unknown errors, assumed uniform): Standard deviation estimates the likely error of any one measurement Uncertainty in the mean is what is quoted:

Propagating Uncertainties Functions of one variable (general formula): Specific cases:

f =Apply equationSimplify Propagating Uncertainties Functions of >1 variable (general formula): Specific cases:

Combining Uncertainties What about if have two or more measurements of the same quantity, with different uncertainties? Obtain combined mean and uncertainty with: Remember we are using the uncertainty in the mean here:

Fitting Often we make measurements of several quantities, from which we wish to 1.determine whether the measured values follow a pattern 2.derive a measurement of one or more parameters describing that pattern (or model) This can be done using curve-fitting E.g. EXCEL function linest. Performs linear least-squares fit

Method of Least Squares This involves taking measurements y i and comparing with the equivalent fitted value y i f Linest then varies the model parameters and hence y i f until the following quantity is minimised: Linest will return the fitted parameter values (=mean) and their uncertainties (in the mean) In this example the model is a straight line y i f = mx+c. The model parameters are m and c never In the second year lab never use the equations returned by ‘Add Trendline’ or linest to estimate your parameters!!!

Weighted Fitting Those still awake will have noticed the least square method does not depend on the uncertainties (error bars) on each point. Q: Where do the uncertainties in the parameters come from? –A: From the scatter in the measured means about the fitted curve Equivalent to: Assumes errors on points all the same What about if they’re not?

Weighted Fitting To take non-uniform uncertainties (error bars) on points into account must use e.g. chi-squared fit. Similar to least-squares but minimises: Enables you to propagate uncertainties all the way to the fitted parameters and hence your final measurement (e.g. derived from gradient). This is what is used by chisquare.xls (download from Second Year web-page)  this is what we expect you to use in this lab!

General Guidelines Always: Calculate uncertainties on measurements and plot them as error bars on your graphs Use chisquare.xls when curve fitting to calculate uncertainties on parameters (e.g. gradient). Propagate uncertainties correctly through derived quantities Quote uncertainties on all measured numerical values Quote means and uncertainties to a level of precision consistent with the uncertainty, e.g: 3.77±0.08 kg, not ± kg. Quote units on all numerical values

General Guidelines Always: Think about the meaning of your results –A mean which differs from an expected value by more than 1-2 multiples of the uncertainty is, if the latter is correct, either suffering from a hidden systematic error (bias), or is due to new physics (maybe you’ve just won the Nobel Prize!) Never: Ignore your possible sources of error: do not just say that any discrepancy is due to error (these should be accounted for in your uncertainty) Quote means to too few significant figures, e.g.: 3.77±0.08 kg not 4±0.08 kg