Introduction to Propagation of Error (a.k.

a Propagation of Uncertainty)
Cell Biology Laboratory 307 [The following presentation is meant to alert you to problems of measurements and to introduce you to considerations for a proper mathematical combination of measurements and associated errors. The ideas presented below may be pursued more formally when or after you take a statistics course. For our Cell Biology Laboratory Course, know how to estimate the error of concentration, such as when someone measures out 100 mg of protein and dissolves it into deionized water to obtain 100 mL of solution.] Whenever you measure a quantity, such as when weighing a substance on a balance, there is always some error associated with the measurement. Usually the scale error is assumed to of the smallest division on the measurement scale. (In addition, there are other possible errors that you may not at first think of: the balance may not be calibrated correctly, the balance may not be level, the substance being weighed may have absorbed water from the atmosphere, the substance may be less pure than the container label indicates (due to contamination when using a dirty spatula), etc. The error of measurement is usually given as the standard deviation when considering individual sample measurements, or as the standard error of the mean when dealing with mean (average) values. The error value may be expressed as absolute error, as relative error, or as percent error. For example, 58.5 mg 0.1 mg would have an absolute error of 0.1 mg, a relative error of (0.1 mg/58.5 mg) = 0.0017 and a percent error of 0.17 %. If you were trying to make 10 ml of a 1.00 mM NaCl solution while using a balance that was accurate to 1 mg, you would need to measure out at least 100 mg of NaCl just to satisfy the weighing aspect of preparing the 1.00 mM solution (because 1.00 mM means a concentration between 0.995 mM and 1.005 mM, and [0.01 mM/1.00 mM] = 1 % error, and similarly, 100 mg 1 mg is a measurement at 1 % error). However, because there is also an error associated with the measurement of the volume of water used for preparing the 1.00 mM NaCL solution, you will want to make a weight measurement that has less than 1 % error! (As well as measuring the volume of water with less than 1 % error.) This leads to the questions of (1) how to mathematically combine the error of weight and the error of volume when trying to estimate the error of concentration in the resultant solution and (2) what is the minimum error values that we ought to aim toward when weighing a mass and measuring a volume, to achieve a given acceptable error in concentration for the final solution. (These considerations are the subject matter of Propagation of Error or Propagation of Uncertainty.) The first thing to remember concerning uncertainty propagation involving a series of errors is that error is additive: when making a solution, measuring a volume within 0.001% will do nothing to reduce the error of final concentration if mass is measured only within 10%. Similarly, measuring the mass within 0.001 % will not help if volume is measured only within 10%. The final composite error will never be less than the least accurate measurement attained; one bad measurement in a series can render all other meticulous measurements a futile exercise. The second thing to remember concerning uncertainty propagation in a series is that error is additive when the error is expressed as variance, in the statistical sense. Be aware that variance is often inaccurately estimated: most statisticians state that variances should be calculated from considering a minimum of 20 individual measurements for the same variable, while the student often stops at 10 (or fewer) measurements. (Students ttest allows small numbers of measurements if the variance is known, usually from similar measurements obtained at another time. This statistical test was devised to use the known variance of previous wine tastings to keep wine assessors from over-indulging during current tastings! and thereby lose their taste accuracy!) The formal approach used to establish the following error formulae can be obtained by browsing the www or Page 1 of 2

reading this link. You will find that propagation of uncertainty can become quite complex with co-variance considerations. Here, we are concerned with introducing a few simplified formulae and considerations that can be used in our laboratory setting, and perhaps useful for your capstone projects. Other aspects can be learned during the study of statistics, or as needed in other courses. With reference to the making of a solution of NaCl, we simplify matters by considering our value of concentration (z) to be a function of two independent variables; mass (x) and volume (y), and thus we eliminate covariance: z = f(x,y) ; concentration = mass / volume; c = m / v ; z = x / y The error perturbations about x and y are considered independent, relatively small, and akin to a fluxion described in calculus. Thus, using the tools of differential calculus, we can infer changes in the above function to be expressed as: = + This, along with the definition of variance, 2 2 2 2 + which can be extended for functions of more than two variables by adding similar terms in the portion. = Using this general formula and evaluating for the sum of two variables, z = x + y ,
2 2 = + 2 2 + or expressed for relative error: = ( + )

()2 eventually (with several skipped steps that are described on p61) leads to the general formula =

2 2 + 2 2 = + or expressed for relative error: = ( ) which can result in a very large relative error if x and y are about the same and (x - y) approaches zero. (Thus, if you seek accuracy, never weigh a small mass on a balance while using a heavy tare!)

Using the general formula and evaluating for the difference of two variables, z = z - y ,

Using the general formula and evaluating for the product of two variables, z = xy ,
2 2 = 2 + 2 or expressed for relative error:

Using the general formula and evaluating for the quotient of two variables z = x/y ,

Thus, a 1 % error in mass and a 1 % error in volume will result in a 2 % = 1.4 % error in concentration. To reduce concentration error to 1 % or less with equal errors in mass and volume, errors will need to less than 0.71 %. Note that absolute error for sum and difference are the same. Note that the relative error for product and quotient are the same. Page 2 of 2

2 2 2 2 2 = 2 + or expressed for relative error: = 2 + 2 4

2 2 = 2 + 2