Gains:
- Ability to distinguish between random and systematic uncertainty and report each measurement result with its uncertainty to the appropriate significant digit.
- Ability to calculate error propagation with the correct rule and independently cross-check the formula derived by artificial intelligence with Monte Carlo
- Ability to decide whether two results are compatible or not, taking into account their uncertainty
No measurement in physics is perfect. There is an uncertainty of ±0.5 mm when measuring length with a ruler, and ±0.1 s when measuring time with a stopwatch. These uncertainties are not a flaw, but the nature of the measurement—and the requirement of scientific integrity is to report each result with its uncertainty. Saying "I found the gravitational acceleration to be 9.7" is incomplete; Saying "I found 9.7 ± 0.3 m/s²" tells you how reliable the result is. In this unit, you will learn how to use artificial intelligence (AI) for uncertainty calculation and error propagation (how measurement uncertainties are carried from those measurements to the calculated result) and how to verify these calculations.
Types of uncertainty and the role of AI
There are two basic types of uncertainty. Random uncertainty is the scattering of results in repeated measurements; Many measurements are taken and reduced by the mean and standard deviation. Systematic uncertainty is an error that always deviates in the same direction (like a miscalibrated balance); It does not decrease with repetition, but is corrected by finding the cause. AI is helpful in explaining these concepts, codifying the calculation of standard deviation and mean, and — especially — deriving complex error propagation formulas. But you must verify every derived formula, because AI often mixes partial derivative signs or square root structure in error propagation formulas.
Status
account
verification
Repeated measurement
Mean ± standard error
Manual standard deviation / √N
Addition/subtraction (z = x ± y)
Uncertainties are summed from the square
Size/unit control
Multiplication/division (z = x·y)
Relative uncertainties are summed from the square
Test with sample number
General function
Propagation by partial derivatives
Cross-checking with Monte Carlo
Basic rules of error propagation
In addition and subtraction, absolute uncertainties are combined by adding their squares: If z = x + y, the uncertainty of z becomes √(δx² + δy²). In multiplication and division, the relative (percentage) uncertainties are summed from their squares: If z = x·y, (δz/z) = √((δx/x)² + (δy/y)²). For more complex functions, the general rule is to use partial derivatives: the effect of each variable on the result is measured by the partial derivative with respect to that variable. AI can derive these formulas, but the safest way to verify the result is a Monte Carlo cross-check: randomly distributing each input to its uncertainty, calculating it thousands of times, and looking at the distribution of the output.
Tip: If you are unsure of the accuracy of an error propagation formula, provide it with Monte Carlo: randomly generate the inputs with their own uncertainty and calculate the function 10,000 times; The standard deviation of the output must match the uncertainty given by the formula. The combination of these two independent methods greatly increases your confidence in the result.
Step by step: an honest uncertainty analysis
1. Determine the uncertainty of each measurement. Instrument resolution, distribution of repeats, or manufacturer data. Write down where the uncertainty comes from.
2. Distinguish between random and systematic. Consider separately what will decrease with repetition and what will not.
3. Calculate the spread with the correct formula. Addition, multiplication, general function — choose the appropriate rule. Have AI derive it, but control it.
4. Cross-check with Monte Carlo. Provide the result in an independent method, especially for complex formulas.
5. Report with appropriate significant digits. Uncertainty is usually rounded to 1-2 significant digits; the result aligns with the last digit of the uncertainty. Write "9.7 ± 0.4 m/s²", not "9.73418 ± 0.4".
three mini cases
Case 1 — Wrong spread formula. A student asked AI about the uncertainty propagation for g = 4π²L/T². AI forgot the coefficient of 2 when deriving the contribution of the uncertainty of T with respect to T (since T² the relative uncertainty should have been 2·δT/T). When the student cross-checked with Monte Carlo, he found that the AI formula estimated the uncertainty by about half and corrected it.
Case 2 — Systematic error concealed. One researcher took a large number of measurements, made the random uncertainty very small, and reported an impressive precision of “±0.01 mm.” But its scale was off-calibre by 0.2 mm — a systematic error. Taking into account this error, which did not decrease with repetition, the true uncertainty was much greater. Lesson: repetition reduces random error, not systematic error.
Case 3 — Monte Carlo saved. A teacher introduced uncertainty in a volume calculus (V = πr²h) and found a result by hand. To be sure, he asked the AI for a Monte Carlo code; He randomly generated r and h with their uncertainties and calculated V 10,000 times. The standard deviation of the distribution agreed with the value found by hand — the analysis was validated.
Four copyable templates
1) Error propagation derivation + verification:
For the expression z = [formula], derive the error propagation from the uncertainties of variables x and y to the uncertainty of z by partial derivatives, STEP BY STEP. Show each partial derivative. Then verify this formula with SympY.Fitting; Mark the step you are not sure about.
2) Monte Carlo cross-check:
Write a Python code that finds the uncertainty of the following calculation using Monte Carlo: z = [formula]. Inputs: x = [value ± uncertainty], y = [value ± uncertainty].Sample each input 10,000 times from a normal distribution, calculate z, print the mean and standard deviation of the output. Compare this with the result of the analytic diffusion formula.
3) Measurement statistics:
Write code that calculates the mean, standard deviation, and standard error of the mean (std/√N) for the following repeated measurement data ([values]). Print the result in the format "value ± standard error (units)" with the appropriate significant digit. Remind us of the distinction between random and systematic uncertainty.
4) Significant digit / reporting check:
Rewrite the results below in the form "value ± uncertainty (unit)", rounding the uncertainty to 1-2 significant digits and aligning the result with the uncertainty. Fix false precision. Results: [here]
Weak prompt / Strong prompt
Weak: "calculate the uncertainty for g."
Result: A number without context, whose formula is not verified, and without a systematic/random distinction.
Strong: "For g = 4π²L/T², derive the uncertainty of g by partial derivatives from measurements of L = 1.00 ± 0.01 m and T = 2.00 ± 0.02 s, then cross-check with Monte Carlo. Write the result in the form 'g = ... ± ... m/s²' with the appropriate significant digit."
Result: A derived, independently verified, correctly reported uncertainty.
Common mistakes
- Not reporting uncertainty at all. A measurement without uncertainty is a scientific half-statement.
- Not verifying the spread formula. YZ frequently mixes partial derivative coefficients and square root structure; Monte Carlo cross-checking is a must.
- Trying to reduce systematic error through repetition. Multi-measurement only reduces random error; calibration error remains.
- Creating false certainty. Writing many more significant digits than uncertainty is a claim of precision that does not exist.
- Collecting uncertainties directly. The independent uncertainties are not linear but squared (into the square root).
Caution: Whether two physical results are "different" can only be decided by looking at their uncertainty. Values 9.7 to 9.81 are compatible if the uncertainty is ±0.3; ±0.02 is contradictory. Ignoring the uncertainty and saying "my result deviates from the theory" is often an erroneous conclusion. Be sure to verify the uncertainty calculation given by the AI before making a decision.
In summary
Uncertainty is integral to physical measurement and every result should be reported with its uncertainty. AI is helpful in coding uncertainty statistics and deriving error propagation formulas; but these formulas contain frequent errors and should be cross-checked by an independent method such as Monte Carlo. Distinguishing random from systematic uncertainty, reporting to the appropriate significant digit, and comparing results to their uncertainty is the basis of honest scientific study. In the next unit we will shift the focus from numerical analysis to symbolic model derivation.
Application task
Choose a calculation that includes at least two measured quantities (e.g. V = πr²h or g = 4π²L/T²). Assign a reasonable uncertainty to each size. Have the AI derive the error propagation with template 1, then have the Monte Carlo cross-check with template 2 and run it. Do the results of the two methods agree? Report the result with the appropriate significant digit. Write down in 5-6 sentences: is there a formula error or mismatch?
checklist
- [ ] I determined the uncertainty and source of each measurement.
- [ ] I distinguished random and systematic uncertainty.
- [ ] I calculated the error propagation with the correct rule.
- [ ] I cross-checked the result independently with Monte Carlo or by hand.
- [ ] I reported the result with the appropriate significant digit, aligned with its uncertainty.
- [ ] I took uncertainty into account when comparing the results.