Linear combinations and unbiased estimates
In a minute
Scaling, shifting and adding changes their and spread in predictable ways. For a , always add. For independent and , , even for a difference. From a , and are of the mean and .
You'll be able to
- Find the mean and variance of from those of , and say what a change of scale or a fixed amount does to each.
- Find the expected value of a linear combination of several random variables, such as a total or a difference.
- Find the variance of a linear combination of independent random variables, and tell a total of separate items from times one item.
- Use as an unbiased estimate of and as an unbiased estimate of , from raw data, a frequency table or a given .
- Explain what unbiased means; seeing why helps understanding but is not examined.
Before you start3 quick questions on what this builds on
Revise first: 1.2 Arithmetic sequences and series, 4.1 Sampling, data quality and outliers, 4.3 Central tendency and dispersion, 4.6 Probability diagrams, rules and independence, 4.7 Discrete random variables and expected value and 4.8 Binomial distribution
Scaling and shifting one random variable
In 4.7 you found the expected value of a random variable. Often you need a new variable built from it, such as a count turned into money. If you know the mean and spread of , you can write down those of straight away.
In an exam is given, or comes from a known distribution, such as for . You won't work it out from the definition.
Multiplying by stretches every distance from the mean by , so every squared distance by . Adding slides every value along together: the mean moves, the spread does not.
Both rules are in the formula booklet, for any constants and . The square root of the second gives the standard deviation, which is never negative.
Try it first, then reveal one step at a time.
- Match to : and .
- Mean: .
- The variance of is the standard deviation squared: .
- Variance: . The plays no part.
- Check: the standard deviation of is , as says.
Try it first, then reveal one step at a time.
- Write the score in terms of . There are wrong answers, so .
- From the binomial distribution: and .
- Mean: .
- Variance: , so the standard deviation is
- Comment: in the long run, pure guessing still averages marks, so this penalty does not remove the benefit. The large standard deviation means any one attempt could score well below or above .
Expected value of a linear combination
Averages behave as you would hope: each mean is scaled by its coefficient, and the means add.
This is in the formula booklet. It holds for any random variables, independent or not. A constant added to the combination is added to its mean, as with .
Try it first, then reveal one step at a time.
- The total is , with every coefficient .
- Add the means: .
- A mean needs no independence, so this holds even if the swimmers affect each other.
Try it first, then reveal one step at a time.
- Define the profit: .
- Use the rule: .
- Substitute: .
- Comment: dollars is a long-run average over many days. It says nothing about the spread, so the profit on a given day may be far from dollars.
Variance of a linear combination
Spread is different. Each independent variable brings its own uncertainty, so the variances add, even for a difference: the gap between two uncertain times is more uncertain than either time.
Independent means knowing one variable's value tells you nothing about the others, as for . In a question, it is an assumption to state.
The booklet gives this for independent variables. Add variances, never standard deviations: find the standard deviation last, as a square root.
Try it first, then reveal one step at a time.
- Mean: minutes.
- Square each standard deviation to get a variance: and .
- and are independent, so the variances add: .
- Standard deviation: minutes. Adding the standard deviations would give , which is too big.
Try it first, then reveal one step at a time.
- Mean: .
- Variance: , so .
- So the standard deviation of is ml.
- Interpret: on average ml of space is left, give or take about ml. The cup overflows when , which is standard deviations below the mean. If the amounts are roughly normal, that is very rare.
If and are also normally distributed, then so is , and you can find on your GDC. That step is in 4.15.
A total of n items or n times one item
A box of eggs holds different eggs, so its total mass is . That is not , one egg's mass multiplied by . Both have the same mean but different variances.
Compare . Separate items vary independently, so heavy and light ones partly balance out; one item scaled by has nothing to balance it.
Try it first, then reveal one step at a time.
- Let , with each an egg's mass.
- The eggs are independent, so .
- Standard deviation: g.
- Interpret: a box's total mass has mean g and typically lies within about g of it. Heavy and light eggs partly cancel, so , with standard deviation g, would overstate the spread.
Try it first, then reveal one step at a time.
- Define , where is the time at stop .
- Mean: minutes.
- Variance: .
- Standard deviation: minutes.
- Writing instead would give and minutes, which loses the marks.
The of a random , with independent , is a linear combination too: . The rules give and . You will meet again in 4.15.
Unbiased estimates of the mean and variance
In practice you rarely know a 's mean or variance . You take a and estimate them. A different sample would give a different estimate.
The sample mean is an unbiased estimate of . The last section showed why: . That demonstration helps understanding but is not examined.
In 4.3 you divided by ; call that . Distances are measured from , the centre of the same sample, so they come out a little small. On average, underestimates .
Dividing by instead makes the estimate slightly larger, which exactly removes the bias. Showing that is not examined.
The first form is in the booklet: use it when you are given . In the second, the distinct values have frequencies ; for raw data every .
Your GDC gives both. On the TI-84, Sx is and x is . The TI-Nspire shows sx := sn-1x and σx := σnx, and the Casio shows sx and σx. Square Sx for the unbiased estimate of the variance.
Try it first, then reveal one step at a time.
- Enter the values and the frequencies , and run one-variable statistics with the frequency list.
- Check the GDC's is , the total frequency.
- Unbiased estimate of : .
- Read Sx and square the stored value:
- By hand, , and , which agrees.
Try it first, then reveal one step at a time.
- The reported value is , so .
- Use the booklet formula: .
- Units: a variance of lengths in cm is in cm².
- Comment: the factor raises the estimate by about . For larger samples is closer to , so the correction matters less.
Common mistakes
- Adding standard deviations. Independent variables with standard deviations and have a sum with standard deviation , not .
- Subtracting variances for a difference. For independent and , .
- Writing for the total of separate independent items. Their total has variance ; wrongly gives .
- Forgetting to square the coefficient, or letting the constant change the variance. : the drops out.
- Giving (the GDC's x), or a standard deviation, when an unbiased estimate of the variance is asked for. Square Sx.
Exam tips
- Find: define the combination first, such as , then write the variance line, such as . That line earns the method mark even if the arithmetic slips.
- State an assumption: for the variance of a sum or difference, the assumption is independence, said in context, such as one stop's time not affecting another's.
- Write down an unbiased estimate of the population variance: give , the square of the GDC's Sx, to 3 s.f. Don't give the standard deviation when the variance is asked for.
Key takeaways
Practise this
13 past-paper questions on Linear combinations of random variables in the AI HL bank, each with a mark scheme and a worked explanation.