Why Order-Statistic Errors Are Easy to Make
Order statistics are ranks, not new independent observations. For a sample \(X_1,\ldots,X_n\), the ordered values are \(X_{(1)}\leq X_{(2)}\leq\cdots\leq X_{(n)}\). A small change in rank or inequality can change an event from “at most one observation is below the threshold” to “at least one is below it.” Other errors arise when a formula that assumes an iid sample is applied to dependent variables, when the ordered support is omitted, or when repeated values are treated as impossible even though the distribution has atoms.
A dependable solution begins by translating the event into words or counts. Then check the assumptions behind the formula, the possible support, and any combinatorial factor. Those checks are particularly valuable in advanced, multi-step order-statistic problems aligned with the SOA Exam P (Probability) syllabus.
Translate the Rank Before Calculating
For a fixed threshold \(t\), define \(N_t=\sum_{j=1}^n I(X_j\leq t)\), the number of observations at or below \(t\). The rank definition gives the exact equivalences
If the observations are independent and identically distributed with cumulative distribution function \(F\), then each indicator has success probability \(F(t)\), and \(N_t\sim\operatorname{Bin}(n,F(t))\). This binomial result does not follow from identical marginal distributions alone: independence is also needed. For a discrete distribution, the definition uses “at most” and includes any mass at \(t\); replacing \(F(t)\) by a strict-threshold probability can give the wrong count.
Worked Example: Two Rank Conditions and Three Regions
Let \(X_1,\ldots,X_5\) be iid Uniform\((0,10)\). Find \(P(X_{(2)}>4,\ X_{(4)}\leq 8)\). State the method and identify the inequality-direction trap.
Method. Divide the support into three disjoint regions: \(A=\{X_j\leq4\}\), \(B=\{4<X_j\leq8\}\), and \(C=\{X_j>8\}\). Their probabilities are \(0.4,0.4,0.2\), respectively.
The condition \(X_{(2)}>4\) says that at most one observation is in \(A\). The condition \(X_{(4)}\leq8\) says at least four observations are at most 8, or equivalently at most one observation is in \(C\). Therefore the possible counts \((N_A,N_B,N_C)\) are \((0,5,0)\), \((0,4,1)\), \((1,4,0)\), and \((1,3,1)\).
Using the multinomial probabilities for these four disjoint count outcomes,
Check. Each listed count has \(N_A\leq1\) and \(N_C\leq1\); these are exactly the two rank restrictions. The probability is \(0.1894\) to four decimals. A common wrong approach is to read \(X_{(2)}>4\) as “at least two observations exceed 4” and then count the wrong event. The correct interpretation is that at most one observation is at or below 4.
When the Binomial Count Does—and Does Not—Apply
The binomial count is a convenient consequence of iid sampling, not a universal order-statistic rule. If the observations are dependent, the count \(N_t\) need not be binomial. Even if each observation separately has the same continuous distribution, dependence can substantially change probabilities involving ranks. Before applying an iid formula, verify both independence and identical distribution.
Worked Example: Identical Marginals Do Not Guarantee Independence
Let \(U\sim\operatorname{Uniform}(0,1)\), and define \(X_1=U\) and \(X_2=1-U\). Each variable is Uniform\((0,1)\), but they are dependent. Find \(P(X_{(1)}>0.3,\ X_{(2)}<0.8)\).
Method. For two observations, \(X_{(1)}>0.3\) requires both observations to exceed 0.3, and \(X_{(2)}<0.8\) requires both to be below 0.8. Together these conditions say both \(U\) and \(1-U\) lie in \((0.3,0.8)\).
The first variable requires \(0.3<U<0.8\). The second requires \(0.3<1-U<0.8\), which rearranges to \(0.2<U<0.7\). Their intersection is \(0.3<U<0.7\), an interval of length \(0.4\). Hence
Check. The answer is the probability that a Uniform\((0,1)\) variable falls in an interval of length \(0.4\), so it is \(0.4\). If the variables were incorrectly treated as independent, the probability would be \((0.8-0.3)^2=0.25\). The marginals alone do not justify that multiplication; the actual pair lies on the line \(x_1+x_2=1\), not throughout the unit square.
Joint Densities: Support and Combinatorial Factors
For iid observations from a continuous distribution with density \(f\) and cumulative distribution function \(F\), the joint density of two selected order statistics has a combinatorial factor. For \(1\leq r<s\leq n\), at \(y<z\),
This formula assumes iid sampling from a continuous distribution. The density is zero outside the parent distribution’s support and the required order \(y<z\). The factor counts how the remaining observations can be allocated: \(r-1\) below \(y\), \(s-r-1\) between \(y\) and \(z\), and \(n-s\) above \(z\). The factors \(f(y)\) and \(f(z)\) account for observations at the two specified ranks.
The observations at ranks \(r\) and \(s\) contribute \(f(y)\) and \(f(z)\).
The three groups have sizes \(r-1\), \(s-r-1\), and \(n-s\), giving the multinomial coefficient shown in the formula.
Require \(y<z\) and that both values lie in the parent support. Otherwise the density is zero.
The coefficient can also be justified directly. There are \(n!/\{(r-1)!(s-r-1)!(n-s)!\}\) ways to allocate observations to the three remaining groups once the two rank observations are specified. For each allocation, independence gives the same product of group probabilities and densities. Summing over the allocations multiplies that product by the coefficient. Omitting some allocations therefore undercounts the density.
Worked Example: Checking Support and the Joint-Density Coefficient
Let \(X_1,\ldots,X_4\) be iid Uniform\((2,6)\). Find the support and the joint-density value of \(X_{(1)}=y\) and \(X_{(3)}=z\) at \(y=3,z=5\).
Method. The observations are continuous and have density \(f(x)=1/4\) for \(2<x<6\), with \(F(x)=(x-2)/4\) on that interval. Since \(y\) is the minimum and \(z\) is the third order statistic, the support is \(2<y<z<6\). The remaining observations consist of zero below \(y\), one between \(y\) and \(z\), and one above \(z\).
For \(r=1,\ s=3,\ n=4\), the combinatorial coefficient is
At \(y=3\), \(F(y)=1/4\) and \(f(y)=1/4\). At \(z=5\), \(F(z)=3/4\) and \(f(z)=1/4\). Thus \(F(z)-F(y)=1/2\) and \(1-F(z)=1/4\). Substitution gives
Check. The answer is \(0.1875\) per square unit, and the point satisfies \(2<3<5<6\). A common coefficient error is to use 6 rather than 24, omitting assignments among the rank observations and the remaining groups; that gives \(6/128=0.046875\), approximately \(0.0469\), which is one quarter of the correct density. A density value is not itself a probability at a point; probabilities come from integrating the density over a region.
Atoms, Ties, and Strict Inequalities
For a continuous distribution, two independent observations tie with probability zero. For a discrete distribution, ties can have positive probability and are included in the ordered sample with their multiplicities. The continuous joint-density formula is not a substitute for a discrete probability calculation. In particular, if \(X_{(r)}=X_{(s)}\), the event may have positive probability when \(r<s\), even though a continuous joint density is supported only where the selected values are strictly ordered.
Worked Example: A Tie at Two Adjacent Ranks
Let \(X_1,\ldots,X_5\) be independent Bernoulli variables with \(P(X_j=1)=0.3\) and \(P(X_j=0)=0.7\). Find \(P(X_{(2)}=X_{(3)}=1)\).
Method. Since all values are zero or one, the second and third ordered values are both one exactly when there is at most one zero, or equivalently at least four ones. If \(K\) is the number of ones, then \(K\sim\operatorname{Bin}(5,0.3)\). Therefore
Check. The event requires four or five ones, and the two probabilities sum to \(0.03078\). Treating the sample as continuous and declaring an adjacent-rank tie impossible would give zero, which is incorrect for this discrete model. The answer is \(0.0308\) to four decimals.
A Practical Error-Checking Routine
Before accepting an order-statistic calculation, use a short sequence of checks. These checks do not replace a derivation, but they often expose an incorrect rank, assumption, or factor before it affects the final answer.
- Translate the rank. For example, \(X_{(r)}\leq t\) means at least \(r\) observations are at or below \(t\).
- State the sampling model. Identify whether observations are iid, merely identically distributed, or dependent.
- Check endpoints and ties. Use the correct strict or inclusive inequality, especially when the distribution has atoms.
- Write the support. For ordered continuous values, impose their ordering as well as the parent support.
- Account for assignments. In a joint-density calculation, verify that every allocation of the remaining observations is counted.
- Check the result. Confirm that probabilities lie in \([0,1]\), that the event’s count interpretation matches the answer, and that a density has the correct units.
The central habit is to describe the event before selecting a formula. Rank definitions determine the event; sampling assumptions determine which probability model applies; and support and combinatorial checks determine whether the calculation has counted exactly the outcomes it should.
Check Your Understanding
For each problem, translate the ranks into counts before calculating.
- Let \(X_1,\ldots,X_6\) be iid Uniform\((0,1)\). Find \(P(X_{(3)}>0.5)\). (Answer: \(0.3438\))
- Let \(X_1,\ldots,X_5\) be iid continuous observations with \(F(a)=0.2\) and \(F(b)=0.7\), where \(a<b\). Find \(P(X_{(2)}\leq a,\ X_{(4)}>b)\). (Answer: \(0.0720\))
- Let \(X_1,\ldots,X_4\) be independent Bernoulli variables with success probability \(0.4\). Find \(P(X_{(2)}=X_{(3)}=1)\). (Answer: \(0.1792\))
- Let \(U\sim\operatorname{Uniform}(0,1)\), \(X_1=U\), and \(X_2=1-U\). Find \(P(X_{(1)}>0.2,\ X_{(2)}<0.9)\). (Answer: \(0.6000\))
- For four iid Uniform\((2,6)\) observations, state the support and find \(f_{X_{(1)},X_{(3)}}(3,5)\). (Answer: \(2<y<z<6\); density \(0.1875\) per square unit)