Markov and Chebyshev inequalities
A ropeway operator may know only the average wait for a cabin, or only the mean and standard deviation of a day’s riders. How large can the probability of a far-off value be, given nothing more than those numbers?
Where you are 4 in Limit theorems
Why this exists
A probability needs a whole distribution, but the information at hand is often one or two numbers: an average from a logbook, a standard deviation from a specification sheet. Expectation, linearity and indicators and Variance and covariance compute those numbers from a distribution. This lesson runs the other way, from the numbers back to probabilities. The answers are bounds rather than values, and each holds for every distribution with the given mean or variance. That uniformity is what makes them tools for proof: the The law of large numbers follows from Chebyshev’s inequality in one line.
Definition
A tail probability of a random variable X is a probability of the form P(X \ge a) or P(X \le a): the chance that X lands at or beyond a level a. If X has a finite mean \mu and k > 0, the two-sided tail P(|X - \mu| \ge k) is the chance that X lands at least k from its mean, on either side.
A tail bound from given information (the mean, or the mean and the variance) is a number that P(X \ge a), or P(|X - \mu| \ge k), cannot exceed for any random variable with that information. A bound is sharp if some variable with that information attains it, so that no smaller number is a bound.
In plain words, the two inequalities of this lesson say: a nonnegative quantity with a small mean is rarely large (Markov), and a quantity with a small variance is rarely far from its mean (Chebyshev).
Intuition
Markov’s inequality is a budget. A nonnegative X with mean 2.5 that put probability p on values of at least 10 would spend at least 10p of its mean there, and the rest of X cannot pay any back, since it is never below 0. So 10p \le 2.5. Only the most wasteful arrangement reaches the bound: all the probability at 0 or at exactly 10 (Result 5).
Chebyshev’s inequality is Markov’s applied to the squared distance (X - \mu)^2, whose mean is the variance. Squaring makes far values expensive, so the bound falls like \displaystyle \frac{1}{k^2} rather than \displaystyle \frac{1}{k}. In units of the Standard deviation (the count a z-score gives), at most \displaystyle \frac14 of any distribution lies two or more standard deviations from its mean, and at most \displaystyle \frac19 three or more.
Figure 1 sets both bounds against one exact tail. Neither is close: that is the price of assuming nothing about the shape. Figure 2 shows a shape that pays the price in full.
Results
Result 1 (Markov’s inequality). Let P(X \ge 0) = 1 and let E(X) be finite. For every a > 0,
Result 2 (Markov through a function). Let g be nonnegative and nondecreasing on an interval I with P(X \in I) = 1, let a \in I with g(a) > 0, and let E(g(X)) be finite. Then
In particular, if r > 0 and E(|X|^r) is finite, then \displaystyle P(|X| \ge a) \le \frac{E(|X|^r)}{a^r} for every a > 0.
Result 3 (Chebyshev’s inequality). Let X have finite mean \mu and finite variance \sigma^2. For every k > 0,
If \sigma > 0, then for every t > 0, \displaystyle P(|X - \mu| \ge t\sigma) \le \frac{1}{t^2} and \displaystyle P(|X - \mu| < t\sigma) \ge 1 - \frac{1}{t^2}.
Result 4 (one-sided bound, Cantelli). Under the hypotheses of Result 3, for every k > 0,
When \sigma > 0 this is smaller than the bound \displaystyle \frac{\sigma^2}{k^2} that Result 3 gives for one side.
Result 5 (the bounds are sharp). (a) Let P(X \ge 0) = 1, E(X) finite and a > 0. Then \displaystyle P(X \ge a) = \frac{E(X)}{a} exactly when P(X \in \{0, a\}) = 1. (b) Let \sigma > 0 and t \ge 1. The variable equal to \mu - t\sigma, \mu and \mu + t\sigma with probabilities \displaystyle \frac{1}{2t^2}, \displaystyle 1 - \frac{1}{t^2} and \displaystyle \frac{1}{2t^2} has mean \mu, standard deviation \sigma and \displaystyle P(|X - \mu| \ge t\sigma) = \frac{1}{t^2}. For 0 < t \le 1 the bound \displaystyle \frac{1}{t^2} \ge 1 says nothing, and X = \mu \pm \sigma, each with probability \displaystyle \frac12, has P(|X - \mu| \ge t\sigma) = 1. So no bound built from \mu and \sigma alone beats Chebyshev’s.
Result 6 (averages). Let X_1, \dots, X_n each have mean \mu and variance \sigma^2, and be pairwise uncorrelated (independent is enough). Their average \displaystyle \bar{X}_n = \frac{1}{n}(X_1 + \dots + X_n) satisfies, for every \varepsilon > 0,
Why it is true
Two facts are used throughout: E(\mathbf{1}_A) = P(A), and monotonicity, E(U) \le E(V) whenever U \le V and both means are finite, proved at the start of Why it is true in Variance and covariance. Changing a variable on an event of probability 0 changes no mean and no probability, so a hypothesis such as P(X \ge 0) = 1 may be read as X \ge 0 at every outcome.
Result 1. Let A = \{X \ge a\}. At every outcome a \mathbf{1}_A \le X: on A this says a \le X, and off A it says 0 \le X. Monotonicity gives a P(A) \le E(X), and dividing by a > 0 finishes.
Result 2. If X \ge a, then g(X) \ge g(a) because g is nondecreasing on I. So \{X \ge a\} \subseteq \{g(X) \ge g(a)\}, and Result 1 applied to the nonnegative variable g(X) at the level g(a) > 0 bounds the larger event. For the moment form, apply this to |X| with g(x) = x^r on I = [0, \infty).
Result 3. The events \{|X - \mu| \ge k\} and \{(X - \mu)^2 \ge k^2\} are the same. Result 1 applied to (X - \mu)^2, which is nonnegative with mean \sigma^2, at the level k^2 gives \displaystyle \frac{\sigma^2}{k^2}. Put k = t\sigma for the second form, and take complements for the third.
Result 4. Let Y = X - \mu, so E(Y) = 0 and E(Y^2) = \sigma^2. For any u \ge 0, if Y \ge k then Y + u \ge k + u > 0, so (Y + u)^2 \ge (k + u)^2. Result 1 applied to (Y + u)^2, with E\big((Y + u)^2\big) = \sigma^2 + 2u E(Y) + u^2 = \sigma^2 + u^2, gives
The right side is smallest at \displaystyle u = \frac{\sigma^2}{k} (its derivative in u has the sign of uk - \sigma^2). There the numerator is \displaystyle \frac{\sigma^2(k^2 + \sigma^2)}{k^2} and the denominator is \displaystyle \frac{(k^2 + \sigma^2)^2}{k^2}, so the bound is \displaystyle \frac{\sigma^2}{\sigma^2 + k^2}. The lower tail is the upper tail of -X, which has mean -\mu and the same variance. Finally \sigma^2 + k^2 > k^2.
Result 5. First a lemma: if Z \ge 0 and E(Z) = 0, then P(Z = 0) = 1. By Result 1, \displaystyle P(Z \ge \frac1m) \le m \cdot 0 = 0 for each m; the events \displaystyle \{Z \ge \frac1m\} grow with m to \{Z > 0\}, so continuity (Result 8 of Probability axioms and sample spaces) gives P(Z > 0) = 0.
(a) With A = \{X \ge a\}, linearity gives
a sum of means of two nonnegative variables. It is 0 exactly when both means are 0, so by the lemma exactly when, with probability 1, X = 0 wherever X < a and X = a wherever X \ge a. That is P(X \in \{0, a\}) = 1. Conversely, if P(X \in \{0, a\}) = 1, then E(X) = aP(X = a) = aP(A).
(b) Let \displaystyle p = \frac{1}{2t^2}; as t \ge 1, \displaystyle p \le \frac12, so the three probabilities are nonnegative and sum to 1. The values are symmetric about \mu, so the mean is \mu, and the variance is 2p(t\sigma)^2 = \sigma^2. The event |X - \mu| \ge t\sigma is \{X \ne \mu\}, of probability \displaystyle 2p = \frac{1}{t^2}. For t \le 1, X = \mu \pm \sigma has mean \mu, variance \sigma^2, and |X - \mu| = \sigma \ge t\sigma always.
Result 6. By linearity E(\bar{X}_n) = \mu. All covariances are 0, so Result 4 of Variance and covariance gives \operatorname{Var}(X_1 + \dots + X_n) = n\sigma^2, and Result 2 there gives \displaystyle \operatorname{Var}(\bar{X}_n) = \frac{\sigma^2}{n}. Result 3 with k = \varepsilon finishes. Independent variables with finite variances are uncorrelated by Result 5 of Variance and covariance.
Worked examples
Waiting for a cabin
Isabel’s logbook gives only the mean wait for a cabin at the valley station of her ropeway: 2.5 minutes. Let W \ge 0 be the wait. Bound P(W \ge 10) and P(W \ge 2), show that the first bound cannot be improved from the mean alone, and compare it with the exact value if W \sim \operatorname{Exponential}(0.4).
SolutionW is nonnegative with a finite mean, so Result 1 applies. With a = 10, \displaystyle P(W \ge 10) \le \frac{2.5}{10} = 0.25. With a = 2 it gives \displaystyle \frac{2.5}{2} = 1.25, which every probability already satisfies: Markov’s bound says something only for a > E(W).
The variable equal to 10 with probability 0.25 and to 0 otherwise has mean 2.5 and P(W \ge 10) = 0.25, so by Result 5 no smaller bound holds for every wait with mean 2.5.
\operatorname{Exponential}(0.4) has mean \displaystyle \frac{1}{0.4} = 2.5, so it is one of the waits covered, and its tail is P(W \ge 10) = e^{-0.4 \cdot 10} = e^{-4} \approx 0.0183, less than a thirteenth of the bound.
Wind at the top station
Luis records the wind speed V, in meters per second, at the top station. It has density \displaystyle f_V(v) = \frac{v}{16} e^{-v/4} for v \ge 0, and 0 otherwise, and the ropeway closes when V \ge 20. Find E(V) and \operatorname{Var}(V), bound P(V \ge 20) by Results 1 and 3, compute it exactly, and find the levels a > 8 at which Chebyshev’s bound on P(V \ge a) is the smaller.
SolutionBy n integrations by parts, \int_0^\infty v^n e^{-v/4} \, dv = n! \, 4^{n + 1}. So
and \operatorname{Var}(V) = 96 - 64 = 32 by Result 1 of Variance and covariance.
Markov: \displaystyle P(V \ge 20) \le \frac{8}{20} = 0.4. Chebyshev: V \ge 20 forces |V - 8| \ge 12, so \displaystyle P(V \ge 20) \le \frac{32}{144} = \frac29 \approx 0.222. Exactly, since \displaystyle -\left(\frac{v}{4} + 1\right) e^{-v/4} has derivative \displaystyle \frac{v}{16} e^{-v/4},
For a > 8 the same steps give the bounds \displaystyle \frac{8}{a} and \displaystyle \frac{32}{(a - 8)^2}. The second is smaller exactly when 4a < (a - 8)^2, that is a^2 - 20a + 64 > 0, or (a - 4)(a - 16) > 0, which for a > 8 means a > 16. Near the mean, the mean alone is the better guide; far out, the variance wins (Figure 1).
Daily riders
Season records give the number X of riders on a day only as mean 2400 and standard deviation 150. Find an interval that holds X with probability at least 0.9. Bound P(X \ge 2850) by Results 3 and 4, and show that the smaller bound cannot be improved from these two numbers.
SolutionBy Result 3, \displaystyle P(|X - 2400| < t\sigma) \ge 1 - \frac{1}{t^2}, which is 0.9 when \displaystyle t = \sqrt{10}. The half width is \displaystyle 150\sqrt{10} \approx 474.3, so with probability at least 0.9, 1925.7 < X < 2874.3: the count is one of 1926, \dots, 2874.
The level 2850 is k = 450 = 3\sigma above the mean. Since X \ge 2850 forces |X - 2400| \ge 450, Result 3 gives \displaystyle P(X \ge 2850) \le \frac19 \approx 0.111. Result 4 gives
Let X be 2850 with probability 0.1 and 2350 with probability 0.9. Its mean is 285 + 2115 = 2400, its variance is 0.1(450)^2 + 0.9(50)^2, which is 20{,}250 + 2250 = 22{,}500 = 150^2, and P(X \ge 2850) = 0.1. So 0.1 is sharp: no smaller bound follows from the mean and standard deviation alone.
Round-trip tickets
Each of the next 192 riders, independently, buys a round-trip ticket with probability 0.75. Bound the probability that the number X of round trips is 12 or more away from its mean, and compare with the exact value.
SolutionX \sim \operatorname{Binomial}(192, 0.75), so by Result 7 of Variance and covariance its mean is 192(0.75) = 144 and its variance is 192(0.75)(0.25) = 36, with \sigma = 6. The distance 12 is 2\sigma, so Result 3 gives \displaystyle P(|X - 144| \ge 12) \le \frac{36}{144} = 0.25; equivalently, P(132 < X < 156) \ge 0.75.
Exactly, summing the binomial probabilities \binom{192}{j} 0.75^j \, 0.25^{192 - j},
The bound is more than four times too large. It had to allow for every distribution with variance 36, including the three-point one of Result 5; the central limit theorem will explain why the binomial’s own tail sits close to a normal one instead.
How many trips to time
Isabel wants the mean \mu of the trip time from valley to summit. Trip times are independent with standard deviation 8 seconds. How many trips n must she time so that Result 6 guarantees P(|\bar{T}_n - \mu| \ge 2) \le 0.05?
SolutionBy Result 6 with \sigma^2 = 64 and \varepsilon = 2,
and \displaystyle \frac{16}{n} \le 0.05 exactly when n \ge 320. So 320 trips suffice, whatever the shape of the trip-time distribution; the proof used only that the times are uncorrelated. The guarantee is generous: the central limit theorem will suggest that about 62 trips give the same assurance approximately.
The median and the mean
A median of X is a number m with \displaystyle P(X \le m) \ge \frac12 and \displaystyle P(X \ge m) \ge \frac12, as for data in Median and mode. Prove that if X has finite variance, then every median satisfies |m - \mu| \le \sigma. Then find |m - \mu| for X \sim \operatorname{Exponential}(\lambda).
SolutionIf m = \mu there is nothing to prove. If m > \mu, let k = m - \mu > 0. By the definition of a median and Result 4,
So \sigma^2 + k^2 \le 2\sigma^2, that is k^2 \le \sigma^2 and k \le \sigma. If m < \mu, let k = \mu - m and use P(X \le m) = P(X - \mu \le -k) with the lower half of Result 4. Result 3 alone would give only \displaystyle \frac12 \le \frac{\sigma^2}{k^2}, that is \displaystyle k \le \sqrt{2}\,\sigma.
For \operatorname{Exponential}(\lambda), \displaystyle P(X \ge m) = e^{-\lambda m} = \frac12 gives \displaystyle m = \frac{\ln 2}{\lambda}, and \displaystyle \mu = \sigma = \frac{1}{\lambda}, so |m - \mu| = (1 - \ln 2)\sigma \approx 0.307\sigma.
Boundaries
Markov needs X \ge 0. If X = \pm 1, each with probability \displaystyle \frac12, then E(X) = 0 but \displaystyle P(X \ge 1) = \frac12. A variable with X \ge c can be shifted: apply Result 1 to X - c.
Chebyshev needs a finite variance. For the density \displaystyle \frac{2}{x^3} on [1, \infty) in the Boundaries of Variance and covariance, the variance is infinite and Result 3 says nothing; the mean is 2, so Markov still gives \displaystyle P(X \ge a) \le \frac{2}{a}, while the true tail is \displaystyle \frac{1}{a^2}.
The bounds cover every distribution at once, so they are loose for any one family. A normal variable has P(|X - \mu| \ge 2\sigma) \approx 0.046 against Chebyshev’s 0.25 (The normal distribution), and an exponential tail falls like e^{-\lambda a} against \displaystyle \frac{1}{a} and \displaystyle \frac{1}{a^2}. Result 5 shows the looseness is not a flaw of the proofs: with only \mu and \sigma known, nothing better holds.
Small t. For t \le 1 Chebyshev’s bound is at least 1, and Result 5 shows that the probability can indeed be 1.
Where this goes next
Letting n \to \infty in Result 6 gives the The law of large numbers: the chance that an average misses its mean by \varepsilon or more goes to 0 (the school lesson The law of large numbers and the gambler's fallacy states it in words). The The central limit theorem, proved replaces Chebyshev’s bound for an average with an approximation. Result 2 with g(x) = e^{sx}, s > 0, gives P(X \ge a) \le e^{-sa} E(e^{sX}), and Moment generating functions choose s to make that bound fall exponentially in a.
Exercises
Bounds from a mean or a variance
- The number Q of people in line at Luis’s ticket window has mean 12. Bound P(Q \ge 30) and P(Q \ge 60) by Markov’s inequality, and find the smallest level q for which Markov’s inequality gives P(Q \ge q) \le 0.05.
- The ride time T of a cabin has mean 9 minutes and standard deviation 0.5 minutes. Bound P(|T - 9| \ge 1.5), and find an interval centered at 9 that holds T with probability at least 0.96.
- The number N of riders on a four-seat chair has P(N = 0) = 0.1, P(N = 1) = 0.2, P(N = 2) = 0.3, P(N = 3) = 0.2 and P(N = 4) = 0.2. Find E(N) and \operatorname{Var}(N). Bound P(N = 4) by Result 1 and by Result 4, bound P(|N - 2.2| \ge 1.8) by Result 3, and compare all three with the exact values.
- A stalled cabin stops at a point X, in meters, uniform along a 1200 meter span, X \sim \operatorname{Uniform}(0, 1200). Find \sigma_X, bound P(|X - 600| \ge 500) by Result 3, and find it exactly.
- The time T, in minutes, between strong gusts is \operatorname{Exponential}(0.2). Bound P(T \ge 15) by Result 1 and by Result 4, and find it exactly.
- The number S of unplanned stops of Isabel’s ropeway in a week is \operatorname{Poisson}(4). Bound the probability of 10 or more stops by Results 1, 3 and 4, and find it exactly.
- The breaking load L of a haul rope has mean 900 kilonewtons and standard deviation 30 kilonewtons. Give a lower bound for P(860 < L < 940).
- X \ge 0 has E(X) = 3 and E(X^2) = 13. Bound P(X \ge 9) by Result 1, by Result 2 with r = 2, by Result 3 and by Result 4. Which is smallest?
Averages and sample sizes
- Luis estimates the proportion p of riders carrying skis by the proportion among n riders, each independently carrying skis with probability p. How large must n be for Result 6 to guarantee an error of less than 0.04 with probability at least 0.9, whatever p is?
- Trip times are independent with standard deviation 6 seconds. Bound the probability that the average of 50 of them is 1.5 seconds or more from the mean, and find how many trips make that bound 0.04 or less.
- A wind gauge’s 16 readings have pairwise uncorrelated errors, each with mean 0 and standard deviation 0.3 meters per second. Find the smallest c for which Result 6 guarantees that the average error is less than c in absolute value with probability at least 0.9.
Proofs
- Let \sigma > 0 and k > 0. Prove that the variable equal to \mu + k with probability \displaystyle \frac{\sigma^2}{\sigma^2 + k^2} and to \displaystyle \mu - \frac{\sigma^2}{k} otherwise has mean \mu, variance \sigma^2 and \displaystyle P(X - \mu \ge k) = \frac{\sigma^2}{\sigma^2 + k^2}, so Result 4 is sharp.
- Let N have mean \mu > 0 and finite variance. Prove that \displaystyle P(N \le 0) \le \frac{\operatorname{Var}(N)}{\mu^2}. For a count, this bounds P(N = 0).
- Let X \ge 0 with 0 < E(X^2) < \infty. Prove that \displaystyle P(X > 0) \ge \frac{E(X)^2}{E(X^2)}.
- Let P(X \le b) = 1, let E(X) be finite, and let a < b. Prove that \displaystyle P(X > a) \ge \frac{E(X) - a}{b - a}.
- Refute the claim that every nonnegative random variable with mean 1 has \displaystyle P(X \ge 1) \ge \frac12.
- Let X_1, \dots, X_n be pairwise uncorrelated with means \mu_1, \dots, \mu_n and variances at most C. Let \displaystyle \bar{\mu} = \frac1n(\mu_1 + \dots + \mu_n). Prove that \displaystyle P(|\bar{X}_n - \bar{\mu}| \ge \varepsilon) \le \frac{C}{n\varepsilon^2} for every \varepsilon > 0.
- Let X have finite variance \sigma^2, and let \displaystyle P(|X - \mu| \ge k) = \frac{\sigma^2}{k^2} for some k > 0. Prove that P(X \in \{\mu - k, \mu, \mu + k\}) = 1.
Answers
- \displaystyle P(Q \ge 30) \le \frac{12}{30} = 0.4 and \displaystyle P(Q \ge 60) \le \frac{12}{60} = 0.2. The bound \displaystyle \frac{12}{q} \le 0.05 holds exactly when q \ge 240.
- 1.5 = 3\sigma, so \displaystyle P(|T - 9| \ge 1.5) \le \frac{0.25}{2.25} = \frac19 \approx 0.111. For \displaystyle 1 - \frac{1}{t^2} = 0.96 take t = 5, a half width of 2.5: P(6.5 < T < 11.5) \ge 0.96.
- E(N) = 0.2 + 0.6 + 0.6 + 0.8 = 2.2 and E(N^2) = 0.2 + 1.2 + 1.8 + 3.2 = 6.4, so \operatorname{Var}(N) = 6.4 - 4.84 = 1.56. Result 1: \displaystyle P(N \ge 4) \le \frac{2.2}{4} = 0.55. Result 4 with k = 1.8: \displaystyle \frac{1.56}{1.56 + 3.24} = 0.325. Result 3: \displaystyle \frac{1.56}{3.24} = \frac{13}{27} \approx 0.481. Exactly, P(N = 4) = 0.2, and |N - 2.2| \ge 1.8 means N = 0 or N = 4, probability 0.3.
- \displaystyle \operatorname{Var}(X) = \frac{1200^2}{12} = 120{,}000, so \sigma_X \approx 346.4 meters. Result 3: \displaystyle \frac{120{,}000}{250{,}000} = 0.48. Exactly, X < 100 or X > 1100, probability \displaystyle \frac{200}{1200} = \frac16 \approx 0.167.
- E(T) = 5 and \operatorname{Var}(T) = 25. Result 1: \displaystyle \frac{5}{15} = \frac13. Result 4 with k = 10: \displaystyle \frac{25}{25 + 100} = 0.2. Exactly, P(T \ge 15) = e^{-3} \approx 0.0498.
- Mean and variance are both 4. Result 1: \displaystyle \frac{4}{10} = 0.4. Result 3: 10 or more stops forces |S - 4| \ge 6, so the bound is \displaystyle \frac{4}{36} = \frac19 \approx 0.111. Result 4: \displaystyle \frac{4}{4 + 36} = 0.1. Exactly, \displaystyle 1 - \sum_{j=0}^{9} e^{-4} \frac{4^j}{j!} \approx 0.00813.
- With k = 40, Result 3 gives \displaystyle P(|L - 900| \ge 40) \le \frac{900}{1600} = 0.5625, so P(860 < L < 940) \ge 0.4375.
- Result 1: \displaystyle \frac39 = \frac13. Result 2: \displaystyle \frac{13}{81} \approx 0.160. \operatorname{Var}(X) = 13 - 9 = 4, and X \ge 9 forces |X - 3| \ge 6, so Result 3 gives \displaystyle \frac{4}{36} = \frac19 \approx 0.111. Result 4 gives \displaystyle \frac{4}{4 + 36} = 0.1, the smallest.
- Each rider’s indicator has variance \displaystyle p(1 - p) \le \frac14, since \displaystyle \frac14 - p(1 - p) = \left(p - \frac12\right)^2 \ge 0. Result 6 gives the bound \displaystyle \frac{p(1 - p)}{n(0.04)^2} \le \frac{1}{0.0064n}, which is at most 0.1 when n \ge 1562.5. So n = 1563 riders.
- \displaystyle \frac{36}{50(1.5)^2} = \frac{36}{112.5} = 0.32. The bound is \displaystyle \frac{36}{2.25n} = \frac{16}{n}, at most 0.04 when n \ge 400.
- The bound is \displaystyle \frac{0.09}{16c^2}, at most 0.1 when c^2 \ge 0.05625, so \displaystyle c = \sqrt{0.05625} \approx 0.237 meters per second.
- Let \displaystyle q = \frac{\sigma^2}{\sigma^2 + k^2}, so \displaystyle 1 - q = \frac{k^2}{\sigma^2 + k^2}. The deviations X - \mu are k and \displaystyle -\frac{\sigma^2}{k}, with mean \displaystyle kq - \frac{\sigma^2}{k}(1 - q) = \frac{k\sigma^2 - \sigma^2 k}{\sigma^2 + k^2} = 0, so E(X) = \mu. The variance is \displaystyle k^2 q + \frac{\sigma^4}{k^2}(1 - q) = \frac{k^2\sigma^2 + \sigma^4}{\sigma^2 + k^2} = \sigma^2. Since \displaystyle -\frac{\sigma^2}{k} < k, the event X - \mu \ge k is \{X = \mu + k\}, of probability q.
- If N \le 0, then \mu - N \ge \mu > 0, so |N - \mu| \ge \mu. Hence \displaystyle P(N \le 0) \le P(|N - \mu| \ge \mu) \le \frac{\operatorname{Var}(N)}{\mu^2} by Result 3 with k = \mu.
- E(X) is finite since X \le 1 + X^2. Let A = \{X > 0\}; as X \ge 0, X\mathbf{1}_A = X. If P(A) = 0, then X = 0 with probability 1 and E(X^2) = 0, which is excluded; so P(A) > 0. For every real s, (X - s\mathbf{1}_A)^2 \ge 0, and expanding with X\mathbf{1}_A = X and \mathbf{1}_A^2 = \mathbf{1}_A gives 0 \le E(X^2) - 2sE(X) + s^2 P(A). With \displaystyle s = \frac{E(X)}{P(A)} this reads \displaystyle 0 \le E(X^2) - \frac{E(X)^2}{P(A)}, which rearranges to the claim.
- Y = b - X satisfies P(Y \ge 0) = 1 and E(Y) = b - E(X). Since X \le a exactly when Y \ge b - a > 0, Result 1 gives \displaystyle P(X \le a) \le \frac{b - E(X)}{b - a}. Taking complements, \displaystyle P(X > a) \ge 1 - \frac{b - E(X)}{b - a} = \frac{E(X) - a}{b - a}.
- Let X be 10 with probability 0.1 and 0 otherwise. Then X \ge 0 and E(X) = 1, but \displaystyle P(X \ge 1) = 0.1 < \frac12.
- By linearity E(\bar{X}_n) = \bar{\mu}. All covariances are 0, so Results 4 and 2 of Variance and covariance give \displaystyle \operatorname{Var}(\bar{X}_n) = \frac{1}{n^2}\sum_{i=1}^{n} \operatorname{Var}(X_i) \le \frac{nC}{n^2} = \frac{C}{n}. Result 3 with k = \varepsilon gives \displaystyle P(|\bar{X}_n - \bar{\mu}| \ge \varepsilon) \le \frac{\operatorname{Var}(\bar{X}_n)}{\varepsilon^2} \le \frac{C}{n\varepsilon^2}.
- Let Y = (X - \mu)^2, so Y \ge 0 and E(Y) = \sigma^2. The hypothesis says \displaystyle P(Y \ge k^2) = \frac{E(Y)}{k^2}, equality in Markov’s inequality at the level k^2. By Result 5(a), P(Y \in \{0, k^2\}) = 1, that is, X - \mu is 0, k or -k with probability 1.