chapter_01.tex (35527B)
1 \title Probability Theory 2 3 \begin{paragraph}{1.4}{Why Conditional Expectations Are Needed} 4 In elementary probability, the conditional probability of an event \(B\) given an event \(A\) is defined as 5 \[ 6 P(B\mid A)=\frac{P(A\cap B)}{P(A)}, 7 \] 8 provided that \(P(A)>0\). In probability and statistics, however, we sometimes need a notion of conditional probability even for events \(A\) with \(P(A)=0\); for example, \(A=\{Y=c\}\), where \(c\in\mathbb{R}\) and \(Y\) is a random variable having a continuous c.d.f. 9 \end{paragraph} 10 11 \begin{paragraph}{1.4}{Scope of Conditional Probability} 12 General definitions of conditional probability, expectation, and distribution are introduced in this section, and they are shown to agree with those defined in elementary probability in special cases. 13 \end{paragraph} 14 15 \begin{definition}{1.6}{Conditional Expectation and Probability} 16 Let \(X\) be an integrable random variable on \((\Omega,\mathcal{F},P)\). 17 18 \begin{itemize} 19 \item Let \(\mathcal{A}\) be a sub-\(\sigma\)-field of \(\mathcal{F}\). The conditional expectation of \(X\) given \(\mathcal{A}\), denoted by \(E(X\mid\mathcal{A})\), is the a.s.-unique random variable satisfying: 20 \begin{itemize} 21 \item \(E(X\mid\mathcal{A})\) is measurable from \((\Omega,\mathcal{A})\) to \((\mathbb{R},\mathcal{B})\); 22 \item \(\int_A E(X\mid\mathcal{A})\,dP=\int_A X\,dP\) for every \(A\in\mathcal{A}\). 23 \end{itemize} 24 The existence of \(E(X\mid\mathcal{A})\) follows from Theorem~1.4. 25 26 \item Let \(B\in\mathcal{F}\). The conditional probability of \(B\) given \(\mathcal{A}\) is \(P(B\mid\mathcal{A})=E(I_B\mid\mathcal{A})\). 27 28 \item Let \(Y\) be measurable from \((\Omega,\mathcal{F},P)\) to \((\Lambda,\mathcal{G})\). The conditional expectation of \(X\) given \(Y\) is 29 \[ 30 E(X\mid Y)=E[X\mid\sigma(Y)]. 31 \] 32 \end{itemize} 33 \end{definition} 34 35 \begin{paragraph}{1.4.1}{Information in a Random Element} 36 Essentially, the \(\sigma\)-field \(\sigma(Y)\) contains ``the information in \(Y\).'' Hence, \(E(X\mid Y)\) is the expectation of \(X\) given the information provided by \(\sigma(Y)\). 37 \end{paragraph} 38 39 \begin{paragraph}{1.4.1}{Representing Conditional Expectations} 40 The following useful result shows that there is a Borel function \(h\), defined on the range of \(Y\), such that \(E(X\mid Y)=h\circ Y\). 41 \end{paragraph} 42 43 \begin{lemma}{1.2}{Measurable Functions of a Random Element} 44 Let \(Y\) be measurable from \((\Omega,\mathcal{F})\) to \((\Lambda,\mathcal{G})\), and let \(Z\) be a function from \((\Omega,\mathcal{F})\) to \((\mathbb{R}^k,\mathcal{B}^k)\). Then \(Z\) is measurable from \((\Omega,\sigma(Y))\) to \((\mathbb{R}^k,\mathcal{B}^k)\) if and only if there is a measurable function \(h\) from \((\Lambda,\mathcal{G})\) to \((\mathbb{R}^k,\mathcal{B}^k)\) such that \(Z=h\circ Y\). 45 \end{lemma} 46 47 \begin{paragraph}{1.4.1}{Conditioning on a Value} 48 The function \(h\) in \(E(X\mid Y)=h\circ Y\) is a Borel function on \((\Lambda,\mathcal{G})\). For \(y\in\Lambda\), define 49 \[ 50 E(X\mid Y=y)=h(y). 51 \] 52 Thus \(h(y)\) is a function on \(\Lambda\), whereas \(h\circ Y=E(X\mid Y)\) is a function on \(\Omega\). 53 \end{paragraph} 54 55 \begin{paragraph}{1.4.1}{Conditional Expectation of a Vector} 56 For a random vector \(X\), \(E(X\mid\mathcal{A})\) is defined as the vector of conditional expectations of the components of \(X\). 57 \end{paragraph} 58 59 \begin{example}{1.21}{Conditioning on a Countable Partition} 60 Let \(X\) be an integrable random variable on \((\Omega,\mathcal{F},P)\). Let \(A_1,A_2,\ldots\) be disjoint events such that \(\bigcup_i A_i=\Omega\) and \(P(A_i)>0\) for every \(i\), and let \(a_1,a_2,\ldots\) be distinct real numbers. Define 61 \[ 62 Y=\sum_{i=1}^{\infty}a_iI_{A_i}. 63 \] 64 Then 65 \[ 66 E(X\mid Y) 67 = 68 \sum_{i=1}^{\infty} 69 \frac{\int_{A_i}X\,dP}{P(A_i)}I_{A_i}. 70 \tag{1.59} 71 \] 72 73 We verify the two conditions in Definition~1.6 with \(\mathcal{A}=\sigma(Y)\). Since \(\sigma(Y)=\sigma(\{A_1,A_2,\ldots\})\), the right-hand side of (1.59) is measurable on \((\Omega,\sigma(Y))\). For every \(B\in\mathcal{B}\), 74 \[ 75 Y^{-1}(B)=\bigcup_{i:a_i\in B}A_i. 76 \] 77 Using properties of integrals, 78 \begin{align*} 79 \int_{Y^{-1}(B)}X\,dP 80 & =\sum_{i:a_i\in B}\int_{A_i}X\,dP \\ 81 & =\sum_{i=1}^{\infty} 82 \frac{\int_{A_i}X\,dP}{P(A_i)} 83 P\bigl(A_i\cap Y^{-1}(B)\bigr) \\ 84 & =\int_{Y^{-1}(B)} 85 \left[ 86 \sum_{i=1}^{\infty} 87 \frac{\int_{A_i}X\,dP}{P(A_i)}I_{A_i} 88 \right]dP. 89 \end{align*} 90 This verifies the second condition and proves (1.59). 91 92 Let \(h\) be a Borel function on \(\mathbb{R}\) satisfying 93 \[ 94 h(a_i)=\frac{\int_{A_i}X\,dP}{P(A_i)}. 95 \] 96 Then \(E(X\mid Y)=h\circ Y\) and \(E(X\mid Y=y)=h(y)\). 97 98 If \(A\in\mathcal{F}\) and \(X=I_A\), then 99 \[ 100 P(A\mid Y)=E(X\mid Y) 101 =\sum_{i=1}^{\infty}\frac{P(A\cap A_i)}{P(A_i)}I_{A_i}. 102 \] 103 For \(\omega\in A_i\), this equals \(P(A\cap A_i)/P(A_i)=P(A\mid A_i)\). Hence Definition~1.6 agrees with the elementary definition of conditional probability. 104 \end{example} 105 106 \begin{proposition}{1.9}{Conditional Expectation from a Joint Density} 107 Let \(X\) be a random \(n\)-vector and \(Y\) a random \(m\)-vector. Suppose that \((X,Y)\) has a joint p.d.f. \(f(x,y)\) with respect to \(\nu\times\lambda\), where \(\nu\) and \(\lambda\) are \(\sigma\)-finite measures on \((\mathbb{R}^n,\mathcal{B}^n)\) and \((\mathbb{R}^m,\mathcal{B}^m)\), respectively. Let \(g(x,y)\) be a Borel function on \(\mathbb{R}^{n+m}\) for which \(E|g(X,Y)|<\infty\). Then 108 \[ 109 E[g(X,Y)\mid Y] 110 = 111 \frac{\int g(x,Y)f(x,Y)\,d\nu(x)} 112 {\int f(x,Y)\,d\nu(x)} 113 \quad\text{a.s.} 114 \tag{1.60} 115 \] 116 \end{proposition} 117 118 \begin{proof}{1.9}{Proof via Fubini's Theorem} 119 Denote the right-hand side of (1.60) by \(h(Y)\). By Fubini's theorem, \(h\) is Borel, and Lemma~1.2 implies that \(h(Y)\) is Borel on \((\Omega,\sigma(Y))\). Also, 120 \[ 121 f_Y(y)=\int f(x,y)\,d\nu(x) 122 \] 123 is the p.d.f. of \(Y\) with respect to \(\lambda\). For \(B\in\mathcal{B}^m\), 124 \begin{align*} 125 \int_{Y^{-1}(B)}h(Y)\,dP 126 & =\int_B h(y)\,dP_Y(y) \\ 127 & =\int_B 128 \frac{\int g(x,y)f(x,y)\,d\nu(x)} 129 {\int f(x,y)\,d\nu(x)} 130 f_Y(y)\,d\lambda(y) \\ 131 & =\int_{\mathbb{R}^n\times B}g(x,y)f(x,y)\,d(\nu\times\lambda) \\ 132 & =\int_{\mathbb{R}^n\times B}g(x,y)\,dP_{(X,Y)} \\ 133 & =\int_{Y^{-1}(B)}g(X,Y)\,dP. 134 \end{align*} 135 The first and last equalities follow from Theorem~1.2, the second and fourth from the definitions of \(h\) and p.d.f.'s, and the third from Fubini's theorem. 136 \end{proof} 137 138 \begin{definition}{1.4.1}{Conditional Density} 139 For a random vector \((X,Y)\) with joint p.d.f. \(f(x,y)\) with respect to \(\nu\times\lambda\), define the conditional p.d.f. of \(X\) given \(Y=y\) by 140 \[ 141 f_{X\mid Y}(x\mid y)=\frac{f(x,y)}{f_Y(y)}, 142 \tag{1.61} 143 \] 144 where 145 \[ 146 f_Y(y)=\int f(x,y)\,d\nu(x) 147 \] 148 is the marginal p.d.f. of \(Y\) with respect to \(\lambda\). For every fixed \(y\) with \(f_Y(y)>0\), \(f_{X\mid Y}(x\mid y)\) is a p.d.f. with respect to \(\nu\). 149 \end{definition} 150 151 \begin{paragraph}{1.4.1}{Expectation under a Conditional Density} 152 Equation~(1.60) can be rewritten as 153 \[ 154 E[g(X,Y)\mid Y] 155 = 156 \int g(x,Y)f_{X\mid Y}(x\mid Y)\,d\nu(x). 157 \] 158 Thus the conditional expectation of \(g(X,Y)\) given \(Y\) equals its expectation with respect to the conditional p.d.f. of \(X\) given \(Y\), agreeing with elementary probability. 159 \end{paragraph} 160 161 \begin{proposition}{1.10}{Properties of Conditional Expectations} 162 Let \(X,Y,X_1,X_2,\ldots\) be integrable random variables on \((\Omega,\mathcal{F},P)\), and let \(\mathcal{A}\) be a sub-\(\sigma\)-field of \(\mathcal{F}\). 163 164 \begin{itemize} 165 \item If \(X=c\) a.s., where \(c\in\mathbb{R}\), then \(E(X\mid\mathcal{A})=c\) a.s. 166 \item If \(X\le Y\) a.s., then \(E(X\mid\mathcal{A})\le E(Y\mid\mathcal{A})\) a.s. 167 \item For \(a,b\in\mathbb{R}\), 168 \[ 169 E(aX+bY\mid\mathcal{A}) 170 =aE(X\mid\mathcal{A})+bE(Y\mid\mathcal{A}) 171 \quad\text{a.s.} 172 \] 173 \item \(E[E(X\mid\mathcal{A})]=EX\). 174 \item If \(\mathcal{A}_0\subset\mathcal{A}\), then 175 \begin{align*} 176 E[E(X\mid\mathcal{A})\mid\mathcal{A}_0] 177 & =E(X\mid\mathcal{A}_0) \\ 178 & =E[E(X\mid\mathcal{A}_0)\mid\mathcal{A}] 179 \quad\text{a.s.} 180 \end{align*} 181 182 \item If \(\sigma(Y)\subset\mathcal{A}\) and \(E|XY|<\infty\), then 183 \[ 184 E(XY\mid\mathcal{A})=YE(X\mid\mathcal{A}) 185 \quad\text{a.s.} 186 \] 187 \item If \(X\) and \(Y\) are independent and \(E|g(X,Y)|<\infty\) for a Borel function \(g\), then 188 \[ 189 E[g(X,Y)\mid Y=y]=E[g(X,y)] 190 \quad\text{a.s. }P_Y. 191 \] 192 \item If \(EX^2<\infty\), then 193 \[ 194 [E(X\mid\mathcal{A})]^2\le E(X^2\mid\mathcal{A}) 195 \quad\text{a.s.} 196 \] 197 \item If \(X_n\ge0\), then the conditional Fatou lemma gives 198 \[ 199 E\left(\liminf_n X_n\mid\mathcal{A}\right) 200 \le\liminf_n E(X_n\mid\mathcal{A}) 201 \quad\text{a.s.} 202 \] 203 \item If \(|X_n|\le Y\) and \(X_n\to X\) a.s., then the conditional dominated convergence theorem gives 204 \[ 205 E(X_n\mid\mathcal{A})\to E(X\mid\mathcal{A}) 206 \quad\text{a.s.} 207 \] 208 209 \end{itemize} 210 \end{proposition} 211 212 \begin{paragraph}{1.4.1}{Further Convergence Properties} 213 Part (vii) of Proposition~1.10 follows from the more general Theorem~1.7(i). Since \(E(X\mid\mathcal{A})\) is defined only for integrable \(X\), conditional monotone convergence---if \(0\le X_1\le X_2\le\cdots\) and \(X_n\to X\) a.s., then \(E(X_n\mid\mathcal{A})\to E(X\mid\mathcal{A})\) a.s.---is a special case of Proposition~1.10(x). 214 \end{paragraph} 215 216 \begin{paragraph}{1.4.1}{Conditional Inequalities} 217 Holder's inequality (1.40), Liapounov's inequality (1.42), Minkowski's inequality (1.43), and Jensen's inequality (1.47) also hold a.s. when \(E\) is replaced by \(E(\,\cdot\mid\mathcal{A})\). 218 \end{paragraph} 219 220 \begin{example}{1.22}{Conditional Expectation as Best Predictor} 221 Let \(X\) be a random variable on \((\Omega,\mathcal{F},P)\) with \(EX^2<\infty\), and let \(Y\) be measurable from \((\Omega,\mathcal{F},P)\) to \((\Lambda,\mathcal{G})\). We wish to predict \(X\) from an observed value of \(Y\). Let 222 \[ 223 \mathcal{H} 224 = 225 \{g:\ g\text{ is Borel and }E[g(Y)^2]<\infty\}. 226 \] 227 A predictor \(g(Y)\) is assessed by its mean squared prediction error \(E[X-g(Y)]^2\). Then \(E(X\mid Y)\) is the best predictor: 228 \[ 229 E[X-E(X\mid Y)]^2 230 = 231 \min_{g\in\mathcal{H}}E[X-g(Y)]^2. 232 \tag{1.62} 233 \] 234 235 Proposition~1.10(viii) implies \(E(X\mid Y)\in\mathcal{H}\). For every \(g\in\mathcal{H}\), 236 \begin{align*} 237 E & [X-g(Y)]^2 = \\ 238 & =E[X-E(X\mid Y)+E(X\mid Y)-g(Y)]^2 \\ 239 & =E[X-E(X\mid Y)]^2+E[E(X\mid Y)-g(Y)]^2 \\ 240 & \quad+2E\{[X-E(X\mid Y)][E(X\mid Y)-g(Y)]\} \\ 241 & =E[X-E(X\mid Y)]^2+E[E(X\mid Y)-g(Y)]^2 \\ 242 & \quad+2E\!\left(E\{[X-E(X\mid Y)][E(X\mid Y)-g(Y)]\mid Y\}\right) \\ 243 & =E[X-E(X\mid Y)]^2+E[E(X\mid Y)-g(Y)]^2 \\ 244 & \quad+2E\{[E(X\mid Y)-g(Y)]E[X-E(X\mid Y)\mid Y]\} \\ 245 & =E[X-E(X\mid Y)]^2+E[E(X\mid Y)-g(Y)]^2 \\ 246 & \ge E[X-E(X\mid Y)]^2. 247 \end{align*} 248 The third equality follows from Proposition~1.10(iv), the fourth from Proposition~1.10(vi), and the last equality before the inequality from Proposition~1.10(i), (iii), and (vi). 249 \end{example} 250 251 \begin{definition}{1.7}{Independence} 252 Let \((\Omega,\mathcal{F},P)\) be a probability space. 253 254 \begin{itemize} 255 \item Let \(\mathcal{C}\) be a collection of sets in \(\mathcal{F}\). Events in \(\mathcal{C}\) are independent if and only if, for every positive integer \(n\) and distinct \(A_1,\ldots,A_n\in\mathcal{C}\), 256 \[ 257 P(A_1\cap\cdots\cap A_n)=P(A_1)\cdots P(A_n). 258 \] 259 \item Collections \(\mathcal{C}_i\subset\mathcal{F}\), \(i\in I\), are independent if and only if the events in every collection \(\{A_i\in\mathcal{C}_i:i\in I\}\) are independent. 260 \item Random elements \(X_i\), \(i\in I\), are independent if and only if the \(\sigma\)-fields \(\sigma(X_i)\), \(i\in I\), are independent. 261 \end{itemize} 262 \end{definition} 263 264 \begin{paragraph}{1.4.2}{Checking Independence} 265 The following result is useful for checking the independence of \(\sigma\)-fields. 266 \end{paragraph} 267 268 \begin{lemma}{1.3}{Independence of Generated Sigma-fields} 269 Let \(\mathcal{C}_i\), \(i\in I\), be independent collections of events. Suppose that every \(\mathcal{C}_i\) is closed under intersections: if \(A,B\in\mathcal{C}_i\), then \(A\cap B\in\mathcal{C}_i\). Then the \(\sigma\)-fields \(\sigma(\mathcal{C}_i)\), \(i\in I\), are independent. 270 \end{lemma} 271 272 \begin{paragraph}{1.4.2}{Independence and Joint Distributions} 273 Lemma~1.3 immediately shows that random variables \(X_i\), \(i=1,\ldots,k\), are independent according to Definition~1.7 if and only if (1.7) holds, with \(F\) the joint c.d.f. of \((X_1,\ldots,X_k)\) and \(F_i\) the marginal c.d.f. of \(X_i\). Thus Definition~1.7(iii) agrees with the notion of independence in Section~1.3.1. 274 \end{paragraph} 275 276 \begin{paragraph}{1.4.2}{Functions Preserve Independence} 277 Definition~1.7 also shows that if \(X\) and \(Y\) are independent random vectors, then \(g(X)\) and \(h(Y)\) are independent for every pair of Borel functions \(g\) and \(h\). This gives a simple proof of Lemma~1.1. 278 \end{paragraph} 279 280 \begin{paragraph}{1.4.2}{Independence as No Information} 281 For events \(A\) and \(B\) with \(P(A)>0\), they are independent if and only if \(P(B\mid A)=P(B)\). Thus \(A\) provides no information about the probability that \(B\) occurs. The next proposition extends this fact. 282 \end{paragraph} 283 284 \begin{proposition}{1.11}{Conditioning with Independent Information} 285 Let \(X\) be a random variable with \(E|X|<\infty\), and let \(Y_i\) be random \(k_i\)-vectors, \(i=1,2\). If \((X,Y_1)\) and \(Y_2\) are independent, then 286 \[ 287 E[X\mid(Y_1,Y_2)]=E(X\mid Y_1) 288 \quad\text{a.s.} 289 \] 290 \end{proposition} 291 292 \begin{proof}{1.11}{Proof on Generating Rectangles} 293 First, \(E(X\mid Y_1)\) is Borel on \((\Omega,\sigma(Y_1,Y_2))\), since \(\sigma(Y_1)\subset\sigma(Y_1,Y_2)\). It remains to show that, for every \(B\in\mathcal{B}^{k_1+k_2}\), 294 \[ 295 \int_{(Y_1,Y_2)^{-1}(B)}X\,dP 296 = 297 \int_{(Y_1,Y_2)^{-1}(B)}E(X\mid Y_1)\,dP. 298 \tag{1.63} 299 \] 300 If \(B=B_1\times B_2\), with \(B_i\in\mathcal{B}^{k_i}\), then 301 \[ 302 (Y_1,Y_2)^{-1}(B)=Y_1^{-1}(B_1)\cap Y_2^{-1}(B_2) 303 \] 304 and 305 \begin{align*} 306 & \int_{Y_1^{-1}(B_1)\cap Y_2^{-1}(B_2)}E(X\mid Y_1)\,dP \\ 307 & \quad=\int I_{Y_1^{-1}(B_1)}I_{Y_2^{-1}(B_2)}E(X\mid Y_1)\,dP \\ 308 & \quad=\left[\int I_{Y_1^{-1}(B_1)}E(X\mid Y_1)\,dP\right] 309 \left[\int I_{Y_2^{-1}(B_2)}\,dP\right] \\ 310 & \quad=\left[\int I_{Y_1^{-1}(B_1)}X\,dP\right] 311 \left[\int I_{Y_2^{-1}(B_2)}\,dP\right] \\ 312 & \quad=\int I_{Y_1^{-1}(B_1)}I_{Y_2^{-1}(B_2)}X\,dP \\ 313 & \quad=\int_{Y_1^{-1}(B_1)\cap Y_2^{-1}(B_2)}X\,dP. 314 \end{align*} 315 The second and fourth equalities use (1.38) and the independence of \((X,Y_1)\) and \(Y_2\); the third uses the definition of \(E(X\mid Y_1)\). 316 317 Thus (1.63) holds for rectangles. The collection 318 \[ 319 \mathcal{H} 320 = 321 \{B\subset\mathbb{R}^{k_1+k_2}:B\text{ satisfies (1.63)}\} 322 \] 323 is a \(\sigma\)-field. Since 324 \[ 325 \mathcal{B}^{k_1}\times\mathcal{B}^{k_2}\subset\mathcal{H}, 326 \] 327 we have 328 \[ 329 \mathcal{B}^{k_1+k_2} 330 = 331 \sigma(\mathcal{B}^{k_1}\times\mathcal{B}^{k_2}) 332 \subset\mathcal{H}, 333 \] 334 which proves the result. 335 \end{proof} 336 337 \begin{paragraph}{1.4.2}{Conditional Probabilities under Independence} 338 Proposition~1.11 remains true with \(X\) replaced by \(h(X)\) for any Borel \(h\). Consequently, if \((X,Y_1)\) and \(Y_2\) are independent, then 339 \[ 340 P(A\mid Y_1,Y_2)=P(A\mid Y_1) 341 \quad\text{a.s. for every }A\in\sigma(X). 342 \tag{1.64} 343 \] 344 If \(Y_1\) is constant and \(Y=Y_2\), this becomes \(P(A\mid Y)=P(A)\) a.s. for every \(A\in\sigma(X)\). Thus \(\sigma(Y)\) supplies no additional information about the stochastic behavior of \(X\), giving another equivalent and more intuitive definition of independence. 345 \end{paragraph} 346 347 \begin{paragraph}{1.4.2}{Conditional Independence} 348 For nonconstant \(Y_1\), \(X\) and \(Y_2\) are conditionally independent given \(Y_1\) if and only if (1.64) holds. Proposition~1.11 therefore says that if \(Y_2\) and \((X,Y_1)\) are independent, then \(X\) and \(Y_2\) are conditionally independent given \(Y_1\). The proposition may fail if \(Y_2\) is independent of \(X\) but not of \((X,Y_1)\); see Exercise~96. 349 \end{paragraph} 350 351 \begin{paragraph}{1.4.3}{Conditional Distributions without Densities} 352 The conditional p.d.f. was introduced in Section~1.4.1 for random variables having p.d.f.'s with respect to specified measures. We now consider conditional distributions when no p.d.f. need exist. 353 \end{paragraph} 354 355 \begin{paragraph}{1.4.3}{Need for a Regular Version} 356 For random vectors \(X\) and \(Y\) on a common probability space, it is natural to consider \(P[X^{-1}(B)\mid Y=y]\) as the conditional distribution of \(X\) given \(Y=y\). Since conditional probability is defined only almost surely, however, for a fixed \(y\) this expression need not be a probability measure. The next theorem gives a version that is a probability measure for every fixed \(y\). A proof is given in Billingsley (1986, pp.~460--461). 357 \end{paragraph} 358 359 \begin{theorem}{1.7}{Existence of Conditional Distributions} 360 \begin{itemize} 361 \item Let \(X\) be a random \(n\)-vector on \((\Omega,\mathcal{F},P)\), and let \(\mathcal{A}\) be a sub-\(\sigma\)-field of \(\mathcal{F}\). There is a function \(P(B,\omega)\) on \(\mathcal{B}^n\times\Omega\) such that: 362 \begin{itemize} 363 \item \(P(B,\omega)=P[X^{-1}(B)\mid\mathcal{A}]\) a.s. for every fixed \(B\in\mathcal{B}^n\); 364 \item \(P(\,\cdot,\omega)\) is a probability measure on \((\mathbb{R}^n,\mathcal{B}^n)\) for every fixed \(\omega\in\Omega\). 365 \end{itemize} 366 367 If \(Y\) is measurable from \((\Omega,\mathcal{F},P)\) to \((\Lambda,\mathcal{G})\), there is a function \(P_{X\mid Y}(B\mid y)\) such that: 368 \begin{itemize} 369 \item \(P_{X\mid Y}(B\mid y)=P[X^{-1}(B)\mid Y=y]\) a.s. \(P_Y\) for every fixed \(B\in\mathcal{B}^n\); 370 \item \(P_{X\mid Y}(\,\cdot\mid y)\) is a probability measure on \((\mathbb{R}^n,\mathcal{B}^n)\) for every fixed \(y\in\Lambda\). 371 \end{itemize} 372 373 Furthermore, if \(g\) is Borel and \(E|g(X,Y)|<\infty\), then 374 \begin{align*} 375 E[g(X,Y)&\mid Y=y] = \\ 376 & =E[g(X,y)\mid Y=y] \\ 377 & =\int_{\mathbb{R}^n}g(x,y)\,dP_{X\mid Y}(x\mid y) 378 \quad\text{a.s. }P_Y. 379 \end{align*} 380 381 \item Let \((\Lambda,\mathcal{G},P_1)\) be a probability space. Suppose \(P_2\) is a function from \(\mathcal{B}^n\times\Lambda\) to \(\mathbb{R}\) such that \(P_2(\,\cdot,y)\) is a probability measure on \((\mathbb{R}^n,\mathcal{B}^n)\) for every \(y\), and \(P_2(B,\,\cdot)\) is Borel for every \(B\in\mathcal{B}^n\). Then there is a unique probability measure \(P\) on \((\mathbb{R}^n\times\Lambda,\sigma(\mathcal{B}^n\times\mathcal{G}))\) such that 382 \[ 383 P(B\times C)=\int_C P_2(B,y)\,dP_1(y) 384 \tag{1.65} 385 \] 386 for \(B\in\mathcal{B}^n\) and \(C\in\mathcal{G}\). 387 388 \hangindent\parindent If \((\Lambda,\mathcal{G})=(\mathbb{R}^m,\mathcal{B}^m)\), and \(X(x,y)=x\) and \(Y(x,y)=y\) are the coordinate random vectors, then \(P_Y=P_1\), \(P_{X\mid Y}(\,\cdot\mid y)=P_2(\,\cdot,y)\), and (1.65) defines the joint distribution of \((X,Y)\). Its joint c.d.f. is 389 390 \begin{align*} 391 F(x,y) & = \int_{(-\infty,y]} P_{X\mid Y}\bigl((-\infty,x]\mid z\bigr)\,dP_Y(z) \ \ (1.66) \\ 392 & x\in\mathbb{R}^n,\ y\in\mathbb{R}^m. 393 \end{align*} 394 Here \((-\infty,a]=(-\infty,a_1]\times\cdots\times(-\infty,a_k]\) for \(a=(a_1,\ldots,a_k)\). 395 \end{itemize} 396 \end{theorem} 397 398 \begin{definition}{1.4.3}{Conditional Distribution} 399 For fixed \(y\), 400 \[ 401 P_{X\mid Y=y}=P_{X\mid Y}(\,\cdot\mid y) 402 \] 403 is called the conditional distribution of \(X\) given \(Y=y\). Under Theorem~1.7(i), if \(Y\) is an \(m\)-vector and \((X,Y)\) has a p.d.f. with respect to \(\nu\times\lambda\), then \(f_{X\mid Y}(x\mid y)\) from (1.61) is the p.d.f. of \(P_{X\mid Y=y}\) with respect to \(\nu\). 404 \end{definition} 405 406 \begin{paragraph}{1.4.3}{The Two-stage Experiment Theorem} 407 Theorem~1.7(ii) says that a distribution on one space and conditional distributions on a second space determine a joint distribution on the product space. If \(Y\in\mathbb{R}^m\) is selected in stage one according to \(P_Y=P_1\), and \(X\) is then selected according to \(P_2(\,\cdot,y)\), the combined experiment produces a jointly distributed pair \((X,Y)\) whose distribution is given by (1.65), with \(P_{X\mid Y=y}=P_2(\,\cdot,y)\). This gives a way to generate dependent random variables. 408 \end{paragraph} 409 410 \begin{example}{1.23}{A Market Survey} 411 A mail survey asks \(N\) randomly selected customers whether a new product is better than the old one. A response is \(1\) if the new product is preferred and \(0\) otherwise; some customers do not respond. Let \(X\) be the number of ones among the returned questionnaires. 412 413 If every customer responds, then \(X\) has the binomial distribution \(\operatorname{Bi}(p,N)\), assuming independent responses, where \(p\in(0,1)\) is the overall preference rate. Let \(Y\) be the number who respond. Suppose customers respond independently with probability \(\pi\in(0,1)\). Then 414 \[ 415 Y\sim\operatorname{Bi}(\pi,N). 416 \] 417 Given \(Y=y\), \(P_{X\mid Y=y}\) is \(\operatorname{Bi}(p,y)\) for \(y\ge1\), and is point mass at \(0\) for \(y=0\). 418 419 Using (1.66) and counting measure, the joint c.d.f. is 420 \begin{align*} 421 F(x,y) 422 & =\sum_{k=0}^{y} 423 P_{X\mid Y=k}((-\infty,x]) 424 \binom Nk\pi^k(1-\pi)^{N-k} \\ 425 & =\sum_{k=0}^{y}\sum_{j=0}^{\min\{x,k\}} 426 \binom kjp^j(1-p)^{k-j} 427 \binom Nk\pi^k(1-\pi)^{N-k}, 428 \end{align*} 429 for \(x=0,1,\ldots,y\) and \(y=0,1,\ldots,N\). 430 431 The marginal c.d.f. is \(F_X(x)=F(x,\infty)=F(x,N)\). Its p.d.f. with respect to counting measure is 432 \begin{align*} 433 f_X(x) 434 & =\sum_{k=x}^{N} 435 \binom{k}{x}p^x(1-p)^{k-x} 436 \binom Nk\pi^k(1-\pi)^{N-k} \\ 437 & =\binom Nx(\pi p)^x(1-\pi p)^{N-x}, 438 \end{align*} 439 for \(x=0,1,\ldots,N\). Thus \(X\sim\operatorname{Bi}(\pi p,N)\). 440 \end{example} 441 442 \begin{paragraph}{1.4.1}{Conditional Expectations from Densities} 443 The next result generalizes Example~1.21 to conditional expectations of random variables having p.d.f.'s. 444 \end{paragraph} 445 446 \begin{paragraph}{1.4.4}{Dependent Sequences} 447 As applications of conditional expectations, we introduce two important types of dependent sequences: Markov chains and martingales. 448 \end{paragraph} 449 450 \begin{definition}{1.4.4}{Markov Chain} 451 A sequence of random vectors \(\{X_n:n=1,2,\ldots\}\) is a Markov chain, or Markov process, if and only if 452 \[ 453 P(B\mid X_1,\ldots,X_n)=P(B\mid X_n) 454 \quad\text{a.s.}, 455 \quad 456 B\in\sigma(X_{n+1}),\ n=2,3,\ldots. 457 \tag{1.67} 458 \] 459 \end{definition} 460 461 \begin{paragraph}{1.4.4}{The Markov Property} 462 Comparing (1.67) with (1.64), \(X_{n+1}\), representing tomorrow, is conditionally independent of \((X_1,\ldots,X_{n-1})\), representing the past, given \(X_n\), representing today. However, \((X_1,\ldots,X_{n-1})\) need not be independent of \((X_n,X_{n+1})\). 463 \end{paragraph} 464 465 \begin{paragraph}{1.4.4}{Independent Sequences Are Markov} 466 Every sequence of independent random vectors is a Markov chain: Proposition~1.11 shows that both sides of (1.67) equal \(P(B)\). The next example gives dependent Markov processes. 467 \end{paragraph} 468 469 \begin{example}{1.24}{First-order Autoregressive Processes} 470 Let \(\varepsilon_1,\varepsilon_2,\ldots\) be independent random variables, let \(X_1=\varepsilon_1\), and define 471 \[ 472 X_{n+1}=\rho X_n+\varepsilon_{n+1}, 473 \quad n=1,2,\ldots, 474 \] 475 where \(\rho\in\mathbb{R}\). This is a first-order autoregressive process. For \(B\in\mathcal{B}\), 476 \begin{align*} 477 P(X_{n+1}\in B\mid X_1,\ldots,X_n) 478 & =P_{\varepsilon_{n+1}}(B-\rho X_n) \\ 479 & =P(X_{n+1}\in B\mid X_n) 480 \quad\text{a.s.}, 481 \end{align*} 482 where \(B-y=\{x\in\mathbb{R}:x+y\in B\}\); hence \(\{X_n\}\) is a Markov chain. 483 484 For \(y\in\mathbb{R}\), 485 \[ 486 P_{\varepsilon_{n+1}}(B-y) 487 =P(\varepsilon_{n+1}+y\in B) 488 =\int I_B(x+y)\,dP_{\varepsilon_{n+1}}(x). 489 \] 490 Fubini's theorem implies that this is Borel in \(y\). Therefore \(P_{\varepsilon_{n+1}}(B-\rho X_n)\) is Borel with respect to \(\sigma(X_n)\), and hence with respect to \(\sigma(X_1,\ldots,X_n)\). 491 492 Let \(B_j\in\mathcal{B}\), \(j=1,\ldots,n\), and 493 \[ 494 A=\bigcap_{j=1}^n X_j^{-1}(B_j). 495 \] 496 Since \(X_{n+1}=\varepsilon_{n+1}+\rho X_n\) and \(\varepsilon_{n+1}\) is independent of \((X_1,\ldots,X_n)\), Theorem~1.2 and Fubini's theorem give 497 \begin{align*} 498 \int_A & P_{\varepsilon_{n+1}}(B-\rho X_n)\,dP= \\ 499 & =\int_{x_j\in B_j,\ j=1,\ldots,n,t\in B-\rho x_n} 500 dP_{\varepsilon_{n+1}}(t)\,dP_X(x) \\ 501 & =\int_{x_j\in B_j,\ j=1,\ldots,n,x_{n+1}\in B} 502 dP_{(X,\varepsilon_{n+1})}(x,t) \\ 503 & =P\bigl(A\cap X_{n+1}^{-1}(B)\bigr), 504 \end{align*} 505 where \(X=(X_1,\ldots,X_n)\), \(x=(x_1,\ldots,x_n)\), and \(x_{n+1}=\rho x_n+t\). The final argument in the proof of Proposition~1.11 now gives 506 \[ 507 P(X_{n+1}\in B\mid X_1,\ldots,X_n) 508 =P_{\varepsilon_{n+1}}(B-\rho X_n) 509 \quad\text{a.s.} 510 \] 511 The proof that the latter equals \(P(X_{n+1}\in B\mid X_n)\) a.s. is similar and simpler. 512 \end{example} 513 514 \begin{proposition}{1.12}{Characterizations of Markov Chains} 515 A sequence \(\{X_n\}\) is a Markov chain if and only if any one of the following conditions holds: 516 517 \begin{itemize} 518 \item For every \(n\ge2\) and every integrable \(h(X_{n+1})\), with \(h\) Borel, 519 \[ 520 E[h(X_{n+1})\mid X_1,\ldots,X_n] 521 =E[h(X_{n+1})\mid X_n] 522 \quad\text{a.s.} 523 \] 524 \item For every \(n\ge1\) and \(B\in\sigma(X_{n+1},X_{n+2},\ldots)\), 525 \[ 526 P(B\mid X_1,\ldots,X_n)=P(B\mid X_n) 527 \quad\text{a.s.} 528 \] 529 \item For every \(n\ge2\), \(A\in\sigma(X_1,\ldots,X_n)\), and \(B\in\sigma(X_{n+1},X_{n+2},\ldots)\), 530 \[ 531 P(A\cap B\mid X_n) 532 =P(A\mid X_n)P(B\mid X_n) 533 \quad\text{a.s.} 534 \] 535 \end{itemize} 536 \end{proposition} 537 538 \begin{proof}{1.12}{Proof of the Characterizations} 539 First, condition (a) clearly implies (1.67). Conversely, (1.67) implies (a) for simple \(h\) by Proposition~1.10(iii). For nonnegative \(h\), choose nonnegative simple functions \(h_1\le h_2\le\cdots\le h\) with \(h_j\to h\), and apply Proposition~1.10(iii) and (x). The general case follows from \(h=h^+-h^-\). 540 541 Condition (b) clearly implies (1.67). Conversely, 542 \[ 543 \sigma(X_{n+1},X_{n+2},\ldots) 544 = 545 \sigma\left(\bigcup_{j=1}^{\infty}\sigma(X_{n+1},\ldots,X_{n+j})\right). 546 \] 547 It therefore suffices to prove (b) for \(B\in\sigma(X_{n+1},\ldots,X_{n+j})\), by induction on \(j\). The case \(j=1\) is (1.67). The induction hypothesis also gives 548 \begin{align*} 549 E[h(X_{n+1},\ldots,X_{n+j})&\mid X_1,\ldots,X_n] = \\ 550 & =E[h(X_{n+1},\ldots,X_{n+j})\mid X_n] \ \ (1.68) 551 \end{align*} 552 for every Borel \(h\). 553 554 For \(B_1\in\sigma(X_{n+j+1})\) and \(B_2\in\sigma(X_{n+1},\ldots,X_{n+j})\), 555 \begin{align*} 556 E & (I_{B_1}I_{B_2}\mid X_1,\ldots,X_n)= \\ 557 & =E[E(I_{B_1}I_{B_2}\mid X_1,\ldots,X_{n+j})\mid X_1,\ldots,X_n] \\ 558 & =E[I_{B_2}E(I_{B_1}\mid X_1,\ldots,X_{n+j})\mid X_1,\ldots,X_n] \\ 559 & =E[I_{B_2}E(I_{B_1}\mid X_{n+j})\mid X_1,\ldots,X_n] \\ 560 & =E[I_{B_2}E(I_{B_1}\mid X_{n+j})\mid X_n] \\ 561 & =E[I_{B_2}E(I_{B_1}\mid X_n,\ldots,X_{n+j})\mid X_n] \\ 562 & =E[E(I_{B_1}I_{B_2}\mid X_n,\ldots,X_{n+j})\mid X_n] \\ 563 & =E(I_{B_1}I_{B_2}\mid X_n) 564 \quad\text{a.s.} 565 \end{align*} 566 Here the first and last equalities use Proposition~1.10(v), the second and sixth use Proposition~1.10(vi), the third and fifth use (1.67), and the fourth uses (1.68). This completes the induction. 567 568 If (b) holds, then for \(A\in\sigma(X_1,\ldots,X_n)\) and \(B\) in the future \(\sigma\)-field, 569 \begin{align*} 570 E(I_AI_B\mid X_n) 571 & =E[E(I_AI_B\mid X_1,\ldots,X_n)\mid X_n] \\ 572 & =E[I_AE(I_B\mid X_1,\ldots,X_n)\mid X_n] \\ 573 & =E[I_AE(I_B\mid X_n)\mid X_n] \\ 574 & =E(I_A\mid X_n)E(I_B\mid X_n), 575 \end{align*} 576 which is (c). 577 578 Conversely, assume (c). Let \(A_1\in\sigma(X_n)\), \(A_2\in\sigma(X_1,\ldots,X_{n-1})\), and \(B\in\sigma(X_{n+1},X_{n+2},\ldots)\). Then 579 \begin{align*} 580 \int_{A_1\cap A_2}E(I_B\mid X_n)\,dP 581 & =\int_{A_1}I_{A_2}E(I_B\mid X_n)\,dP \\ 582 & =\int_{A_1}E[I_{A_2}E(I_B\mid X_n)\mid X_n]\,dP \\ 583 & =\int_{A_1}E(I_{A_2}\mid X_n)E(I_B\mid X_n)\,dP \\ 584 & =\int_{A_1}E(I_{A_2}I_B\mid X_n)\,dP \\ 585 & =P(A_1\cap A_2\cap B). 586 \end{align*} 587 Disjoint unions of events \(A_1\cap A_2\) generate \(\sigma(X_1,\ldots,X_n)\), so \(E(I_B\mid X_n)=E(I_B\mid X_1,\ldots,X_n)\) a.s., proving (b). 588 \end{proof} 589 590 \begin{paragraph}{1.4.4}{Past and Future Given the Present} 591 Condition (b) of Proposition~1.12 says that the past and future are conditionally independent given the present. This is a property of every Markov chain. More discussion and applications appear in Section~4.1.4. 592 \end{paragraph} 593 594 \begin{definition}{1.4.4}{Martingales} 595 Let \(\{X_n\}\) be integrable random variables on \((\Omega,\mathcal{F},P)\), and let 596 \[ 597 \mathcal{F}_1\subset\mathcal{F}_2\subset\cdots\subset\mathcal{F} 598 \] 599 with \(\sigma(X_n)\subset\mathcal{F}_n\). The sequence \(\{X_n,\mathcal{F}_n:n=1,2,\ldots\}\) is a martingale if and only if 600 \[ 601 E(X_{n+1}\mid\mathcal{F}_n)=X_n 602 \quad\text{a.s.},\quad n=1,2,\ldots. 603 \tag{1.69} 604 \] 605 It is a submartingale if equality is replaced by \(\ge\), and a supermartingale if equality is replaced by \(\le\). The sequence \(\{X_n\}\) itself is called a martingale, submartingale, or supermartingale when the filtration is \(\sigma(X_1,\ldots,X_n)\). Proposition~1.10(v) shows that if \(\{X_n,\mathcal{F}_n\}\) has one of these properties, then so does \(\{X_n\}\). 606 \end{definition} 607 608 \begin{paragraph}{1.4.4}{Basic Martingale Properties} 609 For a martingale \(\{X_n,\mathcal{F}_n\}\), 610 \[ 611 E(X_{n+j}\mid\mathcal{F}_n)=X_n 612 \quad\text{and}\quad 613 EX_1=EX_j. 614 \] 615 For a submartingale the corresponding statements are 616 \[ 617 E(X_{n+j}\mid\mathcal{F}_n)\ge X_n 618 \quad\text{and}\quad 619 EX_1\le EX_2\le\cdots. 620 \] 621 These hold for every \(j=1,2,\ldots\). 622 \end{paragraph} 623 624 \begin{paragraph}{1.4.4}{Constructing Martingales} 625 Given \(\mathcal{F}_1\subset\mathcal{F}_2\subset\cdots\subset\mathcal{F}\) and an integrable \(Y\), the sequence \(\{E(Y\mid\mathcal{F}_n)\}\) is a martingale. Another construction uses independent integrable random variables \(\{\varepsilon_n\}\) and 626 \[ 627 X_n=\varepsilon_1+\cdots+\varepsilon_n. 628 \] 629 Since 630 \[ 631 E(X_{n+1}\mid X_1,\ldots,X_n) 632 =X_n+E\varepsilon_{n+1} 633 \quad\text{a.s.}, 634 \] 635 \(\{X_n\}\) is a martingale if \(E\varepsilon_n=0\), a submartingale if \(E\varepsilon_n\ge0\), and a supermartingale if \(E\varepsilon_n\le0\), for every \(n\). Example~1.24 with \(\rho=1\) also shows that \(\{X_n\}\) is a Markov chain. 636 \end{paragraph} 637 638 \begin{paragraph}{1.4.4}{Another Martingale Example} 639 The next example gives another construction of martingales. 640 \end{paragraph} 641 642 \begin{example}{1.25}{Likelihood Ratio Martingale} 643 Let \((\Omega,\mathcal{F},P)\) be a probability space, let \(Q\) be another probability measure on \(\mathcal{F}\), and let \(\mathcal{F}_1\subset\mathcal{F}_2\subset\cdots\subset\mathcal{F}\). Write \(P_n\) and \(Q_n\) for the restrictions of \(P\) and \(Q\) to \(\mathcal{F}_n\). If \(Q_n\ll P_n\) for every \(n\), then 644 \[ 645 X_n=\frac{dQ_n}{dP_n} 646 \] 647 defines a martingale \(\{X_n,\mathcal{F}_n\}\). 648 649 Now suppose \(\mathcal{F}_n=\sigma(Y_1,\ldots,Y_n)\), and there is a \(\sigma\)-finite measure \(\nu_n\) on \(\mathcal{F}_n\) such that \(P_n\ll\nu_n\ll P_n\). Define 650 \[ 651 p_n(Y_1,\ldots,Y_n)=\frac{dP_n}{d\nu_n}, 652 \qquad 653 q_n(Y_1,\ldots,Y_n)=\frac{dQ_n}{d\nu_n}. 654 \] 655 By Proposition~1.7(iii), 656 \[ 657 X_n 658 = 659 \frac{q_n(Y_1,\ldots,Y_n)} 660 {p_n(Y_1,\ldots,Y_n)}, 661 \] 662 which is called a likelihood ratio in statistics. 663 \end{example} 664 665 \begin{paragraph}{1.4.4}{Functions of Martingales} 666 The following results give useful properties of martingales and submartingales. 667 \end{paragraph} 668 669 \begin{proposition}{1.13}{Convex Functions of Martingales} 670 Let \(\phi\) be convex on \(\mathbb{R}\). 671 672 \begin{itemize} 673 \item If \(\{X_n,\mathcal{F}_n\}\) is a martingale and \(\phi(X_n)\) is integrable for every \(n\), then \(\{\phi(X_n),\mathcal{F}_n\}\) is a submartingale. 674 \item If \(\{X_n,\mathcal{F}_n\}\) is a submartingale, \(\phi(X_n)\) is integrable for every \(n\), and \(\phi\) is nondecreasing, then \(\{\phi(X_n),\mathcal{F}_n\}\) is a submartingale. 675 \end{itemize} 676 \end{proposition} 677 678 \begin{proof}{1.13}{Proof by Conditional Jensen} 679 For the first part, 680 \[ 681 \phi(X_n) 682 =\phi(E(X_{n+1}\mid\mathcal{F}_n)) 683 \le E[\phi(X_{n+1})\mid\mathcal{F}_n] 684 \quad\text{a.s.} 685 \] 686 by Jensen's inequality for conditional expectations. For the second, monotonicity and the submartingale property give 687 \[ 688 \phi(X_n) 689 \le\phi(E(X_{n+1}\mid\mathcal{F}_n)) 690 \le E[\phi(X_{n+1})\mid\mathcal{F}_n] 691 \quad\text{a.s.} 692 \] 693 \end{proof} 694 695 \begin{paragraph}{1.4.4}{Examples of Convex Transformations} 696 Proposition~1.13 shows that if \(\{X_n,\mathcal{F}_n\}\) is a submartingale, then so is \(\{X_n^+,\mathcal{F}_n\}\). If it is a martingale, then \(\{|X_n|,\mathcal{F}_n\}\), \(\{|X_n|^p,\mathcal{F}_n\}\) for \(p>1\), and \(\{|X_n|(\log|X_n|)^+,\mathcal{F}_n\}\) are submartingales whenever the displayed random variables are integrable. 697 \end{paragraph} 698 699 \begin{proposition}{1.14}{Doob's Decomposition} 700 Let \(\{X_n,\mathcal{F}_n\}\) be a submartingale. Then 701 \[ 702 X_n=Y_n+Z_n,\quad n=1,2,\ldots, 703 \] 704 where \(\{Y_n,\mathcal{F}_n\}\) is a martingale, 705 \[ 706 0=Z_1\le Z_2\le\cdots, 707 \] 708 and \(EZ_n<\infty\) for every \(n\). If \(\sup_nE|X_n|<\infty\), then 709 \[ 710 \sup_nE|Y_n|<\infty 711 \quad\text{and}\quad 712 \sup_nEZ_n<\infty. 713 \] 714 \end{proposition} 715 716 \begin{proof}{1.14}{Proof of Doob's Decomposition} 717 Define 718 \[ 719 \eta_1=X_1,\qquad \zeta_1=0, 720 \] 721 and, for \(n\ge2\), 722 \begin{align*} 723 \eta_n & =X_n-X_{n-1}-E(X_n-X_{n-1}\mid\mathcal{F}_{n-1}) \\ 724 \zeta_n & =E(X_n-X_{n-1}\mid\mathcal{F}_{n-1}). 725 \end{align*} 726 Then 727 \[ 728 Y_n=\sum_{i=1}^n\eta_i, 729 \qquad 730 Z_n=\sum_{i=1}^n\zeta_i 731 \] 732 satisfy \(X_n=Y_n+Z_n\) and the required conditions. 733 734 If \(\sup_nE|X_n|<\infty\), then \(EY_1=EY_n\) and \(Z_n\le|X_n|-Y_n\), so 735 \[ 736 EZ_n\le E|X_n|-EY_1. 737 \] 738 Hence \(\sup_nEZ_n<\infty\). Since \(|Y_n|\le|X_n|+Z_n\), it also follows that \(\sup_nE|Y_n|<\infty\). 739 \end{proof} 740 741 \begin{paragraph}{1.4.4}{Doob's Convergence Theorem} 742 The following martingale convergence theorem, due to Doob, has many applications; see, for example, Example~1.27 in Section~1.5.1. A proof is given in Billingsley (1986, pp.~490--491). 743 \end{paragraph} 744 745 \begin{proposition}{1.15}{Martingale Convergence} 746 Let \(\{X_n,\mathcal{F}_n\}\) be a submartingale. If 747 \[ 748 c=\sup_nE|X_n|<\infty, 749 \] 750 then \(X_n\to X\) a.s. for a random variable \(X\) satisfying \(E|X|\le c\). 751 \end{proposition}