概率作为一种特殊的测度,满足有限可加性: P ( ⋃ i = 1 n A i ) = ∑ i = 1 n P ( A i ) , A i ∩ A j = ϕ , ∀ i ≠ j P(\bigcup_{i=1}^{n} A_i) = \sum_{i=1}^{n} P(A_i),A_i \cap A_j = \phi,\forall i \ne j P(i=1⋃nAi)=i=1∑nP(Ai),Ai∩Aj=ϕ,∀i=j
A i , A j A_i,A_j Ai,Aj作为样本空间 σ \sigma σ-代数中的元素, A i ∩ A j = ϕ A_i \cap A_j = \phi Ai∩Aj=ϕ的概率解释是 A i , A j A_i,A_j Ai,Aj这两个事件互斥。上式左边的含义是互斥事件 A 1 , A 2 , ⋯ , A n A_1,A_2,\cdots,A_n A1,A2,⋯,An中至少有一个发生的概率,右边的含义是互斥事件 A 1 , A 2 , ⋯ , A n A_1,A_2,\cdots,A_n A1,A2,⋯,An的概率之和。但在具体应用中,事件互斥的假设实在是太强了,我们希望在没有互斥这个条件下,也能有一个计算 P ( ⋃ i = 1 n A i ) P(\bigcup_{i=1}^{n} A_i) P(⋃i=1nAi)的方法。
尝试估计某个量的值之前,我们往往会尝试找到一些上界和下界,然后试图找到更小的上界和更大的下界,逐步逼近准确值。要寻找 P ( ⋃ i = 1 n A i ) P(\bigcup_{i=1}^{n} A_i) P(⋃i=1nAi)的上界和下界,最容易想到的是外测度的有限次可加性: P ( ⋃ i = 1 n A i ) ≤ ∑ i = 1 n P ( A i ) P(\bigcup_{i=1}^{n} A_i) \le \sum_{i=1}^{n} P(A_i) P(i=1⋃nAi)≤i=1∑nP(Ai)
当且仅当 A 1 , A 2 , ⋯ , A n A_1,A_2,\cdots,A_n A1,A2,⋯,An互斥时取等。这个不等式叫做Boole不等式。要证明这个不等式也非常容易,我们构造一个集列 { B i } i = 1 n \{B_i\}_{i=1}^n {Bi}i=1n,其中 B 1 = A 1 , B 2 = A 2 − A 1 , ⋯ , B n = A n − ⋃ i = 1 n − 1 A n B_1 = A_1,B_2 = A_2-A_1,\cdots,B_n = A_n - \bigcup_{i=1}^{n-1}A_n B1=A1,B2=A2−A1,⋯,Bn=An−⋃i=1n−1An,显然 B i ∩ B j = ϕ , ∀ i ≠ j B_i \cap B_j = \phi, \forall i \ne j Bi∩Bj=ϕ,∀i=j,并且 ⋃ i = 1 n B i = ⋃ i = 1 n A i \bigcup_{i=1}^n B_i = \bigcup_{i=1}^{n} A_i i=1⋃nBi=i=1⋃nAi
根据概率的可加性和单调性( B i ⊂ A i ⇒ P ( B i ) ≤ P ( A i ) B_i \subset A_i \Rightarrow P(B_i) \le P(A_i) Bi⊂Ai⇒P(Bi)≤P(Ai)) P ( ⋃ i = 1 n A i ) = P ( ⋃ i = 1 n B i ) = ∑ i = 1 n P ( B i ) ≤ ∑ i = 1 n P ( A i ) P(\bigcup_{i=1}^{n} A_i) = P(\bigcup_{i=1}^n B_i ) = \sum_{i=1}^n P(B_i) \le \sum_{i=1}^{n} P(A_i) P(i=1⋃nAi)=P(i=1⋃nBi)=i=1∑nP(Bi)≤i=1∑nP(Ai)
基于de Morgan律很容易就能得到Boole不等式的另一种形式。 P ( ⋂ i = 1 n A i ) = 1 − P ( ⋂ i = 1 n A i ˉ ) = 1 − P ( ⋃ i = 1 n A ˉ i ) ≥ 1 − ∑ i = 1 n P ( A ˉ i ) = 1 − ∑ i = 1 n [ 1 − P ( A i ) ] = ∑ i = 1 n P ( A i ) − ( n − 1 ) P(\bigcap_{i=1}^n A_i) = 1-P(\bar{\bigcap_{i=1}^n A_i}) = 1-P(\bigcup_{i=1}^n \bar{A}_i) \ge 1 - \sum_{i=1}^{n} P(\bar{A}_i) \\ = 1-\sum_{i=1}^{n} [1-P(A_i)] = \sum_{i=1}^{n} P(A_i) - (n-1) P(i=1⋂nAi)=1−P(i=1⋂nAiˉ)=1−P(i=1⋃nAˉi)≥1−i=1∑nP(Aˉi)=1−i=1∑n[1−P(Ai)]=i=1∑nP(Ai)−(n−1)
Boole不等式给出了非常粗糙的上界和下界,后续的任务就是试图推导出更精确的上界和下界。Gunias不等式和钟开莱-艾尔迪希(Chung-Erdos)不等式分别提供了至少有一个事件发生概率的更精确的上界和下界。
因为现在没有 A 1 , A 2 , ⋯ , A n A_1,A_2,\cdots,A_n A1,A2,⋯,An互斥的假设,所以他们两两之间交集不一定为0,在这 n n n个事件中,总有一个事件与其他事件“重合程度”最大,即 ∃ k ∈ { 1 , 2 , ⋯ , n } \exists k \in \{1,2,\cdots,n\} ∃k∈{1,2,⋯,n}, ∑ i = 1 , i ≠ k P ( A i ∩ A k ) ≥ ∑ i = 1 , i ≠ j P ( A i ∩ A j ) , ∀ j ∈ { 1 , 2 , ⋯ , n } \sum_{i=1,i \ne k}P(A_i \cap A_k) \ge \sum_{i=1,i \ne j}P(A_i \cap A_j),\forall j \in \{1,2,\cdots,n\} i=1,i=k∑P(Ai∩Ak)≥i=1,i=j∑P(Ai∩Aj),∀j∈{1,2,⋯,n}
下面我们构造两个集列 { B i } i = 1 n \{B_i\}_{i=1}^n {Bi}i=1n与 { C i } i = 1 n \{C_i\}_{i=1}^n {Ci}i=1n: B 1 = A k , B 2 = A 1 − A k , ⋯ , B k = A k − 1 − A k , B k + 1 = A k + 1 − A k , ⋯ , B n = A n − A k C 1 = A k , C 2 = A 1 − A k , C 3 = A 2 − A 1 − A k , ⋯ , C k = A k − 1 − ⋃ j = 1 , j ≠ k − 1 k A j C k + 1 = A k + 1 − ⋃ j = 1 k A j , ⋯ , C n = A n − ⋃ j = 1 n − 1 A j B_1 = A_k,B_2=A_1-A_k,\cdots,B_{k} = A_{k-1} - A_k,B_{k+1} = A_{k+1}-A_{k},\cdots,B_n = A_n - A_k \\ C_1 = A_k,C_2 = A_1-A_k,C_3 = A_2 - A_1 - A_k,\cdots,C_{k} = A_{k-1} - \bigcup_{j=1,j \ne k-1}^kA_j \\ C_{k+1} = A_{k+1} - \bigcup_{j=1}^{k} A_j,\cdots,C_n = A_n - \bigcup_{j=1}^{n-1} A_j B1=Ak,B2=A1−Ak,⋯,Bk=Ak−1−Ak,Bk+1=Ak+1−Ak,⋯,Bn=An−AkC1=Ak,C2=A1−Ak,C3=A2−A1−Ak,⋯,Ck=Ak−1−j=1,j=k−1⋃kAjCk+1=Ak+1−j=1⋃kAj,⋯,Cn=An−j=1⋃n−1Aj
显然 C i ⊂ B i C_i \subset B_i Ci⊂Bi,因此 ⋃ i = 1 n C i ⊂ ⋃ i = 1 n B i \bigcup_{i=1}^n C_i \subset \bigcup_{i=1}^n B_i i=1⋃nCi⊂i=1⋃nBi
集列 { C i } i = 1 n \{C_i\}_{i=1}^n {Ci}i=1n与Boole不等式证明中构造的集列类似, ⋃ i = 1 n C i = ⋃ i = 1 n A i \bigcup_{i=1}^n C_i = \bigcup_{i=1}^n A_i i=1⋃nCi=i=1⋃nAi
根据概率的单调性, P ( ⋃ i = 1 n A i ) = P ( ⋃ i = 1 n C i ) ≤ P ( ⋃ i = 1 n B i ) P(\bigcup_{i=1}^n A_i) = P(\bigcup_{i=1}^n C_i) \le P(\bigcup_{i=1}^n B_i) P(i=1⋃nAi)=P(i=1⋃nCi)≤P(i=1⋃nBi)
根据Boole不等式 P ( ⋃ i = 1 n B i ) ≤ ∑ i = 1 n P ( B i ) = P ( A k ) + P ( A 1 − A k ) + ⋯ + P ( A k − 1 − A k ) + P ( A k + 1 − A k ) + ⋯ + P ( A n − A k ) = P ( A k ) + P ( A 1 ) − P ( A 1 ∩ A k ) + ⋯ + P ( A k − 1 ) − P ( A k − 1 ∩ A k ) + P ( A k + 1 ) − P ( A k + 1 ∩ A k ) + ⋯ + P ( A n ) − P ( A n ∩ A k ) = ∑ i = 1 n P ( A i ) − ∑ j = 1 , j ≠ k n P ( A j ∩ A k ) P(\bigcup_{i=1}^n B_i) \le \sum_{i=1}^n P(B_i) = P(A_k) + P(A_1-A_k) + \cdots \\+ P(A_{k-1}-A_k) + P(A_{k+1}-A_k) + \cdots + P(A_n - A_k) \\ = P(A_k) + P(A_1)-P(A_1\cap A_k) + \cdots + P(A_{k-1})-P(A_{k-1} \cap A_k) \\+ P(A_{k+1})-P(A_{k+1}\cap A_k) + \cdots + P(A_n) - P(A_n \cap A_k) \\ = \sum_{i=1}^n P(A_i) - \sum_{j=1,j \ne k}^{n}P(A_j \cap A_k) P(i=1⋃nBi)≤i=1∑nP(Bi)=P(Ak)+P(A1−Ak)+⋯+P(Ak−1−Ak)+P(Ak+1−Ak)+⋯+P(An−Ak)=P(Ak)+P(A1)−P(A1∩Ak)+⋯+P(Ak−1)−P(Ak−1∩Ak)+P(Ak+1)−P(Ak+1∩Ak)+⋯+P(An)−P(An∩Ak)=i=1∑nP(Ai)−j=1,j=k∑nP(Aj∩Ak)
这个就是Gunias不等式,也可以将其写成: P ( ⋃ i = 1 n A i ) ≤ ∑ i = 1 n P ( A i ) − sup k ∈ { 1 , ⋯ , n } ∑ j = 1 , j ≠ k n P ( A j ∩ A k ) P(\bigcup_{i=1}^n A_i) \le \sum_{i=1}^n P(A_i) - \sup_{k \in \{1,\cdots,n\}} \sum_{j=1,j \ne k}^{n}P(A_j \cap A_k) P(i=1⋃nAi)≤i=1∑nP(Ai)−k∈{1,⋯,n}supj=1,j=k∑nP(Aj∩Ak)
Boole不等式给出的至少有一个事件发生的概率的下界非常粗糙,甚至都不能保证下界是非负的,但钟开莱-艾尔迪希不等式给出的下界一定是非负的: P ( ⋃ i = 1 n A i ) ≥ ( ∑ i = 1 n P ( A i ) ) 2 ∑ i , j = 1 n P ( A i ∩ A j ) P(\bigcup_{i=1}^n A_i) \ge \frac{\left( \sum_{i=1}^n P(A_i) \right)^2}{\sum_{i,j=1}^n P(A_i \cap A_j)} P(i=1⋃nAi)≥∑i,j=1nP(Ai∩Aj)(∑i=1nP(Ai))2
证明 当 n = 2 n=2 n=2时,钟开莱-艾尔迪希不等式为 P ( A 1 ∪ A 2 ) ≥ P ( A 1 ) 2 + P ( A 2 ) 2 + 2 P ( A 1 ) P ( A 2 ) P ( A 1 ) + P ( A 2 ) + 2 P ( A 1 ∩ A 2 ) ⇐ [ P ( A 1 ) + P ( A 2 ) + 2 P ( A 1 ∩ A 2 ) ] P ( A 1 ∪ A 2 ) ≥ P ( A 1 ) 2 + P ( A 2 ) 2 + 2 P ( A 1 ) P ( A 2 ) ⇐ P ( A 1 ) [ P ( A 1 ∪ A 2 ) − P ( A 1 ) ] + P ( A 2 ) [ P ( A 1 ∪ A 2 ) − P ( A 2 ) ] + 2 [ P ( A 1 ∩ A 2 ) P ( A 1 ∪ A 2 ) − P ( A 1 ) P ( A 2 ) ] ≥ 0 ⇐ 2 P ( A 1 ) P ( A 2 ) − [ P ( A 1 ) + P ( A 2 ) ] P ( A 1 ∩ A 2 ) + 2 [ P ( A 1 ∩ A 2 ) P ( A 1 ∪ A 2 ) − P ( A 1 ) P ( A 2 ) ] ≥ 0 ⇐ [ 2 P ( A 1 ∪ A 2 ) − ( P ( A 1 ) + P ( A 2 ) ) ] P ( A 1 ∩ A 2 ) ≥ 0 P(A_1 \cup A_2) \ge \frac{P(A_1)^2+ P(A_2)^2 + 2P(A_1)P(A_2)}{P(A_1) + P(A_2) + 2P(A_1 \cap A_2)} \\ \Leftarrow [P(A_1) + P(A_2) + 2P(A_1 \cap A_2)]P(A_1 \cup A_2) \ge P(A_1)^2+ P(A_2)^2 + 2P(A_1)P(A_2) \\ \Leftarrow P(A_1)[P(A_1 \cup A_2)-P(A_1)] + P(A_2)[P(A_1 \cup A_2)-P(A_2)] \\+2[P(A_1 \cap A_2)P(A_1 \cup A_2)-P(A_1)P(A_2)] \ge 0 \\ \Leftarrow 2P(A_1)P(A_2) - [P(A_1)+P(A_2)]P(A_1\cap A_2) \\ +2[P(A_1 \cap A_2)P(A_1 \cup A_2)-P(A_1)P(A_2)] \ge 0 \\ \Leftarrow [2P(A_1\cup A_2)-(P(A_1)+P(A_2))]P(A_1 \cap A_2) \ge 0 P(A1∪A2)≥P(A1)+P(A2)+2P(A1∩A2)P(A1)2+P(A2)2+2P(A1)P(A2)⇐[P(A1)+P(A2)+2P(A1∩A2)]P(A1∪A2)≥P(A1)2+P(A2)2+2P(A1)P(A2)⇐P(A1)[P(A1∪A2)−P(A1)]+P(A2)[P(A1∪A2)−P(A2)]+2[P(A1∩A2)P(A1∪A2)−P(A1)P(A2)]≥0⇐2P(A1)P(A2)−[P(A1)+P(A2)]P(A1∩A2)+2[P(A1∩A2)P(A1∪A2)−P(A1)P(A2)]≥0⇐[2P(A1∪A2)−(P(A1)+P(A2))]P(A1∩A2)≥0
显然 2 P ( A 1 ∪ A 2 ) − ( P ( A 1 ) + P ( A 2 ) ) ≥ 0 2P(A_1\cup A_2)-(P(A_1)+P(A_2)) \ge 0 2P(A1∪A2)−(P(A1)+P(A2))≥0,因此当 n = 2 n=2 n=2时,钟开莱-艾尔迪希不等式成立。
假设 n = k n=k n=k时,钟开莱-艾尔迪希不等式成立,则 n = k + 1 n=k+1 n=k+1时, P ( ⋃ i = 1 k + 1 A i ) = P ( A k + 1 ∪ ⋃ i = 1 k A i ) ≥ P ( A k + 1 ) 2 + P ( ⋃ i = 1 k A i ) 2 + 2 P ( A 1 ) P ( ⋃ i = 1 k A i ) P ( A k + 1 ) + P ( ⋃ i = 1 k A i ) + 2 P ( A k + 1 ∩ ⋃ i = 1 k A i ) P(\bigcup_{i=1}^{k+1} A_i) = P(A_{k+1} \cup \bigcup_{i=1}^k A_i ) \ge \frac{P(A_{k+1})^2+ P(\bigcup_{i=1}^k A_i )^2 + 2P(A_1)P(\bigcup_{i=1}^k A_i )}{P(A_{k+1}) + P(\bigcup_{i=1}^k A_i ) + 2P(A_{k+1} \cap \bigcup_{i=1}^k A_i )} P(i=1⋃k+1Ai)=P(Ak+1∪i=1⋃kAi)≥P(Ak+1)+P(⋃i=1kAi)+2P(Ak+1∩⋃i=1kAi)P(Ak+1)2+P(⋃i=1kAi)2+2P(A1)P(⋃i=1kAi)
根据Boole不等式, P ( A k + 1 ) + P ( ⋃ i = 1 k A i ) + 2 P ( A k + 1 ∩ ⋃ i = 1 k A i ) ≤ P ( A k + 1 ) + ∑ i = 1 k P ( A i ) + 2 ∑ i = 1 k P ( A k + 1 ∩ A i ) ≤ ∑ i , j = 1 k + 1 P ( A i ∩ A j ) P(A_{k+1}) + P(\bigcup_{i=1}^k A_i ) + 2P(A_{k+1} \cap \bigcup_{i=1}^k A_i ) \\ \le P(A_{k+1}) + \sum_{i=1}^k P(A_i) + 2\sum_{i=1}^k P(A_{k+1} \cap A_i)\le \sum_{i,j=1}^{k+1} P(A_i \cap A_j) P(Ak+1)+P(i=1⋃kAi)+2P(Ak+1∩i=1⋃kAi)≤P(Ak+1)+i=1∑kP(Ai)+2i=1∑kP(Ak+1∩Ai)≤i,j=1∑k+1P(Ai∩Aj)
对 P ( ⋃ i = 1 k A i ) P(\bigcup_{i=1}^k A_i ) P(⋃i=1kAi)使用钟开莱-艾尔迪希不等式,说明分子大于 ( ∑ i = 1 k + 1 P ( A i ) ) 2 \left( \sum_{i=1}^{k+1} P(A_i) \right)^2 (∑i=1k+1P(Ai))2,这一步留给读者自己计算。综上,钟开莱-艾尔迪希不等式对所有的自然数都成立。