Apêndice 6.A Seção 2.4 (Exemplos de Modelo Probabilístico em um DAG)

Prova do Lema 2.18.
f(v1,v3|v2)\displaystyle f(v_{1},v_{3}|v_{2}) =f(v1,v2,v3)f(v2)\displaystyle=\frac{f(v_{1},v_{2},v_{3})}{f(v_{2})}
=f(v2)f(v1|v2)f(v3|v2)f(v2)\displaystyle=\frac{f(v_{2})f(v_{1}|v_{2})f(v_{3}|v_{2})}{f(v_{2})}
=f(v1|v2)f(v3|v2)\displaystyle=f(v_{1}|v_{2})f(v_{3}|v_{2})

Prova do Lema 2.19.

Considere que V2Bernoulli(0.02)V_{2}\sim\text{Bernoulli}(0.02). Além disso, V1,V3{0,1}V_{1},V_{3}\in\{0,1\} são independentes dado V2V_{2}. Também, (V1=1|V2=1)=(V3=1|V2=1)=0.9{\mathbb{P}}(V_{1}=1|V_{2}=1)={\mathbb{P}}(V_{3}=1|V_{2}=1)=0.9 e (V1=1|V2=0)=(V3=1|V2=0)=0.05{\mathbb{P}}(V_{1}=1|V_{2}=0)={\mathbb{P}}(V_{3}=1|V_{2}=0)=0.05. Note que, por construção, {\mathbb{P}} é compatível com figur 2. Isto é, P(v1,v2,v3)=(v2)(v1|v2)(v3|v2)P(v_{1},v_{2},v_{3})={\mathbb{P}}(v_{2}){\mathbb{P}}(v_{1}|v_{2}){\mathbb{P}}(% v_{3}|v_{2}). Além disso,

(V1=1)\displaystyle{\mathbb{P}}(V_{1}=1) =(V1=1,V2=1)+(V1=1,V2=0)\displaystyle={\mathbb{P}}(V_{1}=1,V_{2}=1)+{\mathbb{P}}(V_{1}=1,V_{2}=0)
=(V2=1)(V1=1|V2=1)+(V2=0)(V1=1|V2=0)\displaystyle={\mathbb{P}}(V_{2}=1){\mathbb{P}}(V_{1}=1|V_{2}=1)+{\mathbb{P}}(% V_{2}=0){\mathbb{P}}(V_{1}=1|V_{2}=0)
=0.020.9+0.980.05=0.067\displaystyle=0.02\cdot 0.9+0.98\cdot 0.05=0.067

Por simetria, (V3=1)=0.067{\mathbb{P}}(V_{3}=1)=0.067. Além disso,

(V1=1,V3=1)\displaystyle{\mathbb{P}}(V_{1}=1,V_{3}=1) =(V1=1,V3=1,V2=1)+(V1=1,V3=1,V2=0)\displaystyle={\mathbb{P}}(V_{1}=1,V_{3}=1,V_{2}=1)+{\mathbb{P}}(V_{1}=1,V_{3}% =1,V_{2}=0)
=(V2=1)(V1=1|V2=1)(V3=1|V2=1)+(V2=0)(V1=1|V2=0)(V3=1|V2=0)\displaystyle={\mathbb{P}}(V_{2}=1){\mathbb{P}}(V_{1}=1|V_{2}=1){\mathbb{P}}(V% _{3}=1|V_{2}=1)+{\mathbb{P}}(V_{2}=0){\mathbb{P}}(V_{1}=1|V_{2}=0){\mathbb{P}}% (V_{3}=1|V_{2}=0)
=0.020.90.9+0.980.050.05=0.01865\displaystyle=0.02\cdot 0.9\cdot 0.9+0.98\cdot 0.05\cdot 0.05=0.01865

Como (V1=1)(V3=1)=0.0670.0670.00450.01865=(V1=1,V3=1){\mathbb{P}}(V_{1}=1){\mathbb{P}}(V_{3}=1)=0.067\cdot 0.067\approx 0.0045\neq 0% .01865={\mathbb{P}}(V_{1}=1,V_{3}=1), temos que V1V_{1} e V3V_{3} não são independentes. ∎

Prova do Lema 2.20.
f(v3|v1,v2)\displaystyle f(v_{3}|v_{1},v_{2}) =f(v1,v2,v3)f(v1,v2)\displaystyle=\frac{f(v_{1},v_{2},v_{3})}{f(v_{1},v_{2})}
=f(v1)f(v2|v1)f(v3|v2)f(v1)f(v2|v1)\displaystyle=\frac{f(v_{1})f(v_{2}|v_{1})f(v_{3}|v_{2})}{f(v_{1})f(v_{2}|v_{1% })}
=f(v3|v2)\displaystyle=f(v_{3}|v_{2})

Prova do Lema 2.21.

Considere que V1Bernoulli(0.5)V_{1}\sim\text{Bernoulli}(0.5), (V2=1|V1=1)=0.9{\mathbb{P}}(V_{2}=1|V_{1}=1)=0.9, (V2=1|V1=0)=0.05{\mathbb{P}}(V_{2}=1|V_{1}=0)=0.05, (V3=1|V2=1,V1)=0.9{\mathbb{P}}(V_{3}=1|V_{2}=1,V_{1})=0.9, e (V3=1|V2=0,V1)=0.05{\mathbb{P}}(V_{3}=1|V_{2}=0,V_{1})=0.05. Note que (V1,V2,V3)(V_{1},V_{2},V_{3}) formam uma Cadeia de Markov. Note que, por construção, {\mathbb{P}} é compatível com figur 3. Isto é, P(v1,v2,v3)=(v1)(v2|v1)(v3|v2)P(v_{1},v_{2},v_{3})={\mathbb{P}}(v_{1}){\mathbb{P}}(v_{2}|v_{1}){\mathbb{P}}(% v_{3}|v_{2}). Além disso,

(V3=1)\displaystyle{\mathbb{P}}(V_{3}=1) =(V1=0,V2=0,V3=1)+(V1=0,V2=1,V3=1)\displaystyle={\mathbb{P}}(V_{1}=0,V_{2}=0,V_{3}=1)+{\mathbb{P}}(V_{1}=0,V_{2}% =1,V_{3}=1)
+(V1=1,V2=0,V3=1)+(V1=1,V2=1,V3=1)\displaystyle+{\mathbb{P}}(V_{1}=1,V_{2}=0,V_{3}=1)+{\mathbb{P}}(V_{1}=1,V_{2}% =1,V_{3}=1)
=0.50.90.05+0.50.050.9\displaystyle=0.5\cdot 0.9\cdot 0.05+0.5\cdot 0.05\cdot 0.9
+0.50.050.05+0.50.90.9=0.45125\displaystyle+0.5\cdot 0.05\cdot 0.05+0.5\cdot 0.9\cdot 0.9=0.45125

Além disso,

(V1=1,V3=1)\displaystyle{\mathbb{P}}(V_{1}=1,V_{3}=1) =(V1=1,V2=0,V3=1)+(V1=1,V2=1,V3=1)\displaystyle={\mathbb{P}}(V_{1}=1,V_{2}=0,V_{3}=1)+{\mathbb{P}}(V_{1}=1,V_{2}% =1,V_{3}=1)
=0.50.050.9+0.50.90.9=0.40625\displaystyle=0.5\cdot 0.05\cdot 0.9+0.5\cdot 0.9\cdot 0.9=0.40625

Como (V1=1)(V3=1)=0.50.451250.2260.40625=(V1=1,V3=1){\mathbb{P}}(V_{1}=1){\mathbb{P}}(V_{3}=1)=0.5\cdot 0.45125\approx 0.226\neq 0% .40625={\mathbb{P}}(V_{1}=1,V_{3}=1), temos que V1V_{1} e V3V_{3} não são independentes. ∎

Prova do Lema 2.22.
f(v1,v3)\displaystyle f(v_{1},v_{3}) =f(v1,v2,v3)𝑑v2\displaystyle=\int f(v_{1},v_{2},v_{3})dv_{2}
=f(v1)f(v3)f(v2|v1,v3)𝑑v2\displaystyle=\int f(v_{1})f(v_{3})f(v_{2}|v_{1},v_{3})dv_{2}
=f(v1)f(v3)f(v2|v1,v3)𝑑v2\displaystyle=f(v_{1})f(v_{3})\int f(v_{2}|v_{1},v_{3})dv_{2}
=f(v1)f(v3)\displaystyle=f(v_{1})f(v_{3})

Prova do Lema 2.23.

Considere que V1V_{1} e V3V_{3} são independentes e tem distribuição Bernoulli(0.5)\text{Bernoulli}(0.5). Além disso, V2V1+V3V_{2}\equiv V_{1}+V_{3}. Como (V3=1)=0.5{\mathbb{P}}(V_{3}=1)=0.5 e (V3=1|V1=1,V2=2)=1{\mathbb{P}}(V_{3}=1|V_{1}=1,V_{2}=2)=1, conclua que V1⟂̸V3|V2V_{1}\not\perp\!\!\!\!\perp V_{3}|V_{2}. ∎