conjugate - city university of new york
TRANSCRIPT
Conjugate Priors-
Case of Normal,@-uI
Recall Normal density feat e 202
EK3=µ T×= o' x~N(ye, f)
If Xi , Xz, - - - , Xn are independentand Xin N(Mi , E)
then II. aini - NCE airi , Eairi)[ sum of independent normals is normal ]
Consider the problem
f-Calm = re- CETI T is known
µ is unknown
manner. " a.at:1?.;::re7hnfWhen we don't know somethingwe assign a prior to it .
Let f-fu) be our prior .
What can you say about µ if you are observing z?
ffrtnl-zf.gg#fa Difficulty Denominator !
From that point on , you can find
PIM >Nola) , E [Mla] , etc . . .
How to compute the integral ? [it could be hard]
Solution : Choose ffa) to make the math work !
Is this a limitation ?
Justification for solution above :
mnmv_choice of prior does not matter in practice ifwe have a large amount of observations .
- [typical] choose a prior that will produce a
posterior of the same form as the prior . [conjugate]
( update parameters without changing the model )-- =
prior& Flobserration
(original parameters)
Flute) = SCHAALSffafutffndm
say f-(pi ) = 1- Causbyit feltfu-PT)+0
Denominator : ftp.oe-EI?*#.,.jdv..- A
Observation : If Mr N (p , E) o then µ In nNYT)new parameters
since posterior will have same formas prior, we say the prior is conjugate
f-Calm) f- (a)
firm =
Tp&f none•
wForm 2-=- form 1
prior is conjugate to Normed
Nerd is conjugate to Normed
conjugate prior means : you figured out the integral !
false, =t.EE#.f....e-4ttI-
pDrop
constants .
n:::::i e.ee#.e.az#fHntxft4HfH)we say f-Ca ) x god iff f = c. fee)
a"
proportional to"
constant
theres a e- 4¥
.
e- HE22(n-a)'
+ Tcu-pj= e
-
Else'
tu'- 2am) +04µFp2- yup)
= e-#
Zora 2
= e
-
2. OZ 2
-
@7- E)M2- zu (En + Tp)x e s
= e
-
I.
TIZZ
- e
le - 3¥32-#
2 They
= e
-
0422
where m= Entry¥2
VIRZI92+22
mix - NCEEEI.EE. )
More data : Ni , Zz, - - - , Nn n i. i. d . Conditioned onM\- -
- - -- - - - - - - - - - - - -
"
f-feel a , - - - inn) & Ffa , - - - pen In) ffu) i
= II faints- -- - -
-
-
"
xi÷e- EEE. e- a¥
What if we look at E = &?ci/n (average)a- IM ~ N(µ , of) [sum of independent Normals)
ffntx) - Ftalntffu )g kN(pic)Nfa, %)
rhi - NC , )If we work on the problem f-(MIM,. . - -xn)
we will find f- (a) oh , -→ an ) = flu II) .
n-zoo : Hula ) n NCE, %)
Prior"doesn't matter
"
Thu
Observe N- e- no and let y be such that
ly - not → a
trainee -
- ftp.?;f?ffqY-yjAssumeHIFlyin = no) omyaxfucy)> °
> a :c:ti÷i÷=a:Ei÷:i::Yi÷- ÷Chebyshev PC tri -Ml > e) say → o