1
play

1 Bin(10, 0.3), Bin(100, 0.03) vs. Poi(3) Tender (Central) Moments - PDF document

Feb 17, 2024 •140 likes •191 views

Whither the Binomial Binomial in the Limit Recall example of sending bit string over network Recall the Binomial distribution n ! n = 4 bits sent over network where each bit had i n i ( ) ( 1 ) P X i

  1. Whither the Binomial… Binomial in the Limit • Recall example of sending bit string over network • Recall the Binomial distribution n !  n = 4 bits sent over network where each bit had     i n i ( ) ( 1 ) P X i p p  independent probability of corruption p = 0.1 ! ( )! i n i • Let l = np (equivalently: p = l / n )  X = number of bit corrupted. X ~ Bin(4, 0.1)   In real networks, send large bit strings (length n  10 4 )  l  i   l  n i l  l    i n ! ( 1 / ) n n ( n 1 )...( n i 1 ) n        P ( X i ) 1  Probability of bit corruption is very small p  10 -6   l     i i i ! ( n i )! n n n i ! ( 1 / n )  X ~ Bin(10 4 , 10 -6 ) is unwieldy to compute • When n is large, p is small, and l is “moderate”: • Extreme n and p values arise in many cases    ( 1 )...( 1 ) n n n i  l   l   l  n n i 1 ( 1 / ) e ( 1 / n ) 1  # bit errors in file written to disk (# of typos in a book) i n  # of elements in particular bucket of large hash table l  l l i i e  l    • Yielding: P ( X i ) 1 e  # of servers crashes in a day in giant data center i ! 1 i !  # Facebook login requests that go to particular server Poisson Random Variable Sending Data on Network Redux • X is a Poisson Random Variable: X ~ Poi( l ) • Recall example of sending bit string over network  X takes on values 0, 1, 2…  Send bit string of length n = 10 4  and, for a given parameter l > 0,  Probability of (independent) bit corruption p = 10 -6  X ~ Poi( l = 10 4 * 10 -6 = 0.01)  has distribution (PMF): l i  What is probability that message arrives uncorrupted?    l P ( X i ) e l i 0 ( 0 . 01 ) i !    l    0 . 01 P ( X 0 ) e e 0 . 990049834 l l l  l 0 1 2 i ! 0 ! i  l      • Note Taylor series: ... e 0 ! 1 ! 2 ! ! i 4 , 10 -6 ):   Using Y ~ Bin(10 i 0 l l    i i         l   l   l l  P ( Y 0 ) 0 . 990049829 • So: P ( X i ) e e e e 1 i ! i !    Caveat emptor: Binomial computed with built-in function in R software i 0 i 0 i 0 package, so some approximation may have occurred. Approximation are closer to you than they may appear in some software packages. Simeon-Denis Poisson Poisson Random is Binomial in Limit • Simeon-Denis Poisson (1781-1840) was a prolific • Poisson approximates Binomial where n is large, p is small, and l = np is “moderate” French mathematician • Different interpretations of "moderate"  n > 20 and p < 0.05  n > 100 and p < 0.1 • Really, Poisson is Binomial as • Published his first paper at 18, became professor n   and p  0, where np = l at 21, and published over 300 papers in his life  He reportedly said “Life is good for only two things, discovering mathematics and teaching mathematics.” • Definitely did not look like Charlie Sheen 1

  2. Bin(10, 0.3), Bin(100, 0.03) vs. Poi(3) Tender (Central) Moments with Poisson • Recall: Y ~ Bin( n , p )  E[Y] = np  Var(Y) = np (1 – p ) • X ~ Poi( l ) where l = np ( n   and p  0) P(X = k )  E[X] = np = l  Var(X) = np (1 – p ) = l (1 – 0) = l  Yes, expectation and variance of Poisson are same o It brings a tear to my eye…  Recall: Var(X) = E[X 2 ] – (E[X]) 2  E[X 2 ] = Var(X) + (E[X]) 2 = l + l 2 = l(1 + l) k It’s Really All About Raisin Cake CS = Baking Raisin Cake With Code • Hash tables  strings = raisins  buckets = cake slices • Server crashes in data center • Bake a cake using many raisins and lots of batter  servers = raisins • Cake is enormous (in fact, infinitely so…)  list of crashed machines = particular slice of cake  Cut slices of “moderate” size (w.r.t. # raisins/slice) • Facebook login requests (i.e., web server requests)  Probability p that a particular raisin is in a certain slice  requests = raisins is very small ( p = 1/# cake slices)  server receiving request = cake slice • Let X = number of raisins in a certain cake slice # raisins l  • X ~ Poi( l ), where # cake slices Defective Chips Efficiently Computing Poisson • Let X ~ Poi( l ) • Computer chips are produced  p = 0.1 that a chip is defective  Want to compute P( X = i ) for multiple values of i a      Consider a sample of n = 10 chips  E.g., Computing P ( X a ) P ( X i )   What is P(sample contains  1 defective chip)? i 0 • Iterative formulation:  Compute P(X = i + 1) from P(X = i)  Using Y ~ Bin(10, 0.1):    l l   l i 1 P ( X i 1 ) e /( i 1 )!       10 10   l l            i 0 10 1 9 P ( X i ) e / i ! i 1 P ( Y 1 ) ( 0 . 1 ) ( 1 0 . 1 ) ( 0 . 1 ) ( 1 0 . 1 ) 0 . 7361      0   1   Use recurrence relation:  Using X ~ Poi( l = (0.1)(10) = 1) l l 0      l P ( X 0 ) e e 0 1 1 1 0 !         1 1 1 ( 1 ) 2 0 . 7358 P X e e e l     0 ! 1 ! P ( X i 1 ) P ( X i )  i 1 2

  3. Approximately Poisson Approximation Birthday Problem Redux • Poisson can still provide good approximation • What is the probability that of n people, none share even when assumptions “mildly” violated the same birthday (regardless of year)?   • “Poisson Paradigm” n trials, one for each pair of people ( x , y ), x  y    n =    2  • Can apply Poisson approximation when...  Let E x,y = x and y have same birthday (trial success)  “Successes” in trials are not entirely independent  P(E x,y ) = p = 1/365 (note: all E x,y not independent) o Example: # entries in each bucket in large hash table    n 1 ( 1 ) n n    X ~ Poi( l ) where l      Probability of “Success” in each trial varies (slightly)   2 365 730 o Small relative change in a very small p  0 ( n ( n 1 ) / 730 )        n ( n 1 ) / 730 n ( n 1 ) / 730 o Example: average # requests to web server/sec. may fluctuate P ( X 0 ) e e 0 ! slightly due to load on network     Solve for smallest integer n , s.t.: n ( n 1 ) / 730 e 0 . 5         n ( n 1 ) / 730 ln( e ) ln( 0 . 5 ) n ( n 1 ) 730 ln( 0 . 5 ) n 23  Same as before! Poisson Processes Web Server Load • Consider “rare” events that occur over time • Consider requests to a web server in 1 second  Earthquakes, radioactive decay, hits to web server, etc.  In past, server load averages 2 hits/second  Have time interval for events (1 year, 1 sec, whatever...)  X = # hits server receives in a second  Events arrive at rate: l events per interval of time  What is P(X = 5)? • Split time interval into n   sub-intervals • Model  Assume at most one event per sub-interval  Assume server cannot acknowledge > 1 hit/msec.  Event occurrences in sub-intervals are independent  1 sec = 1000 msec. (= large n )  With many sub-intervals, probability of event occurring  P(hit server in 1 msec) = 2/1000 (= small p ) in any given sub-interval is small  X ~ Poi( l = 2) • N(t) = # events in original time interval ~ Poi( l ) 5 2     2 ( 5 ) 0 . 0361 P X e 5 ! Geometric Random Variable Negative Binomial Random Variable • X is Geometric Random Variable: X ~ Geo( p ) • X is Negative Binomial RV: X ~ NegBin( r , p )  X is number of independent trials until first success  X is number of independent trials until r successes  p is probability of success on each trial  p is probability of success on each trial  X takes on values 1, 2, 3, …, with probability:  X takes on values r , r + 1, r + 2 …, with probability:        n 1 ( ) ( 1 ) n 1 P X n p p         r n r P ( X n )   p ( 1 p ) , where n r , r 1 ,...    r 1 Var(X) = (1 – p )/ p 2  E[X] = 1/ p Var(X) = r (1 – p )/ p 2  E[X] = r / p • Examples: • Note: Geo( p ) ~ NegBin(1, p )  Flipping a fair ( p = 0.5) coin until first “heads” appears. • Examples:  Urn with N black and M white balls. Draw balls (with replacement, p = N/(N + M)) until draw first black ball.  # of coin flips until r- th “heads” appears  Generate bits with P(bit = 1) = p until first 1 generated  # of strings to hash into table until bucket 1 has r entries 3

Recommend

More recommend