Hw10 solutions

The important differences between Gillespie's first reaction and direct methods:

In the direct method, we picked one random number (uniformly on (0,1)) to generate the time of the next reaction (with exponential distribution parametrized by a = sum(a_i) ):

   # how long until the next reaction
   r1 = random.random()
   tau = (1.0/(a+eps))*math.log(1.0/r1)

Then we picked a second random number uniformly on (0,a) and identified which reaction's "bin" it fell into (where the width of each reactions's bin is proportional to its current propensity).

   r2a = random.random()*a
   a_sum = 0
   for i in a_i:
       if r2a < (a_i[i]+a_sum):
           mu = i
           break
       a_sum += a_i[i]

Where a_sum is the "right edge" of the current bin you're considering.

In the first reaction method, one picks a random number (uniform on 0,1) and uses it to generate a a time (exponentially distributed with parameter a_i) for each reaction, and selects the reaction with the smallest time to fire at that time. Tau's for all reactions are recalculated at every step, never saved.

   mintau = t_max
   # which reaction will happen first?
   # caluculate each reaction's time based on a different random number
   for rxn in a_i.keys():
       ai = a_i[rxn]
       ri = random.random()
       taui = (1.0/(ai+eps)) * math.log(1.0/ri)
       if taui < mintau: # "sort as you go"
           mu = rxn     # the putative first rxn 
           tau = taui   # the putative first time
           mintau = tau # reset the min

This method of sorting seemed fastest to us, but maybe storing the tau's in a dictionary and sorting the values would have been faster, like this:

   tau_i = {}
   # which reaction will happen first?
   # caluculate each reaction's time based on a different random number
   for rxn in a_i.keys():
       ai = a_i[rxn]
       ri = random.random()
       taui = (1.0/(ai+eps)) * math.log(1.0/ri)
       tau_i[taui] = rxn

   tau = min(tau_i.keys())
   mu = tau_i[tau]

Here's the code implementing the whole first reaction method. http://web.mit.edu/~soniat/Public/firstrxn.py