Dear Aidan,
Aidan Budd wrote on Jul 23
rd, 2008 at 5:03am:
One reason my question was confusing is that I mis-typed something - I meant to say that when looking at other descriptions of the 12-model matrix (e.g. in Swofford et al.), the non-reversible matrix has q_ij = r_ij * pi_j even if r_ji =/= r_ij - and I understood from your message that
(which I thought meant that the decomposition of q_ij into s_ij * pi_j was not possible for these matrix).
I am pretty sure that that decomposition does NOT exist in general for non-reversible models (because rates and frequencies are now interdependent); I'll check the papers you quote to carefully examine their statements. I'm in an airport now and don't have much time to think about this carefully, but I'll come up with an example and post it here.
Quote:The reason why I'm asking along these lines is that I'm keen to get a reasonable idea of the implications (and interpretation) of non-reversible models. In particular, what I don't understand is how the equilibrium state of these models is maintained. For the reversible models, pi_i * q_ij = pi_j * q_ji for all i,j - so at the steady-state/equilibrium, in a given time interval, the same number of transitions occur from all i to all j [e.g. with a two-state model, the same number of 1->0 transitions occur as 0->1 transitions).
Do I understand right when I read that the above model is reversible, and that for a non-reversible model then, at equilibrium (and taking the binary state model), the number of 1->0 transitions is DIFFERENT from the number of 0->1 transitions? i.e. if for example in a binary model q_01 is greater than q_10 (i.e. r_10 > r_01), then at equilibrium the number of changes of state from 0->1 is greater than the number of changes from state 1->0.
Detailed equilibrium is not necessary for stationarity in general (for models with more that two states, i.e. you could for a three state model you can have more 1->2 than 2->1 transitions but that can be compensated for by transitions from the third state). For a two-state model generally the form q_ij = r_ij pi_j AND stationarity imply reversibility. Indeed, for
Q = [[ -r01 pi_1, r01 pi_1][r10 pi_0, -r10 pi_0]]
to have pi_0, pi_1 (=1 - pi_0) the following must hold:
-r01 pi_1 pi_0 + r10 pi_0 pi_1 = 0; in other words r01 = r10 must hold (hence it will be reversible).
For models with more than two states, you can have the q_ij = r_ij pi_ j representation only if there are some conditions or r_ij and pi_j.
HTH,
Sergei