User:Carl Boettiger/Notebook/Stochastic Population Dynamics/2010/02/24

{| width="800"
 * style="background-color: #EEE"|[[Image:owwnotebook_icon.png|128px]] Stochastic Population Dynamics
 * style="background-color: #F2F2F2" align="center"|  |Main project page
 * style="background-color: #F2F2F2" align="center"|  |Main project page


 * colspan="2"|
 * colspan="2"|

Derin & Alan Meeting
1-2pm
 * Discussion of reasons early warning signals fail
 * 1) Insufficient signal-to-noise ratio for early warning
 * 2) Ensemble vs time average -- satisfying ergodicity assumption?
 * 3) Role of bifurcation type.  existing literature focuses mostly on 1D saddle node.
 * 4) Role of chaos (certainly frustrates warning signals, but not required).
 * How general are warning signals? Normal forms of bifurcations.  Fitting a good model vs. general warning signals.  Problem of extrapolation.
 * Classification scheme of warning signals -- TREE paper(?)
 * Mean First passage time from potential

Expected Variance across Ensembles
The expected variance across ensembles can be calculated from the Fokker-Planck Equation derived from the linear noise approximation. This should agree with the variance computed over a time window when the system has reached stationarity. I do these calculations and compare:

Equilibrium population size solves

$$ e n - a = \frac{e K n^2}{n^2+h^2} $$

Equilibrium variance should be given by

$$ \hat \sigma^2 = \frac{d(\hat n)}{b'(\hat n)-d'(\hat n)} = \frac{e \hat n - a}{ 2 e K \hat n (1 - \frac{\hat n}{\hat n^2 + h^2}) - e} $$

Time averaging from simulation doesn't match analytic prediction!
The analytic solution to $$\hat n $$ is the general solution to a cubic so not very pretty, but easy to calculate numerically in order to give the variance estimate; using the values below I confirm that ODE solver and analytic solution above agree ($$ \hat \sigma^2 = 1352$$), which is only about half that computed over the time window (approx 2360)! Parameter values and solutions from analytics are:

Test of higher order corrections
I try the higher order correction to the mean dynamics (accounts for inflation effect of the variance):

$$ \frac{d}{dt} E(n) = \alpha_1(n) + \frac{1}{2} \sigma^2 \alpha_1''(n) $$

Though not surprisingly, the effect is negligible. In this case the average is slightly higher (572.5) and variance slightly lower (1349), so the curvature at equilibrium must be positive (in contrast to the deflation example with logistic or Levins model). Perhaps this is a consequence of the ergodicity assumption failing; I'll check against simply computing ensemble variance rather than the (mean-over-ensembles) time-averaged variance.

Ensemble averaging simulation does match theory, though time averaging doesn't

 * Okay, variance computed over ensembles only matches the theory (averaging over all time over an ensemble of 1000 I'm getting 1362 ensemble variance, while the time averaged (over a sampletime of 100, sample interval of .1, and max time 500) variance averaged over ensembles comes out a steady 2339. So much for ergodicity?  Note: these conditions should match the code found in subversion revision 25.

Reflections on Warning Signals literature

 * Created a separate Mendeley collection for literature on Early Warning Signals, still very underpopulated. Note that my Mendeley library wasn't available online for a couple days due to some stupid syncing mistakes I made, but the Mendeley support team was fantastic in getting things back in working order (Let's just say I may be responsible for the release of 0.9.6.1).  Note that one can subscribe to this collection as an rss feed.


 * I'm bothered to see the phrase "tipping point" used synonymously with bifurcation -- they invoke very different ideas. A tipping point such as the maximum angle you can push a canoe before it flips is not a bifurcation, it is a perturbation to the state directly large enough to push it over the seperatrix, and directly related to the first passage time problem in the stochastic models.  The bifurcations under consideration are changes in a parameter, not the state of interest, which destroys a stable point.  Phase transitions are examples, though often lacking hysteresis.


 * Another annoying quality is the lack of distinguishing bifurcation from rare event. If alternate stable states are present, what is the probability we would simply jump into them by chance?


 * Detrending before fitting AR(1) model to data is also unjustified. Trends influence variance as well!


 * Taking a sliding window size equal to half the size of the available time series is arbitrary and unjustified, as it is neither the most promising way to detect the trend nor should the window size choice depend on such things as knowing the time the crash occurs.

Current Library
(Also a chance to test Mendeley library embedding)   Research collected using Mendeley

Subversion Log
r25 | cboettig | 2010-02-25 00:12:09 -0800 (Thu, 25 Feb 2010) | 1 line

made samplefreq into a double instead of size_t, still co-opting the autocorrelations, seems to show successfully a different ensemble vs time-averaged variance

r24 | cboettig | 2010-02-24 23:46:07 -0800 (Wed, 24 Feb 2010) | 2 lines

co-opted the autoregressive data for the moment to explore direct ensemble mean and variance (without any time-averaging)

r23 | cboettig | 2010-02-24 21:53:12 -0800 (Wed, 24 Feb 2010) | 2 lines

ode integrator seems to accurately compute mean and variance dynamics, though the time-averaged variance doesn't seem to agree with the analytic ensemble variance; will have to compute that directly from simulation to compare. Runge-Kutta solver on adaptive mesh seems significantly slower than my simple Euler scheme; particularly with the (nearly irrelevant) higher order correction to the mean dynamics.


 * }