This repository has been archived on 2021-05-17. You can view files and clone it, but cannot push or open issues or pull requests.
scientificComputing/simulations/lecture/simulations.tex

298 lines
13 KiB
TeX

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\chapter{Simulations}
\label{simulationschapter}
\exercisechapter{Simulations}
The real power of computers for data analysis is the possibility to
run simulations. Experimental data of almost unlimited sample sizes
can be simulated in no time. This allows to explore basic concepts,
the ones we introduce in the following chapters and many more, with
well controlled data sets that are free of confounding pecularities of
real data sets. With simulated data we can also test our own analysis
functions. More importantly, by means of simulations we can explore
possible outcomes of our experiments before we even started the
experiment. We could even explore possible results for regimes that we
cannot test experimentally. How dynamical systems, like for example
predator-prey interactions or the activity of neurons or whole brains,
evolve in time is a central application for simulations. The advent of
computers at the second half of the twentieth century pushed the
exciting field of nonlinear dynamical systems forward. Conceptually,
many kinds of simulations are very simple and are implemented in a few
lines of code.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Random numbers}
At the heart of many simulations are random numbers that we get from
\enterm[random number generator]{random number generators}. These are
numerical algorithms that return sequences of numbers that appear to
be as random as possible. If we draw random numbers using, for
example, the \code{rand()} function, then these numbers are indeed
uniformly distributed and have a mean of one half. Subsequent numbers
are independent of each other, i.e. the autocorrelation function is
zero everywhere except at lag zero. However, numerical random number
generators have a period, after which they repeat the exact same
sequence of numbers. This differentiates them from truely random
numbers and hence they are called \enterm[random number
generator!pseudo]{pseudo random number generators}. In rare cases this
periodicity can induce problems in simulations whenever more random
numbers than the period of the random number generator are
used. Luckily, nowadays the periods of random nunmber generators are
very large. Periods are at least in the range of $2^{64}$, that is
about $10^{18}$, or even larger.
The pseudo randomness of numerical random number generators also has
an advantage. They allow to repeat exactly the same sequence of
random numbers. After defining the state of the generator or setting a
\enterm{seed} with the \code{rng()} function, a particular sequence of
random numbers is generated by subsequent calls of the random number
generator. This way we can not only precisly define the statistics of
noise in our simulated data, but we can repeat an experiment with
exactly the same sequence of noise values. This is useful for plots
that involve some random numbers but should look the same whenever the
script is run.
\begin{exercise}{randomnumbers.m}{randomnumbers.out}
First, read the documentation of the \varcode{rand()} function and
check its output for some (small) input arguments.
Generate three times the same sequence of 10 uniformly distributed
numbers using the \code{rand()} and \code{rng()} functions.
Generate 10\,000 uniformly distributed random numbers and compute
the correlation coefficient between each number and the next one in
the sequence. This is the serial correlation at lag one.
\end{exercise}
\begin{figure}[t]
\includegraphics[width=1\textwidth]{randomnumbers}
\titlecaption{\label{randomnumbersfig} Random numbers.} {Numerical
random number generators return sequences of numbers that are as
random as possible. The returned values approximate a given
probability distribution. Here, for example, a uniform
distribution between zero and one (top left). The serial
correlation (auto-correlation) of the returned sequences is zero
at any lag except for zero (bottom left). However, by setting a
seed, defining the state, or spawning the random number generator,
the exact same sequence can be generated (right).}
\end{figure}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Univariate data}
A basic type of stochastic simulations is to draw random numbers from
a given distribution like, for example, the normal distribution. This
simulates repeated measurements of some quantity (weight of tigers,
firing rate of neurons). Doing so we must specify from which
probability distribution the data should originate from and what are
the parameters of that distribution (mean, standard deviation, shape
parameters). How to illustrate and quantify univariate data, no matter
whether they have been actually measured or whether they have been
simulated as described in the following, is described in
chapter~\ref{descriptivestatisticschapter}.
\subsection{Normally distributed data}
For drawing numbers from a normal distribution
\begin{equation}
p_{norm}(x) = \frac{1}{\sqrt{2\pi}}e^{-\frac{1}{2}x^2}
\end{equation}
with zero mean and unit standard deviation we use the \code{randn()}
function. This function returns normally distributed numbers $\xi_i$
with zero mean and unit standard deviation. For changing the standard
deviation we need to multiply the returned data values with the
required standard deviation $\sigma$. For changing the mean we just
add the desired mean $\mu$ to the random numbers:
\begin{equation}
x_i = \sigma \xi_i + \mu
\end{equation}
\begin{figure}[t]
\includegraphics[width=1\textwidth]{normaldata}
\titlecaption{\label{normaldatafig} Generating normally distributed
data.}{With the help of a computer the weight of 300 tigers can be
measured in no time using the \code{randn()} function (left). By
construction we then even know the population distribution (red
line, right), its mean (here 220\,kg) and standard deviation
(40\,kg) from which the simulated data values were drawn (yellow
histogram).}
\end{figure}
\begin{exercise}{normaldata.m}{normaldata.out}
Write a little script that generates $n=100$ normally distributed
data simulating the weight of Bengal tiger males with mean 220\,kg
and standard deviation 40\,kg. Check the actual mean and standard
deviation of the generated data. Do this, let's say, five times
using a for-loop. Then increase $n$ to 10\,000 and run the code
again. It is so simple to measure the weight of 10\,000 tigers for
getting a really good estimate of their mean weight, isn't it?
Finally plot from the last generated data set of tiger weights the
first 1000 values as a function of their index.
\end{exercise}
\subsection{Other probability densities}
We can draw random numbers not only from normal
distributions. Functions are provided that let you draw random numbers
of almost any probability distribution. They differ in their shape and
in the number of parameters they have. Most probability distributions
are parameterized by a location parameter that usually coincides with
their mean, and a scale parameter that often coincides with their
standard deviation. Some distributions have additional shape
parameters.
For example, the time intervals $t$ between randomly generated action
potentials (a so called Poisson spike train, see
chapter~\ref{pointprocesseschapter}) are exponentially distributed ---
as are the intervals between state transitions of ion channels, or the
intervals between radioactive decays. The exponential distribution is
only defined for positive time intervals:
\begin{equation}
\label{expdistribution}
p_{exp}(t) = \lambda e^{-\lambda t} \; , \quad t \ge 0, \; \lambda > 0
\end{equation}
A single rate parameter $\lambda$ measured in Hertz parameterizes the
exponential distribution. It defines how often per time unit the event
happens. Both the mean interval between the events and the
corresponding standard deviation equal the inverse rate.
\begin{exercise}{exprandom.m}{exprandom.out}
Draw $n=200$ random time intervals from an exponential distribution
with rate $\lambda=50$\,Hz. Calculate the mean and the standard
deviation of the random numbers and compare them with the expected
values.
\end{exercise}
\subsection{Uniformly distributed random numbers}
\begin{figure}[t]
Exponential distribution of ISIs.
\end{figure}
The \code{rand()} function returns uniformly distributed random
numbers between zero and one
\begin{equation}
p_{uni}(x) = \left\{\begin{array}{rcl} 1 & ; & 0 \le x < 1 \\
0 & ; & \text{else} \end{array}\right.
\end{equation}
For random numbers $x_i$ uniformly distributed between $x_{\rm min}$
and $x_{\rm max}$ we multiply by the desired range $x_{\rm max} -
x_{\rm min}$ and add $x_{\rm min}$ to the numbers $u_i$ returned by
\code{rand()}:
\begin{equation}
x_i = (x_{\rm max} - x_{\rm min}) u_i + x_{\rm min}
\end{equation}
\begin{exercise}{unirandom.m}{unirandom.out}
Simulate the times of $n=10\,000$ action potentials occuring
randomly within a time interval of 100\,s using the \code{rand()}
function. Sort the times using \code{sort()}, compute the time
intervals between the action potentials (interspike intervals, ISIs)
using the \code{diff()} function, plot a histogram of the interspike
intervals in milliseconds, and compute the mean and standard
deviation of the interspike intervals, also in milliseconds. Compare
the mean ISI with the firing rate (number of action potentials per
time).
\end{exercise}
\subsection{Simulating probabilities}
Simulating events that happen with some probability $P$ is also
possible. That could be the probability of head showing up when
flipping a coin, getting an action potential within some time, a ion
channel to open, ... For this draw a random number $u_i$ from a
uniform distribution between zero and one (\code{rand()}). If the
random number is lower than $P$, the event happens, if it is larger
the event does not occur.
\begin{exercise}{probability.m}{probability.out}
Let's flip a coin 20 times. The coin is biased and shows head with a
probability of $P=0.6$. Count the number of heads.
\end{exercise}
\subsection{Random walks}
A basic concept for stochastic models is the random walk. A walker
starts at some initial position $x_0$. It then takes steps to the
right, $x_n = x_{n-1} + 1$, with some probability $P_r$ or to the
left, $x_n = x_{n-1} - 1$, with probability $P_l = 1-P_r$.
For a symmetric random walk, the probabilities to step to the left or
to the right are the same, $P_r = P_l = \frac{1}{2}$. The average
position of many walkers is then independent of the iteration step and
equals the initial position $x_0$. The standard deviation of the
walker positions, however, grows with the square root of the number of
iteration steps $n$: $\sigma_{x_n} = \sqrt{n}$.
\begin{figure}[tb]
Have a figure with a few 1D random walks and the increasing standard
deviation as a shaded area behind. And with absorbing boundary.
\end{figure}
The sub-threshold membrane potential of a neuron can be modeled as a
one-dimensional random walk. The thousands of excitatory and
inhibitory synaptic inputs of a principal neuron in the brain make the
membrane potential to randomly step upwards or downwards. Once the
membrane potential hits a firing threshold an action potential is
generated and the membrane potential is reset to the resting
potential. The firing threshold in this context is often called an
``absorbing boundary''. The time it takes from the initial position to
cross the threshold is the ``first-passage time''. For neurons this is
the interspike interval. For symmetric random walks the first-passage
times are again exponentially distributed. ??? Is that so ???
% Gerstein and Mandelbrot 1964
Higher dimensions, Brownian motion.
The gamma distribution (\code{gamrnd()}) phenomenologically describes
various types of interspike interval dsitributions
(chapter~\ref{pointprocesseschapter}). scale and shape. exercise.
\subsection{Random integers}
\code{randi()}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Bivariate data}
\subsection{Static nonlinearities}
\begin{figure}[t]
\includegraphics[width=1\textwidth]{staticnonlinearity}
\titlecaption{\label{staticnonlinearityfig} Generating data
fluctuating around a function.}{The conductance of the
mechontransducer channels in hair cells of the inner ear is a
saturating function of the deflection of their hairs (left, red
line). Measured data will fluctuate around this function (blue
dots). Ideally the residuals (yellow histogram) are normally
distributed (right, red line).}
\end{figure}
Example: mechanotransduciton!
draw (and plot) random functions.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Dynamical systems}
\begin{itemize}
\item iterated maps
\item euler forward, odeint
\item introduce derivatives which are also needed for fitting (move box from there here)
\item Passive membrane
\item Add passive membrane to mechanotransduction!
\item Integrate and fire
\item Fitzugh-Nagumo
\item Two coupled neurons? Predator-prey?
\end{itemize}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Summary}
with outook to other simulations (cellular automata, monte carlo, etc.)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\printsolutions