.. _exact_integration:
Integrating neural models using exact integration
=================================================
The simple integrate-and fire model
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
For the simple integrate-and-fire model the voltage :math:`V` is given as a solution of the equation:
.. math::
C\frac{dV}{dt}=I.
This is just the derivative of the law of capacitance :math:`Q=CV`. When an input current is applied, the membrane voltage increases with time until it reaches a constant threshold :math:`V_{\text{th}}`, at which point a delta function spike occurs.
A shortcoming of the simple integrate-and-fire model is that it implements no time-dependent memory. If the model receives a below-threshold signal at some time, it will retain that voltage boost until it fires again. This characteristic is not in line with observed neuronal behavior.
The leaky integrate-and fire model
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In the leaky integrate-and-fire model, the memory problem is solved by adding a "leak" term :math:`\frac{-1}{R}V` (:math:`R` is the resistance and :math:`\tau=RC`) to the membrane potential:
.. math::
\frac{dV}{dt}=\frac{-1}{\tau}V+\frac{1}{C}I.
:label: membrane
This reflects the diffusion of ions that occurs through the membrane when some equilibrium is not reached in the cell.
Solving a homogeneous linear differential equation
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
To solve :math:numref:`membrane` we start by looking at a simpler differential equation:
.. math::
\frac{df}{dt}=af\text{, where } f:\mathbb{R}\to\mathbb{R} \text{ and } a\in\mathbb{R}.
Here the solution is given by :math:`f(t)=e^{at}`.
Solving a non-homogeneous linear differential equation
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
When you add another function :math:`g` to the right hand side of our linear differential equation,
.. math::
\frac{df}{dt}=af+g
this is now a non-homogeneous differential equation. Things (can) become more complicated.
Solving it with variation of constants
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
This kind of differential equation is usually solved with "variation of constants" which gives us the following solution:
.. math::
f(t)=e^{ct}\int_{0}^t g(s)e^{-cs}ds.
This is obviously not a particularly handy solution since calculating the integral in every step is very costly.
Solving it with exact integration
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
With exact integration, these costly computations can be avoided.
Restrictions to :math:`g`
-------------------------
But only for certain functions :math:`g`! I.e. if :math:`g` satisfies (is a solution of):
.. math::
\left(\frac{d}{dt}\right)^n g= \sum_{i=1}^{n}a_i\left(\frac{d}{dt}\right)^{i-1} g
for some :math:`n\in \mathbb{N}` and a sequence :math:`(a_i)_{i\in\mathbb{N}}\subset \mathbb{R}`.
For example this would be the case for :math:`g=\frac{e}{\tau_{syn}}t e^{-t/\tau_{\text{syn}}}` (an alpha funciton), where :math:`\tau_{\text{syn}}` is the rise time.
Reformulating the problem
^^^^^^^^^^^^^^^^^^^^^^^^^
The non-homogeneous differential equation is reformulated as a multidimensional homogeneous linear differential equation:
.. math::
\frac{d}{dt}y=Ay
where
.. math::
A=\begin{pmatrix}
a_{n} & a_{n-1} & \cdots & \cdots & a_1 & 0 \\
1 & 0 & \cdots & 0 & 0 & 0 \\
0 & \ddots & \ddots & \vdots & \vdots & \vdots \\
\vdots & \ddots & \ddots & 0 & 0 & 0 \\
0 & 0 & \ddots & 1 & 0 & 0 \\
0 & 0 & \cdots & 0 & \frac{1}{C} & -\frac{1}{\tau} \\
\end{pmatrix}
by choosing :math:`y_1,...,y_n` canonically as:
.. math::
\begin{align*}
y_1 &= \left(\frac{d}{dt}\right)^{n-1}g\\
\vdots &= \vdots\\
y_{n-1} &= \frac{d}{dt}g\\
y_{n} &= g\\
y_{n+1} &= f.
\end{align*}
This makes ist very easy to determine the solution as
.. math::
y(t)= e^{At}y_0
and
.. math::
y_{t+h}=y(t+h)=e^{A(t+h)}\cdot y_0=e^{Ah}\cdot e^{At}\cdot y_0=e^{Ah}\cdot y_t.
This means that once we have calculated :math:`A`, propagation consists of multiplications only.
Example: The leaky integrate and fire model with alpha-function shaped inputs (iaf_psc_alpha)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
The dynamics of the membrane potential :math:`V` is given by:
.. math::
\frac{dV}{dt}=\frac{-1}{\tau}V+\frac{1}{C}I
where :math:`\tau` is the membrane time constant and :math:`C` is the capacitance. :math:`I` is the sum of the synaptic currents and any external input:
Postsynaptic currents are alpha-shaped, i.e. the time course of the synaptic current :math:`\iota` due to one incoming spike is
.. math::
\iota (t)= \frac{e}{\tau_{syn}}t e^{-t/\tau_{\text{syn}}}.
The total input :math:`I` to the neuron at a certain time :math:`t` is the sum of all incoming spikes at all grid points in time :math:`t_i\le t` plus an additional piecewise constant external input :math:`I_{\text{ext}}`:
.. math::
I(t)=\sum_{i\in\mathbb{N}, t_i\le t }\sum_{k\in S_{t_i}}\hat{\iota}_k \frac{e}{\tau_{\text{syn}}}(t-t_i) e^{-(t-t_i)/\tau_{\text{syn}}}+I_{\text{ext}}
:math:`S_t` is the set of indices that deliver a spike to the neuron at time :math:`t`, :math:`\tau_{\text{syn}}` is the rise time and :math:`\iota_k` represents the "weight" of synapse :math:`k`.
Exact integration for the iaf_psc_alpha model
---------------------------------------------
First we make the substitutions:
.. math::
\begin{align*}
y_1 &= \frac{d}{dt}\iota+\frac{1}{\tau_{syn}}\iota \\
y_2 &= \iota \\
y_3 &= V
\end{align*}
for the equation
.. math::
\frac{dV}{dt}=\frac{-1}{Tau}V+\frac{1}{C}\iota
we get the homogeneous differential equation (for :math:`y=(y_1,y_2,y_3)^t`)
.. math::
\frac{d}{dt}y= Ay=
\begin{pmatrix}
\frac{1}{\tau_{syn}}& 0 & 0\\
1 & \frac{1}{\tau_{syn}} & 0\\
0 & \frac{1}{C} & -\frac {1}{\tau}
\end{pmatrix}
y.
The solution of this differential equation is given by :math:`y(t)=e^{At}y(0)` and can be solved stepwise for a fixed time step :math:`h`:
.. math::
y_{t+h}=y(t+h)=e^{A(t+h)}y(0)=e^{Ah}e^{At}y(0)=e^{Ah}y(t)=e^{Ah}y_t.
The complete update for the neuron can be written as
.. math::
y_{t+h}=e^{Ah}y_t + x_{t+h}
where
.. math::
x_{t+h}+\begin{pmatrix}\frac{e}{\tau_{\text{syn}}}\\0\\0\end{pmatrix}\sum_{k\in S_{t+h}}\hat{\iota}_k
as the linearity of the system permits the initial conditions for all spikes arriving at a given grid point to be lumped together in the term :math:`x_{t+h}`. :math:`S_{t+h}` is the set of indices :math:`k\in 1,....,K` of synapses that deliver a spike to the neuron at time :math:`t+h`.
The matrix :math:`e^{Ah}` in the C++ implementation of the model in NEST is constructed `here `_.
Every matrix entry is calculated twice. For inhibitory postsynaptic inputs (with a time constant :math:`\tau_{syn_{in}}`) and excitatory postsynaptic inputs (with a time constant :math:`\tau_{syn_{ex}}`).
The update is performed `here `_. The first multiplication evolves the external input. The others are the multiplication of the matrix :math:`e^{Ah}` with :math:`y` (for inhibitory and excitatory inputs).
If synaptic and membrane time constants become very close, :math:`\tau_m\approx \tau_{syn}`, the matrix :math:`e^{Ah}` becomes numerically unstable. NEST handles this gracefully as described in the `IAF Integration Singularity notebook `_.
References
~~~~~~~~~~
.. [1] RotterV S & Diesmann M (1999) Exact simulation of time-invariant linear
systems with applications to neuronal modeling. Biologial Cybernetics
81:381-402. DOI: https://doi.org/10.1007/s004220050570