\section{Introduction} Compartmental models have become important tools for investigating the behaviour of neurons to the extent that a number of packages exist to facilitate their implementation (\emph{e.g.} Hines and Carnevale \cite{Hines97}; Bower and Beeman \cite{Bower97}). These models are constructed by replacing the continuum description of a neuron by a discrete description of the neuron formed by partitioning it into contiguous segments which interact with their nearest neighbours across common boundaries. A compartment is a mathematical representation of the morphological and biophysical properties of a segment, and a compartmental model is the collection of all compartments along with a specification of their connectivity. The efficacy of any formulation of a compartmental model depends on the faithfulness with which it captures the behaviour of the neuron that it represents, and it is in this respect that the new compartmental model developed in this article will be seen to perform better than existing compartmental models with a similar level of complexity. The traditional approach to compartmental modelling (\emph{e.g.}, Rall \cite{Rall64}; Segev and Burke, \cite{Segev98}) assigns a single potential to a compartment. This potential takes its value through an association with the average value of the current density crossing the membrane of the segment, and in a traditional compartmental model is approximated by the membrane potential at the centre of the segment. However a compartment of this type is aesthetically unsatisfactory since it cannot act as the fundamental unit in the construction of a model dendrite, first, because two compartments are required to define axial current flow, and second, because half compartments are required to represent branch points and dendritic terminals. On the other hand, the new approach to compartmental modelling assigns two potentials to a compartment -- one to represent the membrane potential at the proximal boundary of the segment and the other to represent the membrane potential at its distal boundary. The new compartment can exist as an independent entity and can therefore function as the basic building block of a multi-compartmental neuronal model. Another significant difference between a traditional compartmental model and the new compartmental model lies in the novel procedure for the treatment of transmembrane current. In a traditional compartmental model the influence of transmembrane current on a segment is approximated by requiring these currents to act at the centre of the segment with the single potential assigned to the compartment representing the segment, and consequently these models do not reflect accurately the influence of the precise location of point process input\footnote{Following the terminology of Hines and Carnevale (\cite{Hines97}), a point process is taken to mean either synaptic input (voltage-dependent) or an exogenous point current input (voltage-independent).} on the segment. By contrast, the formulation of the new compartmental model makes it more responsive to the influence of the location of point process input to a segment, and in the presence of these inputs, is shown to be an order of magnitude more accurate that a comparable traditional compartmental model. The accuracy of the new and traditional approaches to compartmental modelling is first assessed by calculating the error in the somal potential of a test neuron when each approach is used to calculate this potential ten milliseconds after the initiation of large scale point current input. In a second comparison, the accuracy of the two approaches is assessed by comparing the statistics of the spike train output generated by each type of compartmental model of the test neuron when subjected to large scale synaptic input. \section{Structure of compartmental models} We are concerned with compartmental models of dendrites. In this context, the fundamental morphological unit is the dendritic \emph{section}, defined to be the length of dendrite connecting one branch point to a neighbouring branch point, to the soma or to a terminal. Compartmental modelling begins by subdividing each dendritic section into segments which are typically regarded as uniform circular cylinders (\emph{e.g.} Segev and Burke, \cite{Segev98}) or tapered circular cylinders (Hines and Carnevale, \cite{Hines97}). In the new approach to compartmental modelling, the known membrane potentials at the ends of a segment (rather than its centre) provide the basis for the development of a set of rules which enable the influence of precisely located point process input to be partitioned between the axial current at the proximal and distal boundaries of the segment. The mathematical equations of the compartmental model are constructed by enforcing conservation of axial current at segment boundaries, dendritic branch points and dendritic terminals. \subsection{Model accuracy and the partitioning of point process input}\label{assertion} The benefit in accuracy gained by taking into account the precise placement of point process input on a dendrite is best appreciated by considering how, in the absence of this facility, small variations in the location of segment boundaries exert a large influence on the solution of a traditional compartmental model. Consider, for example, a point process close to a segment boundary. A small change in the position of that boundary may move the assigned location of this point process from the centre of one segment to that of an adjacent segment. With respect to a traditional compartmental model, the location of this point process is therefore determined only to an accuracy of half a segment length, and this indeterminacy will in turn generate a model solution that is particularly sensitive to segment boundaries. Of course, with a small number of point process input, this problem can be avoided in the traditional approach to compartmental modelling by arranging that only one point process falls on a segment, and that the location of this input coincides with the centre of the segment. However, this strategy is not feasible when dealing with large scale point process input. What is required is a procedure which describes the effect of point process input on a dendritic section in a way that is largely insensitive to how that section is represented by segments. The primary sources of error in the construction of a compartmental model are the well-documented effect of discretising a continuous dendrite, and the less well-documented error introduced by the placement of point process input on the dendrite. In the traditional approach based on a compartmental model with $n$ compartments, the first type of error is $O(1/n^2)$ (by analogy with the finite difference representation of derivatives), but it is not widely recognised that the second type of error is $O(1/n)$ whenever the input does not naturally fall at the centre of segments. Since the accuracy of any model is governed by the least accurate contribution to the model, it is clear that \emph{in practice} the traditional approach to compartmental modelling in the presence of point current and synaptic input is $O(1/n)$ accurate. This theoretical observation is supported by the simulation studies of Subsections \ref{sim1} and \ref{sim2}, and by an example provided for us by an anonymous reviewer. This reviewer used the simulator NEURON to calculate the somal potential of the test neuron shown in Figure \ref{TestNeuron} 10msec after the initiation of point current input. The results of this calculation are shown in Table \ref{reviewer1} \begin{table}[!h] \[ \begin{array}{c|cr|cr} \hline \begin{tabular}{c} Segments \\[-2pt] per branch \end{tabular} & \multicolumn{2}{|c|}{\begin{tabular}{c} Point current input at \\[-2pt] centre of nearest segment \end{tabular}} & \multicolumn{2}{|c}{\begin{tabular}{c} Point current input \\[-2pt] divided proportionately \end{tabular}} \\ section & V\,\mbox{(mV)} & \Delta V\,\mbox{(mV)} & V\,\mbox{(mV)} & \Delta V\,\mbox{(mV)} \\ \hline &&\\[-11pt] 1 & 10.2355 & & 10.5692 & \\ 2 & 10.2311 & (-4.4616\times10^{-3}) & 10.3357 & (-2.3352\times10^{-1}) \\ 4 & 10.2367 & ( 5.6256\times10^{-3}) & 10.2725 & (-6.3143\times10^{-2}) \\ 8 & 10.2333 & (-3.4428\times10^{-3}) & 10.2556 & (-1.6908\times10^{-2}) \\ 16 & 10.2470 & ( 1.3754\times10^{-2}) & 10.2519 & (-3.6550\times10^{-3}) \\ 32 & 10.2509 & ( 3.8793\times10^{-3}) & 10.2508 & (-1.1320\times10^{-3}) \\ 64 & 10.2521 & ( 1.1874\times10^{-3}) & 10.2506 & (-2.4666\times10^{-4}) \\ 128 & 10.2530 & ( 8.8765\times10^{-4}) & 10.2505 & (-6.3146\times10^{-5}) \\ 256 & 10.2511 & (-1.9053\times10^{-3}) & 10.2505 & (-1.5181\times10^{-5}) \\ \hline \end{array} \] \centering \parbox{5.7in} {\caption{\label{reviewer1} The somal potential of the test neuron shown in Figure \ref{TestNeuron} is given 10msec after the initiation of point current input. The calculation is done for nine different levels of discretisation and two methods for the placement of exogenous point current input.}} \end{table} The results shown in the middle panel of Table \ref{reviewer1} (traditional compartmental model) are based on placing the exogenous point current input at the centre of its nearest segment, whereas the results shown in the right hand panel (modified compartmental model) are based on the division of the point current input between the centres of adjacent compartments in proportion to the conductance between the location of the input and these centres. Several important differences between the two procedures for allocating the location of point current input are evident from the results set out in Table \ref{reviewer1}. The results based on dividing the current proportionately between the centres of neighbouring compartments converge smoothly and more rapidly to the true potential than those based on the traditional approach in which the current is placed at the centre of the compartment. An extrapolation procedure demonstrates that the potentials generated by the modified approach converge quadratically to the true somal potential as the number of compartments is increased. Moreover, not only does the solution following the traditional approach (middle panel) converge to the true potential more slowly than the modified approach (right hand panel), the former appears to oscillate as it approaches this potential. Finally, further evidence for the superior convergence of the modified approach is clear from the observation that the best estimate of the true potential using the traditional approach with 256 segments per branch section is achieved in the modified approach with approximately 28 segments per branch section. It will be seen in Section \ref{PointInput} that the procedure used by the reviewer to partition point current input is a special case of the general procedure for partitioning point process input. By contrast with the traditional approach, the new approach to compartmental modelling describes the influence of point process input to an accuracy of $O(1/n^2)$, and therefore one would anticipate that it does not degrade the overall accuracy of the model. The validity of this assertion is demonstrated through the simulation studies in Subsections \ref{sim1} and \ref{sim2}.