\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}). The use of compartmental models is motivated by the desire to reduce the mathematical complexity inherent in a continuum description of a neuron. This simplification is achieved by replacing the family of partial differential equations defining the continuum description of a neuron by a compartmental model of that neuron in which the behaviour of the neuron is described in terms of the solution of a family of ordinary differential equations (Rall, \cite{Rall64}). The traditional approach to compartmental modelling, introduced by Rall (\cite{Rall64}), assumes that a ``lump of membrane becomes a compartment; the rate constants governing exchange between compartments are proportional to the series conductance between them". Figure \ref{Rall} illustrates the Rall interpretation of a how a dendrite can be represented in terms of compartments (neuronal membrane) and linking resistances (the axoplasm). \begin{figure}[!h] \centering \includegraphics[ ]{NewCompFig1.eps} \parbox{5in}{\caption{\label{Rall} The Rall segmentation of a length of dendrite into lumped regions. The membrane defines the compartment, and the resistive property of the axoplasm is represented in the model by resistors linking compartments}} \end{figure} %\begin{figure}[!h] %\centerline{\begin{mfpic}[1][1]{0}{250}{-25}{35} %\pen{1pt} %\lines{(30,10),(70,10),(70,-10),(30,-10)} %\dashed\lines{(30,10),(0,10)} %\dashed\lines{(30,-10),(0,-10)} %\rect{(80,10),(150,-10)} %\lines{(200,10),(160,10),(160,-10),(200,-10)} %\dashed\lines{(200,10),(230,10)} %\dashed\lines{(200,-10),(230,-10)} %\arrow\lines{(75,15),(75,5)} %\arrow\lines{(75,-15),(75,-5)} %\arrow\lines{(155,15),(155,5)} %\arrow\lines{(155,-15),(155,-5)} %\pen{2pt} %\headlen7pt %\lines{(40,15),(110,15)} %\lines{(40,-15),(110,-15)} %\lines{(120,15),(190,15)} %\lines{(120,-15),(190,-15)} %\tlabel[bc](75,20){\textsf{Membrane}} %\tlabel[bc](155,20){\textsf{Membrane}} %\end{mfpic}} %\centering %\parbox{5in}{\caption{\label{Rall} %The Rall segmentation of a length of dendrite into lumped regions. %The membrane defines the compartment, and the resistive %property of the axoplasm is represented in the model by resistors %linking compartments}} %\end{figure} This partitioning of a dendrite into repeating units is analogous to the representation of a transmission line as a ladder network of simple electrical circuits. In the case of a neuron, Rall defines a compartment to be the mathematical description of the effect of input acting on a localised region of neuronal membrane, and models this by an electrical circuit. The resistive properties of the axoplasm determine the linking resistances between compartments. Other authors (\emph{e.g.} Segev and Burke, \cite{Segev98}) treat the neuronal segment, including the membrane and axoplasm, as the compartment. However both definitions lead to the same mathematical model simply because the iso-potential property, implicit in Rall (\cite{Rall64}) and explicit in Segev and Burke's (\cite{Segev98}) definition of a compartment, dominates the construction of the underlying family of ordinary differential equations. However, the definition of a compartment as an iso-potential region is unsatisfactory since a compartment defined in this way cannot act as a fundamental unit in the construction of a model dendrite for two good reasons. First, iso-potential compartments must exist in pairs to support axial current flow, and second, half compartments are required to represent branch points and dendritic terminals (\emph{e.g.} Segev and Burke, \cite{Segev98}). What is required is a definition of a compartment in which the compartment exists as an independent unit that can act as a building block for a model of the electrical behaviour of a neuron, where it is understood that ``the function of the model is to represent the necessity that exists in nature by the logical necessity of the model. In the case of a good model one parallels the other'' (Regnier, \cite{Regnier64}). A traditional compartmental model does not satisfy this criterion since its iso-potential structure provides no mechanism to differentiate between input \emph{at different locations within the segment} represented by the compartment. It is precisely through the definition of a compartment, which allows this distinction to be made, that the new compartment model gives superior accuracy and precision to that of a traditional model. The accuracy with which a compartmental model describes the behaviour of a neuron may be assessed from the knowledge that all compartmental models converge to the solution of the continuum model of that neuron as the maximum length of segment approaches zero. To take advantage of this result, a test neuron is constructed for which the continuum description has an exact solution. This test neuron and its exact solution are used as a reference against which the accuracy of a traditional compartmental model and the new compartmental model are compared. \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 section of a dendrite into smaller contiguous units called 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}). The mathematical model of a dendrite is constructed by representing each segment by a compartment, and connecting these in a branching pattern corresponding to that of the dendrite. When joined in this way, each compartment interacts only with compartments representing adjacent segments. Note that some numerical schemes for the solution of the continuum model (\emph{e.g.}, Finite Differences) may have this ``nearest neighbour'' property, but it would be a conceptual error to interpret such equations as a compartmental model. The ``nearest neighbour'' feature of these equations is a contingent property of the numerical algorithm\footnote{For example, a second order central difference approximation for the spatial derivatives of the continuum model will be structurally identical to a compartmental model when the error structure of the discretisation is ignored. However, the ``nearest neighbour'' property of the numerical algorithm is absent for a higher order finite difference scheme.} and vanishes with a different choice of algorithm, whereas the ``nearest neighbour'' feature of a compartmental model is unavoidable. In a traditional compartmental model, the compartment has a single potential which is viewed as the potential at the centre of the segment represented by that compartment. This potential may be thought of as the average potential of that segment. All voltage-regulated input to the segment, independent of its location, acts with this potential. The assumption that all input to a segment acts with a single potential irrespective of location on the segment implies that a traditional compartmental model regards dendritic segments as iso-potential regions of dendrite. Spatial variations in biophysical properties of the dendrite and its morphology are expressed through differences in the properties of compartments and their linking resistors. By contrast, the new compartmental model assigns two potentials to a compartment, one at each boundary of the segment represented by the compartment. Compartments constructed in this way can serve as the basic building blocks of a model dendrite because they sustain axial currents independent of neighbouring compartments. Most importantly, the assumption that transmembrane current acts at the centre of a segment, as in a traditional compartmental model, is now inappropriate and must be replaced in the new compartmental model by a rule to partition transmembrane current between the axial currents flowing at segment boundaries. As with the traditional compartmental model, compartments in the new model are connected together by enforcing conservation of axial current at segment boundaries, dendritic branch points and dendritic terminals.