The Basics of System Dynamics:
Discrete vs. Continuous Modelling of Time 1
Günther Ossimitz 2
Maximilian Mrotzek 3
University of Klagenfurt
Department of Mathematics
Universitätsstrasse 65
9020 Klagenfurt, Austria
Phone +43 463 2700 3132
Fax +43 463 2700 3199
E-mail [Link]@[Link]
Abstract
System Dynamics deals with modelling of processes over time. In this paper we discuss two
ways to model changes over time: finite vs. infinitesimal. This leads to two different concepts
of time: discrete time as a succession of time points and time intervals vs. continuous time.
Although the System Dynamics concept of distinguishing between stocks and flows suggests a
discrete modelling of time, System Dynamics is considered mostly a modelling technique
based on continuous time. In the paper we argue to see System Dynamics modelling
compatible with both the continuous and the discrete concept of time. We will show that this
“hybrid” potential makes System Dynamics a superior technique for modelling time, which
combines the advantages of continuous and discrete time concepts.
Keywords
System Dynamics, Time, Discrete Time, Continuous Time
INTRODUCTION
System Dynamics (SD) models deal with time. In every SD model time is the central
independent variable and the main organizing principle. Despite of its central role, both theory
and practice of SD does not much bother about how to model time. For SD learners the usage
of dt as the symbol for a time-step might seem somewhat confusing. In SD theory dt
represents an infinitesimal time-step and in practical modelling dt stands for the duration of
the finite time-step of the simulation. For experts in SD this ambiguity seems to be no
problem at all.
In this paper we will look “behind the curtains” of how time is used. We will show that there
are two different modes of dealing with time, which are both related to SD: (a) the continuous
concept of time and (b) the discrete concept of time. The discrete concept of time is based
upon the distinction between time-points and time intervals. The continuous concept of time
is closely related to infinitesimal mathematics and does not need any time intervals for
explaining change, as change happens as momentarily change at each point of time.
1
Presented at the International System Dynamics Conference 2008, Athens/Greece
2
Prof. Dr. Günther Ossimitz is Professor at the department of mathematics at the University of Klagenfurt.
3
Maximilian Mrotzek, MSc, is a PhD student at the department of mathematics at the University of Klagenfurt.
Ossimitz/Mrotzek: Discrete vs. Continuous Modelling of Time 1 of 8
MODELLING OF TIME
In this section we will discuss two different modes of how to model time. The main difference
between these modes is how they deal with changes over time, which may be finite or
infinitesimal.
Finite Changes over Time
This mode of dealing with changes over time resembles our everyday experience. It is based
on a very simple principle: In order to notify some change, some time has to pass by. 4 As a
consequence changes are related to time intervals, whereas the state of a variable is specified
for specific points in time. A simple example for this approach is the relation between a
population and changes of population. The number of people is specified for some point in
time. The population changes over time due to births, deaths, immigration and emigration. All
these figures are actually counted as statistical data for finite time-intervals (days, months,
years).
The distinction between time-points and time-intervals yields to two distinct types of data:
data related to time-points and data related to time-intervals. For simplicity let us call the data
related to time-points “stocks” and the data related to time-intervals “flows”. We will discuss
later whether these connotations are identical or just similar to the stock-flow-concept in SD.
Flows are associated to stocks as stocks are exclusively changed over time via flows. Often it
is possible to distinguish two structurally different subtypes of flows: inflows and outflows.
For a population births and immigration are inflows, whereas emigration and deaths are
outflows. To distinguish between births (inflow) and deaths (outflow) is crucial for practical
understanding of populations. It would not make practical sense to define e.g. deaths as
“negative births”.
If we see stocks as state variables for time-points and flows as changes of the stocks for
certain time-intervals, the relation between stocks and flows is trivial arithmetic. For a given
time interval (t0,t1) and given flows for that time interval we can calculate the “new” value of
the stock at the end of the time interval according to the following equation:
stock(t1) = stock(t0) + inflows(t0,t1) - outflows(t0,t1) [1]
This simple procedure of calculating the new value of a stock needs just elementary
arithmetic, without any infinitesimal differentiation or integration process. In some cases the
value of the stock at t0 and t1 are given and we can calculate the (in)flow for the time interval
(t0,t1). For example, a mileage counter in a car shows at the start of a trip 10382 miles, and at
the end of the trip two hours later a value of 10502 miles. The distance travelled in two hours
was 10502 miles - 10382 miles = 120 miles, which implies the car drove at an average speed
of 60 miles per hour.
This example gives us also a clue that the flows as changes of stocks over time intervals are
not precisely the same as flows in SD. In SD the flow would be 60 miles/h, which is a speed.
In the example we have a “flow” of 120 miles distance travelled in 2 hours. Absolute changes
4
This principle holds no matter whether the changes are continuous (like water filling a bathtub via a faucet) or
instantly (like changes of a bank account via a transferral) Even for defining the infinitesimal change at some
point in time we need to know the development over a short time-span around the time-point of interest.
Ossimitz/Mrotzek: Discrete vs. Continuous Modelling of Time 2 of 8
over time intervals and relative changes per time unit are not the same, although the relation
between them is very easy to calculate:
absolute change(t0,t1) = average change per time unit ⋅ (t1 - t0) [2]
Here (t1 – t0) is the length of the time interval. Often [2] is used for calculating the average
change per time unit with the absolute change and the length of the time interval given.
Infinitesimal Changes over Time
The second way to model change over time is to make time intervals infinitesimally short
using a mathematical limes-process. Mathematically is this the first derivative of some state
variable with time as the independent variable. We can do so only if the state function over
time is differentiable. This implies necessarily that there are no discontinuities in the state
function. The derivative function usually is also defined for a continuum of time points that
makes it possible to repeat the differentiation process (provided that the derivative is
differentiable, too), yielding the second derivative function. We see that the concept of
derivatives allows formulating higher order changes for single time points. For the concept of
finite changes over time this is not so easy, since first order changes are defined not for time-
points, but for time intervals. When comparing two changes of adjacent time interval we get
some figure which is related to the overall length of the adjacent intervals.
The concept of continuous changes is typically modelled by differential equations or
infinitesimal accumulation processes (i.e. mathematical integration) over time. It has a strong
foundation in the mathematical field of Calculus and is applied for modelling time-related
processes in many sciences.
What role do changes over time intervals play in this approach? Since all changes are related
to time points via momentarily rates of change, there is no more need for dealing with time-
intervals. This elimination of time-intervals has the advantage of mathematical elegance,
allowing higher derivatives, but has also the consequence that there is no need for discerning
between data related to points in time vs. time-interval-related data, as in the finite approach
of change.
In the finite concept of time, we can discern between stocks and flows in a trivial manner:
stocks are data related to time-points and flows are changes of stocks, related to time-
intervals. Under the infinitesimal concept of time both stocks and flows are related to time-
points, so they cannot be any longer discerned in that simple way as in the finite concept of
time. Actually in the continuous model of time it is often a matter of the standpoint whether a
certain variable is considered to be a stock or a flow. For example considering a linear
motion over time the velocity can be considered as a stock, being changed by acceleration as
the associated flow. Velocity can also be considered as a flow changing the position(place)
as a stock over time. So the stock-flow-concept becomes somewhat arbitrary when we deal
with continuous changes over time.
Two Models of Time: Discrete vs. Continuous
The two modes of modelling changes over time can be formalized to two distinct concepts of
seeing and modelling time: discrete time and continuous time. The concept of discrete time is
based upon a distinction between time-points and time intervals. Typically the time axis is
divided into a number of adjacent time-segments (which usually are of fixed length). Both the
Ossimitz/Mrotzek: Discrete vs. Continuous Modelling of Time 3 of 8
number of time intervals and time points that are specified are finite. Actually the number of
time points = Number of time-intervals +1.
The concept of continuous time models time as a continuum of subsequent time-points. This
implies that data given for some time-span are specified as a continuous function over time.
Typically this function is also assumed to be differentiable, so that changes of the state
variable over time can be modelled as the first derivative of the state function. Time intervals
do not play a specific role as an organizing principle. Of course one can look at any time
intervals, but there are no pre-defined time-steps of finite length, which structure the whole
model.
Figure 1: Continuous Time vs. Discrete Time
Figure 1 sketches the concepts of continuous time on the left and the discrete time model on
the right. The distinction between a continuous and discrete concept of time can be explained
due to different types of time-related data. If the number of data given is finite, the discrete
model of time applies quite naturally, no matter whether the data given belong to time-points
or represent changes over time-intervals. For continuous data, like continuous waves in signal
theory, the continuous concept of time fits naturally.
Fixed vs. Constant Time-steps
Discrete Event Simulations are typically based on fixed iteration. A typical example is a
stochastic process on a discrete Markov Chain with a finite number of states and a transition
matrix describing the probabilities for changing from one state to another during one iteration
step. The overall state of the system after n steps can be calculated by the product of the n
transition matrices or the nth power of a fixed transition matrix. In such a system there is no
way to vary the duration of a time-step, since a time-step is represented just by one iteration
step.
In SD modelling the time-step is a parameter, which can be chosen freely and varied over
different simulation runs. Within one simulation run it is typically kept constant, but this is
just a property of SD simulation software, which traditionally does not adopt strategies of
varying step widths within one simulation run.
TIME IN SYSTEM DYNAMICS
The traditional view
Sterman (2000, p.193f.) describes four equivalent representations of a stock and flow
structure: Hydraulic Metaphor, Stock-Flow Diagram, Integral Equation, and Differential
equation (see Figure 2). In the “Hydraulic Metaphor” the stock is represented through the
water in a bathtub at any time. The amount of water in the bathtub either increases (water
flowing in through tap) or decreases (water flowing out through drain), excluding outside
Ossimitz/Mrotzek: Discrete vs. Continuous Modelling of Time 4 of 8
factors such as evaporation. A "Stock-Flow Diagram" already has an unambiguous
mathematical meaning as a stock accumulates its flows. The stock increases through material
inflow and decreases through material outflow. The “Integral Equation” describes the same
stock-flow principle, as the new Stock(t) is defined through the initial Stock(t0) plus all
Inflow(s) subtracted by all the Outflow(s) between the time t0 and time t. Finally, “Differential
Equation” describe the net rate change of a stock as the Inflow(t) subtracted by the Outflow(t).
Figure 2: Four Representations of Stock-Flow Structure (Sterman, 2000, p.194)
Sterman stresses that these four modes of representing a stock-flow system are equivalent:
“Each representation contains precisely the same information.” (Sterman 2000, p.194). We
agree with Sterman under the assumption that the whole modelling is based on the concept of
continuous time. However, we assume that within a discrete concept of time the four
representations are not necessarily precisely equivalent. Given finite time intervals as
simulation steps there might be flows changing a stock that are not capable with integral
equations or differential equations. We will see that SD models can cope with such types of
flows, too. In such a case the stock-flow-diagram (representing an actual SD model) and the
representation by a differential- or integral equation is not precisely the same, so that
Sterman’s statement “precisely the same information” no longer holds.
The core Elements of System Dynamics and Time
The core idea of SD modelling is the accumulation of flows over finite time intervals of
duration dt. Forrester (1968) argues clearly that the concept of stocks as accumulations is
more appropriate for SD than the concept of flows as derivatives: “Formulation of systems in
terms of differential equations obscures for many students the direction of causality within
systems […]. Representing a system in terms of integral equations gives a more immediate
and evident equivalence between the model and the real world.” (Forrester 1968, p. 6-12.).
Ossimitz/Mrotzek: Discrete vs. Continuous Modelling of Time 5 of 8
Figure 3: Accumulation of Bathtub and Bank Account (Ossimitz and Lapp 2006, p.81)
Moreover, the accumulation process in SD is more powerful than infinitesimal mathematical
integration. It allows accumulating not only smooth flows as the continuous inflow of water in
a bathtub, but also flows that have the character of instant pulses such as incoming payments
on a bank account that instantly change the value of the accumulated money, as illustrated in
Figure 3. This is possible due to the finite time structure of SD-models. In a strictly
infinitesimal integration process a pulse-like instant increase of a stock could not be modelled,
since at the time-points of the instant change of the state variable no derivative can be
specified.
We can illustrate the necessity of accumulation of both continuous and non-continuous flows
also by using Sterman`s bathtub metaphor (Sterman 2000, p.194): There might be not only a
continuous inflow through the faucet, one might also add a certain amount of water in a
moment by pouring a bucket of water into the bathtub. This would give the stock of water in
the bathtub an instant rise which cannot be modelled precisely with infinitesimal changes over
time.
Another advantage of the SD integration process in comparison to differential equations is the
possibility to attach several inflows and outflows to the same state variable (see Figure 4). For
a population we might discern between four distinct flows: births, deaths, immigration and
emigration. The derivative of a state-function at a given point of time is just a single figure. It
represents just the net-change of the corresponding change variable. Taking this into account
the development of different flows over time gives a much more realistic view of why the
state variable behaves in a specific manner.
Ossimitz/Mrotzek: Discrete vs. Continuous Modelling of Time 6 of 8
Figure 4: Using SD or Differential Equations to Model a Population
System Dynamics allows both discrete and continuous time
Altogether we can conclude that SD is compatible both with the continuous and the discrete
concept of time. On the one hand, there is a widely accepted consensus that SD is a modelling
methodology based on continuous time. The time-step can be made as small as one wishes
and all SD modelling software products offer at least a Runge-Kutta-4th order simulation for
good numerical results for continuous models. On the other hand, core aspects of SD
modelling like the strict distinction of stocks and flows combined with the specific
diagramming of stock-flow-diagrams, a finite time-step of constant length dt, and
discontinuous modelling functions (like pulse or ramp functions), are naturally compatible
with a discrete modelling of time in finite time-steps.
Concerning the way SD deals with time we can say that SD is a kind of “hybrid”
methodology, being compatible with both the continuous and the discrete concept of time. A
clear indication for this ambiguity is the usage and meaning of the term dt. In mathematical
calculus dt means a infinitesimally short time-step, a differential. In SD dt means a time-step
of finite length. So dt is both a infinitesimal and a finite time-step, thus allowing us to
associate it with both a continuous and a discrete concept of time.
CONCLUSIONS
One of the biggest advantages of the System Dynamics method is its structural simplicity of
modelling time, based on a clear distinction between stock variables related to time-points and
flows related to time intervals. This concept fits perfectly to our everyday perception of time,
where we discern naturally between time-points and time intervals.
Nevertheless SD has never adopted a clear discrete concept of time. Instead the traditional
mathematical concept of infinitesimal changes at individual time-points was adapted to SD
and is nowadays considered to be the theoretical framework to describe how the SD
methodology deals with time. The numerical integration methods of SD fit perfectly in the
mathematical tradition of numerical methods to integrate differential equations, and dt has
always denoted an infinitesimal interval of time in calculus.
Yet the concept of continuous time has two disadvantages: (1) It needs strong mathematical
assumptions of continuous state-functions and does not work if these assumptions do not
Ossimitz/Mrotzek: Discrete vs. Continuous Modelling of Time 7 of 8
apply. (2) It eliminates de facto the structural difference between points in time and time-
intervals by making time-intervals infinitesimally short, so that they can be treated as time
points. As a consequence the concept of continuous time makes it easier to mix up stocks and
flows, because in this concept the easy structural difference between time-points and time-
intervals is no more relevant.
The empirical research on “Bathtub Dynamics” (Booth-Sweeney and Sterman 2001, Kainz
and Ossimitz 2002, Ossimitz 2002) shows that even highly educated persons have massive
difficulties in discerning between stocks and flows. We can state it as a hypothesis for further
empirical research that this might be due to a poor understanding of the different concepts of
time. Maybe even the way we present SD as a methodology exclusively based on a concept of
continuous time and then doing simulation runs using discrete time steps afterward keeps up
confusion that leads to the poor scores in bathtub dynamics test items.
We see the conflict between a continuous and a discrete concept of time right in the core of
SD methodology. By addressing it explicitly and making both concepts of time clear to SD
learners might help them to understand better both System Dynamics and the world.
References
Booth-Sweeney, L. and J.D. Sterman (2001): Bathtub Dynamics: Initial Results of a Systems
Thinking Inventory. In: System Dynamcis Review, Vol. 16, Issue 4, p.249-286.
Forrester, J. W. (1968): Principles of Systems: Text and Workbook. Cambridge, Mass.: MIT
Press [et al.].
Kainz, D. and G. Ossimitz (2002): Can Students Learn Stock-Flow-Thinking? An Empirical
Investigation. In: Proceedings of System Dynamics 2002, International System Dynamics
Conference, System Dynamics Society, Palermo, Italy, August 2002
Ossimitz, G. (2002): Stock-Flow Thinking and Reading Stock-Flow-Related Graphs: An
Empirical Investigation in Dynamics Thinking Abilities. In: Proceedings of System Dynamics
2002, International System Dynamics Conference, System Dynamics Society, Palermo, Italy,
August 2002.
Ossimitz, G. and C. Lapp (2006): Das Metanoia-Prinzip: Eine Einführung in systemisches
Denken und Handeln. Hildesheim: Franzbecker Verlag.
Sterman, J. D. (2000): Business Dynamics – Systems Thinking and Modelling for a Complex
World. Boston: McGraw-Hill.
Ossimitz/Mrotzek: Discrete vs. Continuous Modelling of Time 8 of 8
