Here is the central thesis of this paper. If time series data are the result of a nonlinear dynamic system, then techniques are available for estimating the number of dynamic variables that produced the data. We recommend an experimental paradigm in which the independent variables are those known to influence performance, and the dependent variables are estimates of dimensionality. This paradigm allows us to investigate whether the independent variable is influencing a dynamic variable, one that changes in time, or is affecting a relatively constant system parameter. We apply this paradigm to some memory data here, but we offer it as a general experimental paradigm. That's the message. The rest of this paper is devoted to details of definition, illustration, and a brief discussion of limitations.What do we mean by a dynamic model? A dynamic model is any model for which some measure x changes over time according to an explicit function. A nonlinear dynamic model is simply one for which the equation specifying the change is not linear. The simplest example of a discrete nonlinear model is the logistic map, for which x changes from time n to time n+1 according to:
The logistic map is often used to introduce chaos concepts because, for certain values of the parameter r, the resulting dynamics display all the features of a chaotic system. It is important to stress that nonlinearity refers not to the behavior of x over n, but to the relation of x_n and x_(n+1). An unattractive feature of nonlinear models is their resistance to solution by formal analysis. On the other hand, recent work has yielded a number of geometric regularities and useful techniques for characterizing these systems and for discriminating them from other systems.
A compelling feature of nonlinear systems is that even simple systems, such as the logistic map, produce behavior that is difficult to distinguish from random behavior. This is illustrated in the Figure 1, which shows a time series produced when r equals 4 and compares it with truly random data.
Figure 1 Behavior of the Logistic map and Random Data. We take this apparently random behavior as a potential challenge to the conventional view of variability. The conventional view is that measures are the linear sum of true scores plus noisy "error", which in turn is the result of the contribution of a very large number of independent factors. The new insight is that seemingly random variability in measures over time can result from very simple deterministic systems.We use the term dynamic variable to refer to variables that change over time, to be distinguished from parameters which do not. In the logistic map example, x is a dynamic variable and r is a parameter. More generally, a dynamic system is a set of functions representing N inputs and N outputs. For example, a discrete system would be represented by a set of N difference equations:
where x_i are dynamic variables and parameters are indexed by p_i. More simply: dynamic variables are indexed by time, parameters are not. There are N variables and K parameters. Notice that the dynamic variables are "coupled", which means that change in one depends on the current value of the other. The number of dynamic variables of a system is known as the dimensionality of the system. The logistic map, for example, has dimensionality equal to one. An example of a two-dimensional dynamic system (Clayton & Frey, in press) is shown next.
This model, similar to one suggested by van Geert (1991), has two dynamic variables, x and y, competing for growth according to growth parameters r_x and r_y but limited in magnitude by carrying capacity K. Here there are two dynamic variables and four parameters. The model also illustrates why we think that memory performance may be profitably modeled by a nonlinear system. Memory typically involves competition, and is limited in capacity, at least in working memory. The result is that the dynamic variables are "coupled" and growth is bounded. In addition to the van Geert work already cited, in recent years dynamic models have been proposed across psychology. See the reference list for a partial listing of this new work. One of the remarkable recent achievements in the study of nonlinear dynamics is the development of techniques for estimating dimensionality from measures on a single variable. Such techniques could be enormously useful to the model builder because they estimate the total number of dynamic variables needed to model the dependent variable.
How is this possible? The major technique involves the embedding dimension, which is defined in terms of successive N-tuples of the time series data. Consider, for example, the first seven values of the logistic map data plotted earlier.
0.10, 0.36, 0.92, 0.30, 0.84, 0.54, 0.99. ... For an embedding dimension of two, the resulting pairs are (0.10, 0.36), (0.36, 0.92), (0.92, 0.30), (0.30, 0.84), (0.84, 0.54), (0.54, 0.99), ...For an embedding dimension of three, the triplets are (0.10, 0.36, 0.92), (0.36, 0.92, 0.30), (0.92, 0.30, 0.84), (0.30, 0.84, 0.54), ...
These successive N-tuples are treated as vectors, or points, in N- dimensional space, and it is the geometric structure of these points that is of interest. To illustrate their utility, suppose we embed the time series from the logistic map in two dimensions and plot the points in 2-space. Figure 2 shows the result. We see that the underlying structure of a system producing seemingly random data is fully revealed by embedding the data in two dimensions.
Techniques for estimating dimensionality, in general, involve distances among the points in the embedding dimensions. And the central idea is that certain properties of these distances will change as the embedding dimension increases, but only so long as the embedding dimension is less than the dimensionality of the data. In the particular technique developed by Grassberger and Procaccio (1983a,b), which rapidly became widely used in other sciences, for each embedding dimension the proportion of distances that are less than a given distance, or radius, r are calculated. This value is called the correlation integral, Cr(I), and the linear portion of the function relating log[Cr(I)] to log(r) is used to estimate the correlation dimension for that embedding dimension.The next figure illustrates the technique applied to data from the logistic map and from data generated as white noise.
Plotted here are dimension estimates as a function of embedding dimension. The logic is this. As long as dimensionality exceeds the embedding dimension, the estimate will equal the embedding dimension. Since white noise has infinite dimensionality, estimates should continuously equal the embedding dimension, as they are seen to do. The logistic map data, however, are the result of a one-dimensional system and should not rise beyond an embedding of one. In general, then, we look to see whether and where this function levels off, or "saturates" as reaching an asymptote is called.Outcome: InterpretationNow, we argue, the ability to estimate the dimensionality of data arising from a given experimental condition invites a new paradigm in which the relationship between independent variables and dimensionality are investigated. The initial logic is this: if the independent variable influences both the dependent variable and the dimensionality estimate, its influence arises from its effect on a dynamic variable. If, however, the independent variable influences the dependent variable but not dimensionality, then its effect is on a parameter. The possible results are summarized in the next table.
One limitation of this approach may already be apparent. If the estimate of dimensionality is very large, or too large to estimate, the suitability of the nonlinear approach would be questioned and the location of the effect of the independent variable it would be unclear.We illustrate all of this with the experiment now to be described.
Experiment Three classification tasks were used, designed to manipulate memory load, so our major experimental question was whether memory load affects a dynamic variable or a parameter.
Method
Twelve subjects were each given three classification tasks that differed only in the response rule. On each trial of each task the stimulus was either an X or an O. In the X-O task, the subject pressed one key for an X, another for an O. In the same-different-1 task the subject pressed one key if the stimulus was the same as the stimulus on the previous trial, otherwise the other key. On the same-different-2 task, the subject pressed one key if the stimulus was the same as the stimulus two trials back, otherwise the other key. Each task had 500 uninterrupted trials, with task order counterbalanced across subjects. We used 500 trials because the planned nonlinear statistics require lengthy time series.
Reaction time was measured on each trial. In addition, since the subject determined the intertrial interval by indicating readiness for the next trial, it occurred to us that the interesting temporal dynamics may be in this subject-determined delay. The subjects were encouraged to develop a strategy to prepare for the next trial, and this interval, too, was measured.
Results
The next figures display the obtained data using the standard plot of dependent variables vs. independent variables, and show the expected effect for both. Both response times and trial-initiation times increased significantly with increased memory load, being fastest for the simple X-O task, slowest for the task requiring memory for the last two trials.
That's the standard analysis. Now we treat the data as a time series and examine the effect of memory load on our new measure, dimensionality. Analyses of variance on Grassberger-Procaccia estimates showed no effect of memory load. That is, there was no evidence that memory load was affecting a dynamic variable, despite its substantial effect on response times. There was, however, also no evidence for low-dimensionality in the response time data. This is shown in the next figure.
This figure shows estimates from the response time data on the left, trial-initiation time on the right. Two functions of slope estimates across embedding dimensions are shown. The solid function is for the obtained data, the dashed line for the same data randomly shuffled. Shuffled data should resemble white noise, so shuffling controls against biased estimations. The technique is similar to the "surrogate data" method. (Theiler, Eubank, Longtin, Galdrikian, & Farmer, 1992; Theiler, Galdrikian, Longtin, Eubank, & Farmer,1992; Kennel & Isabelle, 1993). The response time data show no evidence of saturating and no difference from the shuffled version.The story is somewhat different for the trial initiation data shown in the next figure.
Here there is a significant difference between the obtained and shuffled data. There is also a suggestion that dimensionality estimates of the obtained data levels off, perhaps at a dimensionality of six. Our maximum embedding dimension is ten here because the reliability of dimension estimates decrease as embedding dimension increases, especially for short data sets. So, without using other dimension estimates, which are planned, we tentatively conclude that low-dimensionality in the time-initiation data is possible.In summary, the memory load manipulation produced substantial effects on response and trial initiation times, but no effects on estimates of dimensionality. We have used these data to illustrate the potential for an experimental paradigm that treats dimensionality as a dependenet variable. Provided dimensionality estimates are low, this technique provides evidence about whether the independent variable influences a dynamic varaible or a system parameter.
Finally, nonlinear statistical techniques are very much in the developmental stage Several approaches have been suggested, and recent reviews of the strengths and weaknesses of the techniques are to be found in Abarbanel, Brown, Sidorowich, and Tsimring (1993), Pritchard and Duke (1992), Johnson and Dooley (1994), and Rapp (1995).
Abarbanel, H. D. I., Brown, R., Sidorowich, J. J. & Tsimring, L. S. (1993). The analysis of observed chaotic data in physical systems. Reviews of Modern Physics, 65, 1331-1392.Clayton, K. & Frey, B. (In press). Inter- and intra-trial dynamics in memory and choice. In J. Goldstein & A. Coombs (Eds.), Chaos Psychology: A Reader. Vol 2.
Cooney, J. B. & Troyer, R. (1994). A dynamic model of reaction time in a short-term memory task. Journal of Experimental Child Psychology, 58, 200-226.
Grassberger, P., & Procaccia, I. (1983a). Characterization of strange attractors. Physical Review Letters, 50, 346-349.
Grassberger, P., & Procaccia, I. (1983b). Measuring the strangeness of strange attractors, Physica D, 9, 189-208.
Hoyert, M. S. (1992). Order and chaos in fixed-interval schedules of reinforcement. Journal of the Experimental Analysis of Behavior, 57, 339-363.
Pritchard, W. S. & Duke, D. W. (1992). Measuring chaos in the brain: A tutorial review of nonlinear dynamical EEG analysis. International Journal of Neuroscience, 67, 31-80.
Townsend, J. T. (1992b). Chaos theory: A brief tutorial and discussion. Ch. 4 in A. F. Healy, S. M. Kosslyn, & R. M. Shiffrin (Eds.) From Learning Theory to Connectionist Theory: Essays in Honor of William K. Estes, V.1. Hillsdale NJ: L. Erlbaum Assoc. pp.65-96.
van Geert, P. (1991). A dynamic systems model of cognitive and language growth. Psychological Review, 98, 3-53.