Analog Computation

A recent preprint (Friedrich, 2026) describes the notion of analog computing, and proposes that it be taken seriously as a form of computation that might show up in cognition (alongside, not necessarily replacing, digital computation). He lays out the basic idea, connects it to some empirical literature, and makes the case for it being one of the options in the toolkit. 

I read it after a good discussion on Bluesky, in which I asked how this is different from dynamical systems approaches. 

Real question: what’s the difference between analog computation and dynamical systems theory (eg the watts governor)?

[image or embed]

— Andrew D Wilson (@adw.bsky.social) November 7, 2026 at 1:20 PM

The answers were informative! There is a key difference, and one with implications for whether this approach is useful to the ecological approach. The short answer is no, it's not especially useful, because it's representational. The slightly longer answer seemed worth blogging though. 

Analog Computation

Digital computing transforms a task into symbols, and processes those symbols according to algorithms to produce an outcome; it creates a formal model of the natural system to be understood. Analog computing instead builds a second natural system that implements just the relevant dynamics of the target natural system, and when those run you can compute the behaviour of the other system.

The example is an old device for predicting the tides. Formal models had failed because there were just too many variables and simulating the system took too long. Instead, people made machines that implemented the equations with their activity. The behaviour of each variable was implemented in the behaviour of a set of gears, pulleys, and wires, and the overall activity of the sets of gears led to the pen behaving in a way such that the height of the tides could be read out. 

Figure 1 from the paper

This looks a lot like things like the Watts steam governor or the polar planimeter. However, there is a difference between these dynamical systems and the analog computer. In the former cases, the device is the target system - the Watts steam governor dynamics are the relevant dynamics for the process of regulating the steam flow. In the tide predictor case, the device dynamics have been selected not to achieve a goal (e.g. control the tide) but to model the relevant dynamics of another system. Analog computers are inherently representational. 

Why use them? Well, for the same reason you use any representation - when you don't have the requisite access to the target system. You can't use the actual tides to predict future tides, because their current activity is about their current activity, not their future activity. But you can replace access to that future activity with access to the dynamics of the analog tide predictor device and run it faster than the actual tides, and that gets you the prediction. Until digital computers got fast enough, this analog approach was the better way to compute the tide. 

Is This Notion of Computation Useful for Ecological Psychology?

Friedrich makes the point that anti-computational approaches are all anti-digital computation approaches, and that analog computing shows up in things like embodied robotics and morphological computation (Pfifer and Schier type robots; I talked about them in this paper). So maybe we aren't anti-computational, just anti-digital, and analog computing could be a useful addition to the toolkit. 

For ecological psychology, I think the answer is no. I think we are actually anti-representational, and that this applies regardless of the format of the representation. You only need to compute (digitally or analog) when you don't have access to the dynamics of your target system. But we do - information of, say, affordances, via specification provides the access that enables us to couple our dynamics to those of things that are not us. This access means we can take the Brooks approach, and 'let the world be it's own model'. Friedrich notes the issue of resonance to information but doesn't understand it enough to see it means we don't need the help. 

The other way this fact is revealed is in the cognitive research Friedrich connects his ideas to. There are many specific examples, but the primary thread that he links to is theories of grounded cognition, a la Barsalou, in which representations are not abstract symbols but grounded in perception. The problem here of course is that this is the kind of embodiment that does not fully break with the representational approaches, and thus inherits all the problems ecological psychology aims to solve via specifying information and laws. 

I think expanding the kinds of computation available is probably healthy, and it may be the case that analog computing is the kind of implementation that something like Barsalou-style grounded cognition needs. But analog computing still inherits all the problems intrinsic to representation, and so it's of no use to ecological psychology. We don't need to represent distant dynamics; instead, we can couple to them via perception and action, and representation-free dynamical system approaches remain our most valuable tool.