Prediction or cause? Information theory may hold the key

(PhysOrg.com) -- "A perplexing philosophical issue in science is the question of anticipation, or prediction, versus causality," Shawn Pethel tells PhysOrg.com. "Can you tell the difference between something predicting an event and something actually causing an event?"

Pethel is a scientist working at the Redstone Arsenal in Alabama. Along with Daniel Hahs, he set out to identify a method of distinguishing anticipation from causality using tools from information theory. “Any process that has to react in real time can improve its performance through anticipation, and in studying such processes it is important to find new ways to quantify causality” Pethel says.

The question of anticipation versus causality is one that has real-world application in a number of areas. Pethel points out that this issue has implications in covert operations, as well as in financial areas, especially with regard to the development of bubbles. “Is there a way, from passive measurements, to tell what’s driving what?” Pethel and Hahs try to answer this question in Physical Review Letters: “Distinguishing Anticipation from Causality: Anticipatory Bias in the Estimation of Information Flow.”

“If a system is generating information, it’s like a fingerprint that can be used to figure out where the information is coming from and where it is going,” Pethel explains. “A quantity called transfer entropy has been used since 1990 to measure information flow in experimental data from neuroscience, finance and even music.”

http://www.physorg.com/news/2011-09-theory-key.html

that are self-fullfilling, predictions that are self-cancelling, and other varieties.

In the scenario they give:



y must be using some information to anticipate x. And if y can appear to cause x because it anticipates it - and presumably acts before x; then I don't seem how x can be the sole, direct cause of y. It must be possible that the predictors of x can also cause y.

That is temporally extended processes, not unique events. I'd don't know whether you assumed they were, but just to clear it up. Their simplest manifestation would be as 1D time series.

So x doesn't cause y; x 'prods' y (but y cannot prod back).

If y can predict (to some accuracy) when x is going to 'prod it' (via access to present+past observable states of x), then it can take some preemptive action (on itself not x), to modulate the effect of the prod.

To an outside observer, not privy to the internal 'mind' states of x and y but only to their observable states, if y can very accurately predict when x is going to prod then it will seem very likely that the 'preemptive prod modulation' action of y is actually causing x to prod y (by the virtue of very often proceeding it temporally).