What made you decide to apply statistical physics to social problems?

As a physicist, I have studied many different types of physical system using statistics, and in particular I’ve had a longstanding interest in how magnets become ordered at low temperatures. This ordering is akin to reaching consensus in a socially interacting population, where political preference is analogous to the orientation of a magnetic moment. I realized that there are in fact lots of societal phenomena that can be analysed using statistics, with the interactions between individuals being analogous to the particle interactions in a physical system. The human world provides such a rich laboratory that I can see statistical physics almost everywhere that I look.

But what exactly allows you to make this link between physical and societal phenomena?

I try to understand the collective behaviour of entire populations by postulating just a few simple interactions between individuals

Statistical physics allows you to study the macroscopic properties of a system of many interacting particles without having to track the behaviour of every single particle. For example, to understand the properties of air in a room you don’t have to care about the motion of every single molecule in that room, even though in principle you could do this. You might instead want to know how the temperature in the room can vary, or what determines wind velocity through the room — these kinds of things can be described using statistical physics. The crucial idea here is one of emergence — that from a few simple rules governing the interaction of individual particles you can end up with collective behaviour that is not described by the rules themselves. So in the context of society, I try to understand the collective behaviour of entire populations by postulating just a few simple interactions between individuals and calculating the outcome of millions of such interactions using computer modelling.

How have you used this approach to model the way people vote?

Part of my work is based on the so-called voter model: an idealized model that assumes individuals cannot make up their own minds about how to vote but simply mimic what their neighbours are doing. This involves placing voters, as if they were simply points, on the nodes of a regular lattice. So using a square lattice, for example, everyone has four neighbours with whom to interact. The model evolves by picking any person on the lattice at random and then assigning that person the voting preference of their nearest neighbour. Depending on exactly how you use the model, the initial state may be random or structured.

How does this simple voter model evolve?

When you run the model, you are guaranteed that the system will eventually reach consensus. In other words, given a choice of two parties to vote for, A or B, the whole population will end up voting for either party A or party B. This model is “conservative” in the sense that by running the system many times it will sometimes end up in state A and at other times state B, but the ratio of these two states exactly matches the initial configuration of the system. So, taking the US for example, if we set our system up so that initially 55% intended to vote Republican and 45% sided with the Democrats, then 55% of time the system will reach a Republican consensus, while 45% of the time it will end up with everyone voting Democrat. In other words, on average no one changes their mind.

But in reality surely there are some people who do think for themselves?

Just a few zealots can completely screw up the system

Indeed. To make the model more sophisticated we have looked at what happens when you make some of the voters “zealots”, which means that they never change their mind, and, at the other extreme, others “vacillators”, who find it very hard to make their mind up and need to hear the opinions of several other voters before changing their intention. In fact, it turns out that the introduction of these types of voter is fundamental to the outcome of the system. Just a few zealots, for example, can completely screw up the system. For one thing, a few zealots on either side of the party divide make it impossible to ever reach consensus because by definition they will never change their mind. In fact, just a handful of zealots in an infinite system causes the system to end up in a more or less 50:50 split between the two parties regardless of the initial distribution of voting intentions.

Do you plan to test your models against real election data?

Up to now we have not been able to say too much about real voting because we have been looking at models that we can solve analytically. But in the near future we hope to do some reverse engineering, which means that we will look at past election results and infer how people came to their decision — in other words, we will work out what fraction of people were likely to have been easily swayed and how many never changed their opinion etc.

Could your research be of practical help to people who set the rules for elections?

Statistical physics cannot be used as a predictive tool, but rather it is a descriptive tool

That’s a tricky question. If I were to study the distribution of people’s heights and find, for example, that it was a Gaussian curve peaking at a figure of 5’10”, then this information might be useful for designing doors on subway trains so that people don’t hit their heads too often. With voting, we can hopefully say something about the distribution of certain kinds of election result for certain types of election. For example, the model might be able to tell us what the chances are that a landslide victory will occur. However, it is very important to emphasize that statistical physics cannot be used as a predictive tool, but rather it is a descriptive tool.

What other systems have you investigated using applied statistical mechanics?

I have used it, for example, to look at the class structure in society; to assess the importance of scientific papers; and to study the relationship between climate change and record-breaking temperatures. And recently I have also investigated the statistics of baseball by trying to identify patterns of winning and losing. We have built a model to predict on average how many wins the first-placed team will record, how many the second-placed team will achieve, and so on, and we have found that our predictions very closely match end-of-season data averaged over more than a century. We can also use this same theory to successfully predict the distribution of teams’ winning and losing streaks.

But is there a danger that you can take this reasoning too far? Can’t statistics only tell us so much?

Certainly in baseball people have gone crazy in trying to quantify the game. Score is the most basic statistic, but then people have introduced many secondary and tertiary statistics, such as an individual’s batting performance in a specific type of game situation. But statistical-physics ideas are best suited to the simplest and most fundamental measures, such as the number of wins and losses, or the score.

Have you collaborated with social scientists?

I’ve had discussions with various sociologists and pure statisticians. But these discussions can be difficult because of the technical jargon on both sides and also the fundamental question of what turns people on. The notion of emergence in complex systems is fundamentally important to me as a statistical physicist, whereas some sociologists couldn’t care less about that. Often they are interested in concrete issues such as who is going to win the next election or whether the minimum wage should be increased. I am, however, still trying to make connections with these people.