This post aims to provide context for, and a brief, accessible overview of, my masters research on fractional Brownian motion. It is structured by the contributions of a series of key figures: Robert Brown, Albert Einstein, Norbert Wiener, Andrey Kolmogorov, Benoit Mandelbrot and (tongue-firmly-in-cheek) me.
Brown
Summer, 1827. The Scottish biologist Robert Brown—who spent the years between 1801 and 1805 circumnavigating Australia with Matthew Flinders—is peering down a microscope, at a sample of pollen particles suspended in pond water.
Having observed the phenomena equally in pollen from all the living plants he had examined, Brown hypothesised that the particles were “male”, and the movement was due to them pruriently seeking “female” pollen with which to reproduce. To test the robustness of his theory, he expanded the scope of his investigation.
If the movement was sexual in nature, then it should not persist for long after the death of the plant. But Brown found that drying or soaking the plants in spirits for a few days had no effect on the motion of their pollen, though by that time they would have been well and truly dead. Dried and preserved plants that had been sitting in a local herbarium for over a hundred years also still exhibited the motion.
Additionally, if the movement was sexual then it should not occur in species that reproduced through other means. Brown tested ground up samples of mosses, in which the existence of sexual organs had not been universally admitted. They still moved.
Needing a new hypothesis, Brown thought that perhaps the motion was a fundamental property of organic tissues. He kept investigating and showed that the movement could be observed in ground up animal tissues, both living and dead.
But this hypothesis also collapsed, as he tested substances with increasingly tenuous connections to living things. The movement was present in substances of loosely “vegetable origin”, such as gum resins, fossilised wood and pit coal. But it was even present in “silicified wood”, so could not be limited to organic bodies.
Over the next few months, Brown tested every substance he could think of: window glass, “many simple earths and metals”, “rocks of all ages”, stalactites, lava, obsidian, pumice, “a fragment of the Sphinx”, volcanic ashes, meteorites, manganese, nickel, plumbago, bismuth, antimony, arsenic, asbestos…
But that was as far as he got. The reasons for the movement remained mysterious.
Einstein
1905 was Einstein’s annus mirabilis, the year in which he published four revolutionary papers on relativity, the equivalence of energy and matter (
Liquid water is effectively a soupy pinball machine of microscopic
When a pollen particle is suspended in water, it too is now being bombarded by all the water molecules that are flying around. Each time a water molecule impacts the pollen grain, it exerts a force on the grain in the direction of impact. If the pollen grain is not too big, and if enough water molecules hit the grain in the same direction at the same time, the grain starts moving in that direction. The sheer number of water molecules enveloping the pollen grain mean that it is constantly being jolted in every direction by the incessant barrage of collisions.
The above simulation is only a simplistic demonstration. In reality, the water molecules would be much smaller relative to the pollen grain, and there would be far more of them. No single collision between a water molecule and the pollen grain would result in observable movement. Instead, each observed jolt is caused “by an enormously large number of successive kicks it receives from different fluid particles which, by rare coincidence, give rise to a net displacement in the same direction”[6].
Using the kinetic theory of heat (ie. that temperature
Einstein’s proposed explanation was controversial at the time, because while the existence of atoms had been theorised, it was not universally accepted. Some physicists thought that atoms were merely a useful computational tool, and didn’t correspond to reality. In 1906, one of the chief proponents of atomic theory, Ludwig Boltzmann, felt defeated and committed suicide. But Einstein’s theory provided a means of using the size of observed vibrations to determine the mass and quantity of the “atoms” that had caused them. This experiment was soon carried out by Jean Perrin; the results won him the 1926 Nobel Prize in physics and silenced all remaining skepticism on the existence of atoms.
Wiener
Though Einstein introduced mathematics into the study of Brownian motion, it was only in the context of modelling the physical phenomena of small particles moving erratically in a liquid. Norbert Wiener—a former child prodigy who in the 1940s was already corresponding with union leaders to warn about the risk of intelligent machines replacing human workers[8]—created a mathematical formulation of Brownian motion that was much more abstract, and hence much more widely applicable.
To appreciate Wiener’s contribution, we need to introduce the concept of a stochastic process. (Stochastic is just a fancy word for random.) A stochastic process is a set of random variables, indexed by a parameter which usually denotes ‘time’ or ‘space’. For example, the cumulative number of busses that have arrived at a bus stop across time, the humidity at each point on Earth, and the price of tuna at the Toyosu Fish Market in Tokyo across time can all be thought of as stochastic processes.
The vertical position (or
The Wiener process is an idealisation of physical Brownian motion because it makes some simplifying assumptions. One assumption is that the direction the particle moves in the future is independent of the direction it has drifted in the past. Another assumption is that the paths of the Wiener process are fractal, meaning that you can zoom in as much as you want and the graph will still be infinitely wiggly.
This is not the case for the position of a pollen grain in water because, while there is a truly enormous number of collisions driving the motion, the number is still finite, so there will be some (incredibly tiny) interval of time between each collision during which the pollen drifts in a straight line, unimpeded. Nonetheless, it is a sufficiently close approximation to reality that the model is still useful. And useful for modelling many phenomena beyond the mere jiggling of pollen grains.
Remarkably, Wiener formulated the Wiener process before the modern framework of probability theory was developed. That is, before there were widely accepted definitions for terms like ‘random variable’, ‘expected value’ or ‘probability’, or ‘stochastic process’.
Kolmogorov
Kolmogorov is the guy who made probability respectable.
He also, amongst his prolific output, described (in 1940) a generalisation of the Wiener process that he called the ‘Wiener spiral’, but is today known as fractional Brownian motion (fBm), which was the subject of my thesis. For whatever reason, the generalisation received only limited attention until it was popularised by Mandelbrot 29 years later, so I will explain exactly what fBm is in the following section.
Mandelbrot
Benoit Mandelbrot is perhaps best known for his work on fractals, such as the iconic Mandelbrot set.
While he didn’t discover fBm, he did coin its modern name (ie. ‘fractional Brownian motion’) and was largely responsible for its popularisation, which stemmed from a paper he co-authored in 1968.
Fractional Brownian motion is a generalisation of the Wiener process, or standard Brownian motion. Technically, fBm is not a single process, but a family of related stochastic processes, one for each number between 0 and 1. This number, called the Hurst parameter and usually denoted
The advantage of fBm over Bm is that it loses some of the assumptions that do not realistically apply to many of the real-world processes that people are interested in modelling. In particular, the assumption that how the process has evolved in the past doesn’t affect how it evolves in the future does not apply to many real world processes, where there can be underlying trends that persist for long periods of time. This property loosely corresponds to the mathematical concept of ‘long memory’ or ‘long range dependence’, which fBm is capable of modelling for certain values of the Hurst parameter.
Since Mandelbrot’s paper, fBm (and other stochastic processes that are derived from fBm) have been widely used to model time series or spatial data that exhibit fractal properties. Example fields of application include:
- bioengineering - to model regional blood flow distributions in the heart, lungs and kidneys
- communication theory - to model the arrivals of packets of data on a computer network
- hydrology - to model water levels in rivers and the permeability of soils to water
- finance - to model the prices of stocks and other financial instruments
- genetics - to model stochastic gene expression and DNA walks
- Bayesian machine learning - to induce a prior distribution on the space of regression curves
- computer graphics - to generate natural-looking clouds, terrain and cellular noise
- soft matter physics - to model diffusion in crowded fluids
- fluid dynamics - to model the degree to which each point on a turbulent wave front is ‘out of sync’
Thorburn
And so we arrive at my thesis.
For the record, my name is indubitably the odd-one-out in this line-up. Many other probabilists have made significant contributions in the decades since Mandelbrot’s work, which I don’t have time to go into.
Compared to Bm, fBm has a set of features that make it a more realistic process with which to model many phenomena, but those same features also tend to make it harder to work with mathematically. For example, the fact that fBm can have ‘long-range dependence’ means that when calculating probabilties of certain events in the future, you need to account for the influence of how the process has evolved in the past. This is in contrast to standard Bm, where the future is independent of the past, and so the past does not need to be factored into the calculations.
For these reasons, it is important for mathematicians to have access to a diverse pool of representations for fBm, so that when trying to do computations with fBm, they have multiple avenues of approach they can try. Here, I am using the word ‘representation’ in a very specific sense. Namely, a representation of a stochastic process is a formula that takes in one process (in our case, standard Bm), and spits out another process (in our case, fBm).
Representations thus provide a link between the world of standard Bm, which probabilists are very familiar with and for which there already exists a comprehensive literature, and the world of fBm, which is more complicated and for which there exists fewer known results. When trying to get a result for fBm, it may be possible to use a representation to convert that result from fBm-world back to standard Bm-world, and prove or derive it there in the more familiar setting, where your toolbox is larger.
My thesis investigated aspects of two representation formulae for fBm: the Muravlev representation, and the Molchan-Golosov representation. The properties of the Muravlev representation I investigated are fairly technical, so I point you to my thesis if you would like to know more. But I can say a bit more about my work on the Molchan-Golosov representation.
With a particular abstract definition of a helix, we can show that fractional Brownian motion is a helix. (This is why Kolmogorov originally called it the Wiener spiral.) This helix exists in an infinite-dimensional space, so it is impossible to visualise it perfectly. Nonetheless, to get the basic intuition, picture a helix in the shape of a simple metal spring. Fractional Brownian motion is that helix. When the Hurst parameter
My research suggests that the Molchan-Golosov representation, which turns standard Bm into fBm, has a geometric interpretation. Specifically, it takes the helix corresponding to standard Bm, and either constricts or relaxes it until it reaches the curvature required to be the requested fBm. I was able to prove that this intuition holds in a discrete-time setting, and showed with computer simulation that it appeared to also hold for the Mochan-Golosov representation, which works in a continuous-time setting.
In sum, my contribution to the theory on fractional Brownian motion was to
- further understanding of some technical properties of the Muravlev representation, and
- propose some plausible geometric intuition for the action of the Molchan-Golosov representation.
From Brown’s original paper, A brief account of microscopical observations made in the months of June, July, and August 1827, on the particles contained in the pollen of plants; and on the general existence of active molecules in organic and inorganic bodies. ↩︎
It’s possible the movement I observed was due to something other than the collision of water molecules with the pollen, such as due to vibrations in the whole sample of water caused by the vibrations of large laboratory equipment nearby. It is difficult to know for sure. Other more-successful attempts to film it can be found here, here, here and here. Footage of true Brownian motion as viewed through Robert Brown’s microscope can be found via this page. ↩︎
Ibid. ↩︎
Ibid. ↩︎
This simulation has been adapted from a bl.ocks.org demo created by Mike Skaug. ↩︎
From 100 years of Einstein’s Theory of Brownian Motion: from Pollen Grains to Protein Trains by Debashish Chowdhury. ↩︎
From Einstein’s original paper, Investigations on the Theory of the Brownian Movement. ↩︎
You can see an example of such correspondence here. Reuther was convinced. ↩︎
GIF by Cyp - Own work, CC BY-SA 3.0, Link. ↩︎
From Norbert Wiener: 1894-1964 by Pesi Masani. ↩︎
From Wiener’s Work in Probability Theory by Joseph Doob. ↩︎
Created by Wolfgang Beyer with the program Ultra Fractal 3. CC BY-SA 3.0. ↩︎
If still curious, you can read my thesis here. Thank you to Janet Newman at CSIRO for lending me time on a microscope to try to observe Brownian motion, the Linnean Society of London for showing me Robert Brown’s microscope, and to the many people who helped throughout my research, particularly my supervisor, Kostya Borovkov.