The Strange Math That Predicts (Almost) Anything

Key Concepts:

• Law of Large Numbers: The average outcome of independent trials approaches the expected value as the number of trials increases.
• Independence vs. Dependence: Whether events influence each other. Independent events don't, dependent events do.
• Markov Chain: A sequence of events where the probability of each event depends only on the state attained in the previous event.
• Transition Probabilities: The probabilities of moving from one state to another in a Markov chain.
• Monte Carlo Method: A computational technique that uses random sampling to obtain numerical results.
• PageRank: An algorithm used by Google to rank websites in search results based on the network of links pointing to them.
• Damping Factor: A probability used in PageRank to ensure that a random surfer can jump to any page on the web, preventing them from getting stuck in loops.
• Attention (in Language Models): A mechanism that allows a model to focus on the most relevant parts of the input when making predictions.
• Feedback Loops: Situations where the output of a system influences its input, making it difficult to predict its behavior.

1. The Math Feud in Russia:

• In 1905, Russia was divided between Tsarists (supporting the Tsar) and Socialists (demanding political reform).
• This division extended to mathematicians: Pavel Nekrasov (Tsarist, "Tsar of Probability") vs. Andrey Markov (Socialist, "Andrey The Furious").
• Nekrasov argued math could explain free will and the will of God, while Markov, an atheist, criticized Nekrasov's work as unrigorous.

2. Nekrasov's Argument for Free Will:

• Nekrasov believed the Law of Large Numbers implied independence.
• He observed convergence in social statistics (e.g., Belgian marriages from 1841-1845, averaging around 29,000 per year).
• He reasoned that because these statistics followed the Law of Large Numbers, the decisions causing them (marriage, crime, birth) must be independent acts of free will.

3. Markov's Counter-Argument and Markov Chains:

• Markov aimed to prove that dependent events could also follow the Law of Large Numbers.
• He used "Eugene Onegin" by Alexander Pushkin, analyzing the first 20,000 letters.
• He found that vowel-vowel pairs occurred only 6% of the time, far less than the 18% expected if letters were independent.
• Markov created a "prediction machine" (Markov chain) with states for vowels and consonants and transition probabilities between them.
• This machine demonstrated that even with dependent events, the ratio of vowels to consonants converged to the observed values (43% vowels, 57% consonants).
• Markov concluded that observing convergence in social statistics doesn't prove independence or free will.
• "Thus, free will is not necessary to do probability."
• Markov's work allowed for probability calculations with dependent events, a significant breakthrough.

4. The Manhattan Project and the Monte Carlo Method:

• Stanislaw Ulam, a mathematician, worked on the Manhattan Project.
• He sought to understand neutron behavior inside a nuclear bomb to determine the amount of uranium-235 needed.
• After suffering from encephalitis, Ulam played Solitaire and wondered about the probability of winning a randomly shuffled game.
• He realized he could approximate the answer by playing hundreds of games and counting wins.
• Ulam applied this idea to neutron behavior, simulating random outcomes.
• John von Neumann recognized the power of Ulam's idea but noted that neutron behavior is dependent, requiring a Markov chain.
• They created a simplified Markov chain to model neutron scattering, absorption, and fission.
• This chain was run on the ENIAC, the first electronic computer, to calculate the multiplication factor (k).
• If k > 1, the reaction grows exponentially, leading to a bomb.
• Ulam named the method "Monte Carlo" after his uncle, a gambler, and the Monte Carlo Casino.

5. Google's PageRank Algorithm:

• In the mid-1990s, the internet exploded, creating a need for better search engines.
• Early search engines ranked pages by keyword frequency, which was easily manipulated.
• Sergey Brin and Larry Page at Stanford developed PageRank, modeling the web as a Markov chain.
• Each webpage is a state, and links between pages are transitions.
• The more links a page receives, the higher its rank. However, links from highly ranked pages are worth more.
• A "damping factor" (15%) was introduced to prevent random surfers from getting stuck in loops and to ensure all parts of the web are explored.
• PageRank allowed Google to provide more relevant search results.
• Google was initially called BackRub, then Googol, and finally Google due to a misspelling.

6. Language Models and Markov Chains:

• Claude Shannon, the father of information theory, used Markov chains to predict text.
• He found that using more previous words as predictors improved the accuracy of the predictions.
• Modern language models use Markov chains and "attention" mechanisms to predict the next word in a sequence.
• Attention allows the model to focus on the most relevant parts of the input.
• However, feedback loops, where the output of a model becomes training data for future models, can lead to a "dull, stable state."

7. Limitations and Power of Markov Chains:

• Markov chains don't work well for systems with strong feedback loops, such as global warming.
• However, for many dependent systems, Markov chains offer a way of doing probability.
• Markov chains are "memoryless," meaning they only consider the current state, simplifying complex systems.
• "Problem-solving is often a matter of cooking up an appropriate Markov chain."

8. Card Shuffling and Randomness:

• Riffle shuffling a deck of 52 cards seven times makes it basically random.
• Overhand shuffling requires over 2,000 shuffles to achieve randomness.

9. Conclusion:

• The development of Markov chains stemmed from a math feud between Nekrasov and Markov.
• Markov's work has had a profound impact on various fields, including nuclear physics (Monte Carlo method), search engines (PageRank), and language models.
• Markov chains provide a powerful tool for modeling and predicting the behavior of complex, dependent systems by focusing on the current state and ignoring the past.