Marchenko-Pastur & RMT Lecture Presentation

🗣 In plain English

When you measure how a large group of stocks move together using only a limited amount of history, part of what you measure is real and part is just random noise. Random Matrix Theory gives an exact mathematical boundary for how big that noise can look — anything inside the boundary is treated as static and smoothed away; anything that breaks through it is treated as a genuine signal and left alone.

📖 Where the math actually comes from

Random Matrix Theory wasn't built for finance. It was invented in the 1950s by Eugene Wigner, a nuclear physicist, who was trying to model something that looked impossible to model directly: the energy levels of heavy atomic nuclei, with far too many interacting particles to solve exactly. His insight was that the exact details didn't matter — replace the impossibly complex real system with a large random matrix instead, and the statistical pattern of its eigenvalues turned out to describe the real nucleus's energy spectrum remarkably well.

In 1967, two Soviet mathematicians, Vladimir Marchenko and Leonid Pastur, derived the exact law this page is named after — the precise boundary describing where a large random matrix's eigenvalues will fall purely by chance. Decades later, quantitative finance noticed the same problem hiding in plain sight: a portfolio's correlation matrix, estimated from a limited history, is also part real signal and part pure randomness — and the Marchenko-Pastur law gives an exact, borrowed-from-physics answer for where that line sits.

Lecture Notes: MP Law

The Problem: In high dimensions ($p$ close to $n$), random noise creates fake patterns that fool standard statistical tools.
The MP Bounds: The formula calculates exact thresholds ($\lambda_{\min}$ and $\lambda_{\max}$) where noise is strictly confined.
The Visual: Notice how the blue histogram bars fit neatly beneath the red theoretical curve. Everything inside is pure statistical static.

Lecture Notes: Denoising

Constant Residual Method: We take all eigenvalues trapped inside the noise bulk and flatten them to their average value to preserve total variance.
Preserving Signals: Genuine structural signals burst past the upper $\lambda_{\max}$ threshold and are left completely untouched.
Result: Drastically lowers matrix condition numbers and stops models from chasing ghosts.

Lecture Notes: Portfolio Risk

The Raw Trap (Red): Inverting a noisy covariance matrix tells the optimizer to make erratic, massive long/short bets driven by random coincidence.
The Denoised Solution (Blue): MP clipping smooths out the asset weights into safe, rational allocations.
Real-World Impact: Dramatically reduces out-of-sample volatility and boosts the Sharpe Ratio.