Chapter 15 · Multivariable Stochastic Calculus

The Mathematics of Connections

When two assets move together — and when they don't. The maths of pairs, portfolios, and hidden correlations.
Welcome to the deep end

Everything so far has tracked a single asset. But real trading is rarely about one stock in isolation. A portfolio manager watches NIFTY and BANKNIFTY simultaneously. A fixed income desk monitors interest rates and exchange rates together. A factor investor models momentum interacting with low-volatility. The moment you have two or more random processes, a new mathematical object appears — the cross-variation — and ignoring it causes every multi-asset model to silently fail.

Single asset vs multi-asset tracking — spiderweb of connections between assets Left panel shows a single random path tracked in isolation. Right panel shows a web of connections between five assets, each connected to the others with correlation lines. Chapters 1–14 Chapter 15 One Brownian Motion W(t) Multiple correlated processes

Single-asset calculus is one wire. Multi-asset calculus is a spiderweb — every pair of assets has a connection that must be mathematically accounted for.

1 · Multiple Brownian motions — independent noises

Start with the simplest case: two independent Brownian Motions W₁(t) and W₂(t). Independent means they have no connection — knowing what W₁ did tells you nothing about what W₂ did. Mathematically, this independence shows up in the cross-variation: the product dW₁·dW₂ equals zero.

Two independent Brownian paths — W1 and W2 moving with no connection between them Two independent random paths shown on the same chart — one in purple and one in teal. They sometimes move together, sometimes apart, with no systematic pattern. 0 W₁(t) W₂(t) dW₁ · dW₂ = 0 — independent paths share no information

Two independent Brownian Motions. One goes up while the other goes down — there is no systematic relationship. Their cross-variation is identically zero.

The Multiplication Table from Chapter 8 now extends to cover products between two different Brownian Motions. The independence of W₁ and W₂ adds one new row:

×dW₁dW₂dt
dW₁ dt 0 ← independence! 0
dW₂ 0 ← independence! dt 0
dt 0 0 0

The key new entry: dW₁ · dW₂ = 0 when the two processes are independent. This is the mathematical statement that two dice rolls carry no shared information. When we introduce correlation later in this chapter, this zero will be replaced by ρ·dt — a non-zero cross-variation that captures how much the two noises move together.

2 · The multivariable Itô-Doeblin formula

When a function f depends on two moving assets X and Y simultaneously, the Itô-Doeblin formula from Chapter 11 gains new terms. The most important newcomer is the cross-gamma term f_{xy} dX dY — the interaction between the two assets that has no equivalent in single-asset calculus.

df = fₜ dt + fₓ dX + f_y dY + ½fₓₓ (dX)² + f_{xy} dX dY + ½f_{yy} (dY)²
fₜ dt
Time drift
Same as before — pure time decay of the function's value
fₓ dX
X-Delta
Sensitivity to asset X moving. The slope in X's direction.
f_y dY
Y-Delta
Sensitivity to asset Y moving. The slope in Y's direction.
½fₓₓ (dX)²
X-Gamma
Curvature from X's own volatility. Same as single-asset Itô correction.
f_{xy} dX dY
Cross-Gamma ✨ NEW
The interaction term. How X's movement changes Y's sensitivity. Zero if independent, non-zero if correlated.
½f_{yy} (dY)²
Y-Gamma
Curvature from Y's own volatility. Symmetric counterpart to X-Gamma.
Cross-gamma interaction — two costs rising together creates superlinear total cost increase A 3D surface diagram showing that when two correlated inputs both increase, the total output rises faster than the sum of individual effects — the cross-gamma term captures this extra curvature Without f_{xy} (linear sum) With f_{xy} (cross-gamma) f_{xy} gap Asset X (e.g. HDFC) Asset Y (e.g. ICICI) When both rise together, the portfolio moves faster than the sum of individual moves

The cross-gamma f_{xy} captures the interaction. When two correlated assets both move in the same direction, the total effect is larger than simply adding each asset's individual contribution.

For two independent assets (dW₁·dW₂ = 0), the cross-gamma term f_{xy} dX dY vanishes entirely. The multivariable formula collapses back to two separate single-asset Itô formulas running in parallel. Correlation is what makes f_{xy} survive — and it is the term that catches most multi-asset models by surprise during market crises, when correlations that were near-zero suddenly spike toward 1.

3 · Lévy's theorem — the duck test for Brownian motion

Usually we start with a Brownian Motion and prove properties about it. Lévy's Theorem runs this logic in reverse. It says: if a process satisfies three specific properties, it must be a Brownian Motion — even if you constructed it from complex combinations of other processes. This reverse identification is a powerful tool for verifying that a synthetic strategy behaves like pure randomness.

🦆
Looks like a duck
M(0) = 0, continuous paths
The process starts at zero and has no jumps — it travels continuously like Brownian Motion.
🚶
Walks like a duck
M(t) is a Martingale
The process has no drift — it is a fair game. The expected future value always equals the current value.
🎙️
Quacks like a duck
[M, M](t) = t
The Quadratic Variation accumulates at exactly rate one — the signature property of Brownian Motion.
If continuous + Martingale + [M,M](t) = t → M(t) IS Brownian Motion
Lévy's Theorem: satisfy all three conditions and your process is provably, mathematically, exactly Brownian Motion — regardless of how it was constructed.

Lévy's Theorem is used to verify synthetic strategies. Suppose you build a complex trading strategy M(t) from a combination of momentum signals, volatility adjustments, and cross-asset hedges. If you can prove that M(t) is a continuous martingale with linear quadratic variation, then M(t) behaves exactly like a Brownian Motion — and all of BSM theory applies to it directly. The theorem is a quality certificate for engineered processes.

4 · Correlation — making the wiggles dance together

Two independent Brownian Motions have zero cross-variation. But in the real world, HDFC Bank and ICICI Bank do not move independently — they share common drivers: RBI policy, banking sector sentiment, the broader NIFTY trend. To model this, we build a correlated Brownian Motion from two independent ones.

The Correlation Recipe
1
Start with two completely independent noises W₁(t) and W₂(t). They share no information whatsoever.
2
Choose a correlation coefficient ρ (rho) between −1 and +1. This is how much you want the new process to move with W₁.
3
Mix them using this formula:
dW₃ = ρ · dW₁ + √(1−ρ²) · dW₂
4
The result W₃ is a valid Brownian Motion (check: its variance is ρ²dt + (1−ρ²)dt = dt ✓) AND it has correlation ρ with W₁. The cross-variation dW₁·dW₃ = ρ dt.
Three correlation scenarios — rho equals minus 1 showing opposite paths, rho equals 0 showing independent paths, rho equals plus 1 showing identical paths Three panels. Left shows two paths moving in exactly opposite directions for rho minus 1. Middle shows two paths with no systematic relationship for rho zero. Right shows two paths moving in lockstep for rho plus 1. ρ = −1 ρ = 0 ρ = +1 Perfect opposites A perfect hedge No relationship dW₁·dW₂ = 0 Move in lockstep No diversification

ρ = −1: a perfect hedge — assets cancel. ρ = 0: pure independence — full diversification. ρ = +1: assets move identically — no diversification benefit at all.

The correlation formula dW₃ = ρ·dW₁ + √(1−ρ²)·dW₂ is the mathematical engine behind Cholesky decomposition — the technique used to generate correlated random scenarios for Monte Carlo simulations. When you simulate 500 correlated NIFTY 500 stocks simultaneously, every pair of paths is linked through a correlation matrix, and Cholesky ensures each path is still a valid Brownian Motion.

5 · The correlated multiplication table

Once we introduce correlation, the multiplication table changes. The cross-variation dW₁·dW₃ is no longer zero — it equals ρ dt. This non-zero cross-variation means the interaction term f_{xy} in the multivariable Itô formula survives and must be computed. Ignoring it understates the total risk of a correlated portfolio.

×dW₁dW₃ (corr. ρ with W₁)dt
dW₁ dt ρ dt ← non-zero! 0
dW₃ ρ dt ← non-zero! dt 0
dt 0 0 0

When ρ = 0, the correlated table reduces to the independent table (all cross terms = 0). When ρ = 1, dW₁·dW₃ = dt = dW₁·dW₁, meaning the two processes are identical. The correlation ρ continuously interpolates between independence and perfect co-movement. In a portfolio of N correlated assets, there are N(N−1)/2 cross-variation pairs, each contributing a ρ·dt term to the total P&L equation.

Try it — the correlation simulator

Set the correlation ρ between two assets and simulate their joint paths. Watch how the cross-gamma interaction term changes the total portfolio variance. Compare the variance of the sum X+Y against what you would predict from each asset alone — the difference is the f_{xy} contribution.

0.60
20%
25%
Var(X) + Var(Y)
Cross-term 2ρσ₁σ₂
Var(X+Y) actual
Realised ρ

Purple = Asset X  |  Teal = Asset Y  |  Gold = combined portfolio X+Y  |  The chips show how the cross-term changes total portfolio variance

Chapter 15 — the mathematics of connections:

What comes next: Girsanov's Theorem — the tool that transforms a world with trends into a risk-neutral world where everything is a martingale. This is how we change the laws of financial physics to make pricing tractable.

Next Chapter: Brownian Bridge →