Once a process has dt\sqrt{\dt}-sized kicks and no derivative, calculus can no longer be built from slopes — it must be built from increments. This page introduces the one rule that replaces the missing derivative, (dW)2=dt(\dW)^2 = \dt, and shows why every later formula in stochastic calculus depends on it. The developer should care because this is the moment where Brownian motion stops looking like "a rough function" and starts behaving like "the engine of Itô calculus."

The experiment

We took one Brownian path on [0,T][0, T], sampled it at nn equally spaced times, and computed the following sums from the increments ΔWi=W(ti+1)−W(ti)\Delta W_i = W(t_{i+1}) - W(t_i):

  1. ∑(ΔW)2\sum (\Delta W)^2 — the quadratic variation.
  2. ∑∣ΔW∣\sum \lvert \Delta W \rvert — the total variation (distance travelled).

Then we repeated the above experiment using the same path, read on five grids: n=100, 1,000, 10,000, 100,000, 1,000,000n = 100,\ 1{,}000,\ 10{,}000,\ 100{,}000,\ 1{,}000{,}000.

D2Quadratic variation counter
intervals
16,384
Σ(ΔW)²
0.9898T = 1.00
Σ|ΔW|
101.4→ ∞
Σ(ΔW)² − T
−0.01020.9 sd

Σ(ΔW)² has mean exactly T on every partition and standard deviation T·√(2/n), so it settles rather than drifts: at 16k intervals that spread is 0.0110. Σ|ΔW| has mean √(2nT/π) and multiplies by √2 with every doubling. A path with finite quadratic variation and infinite length is what (dW)² = dt means, and it is why the second-order term in a Taylor expansion cannot be dropped.

·

What happened to path 1 as n increased

One path, two sums, five grids

The first path of the ensemble — 400 paths drawn once on a grid of step 10−610^{-6} over [0,1][0, 1] and coarsened by 10 at a time, so every row is that one path on a different grid. Seed 20260906, from the committed experiments/2_3_quadratic_variation.json (results.grids[*].sum_dW_sq.first_paths[0] and results.grids[*].sum_abs_dW.first_paths[0]).

nndt\dt∑(ΔW)2\sum (\Delta W)^2∑∣ΔW∣\sum \lvert \Delta W \rvert
10010−210^{-2}0.88287.40
1,00010−310^{-3}0.967724.82
10,00010−410^{-4}0.998779.63
100,00010−510^{-5}0.9992252.09
1,000,00010−610^{-6}0.9997798.12
  • The quadratic variation: ∑(ΔW)2\sum (\Delta W)^2 stabilised. As the mesh was refined, the sum of squared increments marched steadily towards TT (towards 1 for this experiment). This is the hallmark of Brownian motion: ∑(ΔW)2→T\sum (\Delta W)^2 \to T.
  • The total variation: ∑∣ΔW∣\sum \lvert \Delta W \rvert blew up. The sum of absolute increments got bigger, and kept growing without bound every time the interval was reduced. By n=106n = 10^6 it was enormous (798) — and still rising.
  • This is the signature fact: Brownian motion has infinite total variation (grows like n\sqrt{n} across five decades in this exercise) but finite quadratic variation.

In plain words:

  • The squared increments converge to a finite number.
  • The absolute increments explode because the path is too rough — every refinement reveals more jaggedness.

This contrast is exactly why classical calculus fails and why stochastic calculus is built on increments and the rule (dW)2=dt(\dW)^2 = \dt, not on derivatives.

Two words: quadratic variation and total variation

Quadratic variation is the total of squared increments over [0,T][0, T], in the limit of a fine grid; total variation is the same total of absolute increments. A smooth curve you could draw on paper has finite total variation and its quadratic variation collapses to 0 as the mesh gets finer, because its increments shrink faster than dt\sqrt{\dt}, as shown with the function sin⁡(2πt)\sin(2\pi t). Brownian motion flips this behaviour: its total variation explodes, but its quadratic variation locks onto TT, giving you the one stable quantity you can build calculus from.

Why the sum settles

A Brownian increment ΔW\Delta W over a step of size dt=T/n\dt = T/n is ΔW∼N(0,dt)\Delta W \sim N(0, \dt). So one squared increment is (ΔW)2=dt⋅Z2(\Delta W)^2 = \dt \cdot Z^2, where Z∼N(0,1)Z \sim N(0, 1). Therefore, every term is dt\dt times a squared standard normal.

A standard normal has variance 1, which makes E[Z2]=1\E[Z^2] = 1, and its fourth moment is 3, which makes Var⁡(Z2)=3−1=2\Var(Z^2) = 3 - 1 = 2. Therefore, Z2Z^2 has mean 1 and variance 2:

  • E[(ΔW)2]=dt\E[(\Delta W)^2] = \dt,
  • Var⁡[(ΔW)2]=2dt2\Var[(\Delta W)^2] = 2\dt^2.

So each term is:

  • on average the size of dt\dt,
  • with a spread of order dt\dt.

If we add nn terms, we sum n=T/dtn = T/\dt independent copies:

  • Mean of the sum is fixed at TT: E[∑(ΔW)2]=n⋅dt=T\E[\sum (\Delta W)^2] = n \cdot \dt = T.
  • Variance of the sum shrinks like 1/n1/n: Var⁡[∑(ΔW)2]=n⋅2dt2=2Tdt=2T2/n\Var[\sum (\Delta W)^2] = n \cdot 2\dt^2 = 2T\dt = 2T^2/n.
  • Therefore, the standard deviation shrinks like 1/n1/\sqrt{n}.

If T=1T = 1, at n=100n = 100, sd⁡=2/n≈0.14\sd = \sqrt{2/n} \approx 0.14, so the sum is typically 1±0.141 \pm 0.14 and we see noticeable wobble. However, at n=106n = 10^6, sd⁡=2/106≈0.0014\sd = \sqrt{2/10^6} \approx 0.0014; the sum is typically 1±0.00141 \pm 0.0014 and the wobble is almost gone.

Each squared increment is wild and lopsided, but it's only dt\dt-sized. As nn increases and dt→0\dt \to 0, the mean stays at 1, and the spread collapses like 1/n1/\sqrt{n}. By the time we reach a million steps, the spread is microscopic, so the sum lands almost exactly on 1 every time. That's why the quadratic variation settles.

The path has no length

For one step of size dt\dt, a Brownian increment is ΔW∼N(0,dt)\Delta W \sim N(0, \dt), so ∣ΔW∣=dt⋅∣Z∣\lvert \Delta W \rvert = \sqrt{\dt} \cdot \lvert Z \rvert with Z∼N(0,1)Z \sim N(0, 1). The average size is E∣ΔW∣=dt⋅E∣Z∣=dt⋅2/π≈0.798⋅dt\E\lvert \Delta W \rvert = \sqrt{\dt} \cdot \E\lvert Z \rvert = \sqrt{\dt} \cdot \sqrt{2/\pi} \approx 0.798 \cdot \sqrt{\dt}. Summing n=T/dtn = T/\dt of them gives:

  • E[∑∣ΔW∣]≈0.798⋅T/dt=0.798⋅nT→∞\E[\sum \lvert \Delta W \rvert] \approx 0.798 \cdot T/\sqrt{\dt} = 0.798 \cdot \sqrt{nT} \to \infty as dt→0\dt \to 0.

That's why a Brownian path has no finite length (its total variation blows up), whereas a smooth curve on paper has a finite ∫∣f′∣\int \lvert f' \rvert. By contrast, a ±dt\pm\sqrt{\dt} coin walk has ∣ΔW∣=dt\lvert \Delta W \rvert = \sqrt{\dt} every step, so ∑∣ΔW∣=n⋅dt=nT\sum \lvert \Delta W \rvert = n \cdot \sqrt{\dt} = \sqrt{nT} exactly, which still diverges but in a perfectly regular way.

The smooth curve has it the other way round

For the smooth curve f(t)=sin⁡(2πt)f(t) = \sin(2\pi t), on the same five grids, the squared increments ∑(Δf)2\sum (\Delta f)^2:

  • Each increment is Δf≈f′(t) dt=2πcos⁡(2πt) dt\Delta f \approx f'(t)\,\dt = 2\pi \cos(2\pi t)\,\dt.
  • So (Δf)2≈4π2cos⁡2(2πt) dt2(\Delta f)^2 \approx 4\pi^2 \cos^2(2\pi t)\,\dt^2.
  • And summing nn of them gives something of order ∫0T4π2cos⁡2(2πt) dt⋅dt→0\int_0^T 4\pi^2 \cos^2(2\pi t)\,\dt \cdot \dt \to 0 as dt→0\dt \to 0.
  • The average of cos⁡2\cos^2 over a period is 12\tfrac12, so ∑(Δf)2≈4π2⋅12⋅dt=2π2⋅dt=19.74/n\sum (\Delta f)^2 \approx 4\pi^2 \cdot \tfrac12 \cdot \dt = 2\pi^2 \cdot \dt = 19.74/n, and the "true" quadratic variation of a smooth ff is zero.

The absolute increments ∑∣Δf∣\sum \lvert \Delta f \rvert:

  • ∣Δf∣≈∣f′(t)∣ dt=2π∣cos⁡(2πt)∣ dt\lvert \Delta f \rvert \approx \lvert f'(t) \rvert\,\dt = 2\pi \lvert \cos(2\pi t) \rvert\,\dt.
  • So ∑∣Δf∣≈∫0T2π∣cos⁡(2πt)∣ dt\sum \lvert \Delta f \rvert \approx \int_0^T 2\pi \lvert \cos(2\pi t) \rvert\,\dt, which is exactly the total rise and fall of the sine wave: 0→1→−1→00 \to 1 \to -1 \to 0 gives 1+2+1=41 + 2 + 1 = 4.
  • As you refine the grid, ∑∣Δf∣\sum \lvert \Delta f \rvert just converges to this finite value.

So:

  • Smooth curve (like sin⁡\sin)
    • ∑∣Δf∣\sum \lvert \Delta f \rvert: finite, equals the total variation (length of the path).
    • ∑(Δf)2\sum (\Delta f)^2: tends to 0 as the grid is refined.
  • Brownian path WW
    • ∑∣ΔW∣\sum \lvert \Delta W \rvert: blows up to ∞\infty (no finite length).
    • ∑(ΔW)2\sum (\Delta W)^2: locks onto TT (finite quadratic variation).

Smooth: finite absolute, zero squared.

Brownian: infinite absolute, finite squared.

The rulebook

On a grid of step dt\dt, a Brownian increment is ΔW=dt⋅Z\Delta W = \sqrt{\dt} \cdot Z:

  • For ∑ΔW⋅dt\sum \Delta W \cdot \dt, one term is ΔW⋅dt=dt⋅Z⋅dt=dt3/2⋅Z\Delta W \cdot \dt = \sqrt{\dt} \cdot Z \cdot \dt = \dt^{3/2} \cdot Z. The terms are signed and cancel, and in fact the sum is exactly dt⋅WT\dt \cdot W_T, which →0\to 0 like 1/n1/n.
  • For ∑(dt)2\sum (\dt)^2, one term is dt2\dt^2. Summing n=T/dtn = T/\dt of them gives n⋅dt2=(T/dt)⋅dt2=T⋅dt→0n \cdot \dt^2 = (T/\dt) \cdot \dt^2 = T \cdot \dt \to 0.

Both sums collapse because every term is "too small" and the power of dt\dt wins.

∑(ΔW)2\sum (\Delta W)^2 does not collapse, as each squared increment is (ΔW)2=dt⋅Z2(\Delta W)^2 = \dt \cdot Z^2 — one term is order dt\dt, and summing n=T/dtn = T/\dt of them gives n⋅dt=Tn \cdot \dt = T. This is the only dt\dt-power that survives refinement.

The rulebook — five limits become the algebra of stochastic calculus:

  • ∑ΔW=WT\sum \Delta W = W_T — survives, random (the endpoint)
  • ∑Δt=T\sum \Delta t = T — survives, deterministic
  • ∑(ΔW)2=T\sum (\Delta W)^2 = T — survives, deterministic — the surprise
  • ∑ΔW⋅Δt=0\sum \Delta W \cdot \Delta t = 0 — dies
  • ∑(Δt)2=0\sum (\Delta t)^2 = 0 — dies

In the conventional form: (dW)2=dt(\dW)^2 = \dt, dW⋅dt=0\dW \cdot \dt = 0, (dt)2=0(\dt)^2 = 0.

This means:

  • "=dt= \dt": when you sum nn terms of size dt\dt, the limit is TT, a finite non-zero number.
  • "=0= 0": when you sum nn terms of size dt3/2\dt^{3/2} or dt2\dt^2, the limit vanishes as the mesh goes to zero.

They are statements about the limiting behaviour of sums, not about any single step. That's the rulebook: only ∑ΔW\sum \Delta W, ∑Δt\sum \Delta t and ∑(ΔW)2\sum (\Delta W)^2 survive; everything else involving extra powers or cross terms of dW\dW or dt\dt dies.

Why ordinary calculus fails

For an example function g(x)=x2g(x) = x^2:

  • Each step is (W+ΔW)2−W2=W2+2W⋅ΔW+(ΔW)2−W2=2W⋅ΔW+(ΔW)2(W + \Delta W)^2 - W^2 = W^2 + 2W \cdot \Delta W + (\Delta W)^2 - W^2 = 2W \cdot \Delta W + (\Delta W)^2 — two terms, one with ΔW\Delta W and the other with (ΔW)2(\Delta W)^2.
  • Sum over the grid: the left side becomes WT2W_T^2 — the path's end point squared.
  • The right side becomes 2∑W⋅ΔW+∑(ΔW)22\sum W \cdot \Delta W + \sum (\Delta W)^2.
  • So WT2=2∑W⋅ΔW+∑(ΔW)2W_T^2 = 2\sum W \cdot \Delta W + \sum (\Delta W)^2.
  • For a smooth path x(t)x(t) instead of WW, xT2=2∑x⋅Δx+∑(Δx)2x_T^2 = 2\sum x \cdot \Delta x + \sum (\Delta x)^2, but since ∑(Δx)2→0\sum (\Delta x)^2 \to 0 as dt→0\dt \to 0, xT2=2∑x⋅Δxx_T^2 = 2\sum x \cdot \Delta x as dt→0\dt \to 0 or, alternatively, xT2=2⋅∫x dxx_T^2 = 2 \cdot \int x \,\dd x. That is the chain rule d(x2)=2x⋅dx\dd(x^2) = 2x \cdot \dd x.
  • But for our Wiener process, ∑(ΔW)2→T\sum (\Delta W)^2 \to T as dt→0\dt \to 0, so WT2=2∑W⋅ΔW+TW_T^2 = 2\sum W \cdot \Delta W + T and the chain rule is short by TT.

For any smooth gg, Taylor gives g(W+ΔW)−g(W)=g′(W)⋅ΔW+12g′′(W)⋅(ΔW)2+16g′′′(W)⋅(ΔW)3+…g(W + \Delta W) - g(W) = g'(W) \cdot \Delta W + \tfrac12 g''(W) \cdot (\Delta W)^2 + \tfrac16 g'''(W) \cdot (\Delta W)^3 + \ldots, and for a smooth path the first term is the only one which survives as dt\dt gets smaller — anything quadratic and above disappears.

For Brownian motion, the second term also survives, since ∑(ΔW)2→T\sum (\Delta W)^2 \to T as dt→0\dt \to 0. So for Brownian motion the chain rule is short by ∑12g′′(W)⋅(ΔW)2=12∑g′′(W)⋅dt\sum \tfrac12 g''(W) \cdot (\Delta W)^2 = \tfrac12 \sum g''(W) \cdot \dt, since (dW)2=dt(\dW)^2 = \dt, or 12∫g′′(W) dt\tfrac12\int g''(W) \,\dt as dt→0\dt \to 0.

The third term does not survive. Averaged over the experiment's 400 paths, ∑∣ΔW∣3\sum \lvert \Delta W \rvert^3 is 0.160 at n=100n = 100 and 0.0016 at n=106n = 10^6: it falls like 1/n1/\sqrt{n}.

The coin has it exactly

For a coin toss with each increment ΔW=±dt\Delta W = \pm\sqrt{\dt}, without any Gaussian component to the magnitude of the step size, (ΔW)2=dt(\Delta W)^2 = \dt, and since dt=T/n\dt = T/n, ∑(ΔW)2=n⋅dt=T\sum (\Delta W)^2 = n \cdot \dt = T. The 2/n\sqrt{2/n} spread in the Gaussian version is the price of the Gaussian steps.

Into 2.4

We now have the rulebook for stochastic calculus: every term in a Taylor expansion can be sized by powers of dt\dt, and only the dt\dt-sized ones survive summation. For Brownian motion that means one extra survivor — the 12g′′(W) dt\tfrac12 g''(W) \,\dt term born from (dW)2=dt(\dW)^2 = \dt — the single place where the chain rule is short.

Section 2.4, the Itô integral, takes that rulebook and applies it to a specific gg, showing how ∫W dW\int W \,\dW acquires exactly that missing term and why stochastic integration has a different algebra from ordinary calculus.

Check

A Brownian path on [0, 1] is chopped into n equal steps and you compute Σ(ΔW)². As n → ∞ it:

Same path, same grids. Σ|ΔW| — the total distance travelled — as n → ∞:

Replace W by the smooth curve f(t) = sin(2πt) on the same grids. As n → ∞, Σ(Δf)² and Σ|Δf|:

On the same grids, Σ ΔW·Δt and Σ(Δt)² as n → ∞:

Expand g(W + ΔW) − g(W) by Taylor and sum over the grid. Which term behaves differently for W than for a smooth path, and why?