When the price itself jitters like Brownian motion, anything built from it — an option, a hedge, even eWe^W — jitters in a more complicated way. The chain rule can't see that. Itô's lemma is the fix: it shows how curvature turns noise into drift, and why 2.4's −12T-\tfrac12 T was just one instance of a general phenomenon. After this page, the reader can write the dynamics of any smooth payoff, not just W2W^2.

What this page shows, and how

The claim is one line: for a smooth gg, dg(Wt)=12g′′(Wt) dt+g′(Wt) dWt\dd g(W_t) = \tfrac12 g''(W_t) \,\dt + g'(W_t) \,\dW_t — the chain rule plus a curvature term. The page shows this in four moves, and the tables come in that order.

First, the algebra. Taylor's expansion, one step at a time and summed over the grid, is an identity with a (ΔW)2(\Delta W)^2 term in it. For g=x2g = x^2 that identity is 2.3's WT2=2∑W⋅ΔW+∑(ΔW)2W_T^2 = 2\sum W \cdot \Delta W + \sum (\Delta W)^2, rearranged in 2.4 into the −12T-\tfrac12 T; the first table shows it holding to 10−1510^{-15}.

Second, the limit. Two things turn that identity into the lemma: (ΔW)2(\Delta W)^2 is replaced by dt\dt, and everything of third order is dropped. Each costs something on a finite grid, and the second table measures both costs for exe^x and x3x^3 — the replacement costs a noise that shrinks like 1/n1/\sqrt{n}, the dropping costs a remainder that shrinks like 1/n1/n. That is the whole error budget of the lemma.

Third, the payoff. The lemma is applied to the one function whose growth can be seen, exe^x. If the chain rule were the whole story, E[eWt]\E[e^{W_t}] would be constant; the lemma's dt\dt term says it grows at rate 12\tfrac12, so E[eWt]=et/2\E[e^{W_t}] = e^{t/2}. The third and fourth tables check that, split the growth into its dW\dW part (zero) and its dt\dt part (all of it), and find the one correction — 12t\tfrac12 t in the exponent — that makes the mean flat. That is the martingale eWt−12te^{W_t - \frac12 t}, and it is where 2.6's −12σ2-\tfrac12\sigma^2 will come from.

Fourth, the controls. A smooth curve gets no dt\dt term, and adding one is a bias that never shrinks; the coin walk has both forms exact. Between the third and fourth moves the general form df=(∂f/∂t+a⋅∂f/∂x+12b2⋅∂2f/∂x2) dt+b⋅∂f/∂x dW\dd f = (\partial f/\partial t + a \cdot \partial f/\partial x + \tfrac12 b^2 \cdot \partial^2 f/\partial x^2) \,\dt + b \cdot \partial f/\partial x \,\dW is assembled from the same pieces, with log⁡S\log S worked as the example 2.6 starts from.

The experiment — what was done

Every number comes from one experiment. First, 400 Brownian paths were drawn once at 10610^6 steps and then read on the five grids n=102…106n = 10^2 \ldots 10^6. On each grid, for three functions — x2x^2, exe^x and x3x^3 — we checked both sides of Itô's lemma in its two forms: the exact (ΔW)2(\Delta W)^2-form and the dt\dt-form, with the third-order term measured on its own. A second run drew 20,000 fresh paths at n=104n = 10^4 to measure E[eWt]\E[e^{W_t}] at five times, splitting the growth into its Itô-sum part and its dt\dt part, with the candidates eW−12te^{W - \frac12 t}, eW−te^{W - t} and eW−te^{W - \sqrt{t}} beside it. The same grids were run on the smooth curve x(t)=t2x(t) = t^2 to show what happens when the quadratic variation vanishes. Finally, the ±dt\pm\sqrt{\dt} coin walk from 2.1 was checked on x2x^2 and x3x^3, five paths per grid, to show the same algebra in a non-Gaussian setting.

·

Reading the x² table — the exact case

Path 1 ends at WT=−1.1702W_T = -1.1702, so the left side is WT2=1.3694W_T^2 = 1.3694 on every grid.

Path 1 for x², five grids

The first path of the ensemble — 400 paths drawn once on a grid of 10610^6 steps over [0,1][0, 1] and coarsened, so every row is that one path read on a different grid. Seed 20260914, from the committed experiments/2_5_ito_lemma.json (grids[*].functions.x^2.first_paths.{ito_sum,qv_term,r_qv,dt_term,r_dt}[0]; the sum column is the first two added, and T=1T = 1).

nn∑2Wi⋅ΔWi\sum 2W_i \cdot \Delta W_i∑(ΔW)2\sum (\Delta W)^2sum(ΔW)2(\Delta W)^2-form residual∑dt=T\sum \dt = Tdt\dt-form residual
1000.33381.03561.3694−2×10−16-2 \times 10^{-16}1+0.0356
1,0000.37690.99251.369401−0.0075
10,0000.37300.99641.36944×10−164 \times 10^{-16}1−0.0036
100,0000.37660.99281.3694−7×10−16-7 \times 10^{-16}1−0.0072
1,000,0000.37110.99831.36946×10−166 \times 10^{-16}1−0.0017
x² dt-form residual, 400 paths

All 400 paths on each grid, from grids[*].functions.x^2.r_dt.{mean,se_of_mean,sd} of the same file. The last column is noise.rms_expected_limit, the formula T2/nT\sqrt{2/n} the spread is compared against, not a measurement.

nnmean of dt\dt-form residual± s.e.sdT2/nT\sqrt{2/n}
100+0.01020.00670.13440.1414
1,000−0.00250.00220.04390.0447
10,000−0.00040.00070.01420.0141
100,000−0.00030.00020.00430.0045
1,000,000+0.00000.00010.00140.0014

Down the x2x^2 table every column settles. The Itô sum ∑2Wi⋅ΔWi\sum 2W_i \cdot \Delta W_i and the quadratic-variation column ∑(ΔW)2\sum (\Delta W)^2 both settle toward their final values, and their sum is WT2=1.3694W_T^2 = 1.3694 on every grid — the left side never moves. The (ΔW)2(\Delta W)^2-form residual is a column of 10−1610^{-16}s: that is the identity WT2=2∑W⋅ΔW+∑(ΔW)2W_T^2 = 2\sum W \cdot \Delta W + \sum (\Delta W)^2 holding to machine precision at every nn. The dt\dt-form residual is exactly ∑(ΔW)2−1\sum (\Delta W)^2 - 1 row by row, because the dt\dt-form replaces ∑(ΔW)2\sum (\Delta W)^2 by T=1T = 1; it is the grid's (ΔW)2(\Delta W)^2-vs-dt\dt noise, not a failure of the lemma. The second table shows what that noise does over many paths: its mean is 0 to within two standard errors at every nn (+0.010 ± 0.007 at n=100n = 100, 0.0000 ± 0.0001 at 10610^6), and its spread matches T2/nT\sqrt{2/n} to within the 400-path noise (0.134 against 0.141 at n=100n = 100, 0.0014 against 0.0014 at 10610^6). So for x2x^2 the dt\dt-form is right on average at every grid size, and right on each individual path only in the limit — and the error is the grid's, not the lemma's.

Reading the eˣ and x³ table — the remainder

Path 1 again (WT=−1.1702W_T = -1.1702). The left side is eWT−1=−0.6897e^{W_T} - 1 = -0.6897 for exe^x and WT3=−1.6025W_T^3 = -1.6025 for x3x^3.

Path 1 for eˣ and x³

The same path 1, for the other two functions, from grids[*].functions.{exp,x^3}.first_paths.{lhs,ito_sum,qv_term,dt_term,r_qv,cubic_term,r_dt}[0] of the same file.

nnggleft side∑g′⋅ΔW\sum g' \cdot \Delta W12∑g′′⋅(ΔW)2\tfrac12\sum g'' \cdot (\Delta W)^212∑g′′⋅dt\tfrac12\sum g'' \cdot \dt(ΔW)2(\Delta W)^2-form residual16∑g′′′⋅(ΔW)3\tfrac16\sum g''' \cdot (\Delta W)^3dt\dt-form residual
100exe^x−0.6897−0.90790.22380.2247−0.0057−0.0064−0.0065
1,000exe^x−0.6897−0.91360.22450.2237−0.00059−0.00065+0.0003
10,000exe^x−0.6897−0.91230.22260.2235−0.000070−0.000075−0.0009
100,000exe^x−0.6897−0.91140.22170.2235−3.6×10−6-3.6 \times 10^{-6}−4.2×10−6-4.2 \times 10^{-6}−0.0018
1,000,000exe^x−0.6897−0.91280.22310.2235−4.3×10−7-4.3 \times 10^{-7}−4.9×10−7-4.9 \times 10^{-7}−0.0004
100x3x^3−1.60251.3047−2.8488−2.6240−0.0584−0.0584−0.2832
1,000x3x^3−1.60250.9832−2.5798−2.6239−0.0060−0.0060+0.0382
10,000x3x^3−1.60251.0188−2.6206−2.6257−0.00067−0.00067+0.0044
100,000x3x^3−1.60251.0098−2.6123−2.6258−0.000019−0.000019+0.0135
1,000,000x3x^3−1.60251.0190−2.6215−2.6258−3.3×10−6-3.3 \times 10^{-6}−3.3×10−6-3.3 \times 10^{-6}+0.0043

Down both tables the pattern is immediate. The (ΔW)2(\Delta W)^2-form residual — the gap between the exact Taylor sum and the left side — shrinks by about a factor of ten each time the grid is refined. That is the third-order term dying like 1/n1/n: for x3x^3 it is 0.058 at n=100n = 100 and 3×10−63 \times 10^{-6} at n=106n = 10^6; for exe^x, 0.006 and 4×10−74 \times 10^{-7}. Beside it, the dt\dt-form residual's spread over the 400 paths only shrinks threefold per row (for exe^x: 0.120, 0.035, 0.010, 0.0045, 0.0013), while on a single path it wanders (path 1: −0.0065, +0.0003, −0.0009, −0.0018, −0.0004): that is the grid's (ΔW)2(\Delta W)^2-vs-dt\dt noise, which dies like 1/n1/\sqrt{n} and is larger at every nn — the ratio of the two spreads is n\sqrt{n} for both functions, about 10 at n=100n = 100 (11 for exe^x, 7 for x3x^3) and about 1,000 at n=106n = 10^6.

For x3x^3 the (ΔW)2(\Delta W)^2-form residual and the third-order term are the same number to every printed digit. That is because the Taylor expansion stops at third order: the (ΔW)2(\Delta W)^2-form is exact algebra, and the only remainder is 16∑g′′′(ΔW)3=∑(ΔW)3\tfrac16\sum g'''(\Delta W)^3 = \sum (\Delta W)^3, since g′′′=6g''' = 6. For exe^x the two columns differ slightly, because the function has fourth-order and higher terms; those extra pieces are about ten times smaller at every nn — the same 1/n1/n rate with a smaller constant — and the table shows them as the tiny difference between the two columns.

The conclusion is that on a grid, the lemma's error does not come from dropping the third-order term — that cost dies like 1/n1/n and is tiny. The error comes from replacing (ΔW)2(\Delta W)^2 by dt\dt — the grid's noise — which dies only like 1/n1/\sqrt{n} and dominates every row.

Why — Taylor summed, and the two forms

Start from one step. Taylor says how much a smooth gg changes when its argument moves from WiW_i to Wi+ΔWiW_i + \Delta W_i:

g(Wi+ΔWi)−g(Wi)=g′(Wi)⋅ΔWi+12g′′(Wi)⋅(ΔWi)2+(third order and beyond).g(W_i + \Delta W_i) - g(W_i) = g'(W_i) \cdot \Delta W_i + \tfrac12 g''(W_i) \cdot (\Delta W_i)^2 + (\text{third order and beyond}).

Add the nn steps. The left sides telescope — every interior g(Wi)g(W_i) cancels — leaving g(WT)−g(0)g(W_T) - g(0). The right side becomes three sums. That exact algebra is the (ΔW)2(\Delta W)^2-form of the lemma: an equality on every grid, for every path, with no limit taken.

Two things turn it into the dt\dt-form. First, replace each (ΔWi)2(\Delta W_i)^2 by dt\dt, using 2.3's rule that ∑(ΔW)2→T\sum (\Delta W)^2 \to T. Second, drop the third-order remainder, which vanishes in the limit. What is left is an arrow, not an equals sign: a limit statement whose grid error is the (ΔW)2(\Delta W)^2-vs-dt\dt noise plus the 1/n1/n remainder, the noise dominating.

Written for one infinitesimal step, the result is the lemma in d-notation, drift first:

dg(Wt)=12g′′(Wt) dt+g′(Wt) dWt.\dd g(W_t) = \tfrac12 g''(W_t) \,\dt + g'(W_t) \,\dW_t.

The chain rule has only the second term. The lemma adds the first — the curvature term — which is exactly the piece the chain rule misses when the path has quadratic variation.

The remainder — why it dies faster than 2.3's bound

The third-order term is guaranteed to vanish because each ∣ΔW∣3\lvert \Delta W \rvert^3 is of size dt3/2\dt^{3/2}, and 2.3 already measured ∑∣ΔW∣3→0\sum \lvert \Delta W \rvert^3 \to 0 like 1/n1/\sqrt{n}. That bound is enough to know the remainder goes away. But the term itself keeps its signs: it is 16∑g′′′(Wi)(ΔWi)3\tfrac16\sum g'''(W_i)(\Delta W_i)^3, and (ΔWi)3(\Delta W_i)^3 is an odd moment with mean zero. The positive and negative cubes cancel. In this run the signed ∑(ΔW)3\sum (\Delta W)^3 has standard deviation 15⋅T3/2/n\sqrt{15} \cdot T^{3/2}/n (0.0387, 0.0039, 0.00039, 0.000039, 0.0000039), measured 0.0388, 0.0036, 0.00038, 0.000038, 0.0000038 down the grids — a factor n\sqrt{n} smaller than the bound from ∑∣ΔW∣3\sum \lvert \Delta W \rvert^3. The coin walk shows the same split cleanly: ∑(ΔW)3=dt⋅WT\sum (\Delta W)^3 = \dt \cdot W_T exactly, which is 1/n1/n, while ∑∣ΔW∣3=dt\sum \lvert \Delta W \rvert^3 = \sqrt{\dt} exactly, which is 1/n1/\sqrt{n}. Same steps, two rates, and the difference is the signs.

The unsigned bound says the remainder cannot survive; the signed sum shows it dies even faster, because the cubes cancel. The cost of dropping the third-order term is tiny — 1/n1/n — and never the dominant error on a grid.

E[e^(W_t)] = e^(t/2) — the growth is the dt term

For g(x)=exg(x) = e^x the lemma gives

d(eWt)=12eWt dt+eWt dWt.\dd(e^{W_t}) = \tfrac12 e^{W_t} \,\dt + e^{W_t} \,\dW_t.

The dWt\dW_t term is an Itô integral, so its mean is zero; if that were all, E[eWt]\E[e^{W_t}] would stay at 1. The dt\dt term is half of eWte^{W_t} itself, so on average the process grows at a constant rate 12\tfrac12, compounding in time — the statement E[eWt]=et/2\E[e^{W_t}] = e^{t/2}.

E[e^(Wₜ)] and its candidates

The second run: 20,000 fresh paths at n=104n = 10^4, read at five times, from expectations.running[*].{exp_W,exp_W_minus_half_t,exp_W_minus_t,exp_W_minus_sqrt_t} of the same file. The et/2e^{t/2} and ete^{\sqrt{t}} columns are the two formulas being compared, not measurements.

ttE[eWt]\E[e^{W_t}] ± s.e.et/2e^{t/2}candidate ete^{\sqrt{t}}E[eWt−12t]\E[e^{W_t - \frac12 t}] ± s.e.E[eWt−t]\E[e^{W_t - t}]E[eWt−t]\E[e^{W_t - \sqrt{t}}]
0.101.0506 ± 0.00241.05131.37190.9993 ± 0.00230.95060.7658
0.251.1378 ± 0.00431.13311.64871.0041 ± 0.00380.88610.6901
0.501.2998 ± 0.00741.28402.02811.0123 ± 0.00580.78840.6409
0.751.4684 ± 0.01091.45502.37741.0092 ± 0.00750.69360.6177
1.001.6546 ± 0.01491.64872.71831.0036 ± 0.00900.60870.6087
Growth split, Itô and dt parts

The same 20,000 paths, from expectations.running[*].growth_split.{exp_W_minus_1,ito_sum_S,dt_sum_D,residual}. The last column is the limit et/2−1e^{t/2} - 1, not a measurement.

ttE[eWt−1]\E[e^{W_t} - 1]=E[S(t)]= \E[S(t)]+E[D(t)]+ \E[D(t)]+ residualet/2−1e^{t/2} - 1
0.100.0506 ± 0.0024−0.0007 ± 0.00240.0513 ± 0.0001−0.000020.0513
0.250.1378 ± 0.0043+0.0042 ± 0.00400.1335 ± 0.0003−0.000010.1331
0.500.2998 ± 0.0074+0.0144 ± 0.00670.2855 ± 0.0009−0.000040.2840
0.750.4684 ± 0.0109+0.0099 ± 0.00940.4586 ± 0.0019−0.000060.4550
1.000.6546 ± 0.0149+0.0014 ± 0.01220.6533 ± 0.0033−0.000050.6487

The tables show that happening numerically. The measured E[eWt]\E[e^{W_t}] at t=0.1,0.25,0.5,0.75,1t = 0.1, 0.25, 0.5, 0.75, 1 tracks et/2e^{t/2} to within a couple of standard errors and is nowhere near the candidate ete^{\sqrt{t}}: at t=1t = 1 the measured 1.65 stands against the candidate's 2.72. The growth-split table writes eWt−1e^{W_t} - 1 as an Itô sum S(t)S(t) plus a dt\dt sum D(t)D(t) plus a tiny residual: E[S(t)]\E[S(t)] is 0 within noise at every time, while E[D(t)]\E[D(t)] matches et/2−1e^{t/2} - 1 to within two standard errors at every time (0.653 ± 0.003 against 0.649 at t=1t = 1). All of the growth is in the dt\dt term; none of it is in the dW\dW term. Here S(t)S(t) is that Itô sum, not a stock price; the price StS_t comes later on this page.

To test the drift, we introduce the process

eWt−cte^{W_t - ct}

and ask which constant cc makes its mean flat. With c=1c = 1, E[eWt−t]\E[e^{W_t - t}] falls like e−t/2e^{-t/2}; with c=1/tc = 1/\sqrt{t} — which is not a constant — the exponent's correction is t\sqrt{t}, and the mean drops from 1 to about 0.61 by t=1t = 1. Only c=12c = \tfrac12 gives a constant mean: the table has E[eWt−12t]≈1\E[e^{W_t - \frac12 t}] \approx 1 at every time, and the martingale test shows the same thing pathwise — the second half of the path does not know the first, with correlation −0.017 and conditional means −0.014 ± 0.010 and −0.006 ± 0.010. That is this page's check passed: eWt−12te^{W_t - \frac12 t} is a martingale, and the entire drift in E[eWt]\E[e^{W_t}] is the lemma's 12g′′ dt\tfrac12 g''\,\dt term.

Why ½ and not √t

For a single coin step of length tt, WtW_t is either +t+\sqrt{t} or −t-\sqrt{t} with equal chance. The instinct is: "WtW_t is about t\sqrt{t} in size, so maybe the mean of eWte^{W_t} is about ete^{\sqrt{t}}." The right way to see why that fails is to expand both cases to second order and look at what cancels:

e+t−1≈+t+12t,e−t−1≈−t+12t.\begin{aligned} e^{+\sqrt{t}} - 1 &\approx +\sqrt{t} + \tfrac12 t, \\ e^{-\sqrt{t}} - 1 &\approx -\sqrt{t} + \tfrac12 t. \end{aligned}

The ±t\pm\sqrt{t} pieces cancel between up and down — they are the "size" of WtW_t, and they average to 0. What adds is the 12t\tfrac12 t term: it is the same in both directions, and it survives in the mean — it is the lemma's 12g′′⋅(ΔW)2\tfrac12 g'' \cdot (\Delta W)^2 with g′′(0)=1g''(0) = 1 and (ΔW)2=t(\Delta W)^2 = t. So the spread of WtW_t (its typical size t\sqrt{t}) controls the fluctuations, but its variance (tt) controls the drift in E[eWt]\E[e^{W_t}]. That is exactly the lognormal fact in another language: the mean of eWte^{W_t} is driven by Var⁡(Wt)=t\Var(W_t) = t, not by the typical magnitude t\sqrt{t}, and that is why the lemma gives et/2e^{t/2}, not ete^{\sqrt{t}}.

The general form, and log S

For a smooth function f(t,Xt)f(t, X_t) of time and a process XtX_t with

dXt=a dt+b dWt,\dd X_t = a \,\dt + b \,\dW_t,

the lemma is "the ordinary chain rule plus one term". Here aa is the drift (the average rate of change of XX) and bb is the volatility (the size of its random kicks). Writing ∂f/∂t\partial f/\partial t, ∂f/∂x\partial f/\partial x, ∂2f/∂x2\partial^2 f/\partial x^2 as "the derivative with respect to the time slot, the XX slot, and the XX slot twice", the three-line route is: start from df=∂f/∂t dt+∂f/∂x⋅dX+12∂2f/∂x2⋅(dX)2\dd f = \partial f/\partial t \,\dt + \partial f/\partial x \cdot \dd X + \tfrac12 \partial^2 f/\partial x^2 \cdot (\dd X)^2 — Taylor with the clock and XX in the slot, the dt2\dt^2 and dt⋅dX\dt \cdot \dd X pieces already dropped — use the two facts about dX\dd X — dX=a dt+b dW\dd X = a \,\dt + b \,\dW and (dX)2=b2 dt(\dd X)^2 = b^2 \,\dt — and collect terms. The result is

df=(∂f∂t+a⋅∂f∂x+12b2⋅∂2f∂x2)dt+b⋅∂f∂x dWt.\begin{aligned} \dd f = {} & \left( \frac{\partial f}{\partial t} + a \cdot \frac{\partial f}{\partial x} + \tfrac12 b^2 \cdot \frac{\partial^2 f}{\partial x^2} \right) \dt \\ & {} + b \cdot \frac{\partial f}{\partial x} \,\dW_t. \end{aligned}

The ordinary chain rule has the ∂f/∂t\partial f/\partial t, a⋅∂f/∂xa \cdot \partial f/\partial x and b⋅∂f/∂xb \cdot \partial f/\partial x pieces — everything except the 12b2⋅∂2f/∂x2\tfrac12 b^2 \cdot \partial^2 f/\partial x^2 term; it appears because the path has quadratic variation and bb enters it squared through (dX)2=b2 dt(\dd X)^2 = b^2 \,\dt.

For log⁡St\log S_t the same pattern runs with

dSt=μSt dt+σSt dWt,\dS_t = \mu S_t \,\dt + \sigma S_t \,\dW_t,

so here a=μSa = \mu S and b=σSb = \sigma S. Put f(S)=log⁡Sf(S) = \log S into the general form: ∂f/∂t=0\partial f/\partial t = 0, ∂f/∂x=1/S\partial f/\partial x = 1/S, ∂2f/∂x2=−1/S2\partial^2 f/\partial x^2 = -1/S^2. The dt\dt term becomes

∂f∂t+a⋅∂f∂x+12b2⋅∂2f∂x2=0+μS⋅(1/S)+12(σS)2⋅(−1/S2)=μ−12σ2,\begin{aligned} & \frac{\partial f}{\partial t} + a \cdot \frac{\partial f}{\partial x} + \tfrac12 b^2 \cdot \frac{\partial^2 f}{\partial x^2} \\ & \quad = 0 + \mu S \cdot (1/S) + \tfrac12(\sigma S)^2 \cdot (-1/S^2) \\ & \quad = \mu - \tfrac12\sigma^2, \end{aligned}

and the dW\dW term is σS⋅(1/S)=σ\sigma S \cdot (1/S) = \sigma. So

d(log⁡St)=(μ−12σ2) dt+σ dWt.\dd(\log S_t) = (\mu - \tfrac12\sigma^2) \,\dt + \sigma \,\dW_t.

The −12σ2-\tfrac12\sigma^2 is exactly the curvature term from the lemma — the correction the ordinary chain rule would miss — and 2.6 takes the lognormal story from here.

The smooth curve — no dt term

Smooth control, x = t²

The smooth control, x(t)=t2x(t) = t^2 on the same grids — one curve, not an ensemble, so there is nothing to average — shown at three of the five grids for exe^x and two for x2x^2 and x3x^3. From smooth.{exp,x^2,x^3}[*].{lhs,chain_sum,chain_residual,chain_residual_times_n,dt_term,residual_with_dt_term}; the constants the last three columns tend to are in smooth_constants.

nnggleft side g(1)−g(0)g(1) - g(0)∑g′(xi)Δxi\sum g'(x_i)\Delta x_ichain-rule residualn×n \times residual12∑g′′(xi)⋅dt\tfrac12\sum g''(x_i) \cdot \dtresidual with the dt\dt term added
100exe^x1.71831.70580.012521.25230.7271−0.7145
10,000exe^x1.71831.71820.0001261.25560.7313−0.7312
1,000,000exe^x1.71831.71831.26×10−61.26 \times 10^{-6}1.25560.7313−0.7313
100x2x^210.986670.013331.33331−0.9867
1,000,000x2x^210.9999991.33×10−61.33 \times 10^{-6}1.33331−1.0000
100x3x^310.976100.023902.38990.9851−0.9612
1,000,000x3x^310.9999982.40×10−62.40 \times 10^{-6}2.40001−1.0000

For the smooth curve x(t)=t2x(t) = t^2 on the same grids, the ordinary chain rule works perfectly and its residual behaves like 1/n1/n. For each smooth gg, the chain-rule residual

g(xT)−g(0)−∑g′(xi)⋅Δxig(x_T) - g(0) - \sum g'(x_i) \cdot \Delta x_i

shrinks down the grids: it is of order 1/n1/n, and nn times that residual tends to a constant — about 1.2556 for g(x)=exg(x) = e^x, 4/3 for g(x)=x2g(x) = x^2, 12/5 for g(x)=x3g(x) = x^3. The reason is that each (Δx)2(\Delta x)^2 is dt2\dt^2-sized: here Δxi≈2ti⋅dt\Delta x_i \approx 2t_i \cdot \dt, so (Δxi)2∼dt2(\Delta x_i)^2 \sim \dt^2, and the sum of dt2\dt^2 terms vanishes as the grid is refined.

If you now add the lemma's dt\dt term anyway — 12∑g′′(xi)⋅dt\tfrac12\sum g''(x_i) \cdot \dt — you introduce a fixed error that does not go away with nn. In the limit it is

12∫01g′′(x(t)) dt,\tfrac12\int_0^1 g''(x(t)) \,\dt,

numerically about 0.7313 for g(x)=exg(x) = e^x, and exactly 1 for g(x)=x2g(x) = x^2 and g(x)=x3g(x) = x^3. That error converges to the constant instead of to zero — for exe^x it is 0.71 at n=100n = 100 and 0.73 at n=106n = 10^6. The one-line moral is: the lemma's extra term is the price of quadratic variation — it is essential on rough paths like WtW_t, but on a smooth curve like x(t)=t2x(t) = t^2 it is the wrong tool and adding it simply injects a non-vanishing bias.

The coin walk — both forms exactly

For the ±dt\pm\sqrt{\dt} coin walk the algebra collapses to its simplest form:

(ΔWi)2=dt(\Delta W_i)^2 = \dt

on every step, every grid, every path. That means the (ΔW)2(\Delta W)^2-form residual and the dt\dt-form residual are literally the same line — there is no randomness left in the square.

For g(x)=x2g(x) = x^2 both residuals are zero on every grid and every path. The worst deviation over five grids × five paths is 1.4×10−131.4 \times 10^{-13}. The lemma is exact because x2x^2 is quadratic and the walk's squared increments are deterministic.

For g(x)=x3g(x) = x^3 both residuals are

∑(ΔWi)3=dt⋅WT\sum (\Delta W_i)^3 = \dt \cdot W_T

exactly. At n=100n = 100 the five paths end at

WT=0.6, 0.4, −0.4, 1.0, −0.6,W_T = 0.6,\ 0.4,\ -0.4,\ 1.0,\ -0.6,

so the residuals are

0.006, 0.004, −0.004, 0.010, −0.006,0.006,\ 0.004,\ -0.004,\ 0.010,\ -0.006,

matching to 4×10−164 \times 10^{-16}. The fourth-order residual is 0.

The coin walk's behaviour is algebraically rigid — the squares are dt\dt and the cubes sum to dt⋅WT\dt \cdot W_T — and every cancellation is exact. Gaussian steps keep the same structure only in expectation: their squares have mean dt\dt but fluctuate, and their cubes have mean zero but wobble. The coin walk shows the template; the Gaussian walk shows the noisy version of it.

Into 2.6 — what the lemma does to the price model

2.6 takes the lemma straight into the price model dS=μS dt+σS dW\dS = \mu S \,\dt + \sigma S \,\dW, with log⁡S\log S as the way in: the lemma turns that SDE into d(log⁡S)=(μ−12σ2) dt+σ dW\dd(\log S) = (\mu - \tfrac12\sigma^2) \,\dt + \sigma \,\dW, and the −12σ2-\tfrac12\sigma^2 becomes the volatility drag — the gap between the mean and the median of a lognormal. From there the section writes down the closed-form sample

ST=S0⋅e(μ−12σ2)T+σWTS_T = S_0 \cdot e^{(\mu - \frac12\sigma^2)T + \sigma W_T}

and sets it against Euler steps to measure how far apart they are and how that gap shrinks with dt\dt. The dt\sqrt{\dt} in the steps is just the Brownian increment, the same ±dt\pm\sqrt{\dt} size that drove every calculation in 2.4 and 2.5.

Check

g is smooth and W is Brownian motion. Itô’s lemma says dg(W_t) equals:

E[e^(W_t)], the average of e^(W_t) over paths, is:

Which of these has constant mean — is a martingale — for Brownian W?

d(W_t³) equals:

x(t) is a smooth curve (say t²) and g is smooth. Which rule does g(x(t)) obey?