2.5
Itô’s lemma
Second-order Taylor plus the rulebook: df = (∂f/∂t + a·∂f/∂x + ½b²·∂²f/∂x²)dt + b·∂f/∂x dW. Worked: W², log S, e^(W−t/2).
When the price itself jitters like Brownian motion, anything built from it — an option, a hedge, even — 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 was just one instance of a general phenomenon. After this page, the reader can write the dynamics of any smooth payoff, not just .
What this page shows, and how
The claim is one line: for a smooth , — 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 term in it. For that identity is 2.3's , rearranged in 2.4 into the ; the first table shows it holding to .
Second, the limit. Two things turn that identity into the lemma: is replaced by , and everything of third order is dropped. Each costs something on a finite grid, and the second table measures both costs for and — the replacement costs a noise that shrinks like , the dropping costs a remainder that shrinks like . 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, . If the chain rule were the whole story, would be constant; the lemma's term says it grows at rate , so . The third and fourth tables check that, split the growth into its part (zero) and its part (all of it), and find the one correction — in the exponent — that makes the mean flat. That is the martingale , and it is where 2.6's will come from.
Fourth, the controls. A smooth curve gets no 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 is assembled from the same pieces, with 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 steps and then read on the five grids . On each grid, for three functions — , and — we checked both sides of Itô's lemma in its two forms: the exact -form and the -form, with the third-order term measured on its own. A second run drew 20,000 fresh paths at to measure at five times, splitting the growth into its Itô-sum part and its part, with the candidates , and beside it. The same grids were run on the smooth curve to show what happens when the quadratic variation vanishes. Finally, the coin walk from 2.1 was checked on and , 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 , so the left side is on every grid.
The first path of the ensemble — 400 paths drawn once on a grid of steps
over 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 ).
| sum | -form residual | -form residual | ||||
|---|---|---|---|---|---|---|
| 100 | 0.3338 | 1.0356 | 1.3694 | 1 | +0.0356 | |
| 1,000 | 0.3769 | 0.9925 | 1.3694 | 0 | 1 | −0.0075 |
| 10,000 | 0.3730 | 0.9964 | 1.3694 | 1 | −0.0036 | |
| 100,000 | 0.3766 | 0.9928 | 1.3694 | 1 | −0.0072 | |
| 1,000,000 | 0.3711 | 0.9983 | 1.3694 | 1 | −0.0017 |
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
the spread is compared against, not a measurement.
| mean of -form residual | ± s.e. | sd | ||
|---|---|---|---|---|
| 100 | +0.0102 | 0.0067 | 0.1344 | 0.1414 |
| 1,000 | −0.0025 | 0.0022 | 0.0439 | 0.0447 |
| 10,000 | −0.0004 | 0.0007 | 0.0142 | 0.0141 |
| 100,000 | −0.0003 | 0.0002 | 0.0043 | 0.0045 |
| 1,000,000 | +0.0000 | 0.0001 | 0.0014 | 0.0014 |
Down the table every column settles. The Itô sum and the quadratic-variation column both settle toward their final values, and their sum is on every grid — the left side never moves. The -form residual is a column of s: that is the identity holding to machine precision at every . The -form residual is exactly row by row, because the -form replaces by ; it is the grid's -vs- 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 (+0.010 ± 0.007 at , 0.0000 ± 0.0001 at ), and its spread matches to within the 400-path noise (0.134 against 0.141 at , 0.0014 against 0.0014 at ). So for the -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 (). The left side is for and for .
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.
| left side | -form residual | -form residual | ||||||
|---|---|---|---|---|---|---|---|---|
| 100 | −0.6897 | −0.9079 | 0.2238 | 0.2247 | −0.0057 | −0.0064 | −0.0065 | |
| 1,000 | −0.6897 | −0.9136 | 0.2245 | 0.2237 | −0.00059 | −0.00065 | +0.0003 | |
| 10,000 | −0.6897 | −0.9123 | 0.2226 | 0.2235 | −0.000070 | −0.000075 | −0.0009 | |
| 100,000 | −0.6897 | −0.9114 | 0.2217 | 0.2235 | −0.0018 | |||
| 1,000,000 | −0.6897 | −0.9128 | 0.2231 | 0.2235 | −0.0004 | |||
| 100 | −1.6025 | 1.3047 | −2.8488 | −2.6240 | −0.0584 | −0.0584 | −0.2832 | |
| 1,000 | −1.6025 | 0.9832 | −2.5798 | −2.6239 | −0.0060 | −0.0060 | +0.0382 | |
| 10,000 | −1.6025 | 1.0188 | −2.6206 | −2.6257 | −0.00067 | −0.00067 | +0.0044 | |
| 100,000 | −1.6025 | 1.0098 | −2.6123 | −2.6258 | −0.000019 | −0.000019 | +0.0135 | |
| 1,000,000 | −1.6025 | 1.0190 | −2.6215 | −2.6258 | +0.0043 |
Down both tables the pattern is immediate. The -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 : for it is 0.058 at and at ; for , 0.006 and . Beside it, the -form residual's spread over the 400 paths only shrinks threefold per row (for : 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 -vs- noise, which dies like and is larger at every — the ratio of the two spreads is for both functions, about 10 at (11 for , 7 for ) and about 1,000 at .
For the -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 -form is exact algebra, and the only remainder is , since . For the two columns differ slightly, because the function has fourth-order and higher terms; those extra pieces are about ten times smaller at every — the same 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 and is tiny. The error comes from replacing by — the grid's noise — which dies only like and dominates every row.
Why — Taylor summed, and the two forms
Start from one step. Taylor says how much a smooth changes when its argument moves from to :
Add the steps. The left sides telescope — every interior cancels — leaving . The right side becomes three sums. That exact algebra is the -form of the lemma: an equality on every grid, for every path, with no limit taken.
Two things turn it into the -form. First, replace each by , using 2.3's rule that . 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 -vs- noise plus the remainder, the noise dominating.
Written for one infinitesimal step, the result is the lemma in d-notation, drift first:
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 is of size , and 2.3 already measured like . That bound is enough to know the remainder goes away. But the term itself keeps its signs: it is , and is an odd moment with mean zero. The positive and negative cubes cancel. In this run the signed has standard deviation (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 smaller than the bound from . The coin walk shows the same split cleanly: exactly, which is , while exactly, which is . 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 — — and never the dominant error on a grid.
E[e^(W_t)] = e^(t/2) — the growth is the dt term
For the lemma gives
The term is an Itô integral, so its mean is zero; if that were all, would stay at 1. The term is half of itself, so on average the process grows at a constant rate , compounding in time — the statement .
The second run: 20,000 fresh paths at , 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 and columns are the two formulas
being compared, not measurements.
| ± s.e. | candidate | ± s.e. | ||||
|---|---|---|---|---|---|---|
| 0.10 | 1.0506 ± 0.0024 | 1.0513 | 1.3719 | 0.9993 ± 0.0023 | 0.9506 | 0.7658 |
| 0.25 | 1.1378 ± 0.0043 | 1.1331 | 1.6487 | 1.0041 ± 0.0038 | 0.8861 | 0.6901 |
| 0.50 | 1.2998 ± 0.0074 | 1.2840 | 2.0281 | 1.0123 ± 0.0058 | 0.7884 | 0.6409 |
| 0.75 | 1.4684 ± 0.0109 | 1.4550 | 2.3774 | 1.0092 ± 0.0075 | 0.6936 | 0.6177 |
| 1.00 | 1.6546 ± 0.0149 | 1.6487 | 2.7183 | 1.0036 ± 0.0090 | 0.6087 | 0.6087 |
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 , not a measurement.
| + residual | |||||
|---|---|---|---|---|---|
| 0.10 | 0.0506 ± 0.0024 | −0.0007 ± 0.0024 | 0.0513 ± 0.0001 | −0.00002 | 0.0513 |
| 0.25 | 0.1378 ± 0.0043 | +0.0042 ± 0.0040 | 0.1335 ± 0.0003 | −0.00001 | 0.1331 |
| 0.50 | 0.2998 ± 0.0074 | +0.0144 ± 0.0067 | 0.2855 ± 0.0009 | −0.00004 | 0.2840 |
| 0.75 | 0.4684 ± 0.0109 | +0.0099 ± 0.0094 | 0.4586 ± 0.0019 | −0.00006 | 0.4550 |
| 1.00 | 0.6546 ± 0.0149 | +0.0014 ± 0.0122 | 0.6533 ± 0.0033 | −0.00005 | 0.6487 |
The tables show that happening numerically. The measured at tracks to within a couple of standard errors and is nowhere near the candidate : at the measured 1.65 stands against the candidate's 2.72. The growth-split table writes as an Itô sum plus a sum plus a tiny residual: is 0 within noise at every time, while matches to within two standard errors at every time (0.653 ± 0.003 against 0.649 at ). All of the growth is in the term; none of it is in the term. Here is that Itô sum, not a stock price; the price comes later on this page.
To test the drift, we introduce the process
and ask which constant makes its mean flat. With , falls like ; with — which is not a constant — the exponent's correction is , and the mean drops from 1 to about 0.61 by . Only gives a constant mean: the table has 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: is a martingale, and the entire drift in is the lemma's term.
Why ½ and not √t
For a single coin step of length , is either or with equal chance. The instinct is: " is about in size, so maybe the mean of is about ." The right way to see why that fails is to expand both cases to second order and look at what cancels:
The pieces cancel between up and down — they are the "size" of , and they average to 0. What adds is the term: it is the same in both directions, and it survives in the mean — it is the lemma's with and . So the spread of (its typical size ) controls the fluctuations, but its variance () controls the drift in . That is exactly the lognormal fact in another language: the mean of is driven by , not by the typical magnitude , and that is why the lemma gives , not .
The general form, and log S
For a smooth function of time and a process with
the lemma is "the ordinary chain rule plus one term". Here is the drift (the average rate of change of ) and is the volatility (the size of its random kicks). Writing , , as "the derivative with respect to the time slot, the slot, and the slot twice", the three-line route is: start from — Taylor with the clock and in the slot, the and pieces already dropped — use the two facts about — and — and collect terms. The result is
The ordinary chain rule has the , and pieces — everything except the term; it appears because the path has quadratic variation and enters it squared through .
For the same pattern runs with
so here and . Put into the general form: , , . The term becomes
and the term is . So
The 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
The smooth control, on the same grids — one curve, not an
ensemble, so there is nothing to average — shown at three of the five grids
for and two for and . 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.
| left side | chain-rule residual | residual | residual with the term added | ||||
|---|---|---|---|---|---|---|---|
| 100 | 1.7183 | 1.7058 | 0.01252 | 1.2523 | 0.7271 | −0.7145 | |
| 10,000 | 1.7183 | 1.7182 | 0.000126 | 1.2556 | 0.7313 | −0.7312 | |
| 1,000,000 | 1.7183 | 1.7183 | 1.2556 | 0.7313 | −0.7313 | ||
| 100 | 1 | 0.98667 | 0.01333 | 1.3333 | 1 | −0.9867 | |
| 1,000,000 | 1 | 0.999999 | 1.3333 | 1 | −1.0000 | ||
| 100 | 1 | 0.97610 | 0.02390 | 2.3899 | 0.9851 | −0.9612 | |
| 1,000,000 | 1 | 0.999998 | 2.4000 | 1 | −1.0000 |
For the smooth curve on the same grids, the ordinary chain rule works perfectly and its residual behaves like . For each smooth , the chain-rule residual
shrinks down the grids: it is of order , and times that residual tends to a constant — about 1.2556 for , 4/3 for , 12/5 for . The reason is that each is -sized: here , so , and the sum of terms vanishes as the grid is refined.
If you now add the lemma's term anyway — — you introduce a fixed error that does not go away with . In the limit it is
numerically about 0.7313 for , and exactly 1 for and . That error converges to the constant instead of to zero — for it is 0.71 at and 0.73 at . The one-line moral is: the lemma's extra term is the price of quadratic variation — it is essential on rough paths like , but on a smooth curve like it is the wrong tool and adding it simply injects a non-vanishing bias.
The coin walk — both forms exactly
For the coin walk the algebra collapses to its simplest form:
on every step, every grid, every path. That means the -form residual and the -form residual are literally the same line — there is no randomness left in the square.
For both residuals are zero on every grid and every path. The worst deviation over five grids × five paths is . The lemma is exact because is quadratic and the walk's squared increments are deterministic.
For both residuals are
exactly. At the five paths end at
so the residuals are
matching to . The fourth-order residual is 0.
The coin walk's behaviour is algebraically rigid — the squares are and the cubes sum to — and every cancellation is exact. Gaussian steps keep the same structure only in expectation: their squares have mean 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 , with as the way in: the lemma turns that SDE into , and the 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
and sets it against Euler steps to measure how far apart they are and how that gap shrinks with . The in the steps is just the Brownian increment, the same 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?