The 2 in Ito’s lemma is not a dimension – a distinct-root census of the tools of stochastic calculus.
Ito’s lemma carries a second-order term, and the corpus is full of twos. This paper writes where this one comes from. NO NEW MATHEMATICAL THEOREM AND NO NEW PHYSICAL LAW IS CLAIMED. Ito (1944), Stratonovich (1966), Wong-Zakai (1965), Lyons (1998) and Mandelbrot-Van Ness (1968) are all standard. This is also a census of the author’s own tools: whether a formula used daily in market analysis shares a root with the corpus’s two.
First, the expansion stops at second order because only the quadratic variation survives. Refining the partition on a Brownian path over the unit interval, three variations were counted. The first variation runs from 23.84 at 1024 steps to 408.00 at 262144 steps, diverging. The second settles at the length of the interval, taking the values 0.895, 1.006, 0.987, 1.003 and 0.997. The third falls from 4.23 times ten to the minus two down to 3.10 times ten to the minus three, going to zero. So the Taylor expansion is not being truncated at second order; the third order and above vanish. Ito’s lemma is an exact identity rather than an approximation, and the one half is a value that emerged in a limit, not a coefficient anyone put there.
Second, and this is the core. That two is not a spatial dimension but the roughness of the path. Computing the quadratic variation for fractional Brownian motion with Hurst exponent H, at H equal to 0.3 the values for 256, 512 and 1024 steps are 9.31, 11.63 and 16.21, diverging; at H equal to 0.7 they are 0.107, 0.083 and 0.060, vanishing; and only at H equal to 0.5 are they stable at 1.05, 0.91 and 1.01. Below one half, second order is not enough; above it, second order is too much. In general the p-variation is finite when p exceeds one over H, so the order to which the expansion must be carried is fixed by one over H. For Brownian motion H is one half, hence two (for paths rougher than that, third-order information is required, as in Lyons’s rough path theory). And Ito’s lemma in d dimensions stops at second order whatever d may be. The two is not a function of the dimension.
Third, the separation from the corpus’s twos can be written. Ito’s two moves with the Hurst exponent and not with the spatial dimension. The n equals two of Paper 01 moves with the spatial dimension (n equals d minus one) and has nothing to do with the Hurst exponent. The two of the Hopf map moves with neither, being the structure of a fibration. The separator is what has to be moved for the number to move, and all three are distinct roots.
Fourth, the extra term itself is a convention. For the same path and the same integrand, where the integrand is read on each subinterval changes the answer. Computing the integral of Brownian motion against itself with two to the twentieth subintervals gives minus 0.2344 at the left endpoint and plus 0.2653 at the midpoint, a difference of exactly half the interval length, 0.5. The closed forms agree: reading at the left gives half the terminal value squared minus half the interval length, and reading at the midpoint gives half the terminal value squared. The Stratonovich convention obeys the ordinary chain rule and carries no extra term. So reading Ito’s lemma as containing an extra term that nature added is not correct: that term is the consequence of choosing the left endpoint.
Fifth, one thing is nevertheless not a convention: that the two conventions disagree at all. That is exactly the statement that the quadratic variation is nonzero. On a smooth path, reading anywhere gives the same value, and the size of the disagreement measures the roughness of the path. This has the shape of Paper 131: how a number is written is convention, and what does not disappear when the writing changes is fact. Here the extra term is the convention and the disagreement is the fact. And the choice between the conventions is not arbitrary. Finance needs Ito, because a position is set before the price moves. Physics needs Stratonovich, because white noise arises as a limit of smooth noise (Wong-Zakai). Being a convention is not the same as being a free choice: what is being modelled fixes the convention uniquely, which is an instance of Paper 95’s point that a convention is not an anything-goes.
Sixth, the phrase ‘average return’ names two different quantities. For geometric Brownian motion with drift 0.10, volatility 0.40 and horizon one, four hundred thousand paths give a mean price of 1.1034 (against a theoretical value of e to the 0.10, namely 1.1052), a mean log price of 0.0188 (against 0.10 minus half of 0.40 squared, namely 0.0200), and a median price of 1.0185 (against 1.0202). The gap is exactly the Ito correction, half the volatility squared, namely 0.08. When one says ‘expected return’ one sometimes means the growth rate of the expected price and sometimes the growth rate of the expected log price. For the same process both are correct, and they are not the same number. And where a typical single path is heading is the median, not the mean: the mean is pulled up by a few paths in the tail. When a distribution is skewed the mean does not describe the usual outcome, which is the same caution Paper 94 raised about fat tails, and the larger the volatility the further apart the two lie.
IMPORTANT (scope): a census of established standard material (Layer 2). Every work cited is standard. What is not claimed: any new mathematical theorem (there is none) / any new physical law (there is none) / that Ito’s lemma is in error (it is not) / that the choice of convention is a matter of taste (it is not; what is being modelled decides it). What is added: separating Ito’s two from any spatial dimension numerically, showing with fractional Brownian motion that the order is fixed by one over H, dividing the extra term (a convention) from the disagreement of the two conventions (a fact), and writing the two roots of ‘average return’. What remains: the same character two is counting three different things, and testing which of them moves when something is moved makes the three come apart. The tools one uses every day are the ones whose roots are most worth checking.
On the making of this work: The ideas and content of this work stem from the author’s own considerations. Assistance from an AI (a large language model) was used for structuring, English translation, and checking the algebra. Any remaining errors or misinterpretations are solely the author’s. Feedback and corrections are sincerely appreciated.
Reference
Yamagishi, Y. (2026). The $2$ in Itô’s Lemma Is Not a Dimension [Paper 132]. Zenodo.
DOI: 10.5281/zenodo.22080358
ORCID: 0009-0006-8587-3450