Smooth Infinitesimal Analysis (a Contrast)

The closing page of the NSA spine, included for contrast. Robinson's hyperreals (ultrapower.md) are not the only rigorous home for infinitesimals. Smooth infinitesimal analysis (SIA) — the elementary face of synthetic differential geometry (SDG) — takes infinitesimals to be nilpotent ( $ε^{2} = 0$ ) rather than invertible, and pays for it by working in intuitionistic logic (intuitionistic-logic.md). It is the framework closest to the physicist's "work to first order in $d x$ , drop $(d x)^{2}$ " (physics-use-of-calculus.md §4.2).

References: Bell, A Primer of Infinitesimal Analysis; Kock, Synthetic Differential Geometry; Moerdijk & Reyes, Models for Smooth Infinitesimal Analysis; Lawvere's program.

1. Nilpotent infinitesimals

SIA posits a set $Δ = {ε : ε^{2} = 0}$ of infinitesimals, the "first-order neighbourhood of $0$ ." These are nilpotent, not invertible: $ε^{2} = 0$ exactly, so one never divides by $ε$ . This is the opposite design choice from the hyperreals, where a nonzero infinitesimal $ε$ has $ε^{2} \neq = 0$ and a genuine (infinite) reciprocal $1/ ε$ (hyperreals.md §1).

Crucially $Δ$ is not ${0}$ — but neither can any of its elements be proved $\neq = 0$ . That tension is exactly what forces the logic to change (§3).

2. The Kock–Lawvere axiom

The single axiom that runs the whole theory:

Kock–Lawvere (KL) axiom. For every function $f : R \to R$ (on the SIA "smooth line" $R$ ) and every $x$ , there is a unique $b \in R$ such that $f (x + ε) = f (x) + b ε for all ε \in Δ.$

The unique $b$ is defined to be the derivative $f^{'} (x)$ . So in SIA the equation

$f (x + ε) = f (x) + f^{'} (x) ε (ε^{2} = 0)$

is exact, not approximate — there is no remainder term to bound, because $ε^{2}$ and all higher powers are literally zero. The physicist's " $df = f^{'} d x$ , neglect $(d x)^{2}$ " is promoted from a heuristic to the founding axiom, with "neglect" meaning "equals zero." Differentiation becomes pure algebra: for $f (x) = x^{2}$ ,

$(x + ε)^{2} = x^{2} + 2 x ε + ε^{2} = x^{2} + 2 x ε,$

so $f^{'} (x) = 2 x$ is read off the coefficient — no limits, no standard part, no quotient.

3. Why the logic must be intuitionistic

The KL axiom is inconsistent with classical logic. If excluded middle held, then for each $ε \in Δ$ either $ε = 0$ or $ε \neq = 0$ ; a nonzero $ε$ with $ε^{2} = 0$ would be a zero divisor, contradicting that $R$ is (KL forces) cancellative in the relevant sense — one can derive that every $ε \in Δ$ is $0$ , collapsing $Δ = {0}$ and killing the axiom's uniqueness clause. The escape is to drop the law of excluded middle (intuitionistic-logic.md):

In SIA one may not assert " $ε = 0$ or $ε \neq = 0$ " for $ε \in Δ$ . The infinitesimals are "not not zero" ( $\neg\neg (> ε = 0)$ ) yet not provably zero. Removing excluded middle removes the case-split that would collapse them.

This is the same phenomenon catalogued in choice-and-excluded-middle.md: constructive logic tolerates distinctions ("nilpotent but not decidably zero") that classical logic flattens. SIA lives in a smooth topos, whose internal logic is intuitionistic; the consistency of the KL axiom is established by building such topos models (Moerdijk–Reyes), so SIA is as rigorous as NSA — it simply relocates the cost from non-constructive choice to non-classical logic.

4. Robinson vs. SIA: a comparison

The two rigorous infinitesimal frameworks are near mirror images:

	Robinson / NSA ( $^{*} R$ )	SIA / SDG
Infinitesimal $ε \neq = 0$	invertible; $ε^{2} \neq = 0$ ; $1/ ε$ infinite	nilpotent; $ε^{2} = 0$ ; no reciprocal
Logic	classical (first-order-logic.md)	intuitionistic (intuitionistic-logic.md)
Foundational cost	non-constructive (BPI/ultrafilter, ultrapower.md §2)	no excluded middle
Derivative	$st (\frac{Δ y}{Δ x})$ — quotient + shadow	coefficient $b$ in $f (x + ε) = f (x) + b ε$ — algebra
Bridge to $R$	transfer principle (transfer-principle.md)	models in a smooth topos
Built for	analysis, measure theory, stochastics (applications.md)	differential geometry, physics field theory
"Drop $(d x)^{2}$ " means	apply $st$ (round away infinitesimals)	it is identically zero

Neither subsumes the other. NSA's invertible infinitesimals are what let it do hyperfinite sums and hence integration and measure theory (nonstandard-integration.md); SIA's nilpotents are what make differential geometry (tangent vectors as maps $Δ \to M$ , forms, jets) effortless and coordinate-free, which is why SDG is attractive for classical field theory and general relativity.

5. Which to reach for

Want to integrate, take limits, do probability, or justify $ε$ – $N$ arguments cheaply? Use NSA — invertible infinitesimals and transfer (applications.md).
Want synthetic differential geometry — vector fields, forms, and connections with " $f (x + ε) = f (x) + f^{'} (x) ε$ " as an identity, and you can live without excluded middle? Use SIA/SDG.
Want to explain physics practice? Both help: SIA legitimizes the first-order-in- $d x$ manipulations of thermodynamics and continuum mechanics; NSA legitimizes the sum-of-slices integrals and infinitesimal quotients (physics-use-of-calculus.md §4).

Where this page is used

Provides the deliberate contrast to the Robinson construction that occupies the rest of this spine.
Closes the loop with physics-use-of-calculus.md §4.2 and the constructive-logic themes of intuitionistic-logic.md and choice-and-excluded-middle.md.

End of the spine. Return to the NSA index or the Mathematics overview.

Keyboard shortcuts