Why the Physicist's Calculus Is Legitimate

Open any physics textbook and the calculus looks nothing like the analysis spine. Nobody defines $\int_{a}^{b} f d x$ as the common value of $sup_{P} L (f, P)$ and $in f_{P} U (f, P)$ (riemann-integral.md). Instead one reads:

"Chop the rod into infinitesimal pieces of length $d x$ . Each piece has mass $d m = λ (x) d x$ . Add them up: $M = \int λ (x) d x$ ."

Here $d x$ is treated as a small length — a genuine, if tiny, quantity you can multiply, cancel, and sum. Derivatives are written $\frac{d y}{d x}$ and manipulated as fractions: physicists cheerfully write $d y = f^{'} (x) d x$ , cancel $d t$ against $d t$ , and separate $\frac{d y}{d x} = g (x) h (y)$ into $\frac{d y}{h ( y )} = g (x) d x$ before integrating both sides. To an analyst this is nonsense on its face: $d x$ is not a number, $\frac{d y}{d x}$ is not a quotient, and "an infinitely small length" was banished from rigorous mathematics by Cauchy and Weierstrass in the 19th century.

And yet it always gives the right answer (for the problems physicists pose). This page explains why — what the informal manipulations really mean, which rigorous theorem each one abbreviates, the several independent frameworks that vindicate infinitesimals outright, and the precise circumstances under which the shortcut can bite.

References: Spivak, Calculus, ch. 13 & epilogue; Robinson, Non-standard Analysis; Bell, A Primer of Infinitesimal Analysis; Tao, Analysis I (on the " $d x$ " heuristic); Courant & John, Introduction to Calculus and Analysis.

1. The apparent scandal

Two habits of physical calculus look, on the surface, incompatible with the $ε$ – $δ$ tradition:

Physicist writes	Analyst objects
$\int f d x =$ "sum of slices $f d x$ "	An integral is a number — a supremum of lower sums, not a sum of infinitesimals (riemann-integral.md §1)
$d x$ is a small length; $f d x$ is a small area	$d x$ names no real number; there is no smallest positive real (real-numbers.md, Archimedean property)
$\frac{d y}{d x}$ is a fraction; $d y = f^{'} (x) d x$	$\frac{d y}{d x}$ is a limit $lim_{h \to 0} \frac{f ( x + h ) - f ( x )}{h}$ , a single indivisible symbol (differentiation.md)
Cancel differentials, " $d u = g^{'} (x) d x$ "	Cancellation presupposes an algebra of $d x$ 's that has not been defined
$d V = d x d y d z$ ; $d A = \overset{n}{^} d A$	Volume and surface elements need Jordan/Lebesgue measure and parametrizations (multiple-integrals.md)

The resolution is not that physicists are sloppy and get lucky. It is that each informal move is a faithful shorthand for a rigorous limit — and, more strongly, that infinitesimals can be given literal, consistent meaning in more than one way. The physicist is computing in a compressed notation whose decompression is a theorem.

2. The core reconciliation: $d x$ is the ghost of $Δ x$

The single most important fact is that Leibniz notation was engineered to make a limit look like an algebraic identity. Start from the honest Riemann sum:

$i = 1 \sum n f (x_{i}) Δ x_{i} mesh \to 0 \int_{a}^{b} f (x) d x .$

As the partition is refined, $Δ x_{i} \to 0$ and the sum tends to the Darboux integral (riemann-integral.md §2.1 shows the tagged Riemann sum agrees with the $sup / in f$ definition). The symbol " $d x$ " is literally the typographic residue of $Δ x_{i}$ after the limit, and the integral sign $\int$ is a stretched " $S$ " for summa. So when a physicist says "add up the slices $f d x$ ," they are naming the Riemann sum and implicitly taking its limit. Nothing is being asserted about an actual infinitesimal; the limit is hidden in the notation.

The heuristic is a mnemonic for a theorem. " $M = \int λ d x$ " unpacks to: partition the rod, approximate each piece's mass by $λ (x_{i}) Δ > x_{i}$ , sum, and take the mesh to zero; the limit exists because $λ$ is continuous (hence integrable, riemann-integral.md §3). Every "sum of infinitesimals" is a limit of finite sums in disguise.

This already dissolves most of the scandal: the integral half of physical calculus is exactly Riemann integration with the limit left tacit.

3. Differentials done rigorously: linear maps and 1-forms

The differential half — treating $d y = f^{'} (x) d x$ as an equation between real things — also has an exact modern meaning, and it is not "a small number." There are two rigorous readings, and physicists unknowingly use both.

3.1 $df$ as the linear part of the change

For a differentiable $f$ , the differential at $x$ is the linear map

$d f_{x} : h ⟼ f^{'} (x) h .$

It is the best linear approximation to the increment: $f (x + h) - f (x) = f^{'} (x) h + o (h)$ (differentiation.md; the multivariable Fréchet derivative in multivariable.md is the same idea). Read this way, " $d y = f^{'} (x) d x$ " is the identity of linear maps $df = f^{'} (x) d x$ , where $d x$ is the identity linear map $h \mapsto h$ . It is exactly true, not approximately — the approximation lives in the $o (h)$ remainder that the differential deliberately discards. The physicist's algebra of differentials is the algebra of these linear maps, which really do multiply by scalars and add.

3.2 $d x$ as a differential form

The deeper reading, and the one that makes $d x d y d z$ and $F \cdot d r$ rigorous, is that $d x$ is a differential 1-form (differential-forms.md). Then:

$df = \frac{\partial f}{\partial x} d x + \frac{\partial f}{\partial y} d y + \dots$ is an exact equation between 1-forms;
the "cancellations" in a change of variables are the pullback of forms;
$d x \land d y \land d z$ is a genuine volume form, so " $d V = d x d y d z$ " is literally correct once $\land$ (which encodes orientation and antisymmetry) is understood;
the flux and circulation elements $d A$ , $d r$ that physicists integrate are surface and line 1-/2-forms (vector-calculus.md).

In this framework the entire apparatus $\int_{\partial M} ω = \int_{M} d ω$ (stokes-general.md) is the machinery physicists manipulate informally as Green's, Stokes', and the divergence theorems (integral-theorems.md). The " $d$ " they push around is the exterior derivative, and $d^{2} = 0$ is why "the boundary of a boundary is empty" — a fact used constantly in gauge theory and electromagnetism.

So the physicist's differentials are not a fiction to be tolerated: they are 1-forms, computed correctly, with the rigorous definitions merely left off-stage.

4. Infinitesimals taken literally: two consistent frameworks

Even the strong reading — " $d x$ is an actual infinitely small quantity you may multiply and cancel" — can be made rigorous and consistent. Twentieth-century logic supplied two independent homes for infinitesimals, both proving that the naive manipulations are theorems, not errors.

4.1 Non-standard analysis (Robinson, 1960s)

Using the compactness / ultrapower machinery of model theory, one builds the hyperreal field $^{*} R$ , a proper ordered-field extension of $R$ containing genuine infinitesimals $ϵ > 0$ smaller than every positive real, and their reciprocals, the infinite numbers. In this field:

$\frac{d y}{d x}$ is a quotient of hyperreals, and the derivative is the standard part $st (\frac{f ( x + ϵ ) - f ( x )}{ϵ})$ ;
$\int f d x$ is an actual (hyperfinite) sum of terms $f (x) d x$ with $d x$ infinitesimal;
"cancel the $d x$ 's" is ordinary field algebra.

The transfer principle — every first-order statement true of $R$ is true of $^{*} R$ and vice versa — guarantees that any result obtained by these infinitesimal computations, once you take standard parts, is a true theorem about the reals. This is precisely the physicist's mode of reasoning, made airtight. Keisler's textbook Elementary Calculus teaches the whole first-year sequence this way. The full development — ultrapower construction, transfer, standard part, and the infinitesimal rebuild of calculus — is the Non-Standard Analysis spine.

Robinson's vindication. The 300-year-old Leibniz/Euler calculus of infinitesimals — the very style physicists never abandoned — was shown in 1966 to be logically consistent and to prove exactly the classical theorems. Physicists kept using it through the whole "crisis of rigor" because it works; non-standard analysis explains why it had to.

4.2 Smooth infinitesimal analysis (synthetic differential geometry)

A second, quite different framework uses nilpotent infinitesimals: quantities $ϵ \neq = 0$ with $ϵ^{2} = 0$ . In a suitable (topos-theoretic, intuitionistic — see intuitionistic logic) setting, every function is infinitesimally linear:

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

which is the physicist's " $df = f^{'} d x$ , drop higher-order terms" promoted to an axiom. Here "neglect $(d x)^{2}$ " is not an approximation — it is exact because $ϵ^{2}$ is literally zero. This is the framework closest to how one actually manipulates differentials in thermodynamics and continuum mechanics ("to first order in $d x$ "). See smooth-infinitesimal.md for the contrast with Robinson's invertible infinitesimals.

The moral of §4: the informal calculus is not merely reconcilable with rigor via limits (§2–3); infinitesimals themselves are respectable objects in at least two consistent theories.

5. A dictionary of physical moves and their theorems

Every standard manipulation corresponds to a named theorem in the analysis spine. The physicist skips the citation; the theorem still does the work.

Physical move	Rigorous justification
"Sum the slices" $\int f d x$	Limit of Riemann sums = Darboux integral (riemann-integral.md)
$u$ -substitution: "let $u = g (x), d u = g^{'} (x) d x$ "	Change-of-variables theorem / chain rule + FTC (riemann-integral.md)
Separation of variables: $\frac{d y}{h ( y )} = g (x) d x$	Chain rule + FTC applied to $\int \frac{d y}{h ( y )} = \int g (x) d x$ (differentiation.md)
Iterated integral $\iint f d x d y$ , "do $x$ then $y$ "	Fubini's theorem (multiple-integrals.md §3)
$d V = d x d y d z$ ; Jacobian factor $∣ \partial (x, y) / \partial (u, v) ∣$	Multivariable change of variables (multiple-integrals.md §4)
Flux $\oint F \cdot d A$ , circulation $\oint F \cdot d r$	Surface/line integrals of forms (vector-calculus.md)
Divergence / Stokes / Green "conservation" laws	The classical integral theorems = special cases of generalized Stokes
Differentiate under the integral, " $\frac{d}{d t} \int = \int \frac{\partial}{\partial t}$ "	Leibniz integral rule (needs dominated convergence / uniform bounds)
$δ (x)$ : "a spike with $\int δ = 1$ "	Distributions (Schwartz): a continuous functional, not a function
$d p d q$ phase-space volume, " $ℏ$ -cells"	Measure on a symplectic manifold; Liouville's theorem

The pattern is uniform: the notation encodes the hypothesis-free skeleton, and a theorem supplies the hypotheses.

6. Why the shortcut is safe in physics: the hypotheses hold silently

There is a second, subtler reason physical calculus rarely goes wrong: the pathologies that motivate the rigorous definitions do not occur for physical fields.

The whole apparatus of $sup / in f$ , uniform continuity, and measurability exists to handle monsters — the Dirichlet function ( $1_{Q}$ , bounded but not integrable, riemann-integral.md §1), the Weierstrass nowhere-differentiable function, sequences whose limit and integral cannot be swapped. These do not describe charge densities, velocity fields, or wavefunctions. Physical quantities are modeled by functions that are:

smooth (or piecewise-smooth with isolated, well-understood singularities),
compactly supported or rapidly decaying (fields die off at infinity),
bounded on bounded regions.

For such functions every theorem in the dictionary of §5 applies with its hypotheses automatically satisfied. Continuity gives integrability; smoothness gives Clairaut's equality of mixed partials (multivariable.md), so $\partial_{x} \partial_{y} = \partial_{y} \partial_{x}$ "just works"; compact support and uniform convergence license the interchange of limits, sums, and integrals. The physicist silently works inside the "nice" regime where rigor and heuristic provably coincide. Choosing a smooth model of a physical system is, implicitly, discharging all the hypotheses at once.

This is the deepest answer to "why is it accepted?": in the domain physics actually inhabits, the informal rules are provably valid, because the functions there never exhibit the behavior the caveats guard against.

7. Where it genuinely bites — and why rigor still earns its keep

The shortcut is not universally safe, and the failures are exactly the places where physics historically stumbled and where mathematical rigor paid off:

Interchanging limits / sums / integrals without uniform convergence (sequences-series.md). Fourier series of a discontinuous signal exhibit the Gibbs phenomenon; termwise differentiation of a convergent series can diverge. The naive "swap freely" rule fails, and one needs uniform (or dominated) convergence.
Divergent integrals and series in QFT. Perturbative quantum field theory produces divergent momentum integrals; making them finite required renormalization, and the perturbation series is typically only asymptotic, not convergent. Naive infinitesimal bookkeeping gives $\infty$ ; the fix is genuine analysis.
The Dirac delta is not a function — treating it as one (" $δ (0) = \infty$ ") leads to contradictions like $δ (x)^{2}$ . Distribution theory is the price of rigor, and it forbids exactly the illegitimate products.
The Feynman path integral has no fully rigorous measure-theoretic definition in $3 + 1$ dimensions to this day; the " $D ϕ$ " is a heuristic awaiting (constructive-QFT) justification.
Order-of-integration swaps when Fubini's hypotheses fail (multiple-integrals.md §3) — a non-absolutely-convergent double integral can give two different answers.

In every case the informal calculus produces a plausible-looking wrong answer, and rigor is what flags the danger. So the analyst's machinery is not pedantry: it is the error-detection layer that tells you when the fast heuristic has left its domain of validity.

8. The two-level picture

The healthy way to hold both together is a division of labor:

A computational layer (the physicist's calculus): infinitesimals, $d x$ as a length, differentials as fractions, $\int$ as a sum. It is fast, dimensionally meaningful, and geometrically vivid. Its manipulations are compressions of limits and its objects are 1-forms / hyperreals under the hood.
A justificatory layer (the analysis spine): $ε$ – $δ$ , $sup / in f$ , measure, uniform convergence. It certifies when the computational layer is valid and diagnoses when it is not.

They are not rivals. The notation is a lossy-but-recoverable compression: for smooth physical data the decompression is guaranteed to succeed (§6), so the physicist may compute in the compact layer and trust the answer; when data is rough or a limit is exchanged, one drops to the justificatory layer to check. This is the same object-level/meta-level relationship that recurs throughout these notes (cf. Models, Truth, and Metalanguage): one layer does the work, another certifies it.

Dimensional analysis reinforces the point: $d x$ carrying units of length, $λ d x$ units of mass, $\int$ accumulating them — the unit bookkeeping is a second, independent check that the infinitesimal manipulation was assembled correctly, and it is one physicists lean on precisely because it is so reliable.

Take-aways

The integral half is Riemann sums with the limit hidden. " $\int f d x$ = sum of slices $f d x$ " is $lim_{mesh \to 0} \sum f (x_{i}) Δ x_{i}$ ; $d x$ is the typographic ghost of $Δ x_{i}$ , $\int$ a stretched summa- $S$ .
The differential half is exact, not approximate. $d y = f^{'} (x) d x$ is a true identity of linear maps, and more deeply of differential 1-forms (differential-forms.md); the whole div/grad/curl/Stokes apparatus is the exterior calculus, computed informally.
Infinitesimals are literally respectable. Robinson's non-standard analysis ( $^{*} R$ + transfer principle) and smooth infinitesimal analysis (nilpotent $ϵ^{2} = 0$ ) both make " $d x$ is a real tiny thing" rigorous and consistent, vindicating the 300-year-old Leibniz style.
Every physical move has a named theorem (change of variables, Fubini, Leibniz rule, Stokes, distributions) — the notation ships the skeleton, the theorem supplies the hypotheses.
It is safe because the hypotheses hold silently: physical fields are smooth / compactly supported, so the monsters ( $1_{Q}$ , Weierstrass) that the $sup / in f$ definitions guard against never appear.
It bites exactly where physics got burned: interchanging limits without uniform convergence, QFT divergences (renormalization), $δ (x)^{2}$ (distributions), the path integral. Rigor is the error-detection layer, not pedantry.

Related discussions: The Riemann Integral (the $sup / in f$ definition and why matching is a theorem), Differential Forms and Exterior Calculus (the rigorous meaning of $d x$ ), The Generalized Stokes Theorem (the machinery behind div/grad/curl), and, for the object-level/meta-level pattern, Models, Truth, and Metalanguage.

Keyboard shortcuts

youyuanwu