Fundamental Lemma of the Calculus of Variations

Every variational principle in physics ends with the same sentence: "since this holds for all variations, the integrand must vanish." That sentence is the fundamental lemma of the calculus of variations, and it does far more work than its throwaway phrasing suggests. When you demand that an action $J[y]=\int_a^b F(x,y,y')\,\mathrm{d}x$ be stationary, integration by parts collapses $\delta J=0$ into a single integral, $\int_a^b g(x)\,\eta(x)\,\mathrm{d}x=0$ , that must hold for every admissible variation $\eta$ . This is a weak statement — an infinite family of averages, one per probe $\eta$ — and by itself it says nothing pointwise. The lemma is the bridge from that weak statement to the strong one, $g(x)=0$ at every point, which is the Euler–Lagrange equation. The mechanism is a detection argument. Think of $\eta$ as a smooth weighting probe you can slide and shape anywhere inside $(a,b)$ . If $g$ were nonzero somewhere — say positive on a little interval — you could park a positive bump exactly there, and the weighted integral would come back strictly positive, contradicting the assumption that it vanishes against all probes. A continuous function has nowhere to hide from an infinite-dimensional family of test functions; the only function invisible to all of them is the zero function. That single idea underwrites Lagrangian mechanics, geodesics, Hamilton's principle, and the field equations of every classical and quantum field theory, and it is the historical seed of distribution theory.

Key takeaways

The lemma states that if a continuous $f$ satisfies $\int_a^b f\,\eta\,\mathrm{d}x=0$ for every smooth $\eta$ vanishing at the endpoints, then $f\equiv 0$ on $[a,b]$ .
It is the logical bridge that turns the stationarity condition $\delta J=0$ — an infinite family of integral identities — into a single pointwise differential equation.
The proof is constructive: assume $f(c)>0$ , use continuity to get a whole interval where $f>\tfrac12 f(c)$ , then park a smooth bump there so the integral is strictly positive — a contradiction.
Continuity of $f$ is essential; without it the conclusion weakens to $f=0$ almost everywhere, since a single isolated spike is invisible to integration.
A family of variants extends its reach: du Bois-Reymond (no differentiability assumed), zero-mean constrained (gives $f=\text{const}$ ), and multidimensional (gives field equations on $\Omega\subset\mathbb{R}^n$ ).
Key idea: an integral identity $\int_a^b f\,\eta\,\mathrm{d}x=0$ that holds for all admissible $\eta$ is vastly stronger than for one $\eta$ — the probes reach every location, so a continuous $f$ can hide nowhere and must vanish identically.

The core insight

An integral identity $\int_a^b f\,\eta\,\mathrm{d}x=0$ holding for all admissible $\eta$ is far stronger than for one $\eta$ : the test functions probe every location, so a continuous $f$ can hide nowhere — it must vanish identically.

Where this lives today

The lemma is invisible in physics only because it is everywhere. Every time an action principle is turned into an equation of motion — the Einstein field equations from the Einstein–Hilbert action, the Yang–Mills equations of the Standard Model — the final "therefore the integrand vanishes" step is the multidimensional fundamental lemma (6). In numerical analysis it is the theoretical backbone of the finite element method: engineers solve the weak form $\int_\Omega(\nabla u\cdot\nabla v - f v)\,\mathrm{d}^n x=0$ for all test functions $v$ , and the lemma is exactly the guarantee that a smooth-enough weak solution satisfies the original strong PDE. The same equivalence powers physics-informed neural networks. And the lemma's deepest legacy is conceptual: "a function is determined by its integrals against all test functions" is the founding axiom of distribution theory, without which quantum field theory could not write down its propagators and delta functions. A live frontier — the regularity theory of variational solutions, part of Hilbert's 19th and 20th problems — is still about how far you can weaken the probe class and the smoothness of $f$ before the clean conclusion frays.

The physical picture

The engine of the lemma is a thought experiment about detection. You are handed a mystery continuous $f$ and told one thing: its weighted average against every smooth, endpoint-vanishing probe $\eta$ is exactly zero. Can $f$ be secretly nonzero? No — because you have an infinite budget of probes. If $f$ humps up positive on any interval, slide a smooth positive bump onto that hump; the product $f\eta$ is then positive everywhere it is nonzero, so its integral cannot be zero. Vanishing against everything is incompatible with $f$ being nonzero anywhere.

Figure

A continuous function f that humps positive somewhere cannot hide from the probes: slide a smooth bump η onto the hump and the product f·η is positive wherever it is nonzero, so ∫ f η dx = +0.072 which is greater than 0 — impossible if the integral must vanish against every η, so f ≡ 0

The probe must be a smooth bump: it and all of its derivatives must die at the edge of its support, so that patching it into the constant zero outside leaves an infinitely differentiable function. The canonical construction — the mollifier $\phi(x)=\exp(-1/(1-x^2))$ — does exactly this. It is strictly positive on $(-1,1)$ , flattens to zero so gently at $x=\pm1$ that every derivative vanishes there, and encloses a definite, computable area. Sliding and rescaling this one bump gives a probe centred anywhere with any width — all the proof ever needs.

Figure

The canonical smooth bump φ(x) = exp(−1/(1−x²)) on the interval (−1, 1): strictly positive inside, with peak value φ(0) = e⁻¹ = 0.368 and enclosed area ∫₋₁¹ φ dx = 0.444 — every derivative vanishes at x = ±1, so it belongs to C^∞ with compact support

The core equations

The lemma itself is a one-line implication, but its content is the quantifier "for all $\eta$ ". Read it as: no continuous function can be orthogonal to every test function unless it is zero. This is the pointwise-from-weak converter.

\int_a^b f(x)\,\eta(x)\,\mathrm{d}x = 0\quad\forall\,\eta\in C^\infty_c(a,b)\ \Longrightarrow\ f(x)\equiv 0 \qquad (1)

where:

symbol	meaning
$f$	a continuous function on $[a,b]$ , the unknown to be pinned down
$\eta$	a smooth test function vanishing with all derivatives near $a$ and $b$
$C^\infty_c(a,b)$	infinitely differentiable functions with compact support in $(a,b)$

Reading: the hypothesis is an infinite bundle of scalar equations (one real number per $\eta$ ); the conclusion pins a whole function to zero. The strength comes entirely from " $\forall\eta$ ". The proof needs one explicit admissible probe, and the standard choice is the mollifier bump — smooth, compactly supported, strictly positive on its interior. Note the sign in the exponent: the argument $-1/(r^2-(x-c)^2)$ tends to $-\infty$ at the support edge, so the bump and all its derivatives decay smoothly to zero there.

\eta_0(x)=\begin{cases}\exp\!\left(-\dfrac{1}{\,r^2-(x-c)^2\,}\right)& |x-c|<r\\[4pt]0&|x-c|\ge r\end{cases}\qquad \int \eta_0\,\mathrm{d}x = 0.44399\,r \qquad (2)

where:

symbol	meaning
$c$	centre of the bump, placed where $f\neq 0$
$r$	radius, small enough that $f$ keeps one sign across $[c-r,c+r]\subset(a,b)$
$0.44399$	the area of the unit mollifier $\int_{-1}^{1}e^{-1/(1-x^2)}\,\mathrm{d}x$ , scaled by $r$

Reading: rescaling the unit bump by width $r$ scales its area linearly, so the enclosed area is always strictly positive — the property that forces the contradiction. The lemma earns its keep at the last step of every variational calculation. Varying $y\to y+\epsilon\eta$ with fixed endpoints and integrating by parts turns $\delta J=0$ into a single integral against $\eta$ ; the boundary term dies because $\eta(a)=\eta(b)=0$ .

\delta J=\int_a^b\left(\frac{\partial F}{\partial y}-\frac{\mathrm{d}}{\mathrm{d}x}\frac{\partial F}{\partial y'}\right)\eta\,\mathrm{d}x=0\quad\forall\,\eta \qquad (3)

where:

symbol	meaning
$J$	the functional $\int_a^b F(x,y,y')\,\mathrm{d}x$ being made stationary
$\eta=\delta y$	the variation, vanishing at both endpoints
$F_y-\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y'}$	the Euler expression that the lemma will force to zero pointwise

Reading: (3) is weak; applying (1) gives the strong Euler–Lagrange equation $F_y-\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y'}=0$ . When the integrand carries $\eta'$ rather than $\eta$ — as it does before you commit to integrating by parts — the relevant tool is the du Bois-Reymond form, which needs no differentiability of the coefficient.

\int_a^b g(x)\,\eta'(x)\,\mathrm{d}x=0\quad\forall\,\eta\in C^\infty_c(a,b)\ \Longrightarrow\ g(x)=\text{const} \qquad (4)

where:

symbol	meaning
$g$	a continuous function multiplying the derivative of the probe
$\text{const}$	a single constant, undetermined by the hypothesis alone

Reading: orthogonality to every $\eta'$ (exactly the zero-mean functions) leaves only the constants — this is what lets one derive Euler–Lagrange without assuming $F_{y'}$ differentiable. Restricting the probes to zero mean weakens the conclusion from $f\equiv 0$ to $f=\text{const}$ — the exact origin of the Lagrange multiplier in isoperimetric problems.

\int_a^b f\,\eta\,\mathrm{d}x=0\ \ \text{for all }\eta\text{ with }\int_a^b\eta\,\mathrm{d}x=0\ \Longrightarrow\ f=\text{const} \qquad (5)

where:

symbol	meaning
$\int_a^b\eta\,\mathrm{d}x=0$	the zero-mean constraint that makes constants invisible to the probes
$\text{const}$	the Lagrange multiplier $\lambda$ of the constrained problem

Finally, replacing the interval by a region promotes the lemma to field theory, converting a varied action into a local field equation at every spacetime point.

\int_\Omega f(\mathbf{x})\,\eta(\mathbf{x})\,\mathrm{d}^n x=0\quad\forall\,\eta\in C^\infty_c(\Omega)\ \Longrightarrow\ f\equiv 0\ \text{ in }\Omega \qquad (6)

where:

symbol	meaning
$\Omega$	an open region in $\mathbb{R}^n$ (space or spacetime)
$\mathrm{d}^n x$	the volume element

The same detection argument with $n$ -dimensional bumps gives the Euler–Lagrange field equations of every action principle.

Derivation AProof of the fundamental lemma by contradiction

The goal is to prove implication (1): given $f$ continuous on $[a,b]$ with $\int_a^b f\eta\,\mathrm{d}x=0$ for every $\eta\in C^\infty_c(a,b)$ , show $f\equiv 0$ . The strategy is contraposition by construction — assume $f$ is nonzero somewhere, manufacture one legitimate test function that makes the integral strictly positive, and read off the contradiction. Continuity is load-bearing: it spreads a single nonzero value into a whole interval where $f$ keeps its sign, exactly the interval the bump exploits.

Step 1. Suppose, for contradiction, that $f$ is not identically zero: there is a point $c\in(a,b)$ with $f(c)\neq 0$ . Without loss of generality take $f(c)>0$ ; the negative case follows by replacing $\eta$ with $-\eta$ . The endpoints need no separate treatment since $\eta$ vanishes there and continuity fixes $f$ at $a,b$ as limits. Why: we only need to defeat the hypothesis at one interior point, and a positive value is enough by symmetry.

Step 2. By continuity of $f$ at $c$ , choose $r>0$ so small that $f(x)>\tfrac12 f(c)>0$ for all $|x-c|<r$ , with $[c-r,c+r]\subset(a,b)$ . Such an $r$ exists because $f$ stays within any prescribed band of $f(c)$ on a small enough neighbourhood. Why: this converts the single positive value $f(c)$ into a positive floor for $f$ over an entire interval — the region the probe will occupy.

Step 3. Take the scaled mollifier bump $\eta_0$ of equation (2), centred at $c$ with radius $r$ . It is $C^\infty$ everywhere (all derivatives vanish at $|x-c|=r$ ), it has compact support $[c-r,c+r]\subset(a,b)$ , and it is strictly positive on the open interval $(c-r,c+r)$ . Hence $\eta_0\in C^\infty_c(a,b)$ is a legitimate probe. Why: the hypothesis is quantified over exactly this class, so $\eta_0$ is a fair test we are entitled to apply.

Step 4. Evaluate the integral against this probe. Outside the support the integrand is zero, and inside it is a product of two positive quantities, so $\int_a^b f\eta_0\,\mathrm{d}x=\int_{c-r}^{c+r} f\,\eta_0\,\mathrm{d}x\ge \tfrac12 f(c)\int_{c-r}^{c+r}\eta_0\,\mathrm{d}x=\tfrac12 f(c)\cdot 0.44399\,r>0$ . Why: bounding $f$ below by its floor and pulling it out gives a strictly positive lower bound, using the computed bump area.

Step 5. But the hypothesis says this same integral is zero, giving $0<\int_a^b f\eta_0\,\mathrm{d}x=0$ , a contradiction. So no such $c$ exists, and $f\equiv 0$ on $[a,b]$ . $\blacksquare$ Why: a strictly positive number cannot equal zero, so the Step 1 assumption is impossible.

Go deeper · Derivation B · 1 minThe du Bois-Reymond form and why differentiability is unnecessary

Derivation BThe du Bois-Reymond form and why differentiability is unnecessary

The goal is to prove implication (4): if a continuous $g$ satisfies $\int_a^b g\,\eta'\,\mathrm{d}x=0$ for all $\eta\in C^\infty_c(a,b)$ , then $g$ is constant — crucially without ever assuming $g$ is differentiable. This matters because when you derive Euler–Lagrange by keeping $F_{y'}$ under the derivative, you do not yet know it is differentiable; the naive proof that integrates by parts first is circular. Du Bois-Reymond's trick is to build a bespoke test function whose derivative equals $g$ minus its own mean, turning the hypothesis into a statement that a non-negative integral vanishes.

Step 1. Define the mean value $\bar g=\frac{1}{b-a}\int_a^b g(x)\,\mathrm{d}x$ , the constant we will prove $g$ equals. Why: any candidate constant must be the average, since averaging both sides of $g=\text{const}$ forces that value.

Step 2. Build the special variation $\eta(x)=\int_a^x\big(g(t)-\bar g\big)\,\mathrm{d}t$ . Then $\eta(a)=0$ trivially, and $\eta(b)=\int_a^b(g-\bar g)\,\mathrm{d}t=0$ by the definition of $\bar g$ , so $\eta$ vanishes at both endpoints and is admissible, with derivative $\eta'=g-\bar g$ . Why: this manufactures a legal test function whose derivative is precisely the deviation of $g$ from its mean.

Step 3. The constant $\bar g$ is invisible to the hypothesis, because $\int_a^b \bar g\,\eta'\,\mathrm{d}x=\bar g\,[\eta]_a^b=0$ since $\eta$ vanishes at both ends. Subtract this from the hypothesis $\int_a^b g\,\eta'\,\mathrm{d}x=0$ to get $\int_a^b(g-\bar g)\,\eta'\,\mathrm{d}x=0$ . Why: shifting $g$ by a constant changes nothing, so we may work with the zero-mean part alone.

Step 4. Substitute $\eta'=g-\bar g$ from Step 2 to obtain $\int_a^b\big(g-\bar g\big)^2\,\mathrm{d}x=0$ . The integrand is continuous and non-negative. Why: choosing $\eta'$ equal to $g-\bar g$ turns the mixed integral into the integral of a perfect square.

Step 5. A continuous non-negative function with zero integral must be identically zero, so $(g-\bar g)^2\equiv 0$ , i.e. $g\equiv\bar g$ is constant. $\blacksquare$ Why: if it were positive anywhere, continuity would give a positive-area interval and a positive integral, contradicting Step 4.

Go deeper · Derivation C · 1 minFrom the weak stationarity condition to the Euler–Lagrange equation

Derivation CFrom the weak stationarity condition to the Euler–Lagrange equation

The goal is to watch the lemma do its job: converting the physicist's shorthand " $\delta J=0$ for all variations" into the pointwise Euler–Lagrange equation. Starting from $J[y]=\int_a^b F(x,y,y')\,\mathrm{d}x$ over curves with fixed endpoints, we perturb by an admissible $\eta$ , extract the first-order term, integrate by parts to isolate $\eta$ , and then — only then — invoke the fundamental lemma to strip the integral. This is where a weak, integrated statement becomes a strong, local differential equation.

Step 1. Replace $y$ by $y+\epsilon\eta$ with $\eta(a)=\eta(b)=0$ and expand $J[y+\epsilon\eta]$ to first order in $\epsilon$ . The first variation is $\delta J=\frac{\mathrm{d}}{\mathrm{d}\epsilon}J[y+\epsilon\eta]\big|_{\epsilon=0}=\int_a^b\big(F_y\,\eta+F_{y'}\,\eta'\big)\,\mathrm{d}x$ . Why: stationarity means this linear term vanishes for every admissible direction $\eta$ , the variational analogue of a zero gradient.

Step 2. Integrate the second term by parts to move the derivative off $\eta$ : $\int_a^b F_{y'}\,\eta'\,\mathrm{d}x=\big[F_{y'}\,\eta\big]_a^b-\int_a^b\frac{\mathrm{d}}{\mathrm{d}x}F_{y'}\,\eta\,\mathrm{d}x$ . Why: the lemma acts on integrals of the form $\int(\cdot)\eta\,\mathrm{d}x$ , so we must get $\eta$ itself, not $\eta'$ , as the common factor.

Step 3. The boundary term $\big[F_{y'}\,\eta\big]_a^b$ vanishes because $\eta(a)=\eta(b)=0$ for fixed-endpoint variations. Collecting the surviving pieces gives $\delta J=\int_a^b\big(F_y-\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y'}\big)\,\eta\,\mathrm{d}x=0$ for all such $\eta$ . Why: this is exactly equation (3) — the weak form, with the Euler expression sitting in the role of $f$ .

Step 4. Apply the fundamental lemma (1) with $f=F_y-\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y'}$ , which is continuous when $F$ is smooth and $y\in C^2$ . Since its integral against every admissible $\eta$ vanishes, it must vanish pointwise: $F_y-\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y'}=0$ on $(a,b)$ . $\blacksquare$ Why: this is the single logical step that upgrades "the average is zero for all probes" to "the integrand is zero everywhere" — no lemma, no differential equation.

Limiting cases

The lemma is really a family, each member tuned to the probe class and the regularity assumed of the integrand.

Case	Result	Meaning
Basic — 1D, continuous $f$	$\int_a^b f\,\eta\,\mathrm{d}x=0\ \forall\eta\in C^\infty_c\ \Rightarrow\ f\equiv 0$	the canonical form that yields the Euler–Lagrange equation
du Bois-Reymond	$\int_a^b g\,\eta'\,\mathrm{d}x=0\ \forall\eta\ \Rightarrow\ g=\text{const}$	no differentiability of $g$ assumed; handles weak solutions
Constrained — zero-mean $\eta$	$\int_a^b f\,\eta\,\mathrm{d}x=0,\ \int_a^b\eta=0\ \Rightarrow\ f=\text{const}$	isoperimetric problems; the constant is the Lagrange multiplier
Multidimensional	$\int_\Omega f\,\eta\,\mathrm{d}^n x=0\ \forall\eta\in C^\infty_c(\Omega)\ \Rightarrow\ f\equiv 0$	field equations from an action over a region $\Omega$
Merely integrable $f$	conclusion weakens to $f=0$ almost everywhere	a measure-zero set of nonzero values is invisible to integration

Common trapThe lemma does not say a zero integral makes the integrand zero

It is tempting to read " $\int_a^b f\eta\,\mathrm{d}x=0$ implies $f=0$ " as a statement about one integral. It is not. A single vanishing integral says almost nothing — $\int_0^1(x-\tfrac12)\,\mathrm{d}x=0$ though $x-\tfrac12$ is almost nowhere zero. The conclusion $f\equiv 0$ requires the integral to vanish against every admissible $\eta$ ; that quantifier is the whole content. Two further cautions ride along. First, continuity of $f$ is essential: drop it and $f$ can be nonzero on a measure-zero set while every integral still vanishes, so the honest conclusion is $f=0$ almost everywhere. Second, the probe class must localize anywhere — restrict $\eta$ to a finite-dimensional family and a nonzero $f$ orthogonal to all of them can survive.

Worked examples

1 · A function that fools one probe but not the lemma

A skeptic claims the continuous function $f(x)=x^2-\tfrac13$ on $[0,1]$ is "invisible" to the lemma, offering as evidence that $\int_0^1 f\,\mathrm{d}x=0$ . Show that a single localized probe already exposes $f$ as nonzero, so the lemma's conclusion is not threatened.

Step 1. Check the skeptic's one datum. $\int_0^1(x^2-\tfrac13)\,\mathrm{d}x=\big[\tfrac{x^3}{3}-\tfrac{x}{3}\big]_0^1=\tfrac13-\tfrac13=\mathbf{0}$ . True — but this is the integral against the single probe $\eta\equiv 1$ , which is not even admissible (it does not vanish at the endpoints). One test function proves nothing.

Step 2. Deploy a proper localized probe where $f$ has a definite sign. Near the left end $f(0)=-\tfrac13<0$ , and $f$ stays negative on $[0,0.2]$ since $f(0.2)=0.04-0.333=-0.293$ . Place the scaled mollifier $\eta_0$ at $c=0.1$ , $r=0.1$ , so its support is exactly $(0,0.2)$ and its area is $0.44399\times 0.1=\mathbf{0.04440}$ .

Step 3. Evaluate the weighted integral. Since $f$ is negative and $\eta_0$ positive across the support, the product is negative: a numerical evaluation gives $\int_0^{0.2}(x^2-\tfrac13)\,\eta_0\,\mathrm{d}x=\mathbf{-0.01429}$ . Dividing by the bump area recovers the bump-weighted average $-0.01429/0.04440=-0.322$ , right next to the central value $f(0.1)=-0.323$ — the probe is reporting $f$ at its centre.

Step 4. Read the verdict. The integral against this admissible $\eta_0$ is $\mathbf{-0.0143\neq 0}$ , so $f$ fails the lemma's hypothesis: it is not orthogonal to all test functions. No contradiction arises — the lemma concludes $f\equiv 0$ only when every probe returns zero, and this one does not. A vanishing integral against one $\eta$ is meaningless; the localized bump unmasks the nonzero $f$ the average had hidden.

2 · The lemma delivers the straight line

Among smooth curves joining $(0,0)$ to $(1,1)$ , which minimizes arc length $J[y]=\int_0^1\sqrt{1+y'^2}\,\mathrm{d}x$ ? Carry the variation through and show where the lemma converts the integral into the answer.

Step 1. Here $F=\sqrt{1+y'^2}$ depends on $y'$ only, so $F_y=0$ and $F_{y'}=\dfrac{y'}{\sqrt{1+y'^2}}$ . After integration by parts with $\eta(0)=\eta(1)=0$ , the first variation is $\delta J=-\int_0^1\dfrac{\mathrm{d}}{\mathrm{d}x}\!\left(\dfrac{y'}{\sqrt{1+y'^2}}\right)\eta\,\mathrm{d}x=0$ for all admissible $\eta$ .

Step 2. Apply the fundamental lemma. The bracket multiplying $\eta$ is continuous, and its integral vanishes against every $\eta$ , so it must vanish pointwise: $\dfrac{\mathrm{d}}{\mathrm{d}x}\!\left(\dfrac{y'}{\sqrt{1+y'^2}}\right)=0$ . This is the step the whole article is about — without it we could only say "the average slope-change is zero", not "the slope-change is zero everywhere".

Step 3. Integrate the pointwise equation. A vanishing derivative means $\dfrac{y'}{\sqrt{1+y'^2}}=\text{const}$ , which forces $y'=\text{const}=:m$ . The endpoints give $m=1$ , so $y(x)=x$ . The minimizer is the straight line $y=x$ , length $\sqrt2$ — the shortest path, recovered because the lemma turned an integral identity into a differential equation.

Physics of the portrait

The lemma is one of the oldest ideas in the calculus of variations and one of the last to be made rigorous. Lagrange used its content freely in the 1750s; it took more than a century, and Paul du Bois-Reymond, to prove it under the mild assumptions physicists actually need.

Origins · 1 minPaul du Bois-Reymond and how this was discovered

Paul du Bois-Reymond

1831–1889 · rigour in the calculus of variations

Paul du Bois-Reymond was a German mathematician — younger brother of the physiologist Emil du Bois-Reymond — whose career was a sustained campaign for rigour in analysis at the moment the subject demanded it. In 1873 he constructed a continuous function whose Fourier series diverges at a point, shattering the belief that continuity guaranteed convergence. He developed an "infinitary calculus" of orders of growth, a precursor of modern asymptotics. And in the calculus of variations he proved the lemma that now bears his name: if $\int g\,\eta'\,\mathrm{d}x=0$ for all test functions $\eta$ , then $g$ is constant — without assuming $g$ is differentiable. That last qualification is the whole point. Euler and Lagrange had derived the Euler–Lagrange equation by blithely differentiating $F_{y'}$ , but nothing in the variational problem guarantees that derivative exists in advance; du Bois-Reymond's argument builds a bespoke test function to reach the conclusion honestly, and it is the version a careful text proves today. His insistence that a continuous function is pinned down by its action on test functions is precisely the duality that Sobolev and Schwartz would later formalize as the theory of distributions.

History

1755

— Lagrange sends Euler his δ-algorithm, founding the calculus of variations

1834

— Hamilton casts mechanics as a stationary-action principle

1879

— du Bois-Reymond proves the lemma with minimal smoothness on the integrand

1900

— Hilbert's Paris problems put variational rigour at centre stage

1936

— Sobolev introduces generalized functions and mollifiers

1945

— Schwartz's theory of distributions formalizes the function–functional duality

Going further

Texts. Gelfand & Fomin, Calculus of Variations, §3 (the lemma and du Bois-Reymond) · Stone & Goldbart, Mathematics for Physics, §1.2 · Arfken & Weber, Mathematical Methods for Physicists, ch. on the calculus of variations · Hassani, Mathematical Physics, ch. on variational methods · Byron & Fuller, Mathematics of Classical and Quantum Physics, ch. 2 · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, ch. 22 · Nakahara, Geometry, Topology and Physics, for the field-theoretic multidimensional version.

Provenance. The statement and the two proofs follow Gelfand & Fomin (§3) and Stone & Goldbart (§1.2); the numbers are re-verified here. A checker should confirm: (1) the mollifier $\phi(x)=e^{-1/(1-x^2)}$ is $C^\infty$ with all derivatives vanishing at $x=\pm1$ , peak $\phi(0)=e^{-1}=0.3679$ , and area $\int_{-1}^{1}\phi\,\mathrm{d}x=0.44399$ ; (2) the worked-example integral $\int_0^{0.2}(x^2-\tfrac13)\eta_0\,\mathrm{d}x=-0.0143$ with $\eta_0$ centred at $c=0.1$ , $r=0.1$ ; (3) the du Bois-Reymond construction yields $\eta(a)=\eta(b)=0$ ; (4) the arc-length example gives $y=x$ . Note a corrected sign relative to some sources: the bump is $\exp(-1/(r^2-(x-c)^2))$ , which decays to zero at the support edge — the form $\exp(-1/((x-c)^2-r^2))$ would diverge there.

Prerequisites. Functionals and Function Spaces · First Variation and Stationarity. Leads to. Euler–Lagrange Equation · Natural and Essential Boundary Conditions · Constrained Variation and Lagrange Multipliers · Hamilton's Principle of Stationary Action.

Ask the tutor: ask me to run the bump-probe contradiction on any continuous $f$ you name, or to push the lemma through a variational problem of your choice to its Euler–Lagrange equation.

Why is it called the "fundamental" lemma?

Because it is the single indispensable step that makes the calculus of variations produce differential equations. Everything upstream — building the functional, forming the first variation, integrating by parts — leaves you with a weak statement: an integral that vanishes for all variations. Only the lemma converts that into the pointwise Euler–Lagrange equation; without it you are stuck with an integral identity that constrains averages but names no local law. Since every action principle in mechanics, optics, and field theory funnels through this one conversion, it earns the name.

What exactly is a test function, and why must it vanish at the endpoints?

A test function $\eta$ is a smooth probe you multiply against the unknown before integrating — in the variational setting, the shape $\delta y$ of an allowed perturbation of the curve. It vanishes at the endpoints because in a fixed-endpoint problem all competing curves share the same boundary values, so the perturbation cannot move the ends; formally $\eta\in C^\infty_c(a,b)$ . That condition also kills the boundary term in the integration by parts, leaving the clean integral $\int(\cdot)\eta\,\mathrm{d}x$ the lemma acts on. Relax it — free endpoints — and the surviving boundary term gives the natural boundary conditions.

Does the bump function

\eta_0

really have all derivatives zero at its edge?

Yes, and that is the whole reason it exists. The mollifier $\exp(-1/(r^2-(x-c)^2))$ approaches the support edge $|x-c|=r$ with an argument tending to $-\infty$ , so the function and every derivative — the function times rational blow-ups — are crushed to zero faster than any polynomial grows. This flatness to infinite order lets you glue the bump to the constant zero outside its support and still have a globally $C^\infty$ function. A raised cosine has a discontinuous second derivative at the edge and would not qualify; the exponential mollifier is the standard fix.

What breaks if

f

is only integrable rather than continuous?

You lose the pointwise conclusion and keep only " $f=0$ almost everywhere". The proof's Step 2 needs continuity to inflate a single nonzero value $f(c)$ into an interval where $f$ keeps its sign; without continuity, $f$ could be nonzero on a measure-zero set while every integral against a smooth $\eta$ still returns zero, because integration cannot see measure-zero sets. This is no technicality: in the $L^2$ theory of PDEs, functions are equivalence classes up to measure-zero changes, and the lemma rightly concludes $f=0$ as an element of $L^2$ — almost everywhere.

How is the constrained version connected to Lagrange multipliers?

Directly. In an isoperimetric problem you minimize $J[y]$ subject to $K[y]=\text{const}$ , which restricts admissible variations to those preserving the constraint — precisely the zero-mean-type condition $\int(\cdot)\eta\,\mathrm{d}x=0$ . The constrained lemma (5) then says the Euler expression of $J$ is forced not to zero but to a constant multiple of the Euler expression of $K$ . That constant is the Lagrange multiplier $\lambda$ , and the effective equation is the Euler–Lagrange equation of $J-\lambda K$ . The catenary and the fixed-perimeter maximum-area problem both fall out this way.

Why is this lemma seen as the seed of distribution theory?

Because it is the first clear statement that a function is completely determined by how it pairs with test functions. The lemma says the only continuous $f$ whose pairings $\int f\eta$ all vanish is $f\equiv 0$ ; equivalently, two continuous functions that pair identically against every $\eta$ are equal. Distribution theory turns this around into a definition: define generalized functions (the Dirac delta, derivatives of discontinuous functions) purely by their pairings with test functions. Sobolev's mollifiers and Schwartz's distributions are the mature form of the duality du Bois-Reymond was already exploiting — which is why the same bump reappears as the mollifier in both subjects.

Fundamental Lemma of the Calculus of Variations

Where this lives today

The physical picture

The core equations

Limiting cases

Worked examples

1 · A function that fools one probe but not the lemma

2 · The lemma delivers the straight line

Physics of the portrait

Going further

Before this

Where this leads