Functionals and Function Spaces

A function maps numbers to numbers; a functional maps whole functions to numbers. The arc length of a curve, the descent time of a bead, the action of a mechanical trajectory, and the bending energy of an elastic beam are each a single real number fixed by an entire profile $y(x)$ — not by any one of its values. Writing such a quantity as $J[y]$ , with square brackets by long convention, signals that its argument is a function drawn from some admissible class $\mathcal{A}$ , and that $J$ lives one storey above ordinary calculus. The calculus of variations is the systematic study of how $J[y]$ responds when the whole curve $y$ is deformed, and of which curves make $J$ stationary — the first variation, the Euler–Lagrange equation, and Hamilton's principle are all this one idea, differentiated. To do calculus on a functional we must first say two things the notation hides: which functions are allowed in, and how we measure when two of them are close. That is the job of the function space — a vector space of functions carrying a norm. The admissible class is almost always an affine slice of such a space: functions of a prescribed smoothness that also satisfy fixed endpoint data $y(a)=y_a$ , $y(b)=y_b$ . A deformation is then $y+\epsilon\eta$ , and the test function $\eta$ must vanish wherever the data are pinned, so that $y+\epsilon\eta$ stays admissible for every small $\epsilon$ . Two choices in that setup do real physical work. The norm — weak ( $C^0$ , heights only) versus strong ( $C^1$ , heights and slopes) — decides which curves count as near and therefore which minima exist at all. And the split of boundary data into essential (built into $\mathcal{A}$ ) and natural (falling out of stationarity) is entirely a statement about how the space was defined. The function space is not decoration; it is part of the problem.

Key takeaways

A functional is a map $J:\mathcal{A}\to\mathbb{R}$ that eats an entire function and returns one real number — arc length, action, energy — with square brackets $J[y]$ marking the functional dependence.
Its domain $\mathcal{A}$ is a function space: a vector space of functions of fixed smoothness and boundary data, carrying a norm so that "nearby function" has a precise meaning.
Most physical functionals are integrals $J[y]=\int_a^b F(x,y,y')\,\mathrm{d}x$ ; this local integral structure is exactly what makes $J$ differentiable — you differentiate $F$ under the integral sign.
The functional derivative is a gradient in infinitely many directions: $\delta J/\delta y = F_y-\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y'}$ , the Euler–Lagrange operator, and setting it to zero is $\nabla J=0$ .
The norm chooses the minima: the sequence $y_n=\tfrac1n\sin(n^2x)$ is $C^0$ -small ( $1/n\to0$ ) yet $C^1$ -large (slope $n\to\infty$ ), which is why weak and strong extrema differ.
Boundary data split into essential (imposed on $\mathcal{A}$ ) and natural (emerging from $\delta J=0$ ) — a distinction that lives entirely in how the space is built.
Key idea: a functional is a real-valued function whose domain is a space of functions; the calculus of variations is ordinary calculus done on that infinite-dimensional domain, with the norm deciding what "small change" and "minimum" even mean.

The core insight

A functional is a real-valued function whose domain is a space of functions; the calculus of variations is just ordinary calculus performed on that infinite-dimensional domain, with the choice of norm deciding what "small change" and "minimum" mean.

Where this lives today

The idea that a physical quantity is a functional, minimised over a function space, is load-bearing across modern computation and theory. The finite element method — the engine behind structural, aerodynamic, and electromagnetic simulation — is Volterra's discretisation heuristic made industrial: an energy functional is minimised over a finite-dimensional subspace of piecewise-polynomial shape functions, turning a PDE into a sparse linear system, with the choice of Sobolev space ( $H^1$ , $H^2$ ) fixing which elements are admissible. In quantum chemistry, density functional theory rests on the Hohenberg–Kohn theorem that the ground-state energy is a functional $E[n]$ of the electron density alone; minimising it is how essentially every molecule and material is now computed (the 1998 Nobel Prize in Chemistry). In quantum field theory the action is a functional over a space of field configurations, and Feynman's path integral sums $e^{iS[\phi]/\hbar}$ across it; the same template drives optimal control, general relativity's Einstein–Hilbert action, and optimal transport. In every case the two questions this article insists on — which functions are admissible, and in what norm — separate a well-posed computation from a divergent one.

The physical picture

Think of an ordinary function $f(x)$ as a machine: drop in a number, read off a number. A functional $J[y]$ is a machine one storey higher: drop in a whole curve and read off a number. The arc-length machine swallows the entire shape of a path and returns its total length; the action machine swallows a trajectory and returns a number whose stationarity is Newton's law. The right mental image is a landscape over the space of admissible curves — each curve is a single point, its $J$ -value is the height above it, and a physical solution sits at the bottom of a valley. The first figure makes that concrete: several competing curves through the same two endpoints, each mapped by the arc-length functional to one number on a line, with the straight segment sitting lowest.

The second thing the picture must show is why the ruler matters. On a function space you can measure how far apart two curves are by their heights alone (the weak, $C^0$ norm) or by their heights and their slopes together (the strong, $C^1$ norm). These disagree violently: a curve can hug another in height while its slope thrashes wildly, so a functional that "sees" slope registers a large change where a height-only ruler sees almost none. The canonical witness is $y_n(x)=\tfrac1n\sin(n^2x)$ — vanishingly small in height, enormous in slope. Whether your functional feels that wiggle is exactly what separates a weak extremum from a strong one, as the second figure shows.

Figure

Weak versus strong distance. The wiggle yₙ(x) = (1/n) sin(n²x) hugs the smooth reference y = 0 in height, ‖yₙ‖_C⁰ = 1/n → 0, yet its slope reaches ‖yₙ′‖ = n → ∞. At n = 10 the height is only 0.1 but the slope climbs to 10 — close in the C⁰ norm, far in the C¹ norm

Admissibility encodes the physics in the same breath. Fixing $y(a)$ and $y(b)$ pins the endpoints of every allowed curve; demanding a continuous slope forbids kinks. Those constraints carve the admissible class $\mathcal{A}$ out of a larger space of functions, and any deformation must respect them — which is precisely why the test functions $\eta$ that build a variation vanish wherever the data are held fixed.

The core equations

A functional is defined by its type: it is a rule that assigns one real number to each function in an admissible class. The square brackets are not decoration — they warn the reader that the slot expects an entire function, not a point value.

J:\mathcal{A}\to\mathbb{R},\qquad y\mapsto J[y] \qquad (1)

where: $J$ — the functional, a real-valued map; $\mathcal{A}$ — the admissible class, a subset of a function space; $y$ — an admissible function, the argument of $J$ ; $J[y]$ — the single real number assigned to $y$ . Reading: $J$ is to functions what an ordinary function is to numbers, one level up. Nearly every functional in physics is local and integral: its value accumulates a density $F$ along the curve. The integrand depends on only three slots — the point, the height, and the slope — and this locality is exactly what makes the theory tractable.

J[y]=\int_a^b F\big(x,\,y(x),\,y'(x)\big)\,\mathrm{d}x \qquad (2)

where: $F$ — the integrand or Lagrangian density, an ordinary function of three arguments; $x$ — the independent variable, $x\in[a,b]$ ; $y(x)$ — the height at $x$ ; $y'(x)$ — the slope $\mathrm{d}y/\mathrm{d}x$ . Reading: to differentiate $J$ you differentiate $F$ under the integral sign — the infinite-dimensional problem collapses to ordinary partial derivatives of $F$ . The domain is not "all functions" but an affine slice: functions smooth enough for $F$ to make sense, obeying the prescribed endpoint data. Those pinned values are the essential boundary conditions.

\mathcal{A}=\{\,y\in C^1[a,b]\;:\;y(a)=y_a,\;y(b)=y_b\,\} \qquad (3)

where: $C^1[a,b]$ — functions with a continuous first derivative on $[a,b]$ ; $y_a,\,y_b$ — prescribed endpoint values (essential boundary conditions). Reading: $\mathcal{A}$ is not a vector space (it does not contain $0$ unless $y_a=y_b=0$ ) but an affine translate of the space of variations that vanish at the ends. To measure proximity we need a norm, and the choice is physical, not cosmetic. The weak norm controls only heights; the strong norm controls heights and slopes together, a strictly finer topology.

\|y\|_{C^0}=\max_{x\in[a,b]}|y(x)|,\qquad \|y\|_{C^1}=\max_{x\in[a,b]}|y(x)|+\max_{x\in[a,b]}|y'(x)| \qquad (4)

where: $\|y\|_{C^0}$ — the uniform (sup) norm, sensitive to height only; $\|y\|_{C^1}$ — the strong norm, sensitive to height and slope. Reading: $C^1$ -close implies $C^0$ -close but never the reverse, which is the whole content of weak-versus-strong extrema. The derivative of a functional along a direction $\eta$ is the first variation, or Gateaux derivative — the infinite-dimensional analogue of a directional derivative. It is an ordinary $\epsilon$ -derivative of the number-valued function $\epsilon\mapsto J[y+\epsilon\eta]$ .

\delta J[y;\eta]\equiv\left.\frac{\mathrm{d}}{\mathrm{d}\epsilon}\,J[y+\epsilon\eta]\right|_{\epsilon=0} \qquad (5)

where: $\delta J[y;\eta]$ — the first variation of $J$ at $y$ in the direction $\eta$ ; $\eta$ — an admissible variation (test function) with $\eta(a)=\eta(b)=0$ at fixed ends. Reading: stationarity means $\delta J[y;\eta]=0$ for every admissible $\eta$ , the infinite-dimensional statement "all directional derivatives vanish". Carrying (5) through the integral (2) and integrating by parts factors the variation into a single density times $\eta$ . Because $\eta$ is arbitrary, that density must vanish pointwise — the functional gradient.

\delta J[y;\eta]=\int_a^b\Big(F_y-\frac{\mathrm{d}}{\mathrm{d}x}F_{y'}\Big)\eta\,\mathrm{d}x\;\;\Longrightarrow\;\;\frac{\delta J}{\delta y}=F_y-\frac{\mathrm{d}}{\mathrm{d}x}F_{y'} \qquad (6)

where: $F_y=\partial F/\partial y$ , $F_{y'}=\partial F/\partial y'$ — partial derivatives of the density; $\delta J/\delta y$ — the functional (variational) derivative. Reading: $\delta J/\delta y$ is the gradient of $J$ over function space, and setting it to zero is the Euler–Lagrange equation — the subject of the very next article.

The physics of it — three derivations

Derivation AThe first variation and the functional derivative

The goal is to turn the abstract map $J[y]$ into something we can differentiate, and to show that its gradient is a single density — the Euler–Lagrange operator — sitting under the integral. We embed the curve in a one-parameter family, reduce $J$ to an ordinary function of the parameter $\epsilon$ , differentiate under the integral sign, integrate by parts to peel every derivative off the arbitrary variation $\eta$ , and invoke the fundamental lemma of the calculus of variations to turn "the integral vanishes for all $\eta$ " into "the density vanishes pointwise". The payoff: stationarity of an infinite-dimensional object becomes a single differential equation.

Step 1. Embed $y$ in the family $y_\epsilon(x)=y(x)+\epsilon\,\eta(x)$ , where $\eta\in C^1$ is a fixed test function with $\eta(a)=\eta(b)=0$ so that $y_\epsilon\in\mathcal{A}$ for every $\epsilon$ . Why: a single real dial $\epsilon$ now sweeps out a curve of admissible functions, so the functional becomes an ordinary function of $\epsilon$ that we already know how to differentiate.

Step 2. Define $\Phi(\epsilon)=J[y_\epsilon]=\int_a^b F(x,\,y+\epsilon\eta,\,y'+\epsilon\eta')\,\mathrm{d}x$ . Because $[a,b]$ is fixed and $F$ is smooth, we may differentiate under the integral sign. Why: the endpoints do not move with $\epsilon$ , so $\mathrm{d}/\mathrm{d}\epsilon$ passes inside the integral and only acts on the integrand.

Step 3. Apply the chain rule to $F$ in its height and slope slots, then set $\epsilon=0$ :

\delta J[y;\eta]=\Phi'(0)=\int_a^b\big(F_y\,\eta+F_{y'}\,\eta'\big)\,\mathrm{d}x.

Why: only the height slot ( $y+\epsilon\eta$ ) and the slope slot ( $y'+\epsilon\eta'$ ) carry $\epsilon$ , contributing $F_y\,\eta$ and $F_{y'}\,\eta'$ respectively — this is the raw first variation.

Step 4. 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\eta\,\frac{\mathrm{d}}{\mathrm{d}x}F_{y'}\,\mathrm{d}x.

The boundary term vanishes because $\eta(a)=\eta(b)=0$ . Why: with the variation off $\eta'$ and onto $F_{y'}$ , both surviving terms multiply the same undifferentiated $\eta$ , and we can factor it out.

Step 5. Collect terms into $\displaystyle\delta J[y;\eta]=\int_a^b\Big(F_y-\frac{\mathrm{d}}{\mathrm{d}x}F_{y'}\Big)\eta\,\mathrm{d}x$ . If this vanishes for every admissible $\eta$ , the fundamental lemma forces the bracket to vanish identically, giving $\delta J/\delta y=F_y-\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y'}=0$ . Why: a continuous function whose weighted integral against all smooth test weights is zero must itself be zero — this is what promotes an integral identity to a pointwise differential equation. *Verified: $\delta J/\delta y=F_y-\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y'}$ , the Euler–Lagrange operator; consistent with Stone & Goldbart §1.2 and Gelfand & Fomin §3.*

Go deeper · Derivation B · 1 minFrom sampled function to functional — the discretisation heuristic

Derivation BFrom sampled function to functional — the discretisation heuristic

The aim is to demystify the functional derivative by exhibiting a functional as the continuum limit of an ordinary function of many variables, so that $\delta J/\delta y$ is literally an infinite-dimensional gradient. We sample the curve on a grid of $N+1$ points, replace the integral by a Riemann sum and slopes by finite differences, so $J$ becomes an ordinary function of the interior samples. Differentiating with respect to one sample — which sits in one height slot and two neighbouring slope differences — and letting the grid spacing vanish, the bracketed difference becomes a derivative and the Euler–Lagrange operator emerges as the gradient component, the same object derived analytically above, now seen as $\nabla J$ .

Step 1. Discretise the interval: $x_k=a+kh$ with $h=(b-a)/N$ and $y_k=y(x_k)$ , approximating slopes by forward differences $y'_k\approx(y_{k+1}-y_k)/h$ . Why: replacing the curve by finitely many numbers turns the functional into a genuine multivariable function whose calculus is elementary.

Step 2. Approximate the integral by a Riemann sum,

J\approx\sum_{k}F\!\left(x_k,\,y_k,\,\frac{y_{k+1}-y_k}{h}\right)h,

so $J$ is now an ordinary function of the $N-1$ free interior values $y_1,\dots,y_{N-1}$ (the endpoints are fixed data). Why: with endpoints pinned, only the interior samples are free — these are the coordinates of our finite-dimensional stand-in for function space.

Step 3. Differentiate with respect to one interior value. The sample $y_k$ appears in the height slot at $x_k$ and inside the slope differences at $x_{k-1}$ and $x_k$ , giving

\frac{\partial J}{\partial y_k}=h\,F_y(x_k,\dots)+\Big[F_{y'}(x_{k-1},\dots)-F_{y'}(x_k,\dots)\Big].

Why: this is just the multivariable chain rule bookkeeping — the two bracketed terms are the fingerprints of the two neighbouring finite-difference slopes that contain $y_k$ .

Step 4. Divide by $h$ and let $N\to\infty$ . The bracket is minus a forward difference of $F_{y'}$ , so $\tfrac1h[F_{y'}(x_{k-1})-F_{y'}(x_k)]\to-\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y'}$ , leaving

\frac{1}{h}\frac{\partial J}{\partial y_k}\;\xrightarrow{\,N\to\infty\,}\;\frac{\delta J}{\delta y(x)}=F_y-\frac{\mathrm{d}}{\mathrm{d}x}F_{y'}.

Why: the continuum limit converts a difference quotient of $F_{y'}$ into its derivative, reproducing exactly the operator of the previous derivation and confirming that $\delta J/\delta y$ is the gradient of $J$ . *Verified: reduces to the Euler–Lagrange operator; the $1/h$ scaling makes $\partial J/\partial y_k$ the discrete gradient whose limit is $\delta J/\delta y$ ; consistent with Stone & Goldbart §1.2.*

Go deeper · Derivation C · 1 minFirst variation of the arc-length functional

Derivation CFirst variation of the arc-length functional

The purpose is to run the general machinery on the oldest functional in geometry — the length of a curve — and watch it deliver the straight line as the unique stationary path, the geodesic of the plane. We take the arc-length density $F=\sqrt{1+y'^2}$ , note that it depends on slope alone, compute its single nonzero partial derivative, assemble the first variation, and integrate by parts. The vanishing boundary term (fixed endpoints) leaves a stationarity condition that forces the slope to be constant — and a curve of constant slope through two given points is the straight segment, closing the loop from abstract functional to concrete, checkable answer.

Step 1. Write the density $F(x,y,y')=\sqrt{1+y'^2}$ . It has no explicit $x$ or $y$ dependence, so $F_y=0$ . Why: length accumulates $\sqrt{1+y'^2}\,\mathrm{d}x=\mathrm{d}s$ , which depends only on how steep the curve is, never on where it is — hence the height slot drops out.

Step 2. Differentiate in the slope slot: $F_{y'}=\dfrac{y'}{\sqrt{1+y'^2}}$ . Why: this is the only surviving derivative, and it is exactly $\sin\theta$ where $\theta$ is the tangent angle — a quantity the straight line will make constant.

Step 3. Assemble the first variation using the raw form from Derivation 1:

\delta J[y;\eta]=\int_a^b\big(F_y\,\eta+F_{y'}\,\eta'\big)\,\mathrm{d}x=\int_a^b\frac{y'\,\eta'}{\sqrt{1+y'^2}}\,\mathrm{d}x.

Why: with $F_y=0$ only the slope term remains, so length responds solely to how the variation changes the slope.

Step 4. Integrate by parts and impose fixed endpoints $\eta(a)=\eta(b)=0$ :

\delta J=\Big[\tfrac{y'\eta}{\sqrt{1+y'^2}}\Big]_a^b-\int_a^b\eta\,\frac{\mathrm{d}}{\mathrm{d}x}\!\Big(\frac{y'}{\sqrt{1+y'^2}}\Big)\mathrm{d}x=-\int_a^b\eta\,\frac{\mathrm{d}}{\mathrm{d}x}\!\Big(\frac{y'}{\sqrt{1+y'^2}}\Big)\mathrm{d}x.

Stationarity for all $\eta$ then forces $\tfrac{\mathrm{d}}{\mathrm{d}x}\big(y'/\sqrt{1+y'^2}\big)=0$ , so $y'/\sqrt{1+y'^2}$ is constant, hence $y'$ is constant — a straight line. Why: constant $y'$ through two fixed points is a single segment, recovering the elementary fact that the shortest plane path is straight, now as a stationarity condition. *Verified: minimiser is the straight segment, the planar geodesic; $F_{y'}=\sin\theta=$ const is the geometric reading; cf. Riley–Hobson–Bence §22.2.*

Limiting cases

Specialising the general integral functional (2) recovers the household functionals of physics and pins down what each demands of its function space.

Case	Result	Meaning
Arc length	$J[y]=\int_a^b\sqrt{1+y'^2}\,\mathrm{d}x$	$F$ depends on slope only; minimisers are straight lines. Needs $y\in C^1$ .
Action of a particle	$S[q]=\int_{t_1}^{t_2}\big(\tfrac12 m\dot q^2-V(q)\big)\,\mathrm{d}t$	$F$ is the Lagrangian $L(q,\dot q,t)$ ; stationarity is Newton's law via Hamilton's principle. Needs $q\in C^2$ .
Dirichlet energy	$J[y]=\tfrac12\int_a^b y'^2\,\mathrm{d}x$	Quadratic functional; Euler–Lagrange equation $y''=0$ . Natural home is the Sobolev space $H^1$ .
Beam bending energy	$E[y]=\tfrac12\int_a^b EI\,y''^2\,\mathrm{d}x$	Second derivative enters; needs $y\in C^2$ and the higher-derivative Euler–Lagrange operator.
Pointwise (degenerate)	$J[y]=y(c),\;\,a<c<b$	A valid linear functional that is not an integral — functionals need not be integrals.
Linear / Riesz form	$J[y]=\int_a^b w(x)\,y(x)\,\mathrm{d}x$	Linear in $y$ ; the exception, since arc length and action are nonlinear in $y'$ .
Several fields	$J[\mathbf{y}]=\int F(x,y_1,\dots,y_n,y_1',\dots)\,\mathrm{d}x$	One Euler–Lagrange equation per component $y_i$ — as for coordinates $q_i(t)$ .

Common trapThe function space is not part of the answer — it is part of the question

The single most common error is to treat "minimise $J[y]$ " as if the integrand alone defined the problem. It does not. The same density $F$ has different extrema — or none — depending on the smoothness demanded, the boundary data imposed, and above all the norm chosen. A curve that minimises in the weak ( $C^0$ ) sense may be a saddle in the strong ( $C^1$ ) sense, because the strong topology admits nearby competitors (wild-slope wiggles like $\tfrac1n\sin n^2x$ ) that the weak topology cannot see. A "solution" that lies outside your admissible class, or is optimal only in a topology you did not intend, is not a solution to your problem. Always state the space before you vary.

Worked examples

1 · Arc length picks the straight line (fully numerical)

Between the points $(0,0)$ and $(1,1)$ , evaluate the arc-length functional $J[y]=\int_0^1\sqrt{1+y'^2}\,\mathrm{d}x$ on the straight line $y=x$ and on the parabola $y=x^2$ (both admissible — both hit the endpoints), and confirm the straight line is shorter, as Derivation 3 predicts.

Step 1. Straight line $y=x$ , so $y'=1$ and the integrand is the constant $\sqrt{1+1}=\sqrt2$ . Hence $J_{\text{line}}=\int_0^1\sqrt2\,\mathrm{d}x=\sqrt2=\mathbf{1.41421}$ , the ordinary Euclidean distance between the endpoints.

Step 2. Parabola $y=x^2$ , so $y'=2x$ and $J_{\text{par}}=\int_0^1\sqrt{1+4x^2}\,\mathrm{d}x$ . The antiderivative is $\tfrac{x}{2}\sqrt{1+4x^2}+\tfrac14\operatorname{arcsinh}(2x)$ ; evaluating at $x=1$ gives $\tfrac12\sqrt5+\tfrac14\operatorname{arcsinh}(2)=1.118034+0.360909=\mathbf{1.47894}$ .

Step 3. Compare: the parabola is longer by $1.47894-1.41421=0.06473$ , about $\mathbf{4.6\%}$ . This matches Derivation 3 — the straight line, the unique curve with $y'=$ const through these endpoints, is the minimising path, and every admissible competitor is strictly longer. The straight line wins: J = √2 ≈ 1.41421 versus 1.47894 for the parabola.

2 · Stationarity of the Dirichlet energy (first variation with numbers)

Take the Dirichlet energy $J[y]=\tfrac12\int_0^1 y'^2\,\mathrm{d}x$ on $\mathcal{A}=\{y\in C^1[0,1]:y(0)=0,\,y(1)=1\}$ . Show that $y(x)=x$ is stationary by computing the first variation in the direction $\eta(x)=\sin\pi x$ , and confirm it is a minimum from the second-order term.

Step 1. The base value: with $y=x$ , $y'=1$ , so $J[x]=\tfrac12\int_0^1 1\,\mathrm{d}x=\mathbf{0.5}$ . The variation $\eta=\sin\pi x$ is admissible because $\eta(0)=\eta(1)=0$ , so $y+\epsilon\eta$ stays in $\mathcal{A}$ .

Step 2. First variation from equation (5): $\delta J=\int_0^1 y'\,\eta'\,\mathrm{d}x=\int_0^1 (1)\,(\pi\cos\pi x)\,\mathrm{d}x=\big[\sin\pi x\big]_0^1=\sin\pi-\sin 0=\mathbf{0}$ . The directional derivative vanishes, so $y=x$ is stationary in this direction — and since $F_y=0,\,F_{y'}=y'$ , the Euler–Lagrange equation $y''=0$ confirms $y=x$ is stationary in every direction.

Step 3. Second-order term: expand fully, $J[x+\epsilon\eta]=\tfrac12\int_0^1(1+\epsilon\pi\cos\pi x)^2\,\mathrm{d}x=\tfrac12\big[1+0+\epsilon^2\pi^2\!\int_0^1\cos^2\pi x\,\mathrm{d}x\big]=0.5+\tfrac{\pi^2}{4}\epsilon^2$ . The coefficient $\pi^2/4\approx\mathbf{2.467}$ is positive, so $J$ curves upward in this direction — a genuine local minimum, not a saddle. δJ = 0 (stationary) and the second variation ½∫η′² = π²/4 · 2 > 0 confirms J[x] = 0.5 is a minimum.

Physics of the portrait

Behind the word functional stands the man who first made "a number that depends on an entire function" a legitimate mathematical object, and in doing so opened the door to function spaces and functional analysis.

Origins · 1 minVito Volterra and how this was discovered

Vito Volterra

1860–1940 · founder of the theory of functionals and functional analysis

Vito Volterra, born in Ancona in 1860 into a poor Jewish family and a mathematical prodigy by his teens, gave the calculus of variations the object it had been using without naming. In a series of papers from 1887 he introduced what he called funzioni che dipendono da altre funzioni — functions of lines — quantities assigning a number to an entire curve or field. This was the birth of the functional as an object of study in its own right, not merely the integrand of a variational problem; Jacques Hadamard later coined the French term fonctionnelle for exactly Volterra's construction. Volterra pushed the analogy relentlessly: if a functional is a function of infinitely many variables (the values $y(x)$ at every point), then it should have derivatives, differentials, and a Taylor expansion — and he built precisely that calculus, the direct ancestor of the functional derivative $\delta J/\delta y$ . His ideas fed the twentieth-century abstraction of function spaces — Fréchet's metric spaces, Riesz's and Banach's normed spaces, and ultimately the Hilbert and Sobolev spaces that host modern variational analysis. A committed anti-fascist, Volterra refused Mussolini's loyalty oath in 1931 and was stripped of his chair; he died in Rome in 1940, but the object he named is now everywhere, from Lagrangian mechanics to quantum field theory.

History

1696

— Johann Bernoulli poses the brachistochrone, the first genuine variational problem

1744

— Euler's Methodus inveniendi systematises the search for extremal curves

1755

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

1887

— Volterra introduces functions of lines, the functional as an object of study

1906

— Fréchet's thesis abstracts metric and function spaces

1932

— Banach's Théorie des opérations linéaires codifies normed function spaces

Going further

Texts. Stone & Goldbart, Mathematics for Physics, Ch. 1 (Calculus of Variations) · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., Ch. 22 · Hassani, Mathematical Physics, Ch. 33 (Calculus of Variations) · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., Ch. 22 · Gelfand & Fomin, Calculus of Variations, Ch. 1 · Courant & Hilbert, Methods of Mathematical Physics, Vol. 1, Ch. IV.

Provenance. The definition and integral form follow Stone & Goldbart Ch. 1, Arfken–Weber–Harris Ch. 22, and Hassani Ch. 33; the weak/strong topology discussion follows Gelfand & Fomin Ch. 1. A verifier should recompute: (1) the first variation $\delta J=\int(F_y-\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y'})\eta\,\mathrm{d}x$ and the fundamental-lemma step; (2) the discretisation limit reproducing the Euler–Lagrange operator; (3) the arc-length variation giving $y'=$ const; (4) the numbers $\sqrt2=1.41421$ and $\int_0^1\sqrt{1+4x^2}\,\mathrm{d}x=1.47894$ (≈4.6% longer); (5) the Dirichlet stationarity $\delta J=0$ with second-order coefficient $\pi^2/4$ ; (6) the $C^0/C^1$ witness $y_n=\tfrac1n\sin(n^2x)$ with sup $1/n$ and slope sup $n$ . All results are standard textbook identities; no empirical constants are quoted.

Prerequisites. Multivariable calculus and partial derivatives · Riemann integration · Vector spaces and norms. Leads to. First Variation and Stationarity · The Euler–Lagrange Equation · Natural and Essential Boundary Conditions · Hamilton's Principle of Stationary Action.

Ask the tutor: ask me to vary any functional you like — arc length, the brachistochrone descent time, or a beam's bending energy — and I will walk you from $\delta J$ to its Euler–Lagrange equation step by step.

What is the difference between a function and a functional?

A function takes a number and returns a number: feed $f$ the input $3$ and it hands back $f(3)$ . A functional takes an entire function and returns a number: feed the arc-length functional the whole curve $y(x)$ and it hands back one length. The square brackets in $J[y]$ are the flag — they announce that the slot expects a function, not a point. The distinction matters because the domain of a functional is an infinite-dimensional space of functions, so its calculus — the calculus of variations — is genuinely richer than ordinary calculus, even though it is built from the same chain rule.

Why do we bother with the function space and its norm — isn't the integrand enough?

No, and this is the deepest point of the subject. The same integrand $F$ can have a minimiser, a maximiser, a saddle, or nothing at all, depending on which functions you allow and how you measure closeness. Change the smoothness (does $y''$ exist?), the boundary data (which endpoints are pinned?), or the norm ( $C^0$ versus $C^1$ ), and the answer changes — a curve minimising length in the weak sense can fail to in the strong sense, because the strong topology admits nearby wild-slope competitors. So the function space is not scaffolding around the problem — it is half the problem.

What exactly is the first variation, and how is it like a derivative?

The first variation $\delta J[y;\eta]$ is the directional derivative of $J$ at the "point" $y$ in the "direction" $\eta$ : you form the ordinary function $\Phi(\epsilon)=J[y+\epsilon\eta]$ and take $\Phi'(0)$ . It answers "how fast does the number $J$ change as I nudge the whole curve $y$ along the shape $\eta$ ?" Stationarity means every such directional derivative vanishes — $\delta J[y;\eta]=0$ for all admissible $\eta$ — which is the infinite-dimensional analogue of setting all partial derivatives of an ordinary function to zero. Integrating by parts collapses that infinite family of conditions into a single equation, the Euler–Lagrange equation.

What do "essential" and "natural" boundary conditions mean?

Essential boundary conditions are the ones you build into the admissible class $\mathcal{A}$ by hand — for example $y(a)=y_a$ — and every admissible curve, and every variation, must respect them (which is why $\eta$ vanishes there). Natural boundary conditions are the ones you do not impose; instead they fall out of the stationarity calculation as the price of leaving an endpoint free. When you integrate by parts, the boundary term $[F_{y'}\eta]_a^b$ survives at a free end, and forcing it to vanish for arbitrary $\eta(a)$ gives a condition like $F_{y'}=0$ there. Which conditions are essential and which are natural is decided entirely by how you defined the function space.

Is every functional an integral?

No — that is a useful misconception to shed early. Point evaluation $J[y]=y(c)$ assigns a real number to each function and is perfectly linear, yet it is not an integral of a smooth density; it is a distribution (a delta functional). Integral functionals dominate physics because most physical quantities accumulate a local density along the curve, and that local structure is exactly what makes the first variation computable by differentiating under the integral sign. But the definition of a functional asks only for a rule mapping functions to numbers, and non-integral functionals — point evaluations, suprema, boundary values — are genuine and occasionally important.

How does minimising over a function space connect to real computation?

Directly, through discretisation. The heuristic in Derivation 2 — sample the curve, turn the functional into an ordinary function of the samples, and minimise — is precisely the finite element method, the workhorse of engineering simulation. In practice you minimise the energy functional over a finite-dimensional subspace of simple shape functions (piecewise polynomials), converting a variational problem into an algebraic system a computer can solve; density functional theory does the same with the electron-density functional $E[n]$ . Convergence as the grid refines is guaranteed by working in a complete function space (a Sobolev space), which is why completeness, not just the integrand, is part of the modern theory.