Functional Derivatives and Field Variations

Every calculation in the last two articles secretly used a calculus that was never formally introduced. "Vary the action, integrate by parts, read off the coefficient of $\delta\phi$ " — that three-step incantation is really the computation of a derivative: the derivative of a number-valued map (the action) with respect to a function-valued argument (the field), evaluated axis by axis in the infinite-dimensional configuration space of the discrete-to-fields article. This article introduces that calculus properly. It is called the functional derivative, written $\delta F/\delta\phi(x)$ , and it is the exact continuum limit of the partial derivative $\partial F/\partial q_i$ — with the bead index $i$ replaced, as always in this chapter, by the spatial label $x$ . The payoff for formalizing what was already being done is threefold. Compactness: the entire content of the field equations becomes one line, $\delta S/\delta\phi(x) = 0$ — "the gradient of the action vanishes" — with every integration-by-parts subtlety absorbed into systematic rules (a product rule, a chain rule, and one fundamental identity, $\delta\phi(y)/\delta\phi(x) = \delta(x - y)$ , which is nothing but the continuum Kronecker delta). Reach: the calculus applies to any functional, not just actions — energies, charges, entropies, constraints — and differentiating them is how one asks any "how does this quantity respond if the field changes here?" question. Depth: the second functional derivative opens the door the first one cannot: it is the stability operator of a solution, its eigenvalues are the squared frequencies of normal modes, and its inverse is the Green function — the response of the field to a poke, and the classical object that quantum field theory will rename the propagator. The Hamiltonian formulation next door is written entirely in this language — Hamilton's field equations are functional gradients, Poisson brackets are built from $\delta/\delta\phi$ and $\delta/\delta\pi$ — so this article is the last piece of pure infrastructure before the chapter's machinery is complete.

Key takeaways

A functional $F[\phi]$ maps a whole function to a number; its functional derivative is defined by the first-order response $\delta F = \int \mathrm{d}^4x\, \frac{\delta F}{\delta\phi(x)}\,\delta\phi(x)$ — the continuum gradient, one component per point.
It is the honest $a \to 0$ limit of partial derivatives on the bead lattice: $\frac{\delta F}{\delta\phi(x_i)} = \lim \frac{1}{a^d}\frac{\partial F}{\partial q_i}$ — the measure weight is what distinguishes it from a plain partial.
Fundamental identity: $\frac{\delta\phi(y)}{\delta\phi(x)} = \delta(x - y)$ — the Kronecker delta $\partial q_j/\partial q_i = \delta_{ij}$ gone continuous; product and chain rules follow the familiar patterns with integrals over shared arguments.
For a local action, $\frac{\delta S}{\delta\phi(x)} = \frac{\partial\mathcal{L}}{\partial\phi} - \partial_\mu\frac{\partial\mathcal{L}}{\partial(\partial_\mu\phi)}$ : the Euler-Lagrange equations are exactly $\delta S/\delta\phi(x) = 0$ — stationarity as a vanishing gradient.
Klein-Gordon showcase: $\delta S/\delta\phi(x) = -(\Box + m^2)\,\phi(x)$ — the field equation read as the action's gradient.
The second functional derivative $\frac{\delta^2 S}{\delta\phi(x)\delta\phi(y)}$ is the stability/fluctuation operator: eigenvalues = normal-mode frequencies squared; its inverse is the Green function, the field's response to a localized source.
The canonical momentum is a functional derivative too: $\pi(x) = \delta L/\delta\dot\phi(x)$ — the definition the Hamiltonian article runs on.
Key idea: the functional derivative is the partial derivative of mechanics carried to infinitely many coordinates — one per point of space — turning field equations into vanishing gradients, stability into second derivatives, and response into inverse operators.

The core insight

The functional derivative $\delta F/\delta\phi(x)$ — defined by $\delta F = \int \frac{\delta F}{\delta\phi(x)}\delta\phi(x)\,\mathrm{d}^4x$ and anchored by $\delta\phi(y)/\delta\phi(x) = \delta(x-y)$ — is the gradient of a functional in the field's infinite-dimensional configuration space; field equations are its zeros and Green functions the inverse of its second derivative.

Where this lives today

Functional derivatives are the working syntax of more fields than any other single piece of this chapter. Quantum field theory is written in them: the generating functional $Z[J]$ yields every correlation function by repeated $\delta/\delta J(x)$ , the effective action's second functional derivative is the full propagator's inverse, and the Schwinger–Dyson equations are just " $\delta S/\delta\phi = 0$ " promoted to operator statements — the classical calculus of this article with $\hbar$ turned on. Density functional theory, quantum chemistry's workhorse (and the most-cited framework in physical science), is named for exactly this operation: the ground-state energy is a functional $E[n]$ of the electron density, and the Kohn–Sham equations are $\delta E/\delta n(x) = \mu$ — a stationarity condition solved millions of times daily in materials and drug design. Machine learning's continuous limits speak it too: training a neural network by gradient descent on a loss functional of the network function is $\delta \mathcal{L}/\delta f(x)$ in disguise, and the "neural tangent kernel" describing wide networks is a second functional derivative — a Green function of training dynamics. Optimal control and variational data assimilation (the algorithms steering rockets and weather forecasts) compute $\delta(\text{cost})/\delta(\text{trajectory})$ by adjoint methods — the integration-by-parts of Derivation A, industrialized. Wherever a number depends on a function and someone wants the best function, Volterra's derivative is the tool doing the wanting. Within the chapter: the calculus is complete, and the Hamiltonian article now uses it verbatim — $\dot\phi = \delta H/\delta\pi$ , $\dot\pi = -\delta H/\delta\phi$ , brackets from paired functional gradients — to give the field its canonical mechanics and open the door the capstone walks through.

The physical picture

Return to the landscape image from the covariant-equations article, and now put coordinates on it. For $N$ beads, a functional of the configuration — say the total potential energy — is an ordinary function of $N$ variables, and its gradient has $N$ components $\partial F/\partial q_i$ : "how much does $F$ respond if I nudge bead $i$ alone?" As the beads densify, the questions densify with them: "how does $F$ respond if I nudge the field here?" — one question per point of space. The functional derivative $\delta F/\delta\phi(x)$ is the organized answer sheet: a function of $x$ , each value one component of an infinite-dimensional gradient. When that gradient vanishes identically, the configuration is a stationary point of the landscape — which is precisely what a solution of the field equations is.

Figure

The continuum gradient, computed — lattice partial derivatives densifying onto the functional-derivative kernel

One subtlety separates the continuum gradient from a naive limit of partials, and it is the same subtlety that separated the Lagrangian density from the Lagrangian. Nudging a single bead changes $F$ by an amount proportional to the bead's share of the system — proportional to the lattice cell $a$ — so the bare partial $\partial F/\partial q_i$ dies as the lattice refines: individual points carry no weight in a continuum. The functional derivative divides that vanishing response by the vanishing cell volume, producing a finite density of response — response per unit volume of nudge. This is why the definition pairs $\delta F/\delta\phi(x)$ with an integral: only integrated against a finite-volume variation $\delta\phi(x)$ does it yield a finite change. The operational picture to keep: press a small bump of unit area into the field at $x$ and read how much $F$ moves — that reading is $\delta F/\delta\phi(x)$ , and the computed figure below performs exactly this experiment. The second derivative has an equally concrete reading. Sit at a stationary configuration (gradient zero) and ask about the landscape's curvature: nudge at $x$ , and watch how the gradient at $y$ responds. That matrix-with-continuous-indices, $\delta^2 S/\delta\phi(x)\delta\phi(y)$ , curves upward in stable directions and downward in unstable ones; its eigenvectors are the normal modes of oscillation about the solution, its eigenvalues their squared frequencies — the bead-chain diagonalization of the oscillations chapter, reborn in function space. And because responding to a poke is solving a linear system, the inverse of this curvature operator is the response function: poke the field with a source at $y$ and the displacement at $x$ is the inverse's $(x, y)$ entry — the Green function, the single most reused object in field theory's later life.

The calculus, built from the lattice

The definition packages the response-density idea; everything else follows from it mechanically. For a functional $F[\phi]$ — any rule turning field configurations into numbers — the first-order change under an arbitrary small variation defines the derivative as the integral kernel of the response:

\delta F \;=\; F[\phi + \delta\phi] - F[\phi] \;=\; \int \mathrm{d}^4x\; \frac{\delta F}{\delta\phi(x)}\,\delta\phi(x) \;+\; O(\delta\phi^2) \qquad (1)

where: $\delta\phi(x)$ — an arbitrary small variation (smooth, vanishing wherever boundary bookkeeping requires); $\delta F/\delta\phi(x)$ — the functional derivative, a function of $x$ determined uniquely by demanding (1) hold for every $\delta\phi$ (the fundamental lemma of the calculus of variations, imported wholesale from the Lagrangian chapter). Dimensional note, often the fastest error-catch: the derivative carries the dimensions of $F$ per (field × volume) — the "per volume" is the density-of-response fingerprint.

Derivation AFrom partial derivatives to the functional derivative — the lattice construction

We build $\delta F/\delta\phi(x)$ as the honest continuum limit of $\partial F/\partial q_i$ on the bead lattice, deriving the measure weight, the definition (1), and the fundamental identity in one pass. The construction settles once and for all what kind of object the functional derivative is — a density dual to variations, not a plain rate — and why the continuum Kronecker delta is a Dirac delta. Everything in this chapter's variational practice descends from this five-step limit.

Step 1. Discretize: field values on a lattice of cell volume $a^d$ , $q_i = \phi(x_i)$ . A functional becomes a function of many variables, $F[\phi] \to F(q_1, \dots, q_N)$ , and its total differential is elementary calculus:

\delta F = \sum_i \frac{\partial F}{\partial q_i}\,\delta q_i

Step 2. Force the sum into the chapter's canonical form — a Riemann sum with the cell volume made explicit:

\delta F = \sum_i a^d \left[\frac{1}{a^d}\frac{\partial F}{\partial q_i}\right] \delta q_i \;\xrightarrow[a \to 0]{}\; \int \mathrm{d}^dx\, \left[\lim_{a\to0}\frac{1}{a^d}\frac{\partial F}{\partial q_i}\right] \delta\phi(x)

Comparing with (1) identifies the bracket:

\frac{\delta F}{\delta\phi(x_i)} = \lim_{a \to 0}\; \frac{1}{a^d}\,\frac{\partial F}{\partial q_i} \qquad (2)

Why the $1/a^d$ weight is unavoidable: nudging one bead moves $F$ by that bead's share $\sim a^d$ of the whole; the bare partial vanishes in the limit, and only response-per-cell-volume survives finite. This is the same bookkeeping that made $\mathcal{L}$ (a density) fundamental rather than $L$ .

Step 3. Fundamental identity. Take $F = \phi(y_j)$ , the evaluation functional. On the lattice $\partial q_j/\partial q_i = \delta_{ij}$ , so by (2):

\frac{\delta\phi(y)}{\delta\phi(x)} = \lim_{a\to0}\frac{\delta_{ij}}{a^d} = \delta^{(d)}(x - y) \qquad (3)

where: the Kronecker-over-volume limit is exactly the defining sequence of the Dirac delta: zero for $i \neq j$ , total weight $\sum_i a^d(\delta_{ij}/a^d) = 1$ . The continuum's "independent coordinates" are independent in the delta-function sense — the identity used silently every time a variation was "read off."

Step 4. Rules by inheritance. Because (2) is a limit of partials, the algebra of partials survives verbatim: linearity; the product rule $\frac{\delta(FG)}{\delta\phi(x)} = \frac{\delta F}{\delta\phi(x)}G + F\frac{\delta G}{\delta\phi(x)}$ ; and the chain rule through an intermediate function $g(y)$ picks up an integral over the shared continuous index — $\frac{\delta F[g[\phi]]}{\delta\phi(x)} = \int\mathrm{d}^dy\, \frac{\delta F}{\delta g(y)}\frac{\delta g(y)}{\delta\phi(x)}$ — sums over repeated indices becoming integrals, as always.

Step 5. First harvest: for a local action $S = \int\mathcal{L}(\phi, \partial_\mu\phi)\,\mathrm{d}^4x$ , applying the rules (differentiate under the integral with (3); the gradient term's chain rule produces $\partial_\mu\delta(x-y)$ , which integrates by parts onto the coefficient) delivers

\frac{\delta S}{\delta\phi(x)} = \frac{\partial\mathcal{L}}{\partial\phi}(x) - \partial_\mu\!\left(\frac{\partial\mathcal{L}}{\partial(\partial_\mu\phi)}\right)\!(x) \qquad (4)

— the Euler-Lagrange expression, now recognized as the gradient component at $x$ , with the field equations the statement $\delta S/\delta\phi(x) = 0$ : stationary points of the landscape, exactly as pictured. Verified: construction per Goldstein ch. 13-context and Gelfand–Fomin; the delta-sequence in Step 3 is the standard box-function limit; (4) reproduces the covariant-equations article's crank, as it must.

The showcase computation of the calculus — worth doing once in full, since every field theory text does it in shorthand forever after:

Go deeper · Derivation B · 1 minThe Klein-Gordon action's gradient

Derivation BThe Klein-Gordon action's gradient

We compute $\delta S/\delta\phi(x)$ for the free scalar action directly from the definition, without invoking the Euler-Lagrange formula — the point being to watch the calculus operate: expand, integrate by parts, strip the arbitrary variation. The result, $-(\Box + m^2)\phi$ , identifies the field equation as the literal gradient of the action, and its linearity in $\phi$ sets up the second-derivative story that follows.

Step 1. Expand the action at $\phi + \delta\phi$ and keep first order:

\delta S = \int \mathrm{d}^4x \left[ \partial^\mu\phi\;\partial_\mu(\delta\phi) - m^2\phi\,\delta\phi \right]

(The kinetic term's two factors each donate one variation; symmetry gives the factor 2 cancelling the $\tfrac12$ .)

Step 2. Integrate the kinetic piece by parts, discarding the boundary term (variations vanish there — or fields die at infinity):

\int \mathrm{d}^4x\; \partial^\mu\phi\,\partial_\mu(\delta\phi) = -\int \mathrm{d}^4x\; (\Box\phi)\,\delta\phi

Step 3. Collect into the defining form (1):

\delta S = \int \mathrm{d}^4x\; \big[-\Box\phi - m^2\phi\big]\,\delta\phi \quad\Longrightarrow\quad \frac{\delta S}{\delta\phi(x)} = -\big(\Box + m^2\big)\phi(x) \qquad (5)

where: the operator $-(\Box + m^2)$ — the Klein-Gordon operator with a sign — acts on the configuration to give the gradient. Setting it to zero recovers the field equation; more interestingly, not setting it to zero is meaningful too: for the source-coupled action $S_J = S + \int J\phi$ , the gradient is $-(\Box + m^2)\phi + J$ , and stationarity gives the sourced equation $(\Box + m^2)\phi = J$ — the Yukawa problem of the covariant-equations article, now read as "gradient = 0 shifted by a tilt $J$ of the landscape." Verified: matches (4) applied to the KG density, and the sourced form reproduces the worked Yukawa example of the covariant-equations article.

Linearity of the gradient (5) in $\phi$ means the landscape is exactly quadratic — a paraboloid in function space — and its curvature is configuration-independent. That curvature is the calculus's second story:

Go deeper · Derivation C · 2 minThe second derivative — stability operator, normal modes, Green function

Derivation CThe second derivative — stability operator, normal modes, Green function

We differentiate the gradient once more to obtain the second functional derivative, interpret it as the curvature of the action landscape, and extract its two working identities: eigenvalues = squared mode frequencies (stability), and inverse = Green function (response). This is where the functional calculus stops re-deriving known equations and starts producing the objects the rest of physics is built from; the free scalar makes everything explicit.

Step 1. Differentiate (5) with respect to $\phi(y)$ , using the fundamental identity (3):

\frac{\delta^2 S}{\delta\phi(y)\,\delta\phi(x)} = -\big(\Box_x + m^2\big)\,\delta^{(4)}(x - y)

where: the object is a "matrix" with two continuous indices, diagonal (delta-supported) because the action is local — curvature couples each point only to its infinitesimal neighbours through $\Box$ . For a general action the same computation gives $-\big(\Box + V''(\phi(x))\big)\delta(x-y)$ : curvature now depends on the configuration — the landscape is no longer a paraboloid, and stability becomes solution-by-solution business.

Step 2. Stability reading. Expand the action about a stationary point $\bar\phi$ : the increment is $\Delta S = \tfrac12\iint \eta(x)\, \frac{\delta^2S}{\delta\phi\delta\phi}\, \eta(y) + O(\eta^3)$ — a quadratic form in the fluctuation $\eta$ . Diagonalize: eigenfunctions of $-\Box - V''(\bar\phi)$ (in the static problem: of $-\nabla^2 + V''(\bar\phi)$ ) are the normal modes about $\bar\phi$ ; eigenvalues are $\omega^2$ of small oscillations. All positive: $\bar\phi$ is a stable minimum. A negative eigenvalue: an unstable direction — the field rolls away along that mode. For $\bar\phi = 0$ in the free theory, the modes are plane waves and the eigenvalues $k^2 + m^2 > 0$ : the vacuum is stable, and the mode frequencies are exactly the dispersion of the density article — same physics, third derivation route. Why this is the bead-chain story verbatim: for $N$ beads, stability of an equilibrium is the positivity of the Hessian $\partial^2V/\partial q_i\partial q_j$ , and normal modes diagonalize it. The second functional derivative IS that Hessian with $i, j \to x, y$ — the chapter's dictionary applied to second derivatives.

Step 3. Response reading. Tilt the landscape with a source and re-solve: $(\Box + m^2)\phi = J$ . Linearity means solve-once-superpose-forever: define the Green function as the inverse of the curvature kernel,

\big(\Box_x + m^2\big)\,G(x, y) = \delta^{(4)}(x - y), \qquad \phi(x) = \int \mathrm{d}^4y\; G(x, y)\,J(y) \qquad (6)

where: $G(x,y)$ — the field at $x$ produced by a unit point-poke at $y$ : the landscape's compliance matrix. For the static problem in three dimensions $G$ is the Yukawa profile $e^{-m r}/4\pi r$ already met; in one dimension, $G = e^{-m|x - y|}/2m$ (checked in the worked example). Everything about the free field's interaction with the world — how sources dress themselves, how influences propagate, what "force between two sources" means — is stored in $G$ ; and when the quantum capstone arrives, this same inverse-curvature object, with the right boundary conditions in time, is the propagator. Verified: second derivative by direct differentiation with (3); stability criterion per Gelfand–Fomin's second-variation theory; 1D Green function checked by substitution in the worked example below.

Figure

Derivative and inverse curvature, computed — the bump test and the Green-function response of a massive field

One more functional derivative deserves its formal introduction here, because the next article is built on it. Freeze time and regard the Lagrangian $L = \int\mathcal{L}\,\mathrm{d}^3x$ as a functional of the instantaneous field and its velocity; the canonical momentum density is its gradient with respect to the velocity:

\pi(x) \;\equiv\; \frac{\delta L}{\delta\dot\phi(x)} \;=\; \frac{\partial\mathcal{L}}{\partial\dot\phi}(x) \qquad (7)

where: the middle expression is the definition (spatial functional derivative at fixed time — the $d = 3$ version of the calculus); the right equality holds for local densities and is the working formula. For the Klein-Gordon field, $\pi = \dot\phi$ — each point-oscillator's momentum, exactly as the bead picture demands. The Hamiltonian article takes this object and completes classical mechanics' program for fields: Legendre transform, Hamilton's equations as paired functional gradients, Poisson brackets from $\delta/\delta\phi$ and $\delta/\delta\pi$ .

Limiting cases

Case	Result	Meaning
$F$ local, no derivatives: $\int f(\phi)\,\mathrm{d}^4x$	$\delta F/\delta\phi(x) = f'(\phi(x))$	differentiation under the integral — the easy case
$F$ with gradients	extra $-\partial_\mu(\partial f/\partial(\partial_\mu\phi))$	integration by parts, mechanized
evaluation functional $\phi(y)$	$\delta(x - y)$	the continuum Kronecker delta
$N$ beads, finite $a$	$\delta F/\delta\phi \approx a^{-d}\,\partial F/\partial q_i$	the lattice weight; numerics uses exactly this
quadratic action	curvature independent of $\phi$ ; exact Green-function solution	free fields are paraboloids in function space
$V''(\bar\phi) < 0$ somewhere	negative curvature directions	instability; $\bar\phi$ a saddle of the landscape
higher-derivative $\mathcal{L}$	$\delta S/\delta\phi$ gains $+\partial_\mu\partial_\nu(\cdots)$ terms, signs $(-1)^n$	Ostrogradsky pattern — see covariant-equations article's caution

Common trapA functional derivative is a density — five ways the bookkeeping bites

The calculus is mechanical, but its density character generates a standard list of stumbles. (1) Dimensions: $\delta F/\delta\phi(x)$ has the dimensions of $F$ /(field × volume) — if your "functional derivative" has the dimensions of $F$ /field, you computed a partial derivative and forgot the measure (the lattice $1/a^d$ ). (2) Deltas are normal: results proportional to $\delta(x-y)$ or its derivatives are not errors but the signature of locality; only integrated quantities need to be finite. (3) Boundary honesty: formula (4) assumed the surface term died; on finite regions the derivative acquires boundary-supported pieces, and "natural boundary conditions" are exactly the demand that those vanish (spelled out in the covariant-equations article's FAQ). (4) Fixed-time versus spacetime: $\delta L/\delta\dot\phi(x)$ (three-dimensional, fixed $t$ ) and $\delta S/\delta\phi(x)$ (four-dimensional) are different derivatives with different deltas — mixing them is the classic Hamiltonian-formalism error, and the reason texts fuss about $\delta^{(3)}$ versus $\delta^{(4)}$ . (5) The variation must be arbitrary: reading off the kernel in (1) is only legitimate if $\delta\phi$ ranges over all functions; constrained variations (fixed norm, gauge conditions) leave the derivative determined only up to the constraint's annihilator — Lagrange multipliers return, precisely as in the constrained-mechanics chapters.

Worked examples

1 · The bump test — a functional derivative measured (confidence builder)

Let $F[\phi] = \int_0^1 \phi^2(x)\,\mathrm{d}x$ and $\phi(x) = \sin(\pi x)$ . (i) Compute $F$ and the functional derivative. (ii) Predict the change in $F$ when a narrow bump of area $0.010$ is added at $x_0 = 0.25$ , and compare with a direct evaluation.

Step 1. $F = \int_0^1 \sin^2(\pi x)\,\mathrm{d}x = \mathbf{0.500}$ . The derivative, by the local rule: $\delta F/\delta\phi(x) = 2\phi(x) = 2\sin(\pi x)$ — a function, one gradient component per point.

Step 2. Prediction. At $x_0 = 0.25$ : $2\sin(\pi/4) = 2(0.7071) = \mathbf{1.4142}$ . For a bump $\delta\phi$ concentrated near $x_0$ with $\int\delta\phi = 0.010$ , the definition (1) gives $\delta F \approx 1.4142 \times 0.010 = \mathbf{0.01414}$ .

Step 3. Direct check. Take $\delta\phi$ = a rectangle of width $w = 0.02$ and height $0.5$ (area $0.010$ ) centred at $0.25$ : $\delta F = \int[(\phi + \delta\phi)^2 - \phi^2] = \int[2\phi\,\delta\phi + \delta\phi^2]$ . First piece: $2 \times 0.5 \int_{0.24}^{0.26}\sin(\pi x)\mathrm{d}x = 0.5 \times 2 \times 0.02 \times \sin(\pi \cdot 0.25) \approx 0.01414$ ✓. Second piece: $0.25 \times 0.02 = 0.005 \times ... = 0.0050 \times$ — evaluate: $\int\delta\phi^2 = 0.5^2 \times 0.02 = \mathbf{0.0050}$ — the $O(\delta\phi^2)$ error, 35% here because the bump is tall, shrinking linearly as the bump flattens at fixed area. The experiment displays both the derivative and the meaning of "first order."

2 · Response from the inverse curvature — a 1D Green function with numbers (exam level)

For the static 1D energy functional $E[\phi] = \int \big[\tfrac12(\phi')^2 + \tfrac12 m^2\phi^2 - \rho\phi\big]\mathrm{d}x$ with $m = 2.0$ : (i) find the stationarity condition and the Green function; (ii) compute the field at distance $0.5$ from a point source of strength $q = 0.10$ ; (iii) verify the stability of the solution.

Step 1. Gradient: $\delta E/\delta\phi(x) = -\phi'' + m^2\phi - \rho$ . Stationarity: $-\phi'' + m^2\phi = \rho$ — the static sourced equation. Curvature kernel: $(-\mathrm{d}^2/\mathrm{d}x^2 + m^2)\,\delta(x-y)$ ; its inverse solves $(-G'' + m^2 G) = \delta(x - y)$ , giving $G(x,y) = \frac{1}{2m}e^{-m|x-y|}$ — check: away from $x = y$ , $G'' = m^2G$ ✓; the kink at $x = y$ has jump $G'|^+_- = -1$ ✓ (slope $\mp\tfrac12 e^{-m|x-y|}$ on each side, difference $-1$ at coincidence).

Step 2. Response: $\phi(x) = qG = \frac{0.10}{2(2.0)}\,e^{-2.0 \times 0.5} = 0.025 \times e^{-1} = 0.025 \times 0.3679 = \mathbf{0.0092}$ . Compare the massless limit: $m \to 0$ makes $G$ diverge — in 1D a massless field cannot forget a source at any distance; the mass term's well is what localizes the response within $1/m = 0.5$ .

Step 3. Stability: the curvature operator $-\mathrm{d}^2/\mathrm{d}x^2 + m^2$ has spectrum $k^2 + m^2 \ge m^2 = 4.0 > 0$ — every fluctuation direction curves upward, the solution is a true minimum, and the lowest normal mode about it oscillates at $\omega = m = 2.0$ : the mass gap, appearing for the third time in three articles, now as the floor of a Hessian spectrum.

Origins · 1 minVito Volterra and how this was discovered

Vito Volterra

1860–1940 · Rome

Volterra invented the derivative this article teaches. In 1887, at twenty-seven, he began publishing on "functions that depend on other functions" — his funzioni di linee, functions of lines — defining their variation exactly as equation (1) does and developing the product and chain rules, decades before "functional analysis" had a name (the very word "functional" was coined, by Hadamard, for Volterra's objects). His motivations were physical from the start: elasticity with memory (materials whose stress depends on the whole history of strain — hereditary mechanics, today's viscoelasticity) and the integro-differential equations that bear his name. The Lotka–Volterra predator-prey equations, produced when his zoologist son-in-law brought him Adriatic fishery data, made him a founder of mathematical biology almost as a side effect. Volterra was also, institutionally, Italian science: senator, founder of the national research council — until 1931, when he became one of only twelve Italian professors (of some 1,250) who refused Mussolini's loyalty oath. Stripped of his chair and academies, he spent his last decade working abroad and in internal exile. The calculus he built for stretched membranes and remembering materials became, in other hands, the language of quantum field theory — every $\delta/\delta\phi(x)$ in a physics paper is Volterra's line-function derivative at work.

History

1755

— Lagrange's

\delta

-calculus gives variations an algebra; the derivative behind it stays implicit for a century.

1887

— Volterra defines functions of lines and their derivatives — the functional calculus proper begins.

1903

— Hadamard coins "functional"; Fréchet and others make the infinite-dimensional calculus rigorous.

1927

— Volterra's Theory of Functionals systematizes the field, hereditary elasticity included.

1948

— Feynman's path integral and Schwinger's variational principle put functional derivatives at the center of quantum field theory, where they remain.

Going further

References. Gelfand & Fomin, Calculus of Variations (the standard text for first and second variations) · Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., ch. 13 · Landau & Lifshitz, Mechanics §§1–2 and The Classical Theory of Fields §32 (variational practice) · Volterra, Theory of Functionals and of Integral and Integro-Differential Equations (1930)

Provenance. Equations are standard variational results verifiable in Goldstein Ch. 13, Landau–Lifshitz Vol. 1 §§1–2 and Vol. 2 §32, and Gelfand–Fomin. A verifier should: (1) reproduce the bulk+surface split via integration by parts and the 4D divergence theorem; (2) confirm the Klein–Gordon kernel $\delta S/\delta\phi=-(\Box+m^2)\phi$ ; (3) check the Ostrogradsky sign pattern $(-1)^n$ ; (4) verify $\delta\phi(y)/\delta\phi(x)=\delta^{(4)}(x-y)$ from the delta sifting property. No empirical constants are used; all relations are mathematical identities. Additionally for this upgrade: the lattice construction $\delta F/\delta\phi = \lim a^{-d}\partial F/\partial q_i$ per Gelfand–Fomin's finite-difference motivation; the second-variation/normal-mode correspondence per Gelfand–Fomin ch. 5; the 1D Green function $e^{-m|x-y|}/2m$ checked by jump condition; worked-example numbers recomputed directly ( $F = 0.500$ , kernel value $1.4142$ , predicted $\delta F = 0.01414$ with $O(\delta\phi^2)$ residue $0.0050$ ; response $0.025\,e^{-1} = 0.0092$ ).

Builds on: From Discrete Systems to Continuous Fields · The Lagrangian Density and Field Action · Calculus of Variations Leads to: Covariant Euler-Lagrange Equations for Fields · Hamiltonian Formulation for Fields · Poisson Brackets for Fields · Noether's Theorem for Continuous Symmetries

Is δF/δφ(x) a genuine derivative or just notation for "whatever multiplies δφ after integration by parts"?

Both, and the equivalence is the point. Rigorously, it is the density of the Fréchet (or Gâteaux) derivative: $F$ is differentiable at $\phi$ if a linear map $DF_\phi$ exists with $F[\phi + \eta] = F[\phi] + DF_\phi[\eta] + o(\|\eta\|)$ , and when that linear map is representable as an integral against a function, that function is $\delta F/\delta\phi(x)$ . The practical recipe — expand to first order, integrate by parts until $\delta\phi$ stands bare, read the coefficient — is precisely the computation of this representing function, so the folklore procedure and the rigorous object coincide whenever the procedure terminates. The caveats live where representation fails: functionals like $F = \phi'(0)$ (evaluation of a derivative at a point) have derivatives that are distributions ( $-\delta'(x)$ ) rather than functions — still usable, integrated against smooth variations; and genuinely non-smooth functionals (suprema, obstacle problems) need one-sided or sub-differential notions. For the polynomial-in-field-and-derivatives functionals this chapter uses, none of these pathologies arise, and the working rule is safe: the functional derivative is the ordinary derivative of $F$ along the direction of a delta-bump, normalized per unit bump area — exactly what the computed bump test measures.

Why do physicists happily differentiate WITH RESPECT TO a function evaluated at a point, when a point-supported variation isn't even a nice function?

Because the point-indexed derivative is dual bookkeeping, not a claim that point-bumps are admissible variations. The honest statements always involve smearing: what is defined is the response to smooth variations, $\delta F = \int(\delta F/\delta\phi)\,\delta\phi$ , and " $\delta/\delta\phi(x)$ " extracts the kernel of that pairing — a function (or distribution) of $x$ whose values are only probed through integrals. The situation is identical to the Dirac delta itself: $\delta(x)$ is meaningless pointwise yet indispensable as a kernel, and every formula of this article ( $\delta\phi(y)/\delta\phi(x) = \delta(x-y)$ above all) is an identity between kernels, verified by smearing both sides against test variations. This duality view also explains the calculus's robustness in the presence of the singular results the pitfall box mentioned: second derivatives proportional to $\delta(x-y)$ , currents with delta-supported responses, and the like are all legitimate kernels awaiting their smearing. When the formal manipulations do outrun safety — products of distributions at coincident points, the classic quantum-field-theory divergences — the breakdown is genuinely informative: it is the ultraviolet mode-counting of the discrete-to-fields article resurfacing, and the cure (regularize: put the lattice back, smear the points) is the same dictionary run in reverse.

How do functional derivatives actually get computed in numerics — the delta functions can't go on a computer?

By running Derivation A's limit backwards: discretize, take honest partial derivatives, remember the measure. On a grid with cell volume $h^d$ , the working replacement is $\delta F/\delta\phi(x_i) \approx (1/h^d)\,\partial F/\partial q_i$ — and forgetting the $1/h^d$ is the single most common bug in home-rolled field optimizers (symptom: results that change systematically under grid refinement). Three industrial-strength refinements. (1) Adjoint methods: when $F$ is defined through the solution of a PDE (cost of a flow, misfit of a forecast), differentiating through the solver naively costs one solve per grid point; the adjoint trick — integration by parts at the level of the discrete equations — delivers the entire gradient for the cost of ONE extra (adjoint) solve, which is why weather-model data assimilation and airplane-shape optimization are feasible at all. (2) Automatic differentiation: modern frameworks differentiate the discretized functional exactly (to machine precision) by chain-ruling the code itself; the functional-calculus structure survives as the continuum limit the results must converge to. (3) Mode-space evaluation: for quadratic functionals, differentiate in the eigenbasis of the curvature operator (FFT for translation-invariant kernels) — the Green-function solve of the worked example becomes a division per mode. In all three, the continuum calculus is the specification and the sanity check: discrete gradients must converge, under refinement, to the smooth kernel this article computes analytically.

What exactly does the second functional derivative's POSITIVITY buy — is "stability" here the same word as in the fixed-point chapters?

Same concept, upgraded arena, with one physically crucial refinement. In the fixed-point chapters, stability of an equilibrium was decided by the Hessian of the potential (or the Jacobian's spectrum); here the equilibrium is a field configuration and the Hessian is $\delta^2E/\delta\phi\delta\phi = -\nabla^2 + V''(\bar\phi)$ — an operator whose spectrum plays the eigenvalue role. Positivity of the full spectrum ⟹ every small disturbance oscillates rather than grows ⟹ the configuration is dynamically stable, with the eigenfunctions as its normal modes: the bead-chain story with matrices promoted to operators. The refinements: (i) continuous spectrum — an infinite system's Hessian has bands, not just discrete levels, and "positivity" must hold down to the band edge (the free vacuum's edge is $m^2$ — the gap); (ii) zero modes — symmetries produce exact zero eigenvalues that threaten nothing (translating a soliton costs no energy: the kink's zero mode is its own gradient), and stability analysis must quotient them out before judging; (iii) negative modes as physics — a single negative eigenvalue does not always mean "wrong answer": for tunneling and nucleation problems the relevant saddle configuration (the "bounce") is required to have exactly one negative mode, and counting them is part of the calculation. So: same word, same quadratic-form test, but spectra replace eigenvalue lists, and the exceptional eigenvalues carry named physics.

The chain rule sprouted an integral — can I see a concrete case where that integral does real work?

The cleanest physical case: differentiating through a solved equation — "how does an observable depend on a parameter function?" Suppose the field $\phi$ solves $(\!-\nabla^2 + m^2(x))\phi = \rho$ with a position-dependent mass (an inhomogeneous medium), and you want the sensitivity of a measurement $F[\phi] = \int w\,\phi$ to the medium profile $m^2(y)$ — the setup of every imaging and inverse problem. The chain rule reads $\frac{\delta F}{\delta m^2(y)} = \int\mathrm{d}x\, \frac{\delta F}{\delta\phi(x)}\,\frac{\delta\phi(x)}{\delta m^2(y)}$ , and the inner factor comes from differentiating the equation itself: perturbing $m^2$ by a bump at $y$ perturbs the solution by $\delta\phi(x) = -G(x, y)\,\phi(y)\,\delta m^2(y)$ (the Green function propagating the local disturbance to the observation point — differentiate the equation, solve for $\delta\phi$ with the same operator). The chain-rule integral then assembles the sensitivity map $\frac{\delta F}{\delta m^2(y)} = -\phi(y)\int w(x)G(x,y)\mathrm{d}x$ — incident field times back-propagated detector sensitivity, evaluated at each candidate scatterer location $y$ . This single formula, in its various dresses, is: the Born approximation of scattering theory, the sensitivity kernel of seismic tomography ("banana-doughnut" kernels mapping which Earth regions a travel time feels), and the backpropagation step of adjoint optimization. The integral over the shared argument is doing exactly what the discrete chain rule's sum over intermediate indices always did — routing influence through every path from perturbation to observable.

What changes in this calculus when I move to the Hamiltonian picture next door?

Mechanically, only the arena: instead of one field history $\phi(x, t)$ over spacetime, the state is a pair of fields $(\phi(\mathbf{x}), \pi(\mathbf{x}))$ on a time slice, and the calculus runs with three-dimensional deltas at fixed time — $\delta\phi(\mathbf{y})/\delta\phi(\mathbf{x}) = \delta^{(3)}(\mathbf{x} - \mathbf{y})$ , with $\phi$ and $\pi$ mutually independent slots ( $\delta\pi/\delta\phi = 0$ ). Functionals of interest become phase-space functionals $F[\phi, \pi]$ — the Hamiltonian, momenta, charges — each with two gradients. The payoff structures are then direct transcriptions of finite mechanics: Hamilton's equations pair the gradients ( $\dot\phi = \delta H/\delta\pi$ , $\dot\pi = -\delta H/\delta\phi$ ), and the Poisson bracket is the antisymmetric pairing $\{F, G\} = \int\mathrm{d}^3x\,\big[\frac{\delta F}{\delta\phi}\frac{\delta G}{\delta\pi} - \frac{\delta F}{\delta\pi}\frac{\delta G}{\delta\phi}\big]$ — sums over $i$ become integrals over $\mathbf{x}$ , as the dictionary demands. One genuinely new subtlety appears for gauge fields: constraints among the $(\phi, \pi)$ mean not all functional gradients are independent, and the bracket must be upgraded (Dirac brackets) — the field-theory version of the constrained-variation caveat this article has already flagged twice. That, plus the mode decomposition turning each $\mathbf{k}$ into one bead-like oscillator, is the entire content of the next article's formalism.