Natural and Essential Boundary Conditions

A variational problem is only half-specified by its functional. Writing $J[y]=\int_a^b F(x,y,y')\,\mathrm{d}x$ tells you what to extremize, but not what the competing curves may do at the two ends — and until that is fixed, the problem has no unique answer. There are exactly two kinds of endpoint statement. Essential (geometric, Dirichlet) conditions fix the value of $y$ at an endpoint; they are wired into the admissible space, so every trial function already obeys them and the variation vanishes there. Natural (dynamical, Neumann-type) conditions are the opposite in spirit — you do not impose them, you discover them, because they fall out of the stationarity requirement wherever an endpoint is left free. The remarkable thing is that both come from a single act: the same integration by parts that produces the Euler–Lagrange equation. Moving one derivative off the perturbation splits $\delta J$ into an interior integral — which the Fundamental Lemma turns into the field equation — and a leftover boundary term $[F_{y'}\,\delta y]_a^b$ that must vanish on its own. At a pinned end $\delta y=0$ does this automatically; at a free end $\delta y$ is arbitrary, so its coefficient $F_{y'}$ must vanish instead. That is the natural boundary condition — a force balance the variational principle hands you for free. This article is the accounting rule that keeps every calculus-of-variations problem well posed: one condition per endpoint, essential or natural, never both, never neither.

Key takeaways

Extremizing $J[y]=\int_a^b F(x,y,y')\,\mathrm{d}x$ , one integration by parts in $\delta J$ leaves a bulk term (Euler–Lagrange) and a boundary term $[F_{y'}\,\delta y]_a^b$ that must vanish separately.
An essential (Dirichlet) condition prescribes $y$ at an endpoint; it lives in the function space, forcing $\delta y=0$ there so the boundary term dies automatically.
A natural (Neumann-type) condition is what stationarity forces at a free endpoint: $\delta y$ is arbitrary, so its coefficient must vanish, $F_{y'}=0$ — a result, not an input.
Each endpoint needs exactly one condition. A second-order Euler–Lagrange equation wants two total; a fourth-order (higher-derivative) one wants four, two per end.
A free-ended beam obeys the natural conditions $EI\,y''=0$ (zero moment) and $EI\,y'''=0$ (zero shear); a geodesic meeting a free boundary satisfies transversality and crosses it orthogonally.
Adding an endpoint energy $g(y(b))$ turns the natural condition into the Robin relation $F_{y'}+g'(y)=0$ without touching the bulk equation.
Key idea: Integration by parts in $\delta J$ generates the boundary term $[F_{y'}\,\delta y]_a^b$ . A prescribed endpoint sets $\delta y=0$ (essential condition); a free endpoint forces $F_{y'}=0$ there (natural condition). Each endpoint needs exactly one.

Where this lives today

The essential/natural split is the organizing principle of the finite element method (FEM), the numerical workhorse of structural, thermal, and electromagnetic engineering. In an FEM code, essential (Dirichlet) conditions are imposed on the trial functions by fixing nodal values, while natural (Neumann) conditions — tractions, heat fluxes, surface charges — are never imposed at all: they emerge from the weak form's boundary integral exactly as $F_{y'}$ emerges here, which is precisely why "natural" is the numerical-analysis name for them. The same accounting governs field theory: the Einstein–Hilbert action of general relativity is variationally ill-posed until one adds the Gibbons–Hawking–York surface term, whose job is to fix the boundary term so the metric can be held fixed (essential) at the boundary — a gravitational instance of adding $g(y(b))$ to $J$ . Transversality reappears as the terminal condition of Pontryagin's maximum principle in optimal control; and the free-edge conditions of membranes, soap films, and lipid bilayers are natural boundary conditions of a bending-energy (Helfrich) functional, read off exactly as the beam conditions were here.

The physical picture

Picture the trial curve $y(x)$ between the two verticals at $x=a$ and $x=b$ , and wiggle it by a small perturbation $\delta y(x)$ . The interior wiggle tells you the field equation: demand that no interior wiggle changes $J$ to first order and you get Euler–Lagrange. But the perturbation also acts at the ends, and that is where boundary conditions are born. If a peg holds an endpoint in place you cannot wiggle there — $\delta y=0$ — an essential condition imposed by hand. If the endpoint is loose you can wiggle it, and stationarity insists even that end wiggle produce no first-order change; the only way is for the coefficient $F_{y'}$ to vanish there. The first figure freezes this contrast: a pinned left end, a free right end, and the one boundary term both must kill.

Figure

At a pinned end the value is fixed so the variation vanishes there, δy(a) = 0 (essential); at a free end the variation is arbitrary so stationarity forces ∂F/∂y′ = 0 at that end (natural) — the boundary term ∂F/∂y′ · δy must vanish at every endpoint

The same accounting has teeth the moment you put real physics behind $F$ . Minimizing the bending energy of an elastic beam delivers not only the Euler–Bernoulli equation in the bulk but, at any unsupported end, the two conditions a structural analyst would otherwise quote from a handbook — zero bending moment and zero shear. Nobody imposes them; the variation coughs them up. The second figure makes this concrete: a uniformly loaded cantilever whose clamped root carries the essential data and whose free tip carries the natural data, with a tip deflection you can compute to the millimetre.

Figure

A uniformly loaded cantilever, clamped at the wall and free at the tip, with EI = 8×10⁵ N·m², load w = 1000 N/m, length L = 2 m. The clamp imposes essential conditions y(0) = 0 and y′(0) = 0; the free tip yields the natural conditions M = EI y″(L) = 0 and V = EI y‴(L) = 0. Solving EI y⁗ = w gives tip deflection y(L) = wL⁴/(8EI) = 2.5 mm

The core equations

Everything begins with the first variation — the linear-in- $\delta y$ part of the change in $J$ . No derivative has yet moved, so the perturbed slope $\delta y'$ still appears explicitly and nothing distinguishes interior from boundary.

\delta J=\int_a^b\left(F_y\,\delta y+F_{y'}\,\delta y'\right)\mathrm{d}x \qquad (1)

where: $J[y]$ — the functional $\int_a^b F(x,y,y')\,\mathrm{d}x$ being extremized; $F_y,\,F_{y'}$ — the partial derivatives $\partial F/\partial y$ and $\partial F/\partial y'$ ; $\delta y,\,\delta y'$ — the variation of $y$ and of its slope, related by $\delta y'=(\delta y)'$ . A curve is stationary when this vanishes for every admissible $\delta y$ . Integrating the $\delta y'$ term by parts transfers its derivative onto $F_{y'}$ — the hinge of the whole subject. It converts equation (1) into an interior integral multiplying $\delta y$ plus a pure boundary term.

\delta J=\int_a^b\left(F_y-\frac{\mathrm{d}}{\mathrm{d}x}F_{y'}\right)\delta y\,\mathrm{d}x+\Big[F_{y'}\,\delta y\Big]_a^b \qquad (2)

where: $\frac{\mathrm{d}}{\mathrm{d}x}F_{y'}$ — the total derivative of $F_{y'}$ along the curve; $[\,\cdot\,]_a^b$ — evaluation at $b$ minus at $a$ , the boundary term. The two pieces are logically independent and must each vanish on their own. The interior integral must vanish for all variations free in the interior, so the Fundamental Lemma forces its integrand to zero everywhere — the Euler–Lagrange equation.

F_y-\frac{\mathrm{d}}{\mathrm{d}x}F_{y'}=0 \qquad (3)

where: $=0$ — forced pointwise by the Fundamental Lemma because $\delta y$ is arbitrary on the open interval $(a,b)$ . This is the field equation; the boundary term below is a separate set of statements. With the interior integral gone, whatever is left of $\delta J$ is the boundary term, which stationarity demands vanish by itself at each end.

\Big[F_{y'}\,\delta y\Big]_a^b=F_{y'}(b)\,\delta y(b)-F_{y'}(a)\,\delta y(a)=0 \qquad (4)

where: $\delta y(a),\,\delta y(b)$ — the endpoint variations, zero if the endpoint value is prescribed and arbitrary if it is free. Each endpoint contributes one term, and there are exactly two ways to make a term vanish. The first way is to fix the value. Then the trial space contains only curves through the prescribed point, so $\delta y=0$ there by construction — an essential (Dirichlet, geometric) condition.

y(a)=y_a\qquad\Longrightarrow\qquad \delta y(a)=0\quad(\text{essential}) \qquad (5)

where: $y_a$ — the prescribed endpoint value, wired into the definition of the admissible functions. Essential conditions are inputs you supply; they never appear as extra equations. The second way is to leave the endpoint free. Then $\delta y$ is arbitrary, so the only escape is for its coefficient to vanish — a natural (Neumann-type) condition, produced by the variation rather than imposed.

\left.\frac{\partial F}{\partial y'}\right|_{x=a}=0\qquad(\text{natural}) \qquad (6)

where: $\partial F/\partial y'$ — the conjugate momentum or generalized force associated with $y$ ; its vanishing is a force-balance statement at the free end. Natural conditions are outputs; you read them off, you do not choose them. When $F$ depends on higher derivatives the procedure just runs longer. For $F(x,y,y',y'')$ two integrations by parts expose two boundary contributions per end — one multiplying $\delta y$ , one multiplying $\delta y'$ — so each end carries two conditions.

\delta J=\int_a^b\!\Big(F_y-\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y'}+\tfrac{\mathrm{d}^2}{\mathrm{d}x^2}F_{y''}\Big)\delta y\,\mathrm{d}x+\Big[\Big(F_{y'}-\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y''}\Big)\delta y+F_{y''}\,\delta y'\Big]_a^b \qquad (7)

where: $F_{y''}=\partial F/\partial y''$ ; the bulk equation is now the fourth-order $F_y-\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y'}+\tfrac{\mathrm{d}^2}{\mathrm{d}x^2}F_{y''}=0$ ; the two boundary coefficients give the natural conditions $F_{y''}=0$ and $\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y''}-F_{y'}=0$ at a free end. A fourth-order equation needs four conditions, two per end, matched by the two variations $\delta y$ and $\delta y'$ . Two variants round out the toolkit. Adding an endpoint energy $g(y(b))$ to $J$ contributes $g'(y(b))\,\delta y(b)$ , converting the natural condition into a Robin (mixed) condition without disturbing Euler–Lagrange. And if the endpoint may move horizontally along a target curve, the changing limit contributes an extra $F\,\delta x$ and the natural condition generalizes to transversality.

F_{y'}\big|_b+g'\!\big(y(b)\big)=0,\qquad\qquad \Big[F+(\phi'-y')\,F_{y'}\Big]_{x=x_b}=0 \qquad (8)

where: $g(y(b))$ — an added endpoint contribution such as a spring or surface energy, $g'$ its derivative; $y=\phi(x)$ — the target curve the free end rides on, $\phi'$ its slope. The left relation is Robin; the right is transversality, collapsing to $F_{y'}=0$ when the target is a vertical line.

Derivation ASplitting the first variation into bulk and boundary

The goal is the single integration by parts that simultaneously delivers the Euler–Lagrange equation and reveals which endpoint data are essential and which are natural. We vary the curve by $y\to y+\varepsilon\eta$ with $\eta$ a smooth perturbation that is NOT required to vanish at the ends, demand that $\mathrm{d}J/\mathrm{d}\varepsilon$ vanish at $\varepsilon=0$ , and watch the derivative move off $\eta$ and onto $F_{y'}$ , leaving behind a boundary term that carries all the endpoint physics.

Step 1. Differentiate $J[y+\varepsilon\eta]$ under the integral at $\varepsilon=0$ : the chain rule on $F(x,y+\varepsilon\eta,y'+\varepsilon\eta')$ gives $\tfrac{\mathrm{d}J}{\mathrm{d}\varepsilon}\big|_0=\int_a^b(F_y\,\eta+F_{y'}\,\eta')\,\mathrm{d}x$ . Why: this is the linear response of $J$ , and it is where the slope perturbation $\eta'=(\eta)'$ still appears explicitly.

Step 2. Integrate the second term by parts, $\int_a^b F_{y'}\eta'\,\mathrm{d}x=[F_{y'}\eta]_a^b-\int_a^b(\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y'})\eta\,\mathrm{d}x$ . Why: this transfers the derivative off the arbitrary $\eta$ onto the known $F_{y'}$ , exposing a common factor of $\eta$ — and the price is the boundary term.

Step 3. Collect into $\delta J=\int_a^b\big(F_y-\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y'}\big)\eta\,\mathrm{d}x+[F_{y'}\eta]_a^b$ . Why: the interior integral and the boundary term are independent, because $\eta$ on $(a,b)$ and $\eta$ at the ends can be chosen separately.

Step 4. Restrict first to perturbations vanishing at both ends. Then $[F_{y'}\eta]_a^b=0$ trivially, and the Fundamental Lemma forces $F_y-\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y'}=0$ . Why: end-vanishing perturbations are a legitimate subset, so $\delta J=0$ for them isolates the field equation.

Step 5. Now widen to perturbations free at the ends. The interior integral already vanishes, so stationarity reduces to $[F_{y'}\eta]_a^b=0$ ; at each end either $\eta=0$ (prescribed — essential) or $F_{y'}=0$ (free — natural). Why: with the bulk annihilated, the survivor is purely a boundary statement — the two ways a product can vanish. *Setting $\eta(a)=\eta(b)=0$ recovers the fixed-endpoint derivation, and the boundary term matches equation (2).*

Go deeper · Derivation B · 1 minNatural conditions for a free-ended elastic beam

Derivation BNatural conditions for a free-ended elastic beam

The goal is to show that the two conditions an engineer quotes for the free tip of a cantilever — zero bending moment and zero shear — are not assumptions but natural conditions produced by minimizing the bending energy. We take the Euler–Bernoulli energy, a genuine higher-derivative functional, apply the two-integration-by-parts boundary structure of equation (7), leave the end free so both $\delta y$ and $\delta y'$ survive, and read off the coefficients that must vanish.

Step 1. Write the energy $J[y]=\int_0^L\big(\tfrac12 EI\,y''^2-q\,y\big)\mathrm{d}x$ and identify $F_{y''}=EI\,y''$ , $F_{y'}=0$ , $F_y=-q$ . Why: the $y''$ dependence promotes the problem to the higher-derivative case, so equation (7) is the correct template.

Step 2. Insert into the boundary term of equation (7), which with $F_{y'}=0$ collapses to $[-EI\,y'''\,\delta y+EI\,y''\,\delta y']_0^L$ . Why: the surviving coefficients are the shear-like $EI y'''$ multiplying $\delta y$ and the moment-like $EI y''$ multiplying $\delta y'$ .

Step 3. At a free tip the end may both displace and rotate, so $\delta y(L)$ and $\delta y'(L)$ are both arbitrary. Why: nothing pins either one, unlike the clamped root where both are held to zero.

Step 4. Set each free coefficient to zero: $EI\,y''(L)=0$ and $EI\,y'''(L)=0$ . Why: independent arbitrary variations vanish only if every coefficient does — giving $M=EI y''=0$ (no moment) and $V=EI y'''=0$ (no shear) at the tip.

Step 5. The bulk equation from (7) is $\tfrac{\mathrm{d}^2}{\mathrm{d}x^2}(EI y'')-q=0$ , i.e. $EI\,y''''=q$ . Why: the same variation yields the field equation, confirming the four required conditions (two essential at the clamp, two natural at the tip) match its fourth order. *The constants fixed by $y(0)=y'(0)=0$ and $y''(L)=y'''(L)=0$ give the tip deflection $y(L)=wL^4/(8EI)$ of Worked example 1.*

Go deeper · Derivation C · 1 minThe free terminal point and the transversality condition

Derivation CThe free terminal point and the transversality condition

The goal is to handle the richest free-endpoint case, in which the terminal point may slide along a target curve $y=\phi(x)$ so that even the abscissa $x_b$ is free. We track the two ways the endpoint can move — the changing integration limit and the changing endpoint value — combine them, and extract the transversality condition, which reduces to the plain natural condition when the target is vertical.

Step 1. Let the endpoint slide by $\delta x_b$ along $\phi$ ; the vertical change of the extremal endpoint is $\delta y(x_b)=(\phi'-y')\,\delta x_b$ . Why: moving along $\phi$ raises the point by $\phi'\,\delta x_b$ , following the extremal by $y'\,\delta x_b$ ; the admissible variation is their difference.

Step 2. Assemble the endpoint part of $\delta J$ from the boundary term $F_{y'}\,\delta y(x_b)$ and the term $F\,\delta x_b$ from the moving upper limit. Why: a varying limit of integration contributes the integrand $F$ at the endpoint times the shift $\delta x_b$ , on top of the usual boundary term.

Step 3. Substitute Step 1 and factor: $\delta J\big|_{\text{end}}=\big[F+F_{y'}(\phi'-y')\big]\,\delta x_b$ . Why: both contributions are now proportional to the single free parameter $\delta x_b$ .

Step 4. Because $\delta x_b$ is arbitrary, the bracket vanishes: $F+(\phi'-y')F_{y'}=0$ at $x=x_b$ — the transversality condition. Why: stationarity against every slide forces the coefficient to zero, replacing the missing endpoint datum.

Step 5. Check limits. A vertical target ( $\phi'\to\infty$ ) gives $F_{y'}=0$ , the ordinary natural condition; the arc-length integrand $F=\sqrt{1+y'^2}$ gives $1+\phi'y'=0$ , so the extremal meets $\phi$ at a right angle. Why: recovering both special cases confirms transversality is the correct generalization, not a new axiom. *The orthogonality result $y'\phi'=-1$ is the shortest-path-to-a-curve condition of Worked example 2.*

Limiting cases

How the single boundary term $[F_{y'}\,\delta y]_a^b$ resolves in the standard endpoint scenarios:

Case	Result	Meaning
Both ends fixed (Dirichlet)	$y(a)=y_a,\ \ y(b)=y_b$	$\delta y(a)=\delta y(b)=0$ , so the boundary term vanishes identically; only Euler–Lagrange survives. Two essential conditions.
One end fixed, one free	$y(a)=y_a,\ \ F_{y'}\big\|_{b}=0$	Essential at $a$ , natural at $b$ . The free end self-selects the slope making $F_{y'}$ vanish.
Both ends free	$F_{y'}\big\|_{a}=0,\ \ F_{y'}\big\|_{b}=0$	Two natural conditions; with the second-order equation they generically fix the two constants.
Endpoint energy added	$F_{y'}\big\|_{b}+g'\!\big(y(b)\big)=0$	An added $g(y(b))$ turns the natural condition into a Robin condition — elastic supports, surface tension.
Endpoint free in $x$ too	$\big[F+(\phi'-y')F_{y'}\big]_{x_b}=0$	Transversality: the terminal point rides a target curve, crossing it at the transversal (for arc length, orthogonal) angle.
Higher-derivative $F(x,y,y',y'')$	$F_{y''}=0,\ \ \tfrac{\mathrm{d}}{\mathrm{d}x}F_{y''}-F_{y'}=0$	Two natural conditions per free end (moment and shear for a beam), matching the fourth-order equation.

Worked examples

1 · The uniformly loaded cantilever (fully worked)

A steel cantilever of length $L=2\ \mathrm{m}$ and flexural rigidity $EI=8\times10^5\ \mathrm{N\,m^2}$ carries a uniform load $w=1000\ \mathrm{N/m}$ , clamped at $x=0$ and free at $x=L$ . Find the deflection $y(x)$ and the tip deflection $y(L)$ .

Step 1. The bending energy $J[y]=\int_0^L(\tfrac12 EI\,y''^2-w\,y)\,\mathrm{d}x$ has the Euler–Lagrange equation $EI\,y''''=w$ (Derivation 2). Integrating four times: $y(x)=\dfrac{w}{24EI}x^4+\dfrac{C_1}{6}x^3+\dfrac{C_2}{2}x^2+C_3x+C_4$ .

Step 2. Essential conditions at the clamp fix position and slope, $y(0)=0$ , $y'(0)=0$ , giving immediately $C_4=0$ , $C_3=0$ — inputs on the trial space.

Step 3. Natural conditions at the free tip: $y'''(L)=0$ gives $C_1=-\dfrac{wL}{EI}$ , $y''(L)=0$ gives $C_2=\dfrac{wL^2}{2EI}$ . These were never assumed — the variation produced them.

Step 4. Assemble $y(x)=\dfrac{w}{24EI}\big(x^4-4Lx^3+6L^2x^2\big)$ and evaluate at the tip: $y(L)=\dfrac{w}{24EI}(1-4+6)L^4=\boxed{\dfrac{wL^4}{8EI}}$ .

Step 5. Numbers: $y(L)=\dfrac{(1000)(2)^4}{8\,(8\times10^5)}=\dfrac{16000}{6.4\times10^6}=2.5\times10^{-3}\ \mathrm{m}$ . The tip drops 2.5 mm, with the free-end conditions $M=EI y''(L)=0$ , $V=EI y'''(L)=0$ coming entirely from the natural boundary term.

2 · Shortest path to a target curve (transversality)

Find the shortest curve from $(0,0)$ to the vertical line $x=3$ , then generalize to a target curve $y=\phi(x)$ .

Step 1. The length functional $J[y]=\int_0^{x_b}\sqrt{1+y'^2}\,\mathrm{d}x$ has $F=\sqrt{1+y'^2}$ and $F_{y'}=\dfrac{y'}{\sqrt{1+y'^2}}$ . Euler–Lagrange gives $y'=\text{const}$ : every extremal is a straight line.

Step 2. The start is fixed (essential at $x=0$ ), but the end slides freely along $x=3$ — a free endpoint in $y$ , so the natural condition applies: $F_{y'}\big|_{x=3}=0\Rightarrow y'=0$ .

Step 3. The straight line through $(0,0)$ with $y'=0$ is the horizontal segment to $(3,0)$ , of length $\mathbf{3}$ . It meets the vertical target at a right angle — the natural condition is literally "hit the boundary perpendicularly".

Step 4. Generalize: for a target curve $y=\phi(x)$ , transversality $F+(\phi'-y')F_{y'}=0$ simplifies to $\dfrac{1+\phi'y'}{\sqrt{1+y'^2}}=0$ , hence $\mathbf{1+\phi'y'=0}$ , i.e. $y'\phi'=-1$ .

Step 5. Product of slopes $-1$ means the extremal crosses the target orthogonally — the same as Step 3 for a vertical line ( $\phi'\to\infty\Rightarrow y'\to0$ ). Geodesics meet a free boundary at right angles, a result that carries over to Fermat's principle.

Physics of the portrait

Natural boundary conditions are the fingerprint of one idea — that a physical law and its boundary behaviour can be squeezed from a single variational principle. The clean algorithm that turns a functional into a field equation plus its endpoint conditions is the work of Joseph-Louis Lagrange, who at nineteen sent Euler a method that reorganized all of mechanics around it.

Origins · 1 minJoseph-Louis Lagrange and how this was discovered

Joseph-Louis Lagrange

1736–1813 · founder of the δ-calculus of variations and analytical mechanics

Lagrange was born Giuseppe Lodovico Lagrangia in Turin in 1736 and, essentially self-taught, was a professor there by nineteen. In 1755 he wrote to Leonhard Euler describing a purely analytic method for extremal problems: instead of Euler's geometric polygon constructions, Lagrange introduced the variation operator $\delta$ — a formal "small change" that commutes with differentiation and integration and reduces the whole subject to symbol-pushing. Euler was so impressed he held back his own work so the young Lagrange could publish first, and coined the name "calculus of variations". It was Lagrange's integration by parts of the first variation that first isolated the boundary term $[F_{y'}\,\delta y]_a^b$ and made explicit that a free endpoint imposes its own condition. Called to Berlin in 1766 to succeed Euler, and later to Paris, he crowned the programme with the Mécanique analytique (1788), a treatise so committed to the analytic ideal that he boasted it contained not a single diagram. The essential-versus-natural terminology was standardized much later, in Courant and Hilbert's 1924 Methoden der mathematischen Physik, but the machinery behind the distinction is Lagrange's.

History

1744

— Euler's Methodus inveniendi lays the geometric foundations of the calculus of variations

1755

— the nineteen-year-old Lagrange sends Euler the δ-algorithm, mechanizing the first variation

1786

— Legendre analyses the second variation and the question of sufficiency

1788

— Lagrange's Mécanique analytique recasts all mechanics as a variational principle

1877

— du Bois-Reymond gives the first rigorous proof of the fundamental lemma

1924

— Courant and Hilbert codify the essential versus natural boundary-condition terminology

Going further

Texts. Gelfand & Fomin, Calculus of Variations, §§1–2, 6, 14 (the cleanest treatment of natural and transversality conditions) · Stone & Goldbart, Mathematics for Physics, Ch. 1 · Arfken & Weber, Mathematical Methods for Physicists, calculus-of-variations chapter · Hassani, Mathematical Physics, variational-methods chapter · Courant & Hilbert, Methods of Mathematical Physics, Vol. I, Ch. IV (origin of the terminology and the beam application) · Riley, Hobson & Bence, Ch. 22. Every equation here is a standard calculus-of-variations result. The key checkpoints are: (1) the integration-by-parts split $\delta J=\int_a^b(F_y-\mathrm{d}_xF_{y'})\eta\,\mathrm{d}x+[F_{y'}\eta]_a^b$ ; (2) the higher-derivative boundary terms $F_{y''}=0$ and $\tfrac{\mathrm{d}}{\mathrm{d}x}F_{y''}-F_{y'}=0$ , giving the beam conditions $EI y''=0$ and $EI y'''=0$ , with tip deflection $y(L)=wL^4/(8EI)=2.5\ \mathrm{mm}$ for $w=1000\ \mathrm{N/m}$ , $L=2\ \mathrm{m}$ , $EI=8\times10^5\ \mathrm{N\,m^2}$ ; (3) the transversality condition and its orthogonality reduction $1+\phi'y'=0$ for $F=\sqrt{1+y'^2}$ . Material constants are symbolic; the numerical example is reproducible from the boxed formula.

Prerequisites. First Variation and Stationarity · Euler–Lagrange Equation · Fundamental Lemma of the Calculus of Variations · Functionals and Function Spaces. Leads to. Functionals of Several Functions and Higher Derivatives · Hamilton's Principle of Stationary Action · The Brachistochrone Problem · Geodesics as Variational Problems.

Ask the tutor: Hand me any functional and say which ends are pinned and which are free, and I will walk you through the integration by parts and the essential and natural conditions it forces — step by step.

What is the one-line difference between essential and natural conditions?

An essential condition is something you impose; a natural condition is something you derive. Essential (Dirichlet) conditions prescribe $y$ at an endpoint and are built into the admissible trial space, so every competing curve already obeys them and $\delta y$ vanishes there. Natural (Neumann-type) conditions are not imposed: at a free endpoint the variation is arbitrary, so stationarity forces $F_{y'}=0$ . Essential conditions constrain the function space, natural conditions constrain the variational equation, and each endpoint carries exactly one.

Why are they called "natural" conditions?

Because they arise naturally — automatically — from the variational principle, without being added by hand. When you integrate $\delta J$ by parts, the boundary term $[F_{y'}\,\delta y]_a^b$ appears on its own; at any free endpoint the requirement that it vanish for arbitrary $\delta y$ forces $F_{y'}=0$ with no further input. The name became standard through Courant and Hilbert (natürliche Randbedingungen), and it carried into the finite element method, where "natural" conditions (fluxes, tractions) are enforced automatically by the weak form, unlike "essential" ones imposed on the trial space.

How many boundary conditions does my variational problem need?

Count the order of the Euler–Lagrange equation. A functional depending on $y'$ gives a second-order equation needing two conditions, one per end; one depending up to $y''$ gives a fourth-order equation needing four, two per end. The variation delivers exactly the right number automatically: each integration by parts exposes one boundary coefficient per end, and each is one datum you may prescribe (essential) or leave to the natural condition. If your count does not match the order of the equation, you have over- or under-specified it.

Can I trade a natural condition for an essential one?

Not freely — but you can reshape a natural condition by editing the functional. Adding a boundary energy $g(y(b))$ contributes $g'(y(b))\,\delta y(b)$ , converting the natural condition $F_{y'}=0$ into the Robin condition $F_{y'}+g'(y)=0$ , without changing the bulk equation. This is how an elastic support, surface tension, or the Gibbons–Hawking–York term in gravity enters. What you cannot do is impose an essential and the natural condition at the same end: prescribing $y$ already kills the boundary term via $\delta y=0$ , so demanding $F_{y'}=0$ on top over-determines the problem.

What is transversality and when do I need it?

Transversality is the generalized natural condition you need when a free endpoint can move in $x$ as well as $y$ — when the terminal point must land on a target curve $y=\phi(x)$ rather than a fixed vertical line. The moving upper limit adds $F\,\delta x_b$ to the boundary variation, giving $F+(\phi'-y')F_{y'}=0$ at the endpoint. For a vertical target it collapses to $F_{y'}=0$ ; for the arc-length integrand it says the extremal meets the target at a right angle. It is the calculus-of-variations ancestor of the terminal conditions in optimal control.

Is the essential/natural distinction just for one-dimensional problems?

No — it generalizes directly to fields in higher dimensions and is arguably more important there. For a functional over a region with boundary $\partial\Omega$ , integration by parts becomes the divergence theorem and the boundary term is a surface integral over $\partial\Omega$ . Prescribing the field on the boundary (essential/Dirichlet) makes the surface variation vanish; leaving it free forces a natural condition on the normal derivative or flux (Neumann). This is the structure of electrostatics (potential vs. surface charge), heat conduction (temperature vs. flux), and elasticity (displacement vs. traction) — the backbone of the finite element method in every dimension.