Conditions for Extrema: Legendre, Jacobi and Weierstrass
CSIR NET calculus of variations: second variation, Legendre and Jacobi conditions, conjugate points, Weierstrass E-function, convexity and corner conditions.
- 23 min
On this page · 19 sections
- Prerequisites
- The Hook: The Same Extremal, Two Different Answers
- Recap: Weak and Strong Extrema
- The Second Variation
- The Legendre Condition
- The Jacobi Condition and Conjugate Points
- The Weierstrass Condition: Strong Extrema
- Fields of Extremals (in one paragraph)
- Convexity: The Fastest Sufficient Condition
- Broken Extremals: Weierstrass–Erdmann Corner Conditions
- When the Minimum Does Not Exist
- Summary of Conditions
- Memory Tricks
- Don’t Confuse
- Common Mistakes
- Practice
- Quick Revision Table
- Key Takeaways
- What’s Next
Prerequisites
- Functionals, First Variation and the Euler–Lagrange Equation: weak and strong extrema, extremals, the first variation.
- Continuity, Differentiability and Bounded Variation: Taylor’s theorem with remainder, which is how the second variation and the Weierstrass function are derived.
- Metric Spaces and Several Variables: the second-derivative (Hessian) test in \( \mathbb{R}^n \), whose infinite-dimensional analogue this page develops.
The Hook: The Same Extremal, Two Different Answers
Recap: Weak and Strong Extrema
| Weak extremum | Strong extremum | |
|---|---|---|
| Competitors \( \hat y \) | \( \lvert \hat y - y \rvert \) and \( \lvert \hat y' - y' \rvert \) small | only \( \lvert \hat y - y \rvert \) small; slopes may differ wildly |
| Norm | \( C^1 \) norm | \( C^0 \) (sup) norm |
| Relation | strong \( \Rightarrow \) weak (more competitors to beat) | |
| Necessary conditions | E–L, Legendre, Jacobi | E–L, Legendre, Jacobi and Weierstrass |
The Second Variation
For an extremal \( y \) and a variation \( h \) with \( h(a) = h(b) = 0 \), Taylor’s theorem in \( \varepsilon \) gives
$$J[y + \varepsilon h] = J[y] + \varepsilon\,\delta J[h] + \frac{\varepsilon^2}{2}\,\delta^2 J[h] + o(\varepsilon^2),$$with \( \delta J[h] = 0 \) and
$$\delta^2 J[h] = \int_a^b \big( F_{yy}h^2 + 2F_{yy'}hh' + F_{y'y'}h'^2 \big)\, dx.$$Write \( 2hh' = (h^2)' \) and integrate the middle term by parts (the boundary term vanishes):
$$\boxed{\delta^2 J[h] = \int_a^b \big( P\,h'^2 + Q\,h^2 \big)\, dx, \qquad P = F_{y'y'}, \quad Q = F_{yy} - \frac{d}{dx}F_{yy'},}$$all evaluated along the extremal. Necessary condition for a minimum: \( \delta^2 J[h] \ge 0 \) for every admissible \( h \). (For a maximum, \( \le 0 \).)
The Legendre Condition
A variation \( h \) can be small while \( h' \) is large: think of \( h = \frac{1}{n}\sin(n^2 x) \) localised near a point. For such \( h \) the \( Ph'^2 \) term dominates \( \delta^2 J \). So:
Legendre hoped that \( P > 0 \) would also be sufficient. The hook shows it is not: \( P = 2 > 0 \) for \( \int (y'^2 - y^2) \) on every interval, yet \( y \equiv 0 \) fails to be a minimum once \( b > \pi \). The missing ingredient is Jacobi’s.
The Jacobi Condition and Conjugate Points
The second variation is itself a quadratic functional in \( h \). Its Euler–Lagrange equation is the Jacobi (accessory) equation
$$\boxed{-\frac{d}{dx}\big(P u'\big) + Q u = 0.}$$A shortcut. If \( y(x, \alpha) \) is a one-parameter family of extremals all passing through \( (a, A) \), then \( u = \frac{\partial y}{\partial\alpha} \) solves the Jacobi equation with \( u(a) = 0 \). So conjugate points are where neighbouring extremals from the same starting point cross the given one again. For \( \int (y'^2 - y^2) \), the extremals through the origin are \( \alpha\sin x \), \( u = \sin x \), and the first conjugate point is \( \pi \): every extremal from the origin meets again at \( (\pi, 0) \).
Why it works: if \( \tilde a < b \) is conjugate, the function equal to \( u \) on \( [a, \tilde a] \) and 0 on \( [\tilde a, b] \) makes \( \delta^2 J = 0 \) with a corner, and smoothing the corner makes it negative.
Sufficient conditions for a weak minimum
graph TD
A["Extremal: solves E-L and the boundary values"] --> B{"Strengthened Legendre: F_pp positive on the whole interval?"}
B -->|no| X["Sufficiency test fails: check the necessary Legendre condition"]
B -->|yes| C{"Strengthened Jacobi: no conjugate point up to and including b?"}
C -->|no| Y["Not a minimum if a conjugate point lies strictly inside"]
C -->|yes| D["Weak minimum"]
D --> E{"Weierstrass: E non-negative near the extremal for all slopes w?"}
E -->|yes| F["Strong minimum"]
E -->|fails on the extremal| G["Weak but not strong minimum"]The Weierstrass Condition: Strong Extrema
A strong competitor may have slopes far from \( y' \), so a test that only looks at small slope changes (Legendre) cannot be enough. Weierstrass compares the value of \( F \) at an arbitrary slope \( w \) with its tangent-line approximation at the extremal’s slope \( p \):
$$\boxed{E(x, y, p, w) = F(x, y, w) - F(x, y, p) - (w - p)\,F_{y'}(x, y, p).}$$\( E \ge 0 \) says that, as a function of the slope, \( F \) lies above its tangent line at \( p \).
By Taylor’s theorem in the slope, \( E = \frac{(w - p)^2}{2}F_{y'y'}(x, y, q) \) for some \( q \) between \( p \) and \( w \). Hence:
- If \( F_{y'y'}(x, y, w) \ge 0 \) for all \( (x, y) \) near the extremal and all \( w \) (\( F \) convex in \( y' \)), then \( E \ge 0 \) automatically. For \( F = y'^2 - y^2 \), \( E = (w - p)^2 \ge 0 \), which is why worked example 1 gives a strong minimum when \( b < \pi \).
- Letting \( w \to p \) in the necessary condition recovers Legendre’s \( F_{y'y'} \ge 0 \).
Fields of Extremals (in one paragraph)
A field of extremals on a region \( R \) is a one-parameter family of extremals such that exactly one passes through each point of \( R \); its slope function is \( p(x, y) \). A central field is a pencil of extremals through one point. The strengthened Jacobi condition is exactly what guarantees that the given extremal can be embedded in a field (a central field from a point slightly to the left of \( a \)). On a field, \( J \) of the field’s extremal can be compared with \( J \) of any competitor through Hilbert’s invariant integral, and the difference is the integral of \( E \). That is the proof of the Weierstrass sufficiency theorem and the reason the conditions fit together.
Convexity: The Fastest Sufficient Condition
Proof. Convexity gives \( F(x, y + h, y' + h') \ge F(x, y, y') + F_y h + F_{y'}h' \). Integrate: \( J[y + h] \ge J[y] + \delta J[h] = J[y] \).
Test for convexity (\( C^2 \) case): \( F_{yy} \ge 0 \), \( F_{y'y'} \ge 0 \) and \( F_{yy}F_{y'y'} - F_{yy'}^2 \ge 0 \) for all \( (y, p) \), i.e. the \( 2 \times 2 \) Hessian is positive semi-definite (see Quadratic Forms). Every quadratic functional \( \int (py'^2 + qy^2 - 2fy) \) with \( p > 0 \), \( q \ge 0 \) qualifies. This is the test to use first in Part B: it needs no Jacobi equation.
Broken Extremals: Weierstrass–Erdmann Corner Conditions
If the admissible class allows piecewise \( C^1 \) curves, a minimiser may have corners. At a corner \( c \), applying the du Bois-Reymond argument on both sides gives the Weierstrass–Erdmann conditions:
$$F_{y'}\big|_{c^-} = F_{y'}\big|_{c^+}, \qquad \big(F - y'F_{y'}\big)\big|_{c^-} = \big(F - y'F_{y'}\big)\big|_{c^+}.$$If \( F_{y'y'} \ne 0 \) for all slopes, \( F_{y'} \) is strictly monotone in \( y' \) and the first condition forces equal one-sided slopes: no corners. Corners need an integrand that is not strictly convex in \( y' \).
Example. \( F = (y'^2 - 1)^2 \). At slopes \( \pm 1 \): \( F_{y'} = 4y'(y'^2 - 1) = 0 \) and \( F - y'F_{y'} = 0 \) on both sides, so a switch between slopes \( 1 \) and \( -1 \) satisfies both conditions. With \( y(0) = y(2) = 0 \), the zig-zags \( 1 - \lvert x - 1 \rvert \), \( \lvert x - 1 \rvert - 1 \) and any saw-tooth with slopes \( \pm 1 \) give \( J = 0 \), the minimum, while the smooth extremal \( y \equiv 0 \) gives \( J = 2 \).
When the Minimum Does Not Exist
Lower bounds do not guarantee a minimiser. Weierstrass’s example:
$$J[y] = \int_{-1}^{1} x^2y'^2\, dx, \qquad y(-1) = -1, \quad y(1) = 1.$$\( J \ge 0 \). The curves \( y_\varepsilon = \frac{\arctan(x/\varepsilon)}{\arctan(1/\varepsilon)} \) are admissible and \( J[y_\varepsilon] \le \frac{\pi\varepsilon}{\arctan^2(1/\varepsilon)} \to 0 \), so \( \inf J = 0 \). But \( J[y] = 0 \) forces \( y' = 0 \) for \( x \ne 0 \), so \( y \) is constant on each side and cannot be continuous with \( y(\pm 1) = \pm 1 \). The infimum is not attained. The E–L equation \( (x^2y')' = 0 \) gives \( y = d - \frac{c}{x} \), with no \( C^1 \) solution on \( [-1, 1] \) meeting the boundary values. Note that \( P = 2x^2 \) vanishes at \( x = 0 \): the strengthened Legendre condition fails there.
Summary of Conditions
| Condition | Statement (for a minimum) | Type |
|---|---|---|
| Euler–Lagrange | \( F_y - \frac{d}{dx}F_{y'} = 0 \) | necessary, weak and strong |
| Legendre | \( F_{y'y'} \ge 0 \) along the extremal | necessary, weak and strong |
| Jacobi | no conjugate point in \( (a, b) \) | necessary (when \( P > 0 \)) |
| Weierstrass | \( E(x, y, y', w) \ge 0 \) for all \( w \) | necessary for strong |
| Weierstrass–Erdmann | \( F_{y'} \) and \( F - y'F_{y'} \) continuous at corners | necessary, broken extremals |
| E–L + strengthened Legendre + strengthened Jacobi | sufficient for weak | |
| above + field + \( E \ge 0 \) near the extremal | sufficient for strong | |
| E–L + \( F \) convex in \( (y, y') \) | sufficient for global |
For a maximum, reverse the inequalities in Legendre and Weierstrass; the Jacobi condition is unchanged.
Memory Tricks
Don’t Confuse
| This | vs | That |
|---|---|---|
| Legendre: \( F_{y'y'} \ge 0 \) (necessary) | strengthened Legendre: \( F_{y'y'} > 0 \) (part of sufficiency) | |
| Jacobi necessary: no conjugate point in \( (a, b) \) | strengthened: none in \( (a, b] \) | |
| \( E \ge 0 \) along the extremal (necessary for strong) | \( E \ge 0 \) near it, with a field (sufficient for strong) | |
| weak minimum (\( \int y'^3 \) line) | strong minimum (\( \int (y'^2 - y^2) \), \( b < \pi \)) | |
| infimum (may not be attained: \( \int x^2y'^2 \)) | minimum (attained) | |
| \( Q = F_{yy} - \frac{d}{dx}F_{yy'} \) | \( F_{yy} \) alone (wrong unless \( F_{yy'} \) is constant along the extremal) |
Common Mistakes
Practice
All questions below are original practice questions written for this page. They are not from any CSIR NET paper. In Level 3, as in Part C, one or more options may be correct.
Level 1: Foundation
- State Legendre’s necessary condition for a minimum.
- Define a point conjugate to \( a \).
- Compute the Weierstrass function \( E \) for \( F = y'^2 \).
Answers
- \( F_{y'y'}(x, y(x), y'(x)) \ge 0 \) for all \( x \in [a, b] \) along the extremal.
- A point \( \tilde a > a \) where the solution of the Jacobi equation \( -(Pu')' + Qu = 0 \) with \( u(a) = 0 \), \( u'(a) = 1 \) vanishes.
- \( E = w^2 - p^2 - 2p(w - p) = (w - p)^2 \ge 0 \).
Level 2: Part B Style
Practice Q1. For \( J[y] = \int_0^b (y'^2 - 4y^2)\, dx \) along the extremal \( y \equiv 0 \), the first point conjugate to \( x = 0 \) is:
(a) \( \frac{\pi}{4} \) (b) \( \frac{\pi}{2} \) (c) \( \pi \) (d) \( 2\pi \)
Answer: (b)
\( P = 2 \), \( Q = -8 \). Jacobi: \( -2u'' - 8u = 0 \), i.e. \( u'' + 4u = 0 \), with \( u(0) = 0 \), \( u'(0) = 1 \): \( u = \frac{1}{2}\sin 2x \), first zero at \( \frac{\pi}{2} \). Option (a) is the zero of \( \cos 2x \), from the wrong initial data.
Practice Q2. For \( J[y] = \int_0^1 (1 + x)y'^2\, dx \), \( y(0) = 0 \), \( y(1) = 1 \), the extremal gives:
(a) a minimum, with value \( \frac{1}{\ln 2} \) (b) a minimum, with value \( \ln 2 \) (c) a maximum, with value \( \frac{1}{\ln 2} \) (d) a minimum, with value 2
Answer: (a)
\( F_{y'} = 2(1 + x)y' = C \), so \( y' = \frac{c}{1 + x} \), \( y = c\ln(1 + x) \); \( y(1) = 1 \Rightarrow c = \frac{1}{\ln 2} \). Then \( J = \int_0^1 (1 + x)\frac{c^2}{(1 + x)^2}\, dx = c^2\ln 2 = \frac{1}{\ln 2} \). \( F \) is convex in \( y' \) (and has no \( y \)), so this is the global minimum.
Practice Q3. For \( J[y] = \int_0^1 y'^3\, dx \), \( y(0) = 0 \), \( y(1) = -1 \), the extremal \( y = -x \) gives:
(a) a strong minimum (b) a weak minimum that is not strong (c) a weak maximum that is not a strong maximum (d) a strong maximum
Answer: (c)
Slope \( p = -1 \). \( F_{y'y'} = 6p = -6 < 0 \): strengthened Legendre for a maximum. Jacobi: \( u'' = 0 \), \( u = x \), no conjugate point. So a weak maximum. Weierstrass: \( E = (w - p)^2(w + 2p) = (w + 1)^2(w - 2) \), which is positive for \( w > 2 \); a strong maximum needs \( E \le 0 \) for all \( w \). Not strong.
Practice Q4. For \( J[y] = \int_{-1}^{1} x^2y'^2\, dx \) over \( C^1 \) functions with \( y(-1) = -1 \), \( y(1) = 1 \):
(a) the minimum is attained at \( y = x \) (b) \( \inf J = 0 \) and it is not attained (c) \( \inf J = \frac{2}{3} \) (d) \( J \) is unbounded below
Answer: (b)
\( J \ge 0 \), and \( y_\varepsilon = \frac{\arctan(x/\varepsilon)}{\arctan(1/\varepsilon)} \) gives \( J \to 0 \). \( J = 0 \) would need \( y' = 0 \) off \( x = 0 \), impossible for a continuous \( y \) with these end values. \( J[x] = \int_{-1}^1 x^2\, dx = \frac{2}{3} \) is only an upper bound.
Level 3: Part C Style
Practice Q5. Let \( J[y] = \int_0^b (y'^2 - y^2)\, dx \) with \( y(0) = y(b) = 0 \). Which of the following are true?
(a) If \( b = 3 \), then \( y \equiv 0 \) gives a strong minimum. (b) If \( b = 4 \), then \( \inf J = -\infty \). (c) If \( b = \pi \), then \( y \equiv 0 \) is the unique minimiser. (d) The strengthened Legendre condition holds along \( y \equiv 0 \) for every \( b > 0 \).
Answer: (a), (b) and (d)
- (a) True. \( 3 < \pi \): no conjugate point in \( (0, 3] \), \( F_{y'y'} = 2 > 0 \), and \( E = (w - p)^2 \ge 0 \) for all slopes; with a field of extremals this is a strong minimum.
- (b) True. For \( h = \sin\frac{\pi x}{4} \), \( J[ch] = c^2\frac{\pi^2 - 16}{8} \to -\infty \) as \( c \to \infty \).
- (c) False. \( J[c\sin x] = c^2\int_0^\pi(\cos^2x - \sin^2x)\, dx = 0 \) for every \( c \): infinitely many minimisers.
- (d) True. \( F_{y'y'} = 2 > 0 \) regardless of \( b \). That is exactly why Legendre alone is not sufficient.
Practice Q6. Let \( J[y] = \int_0^2 (y'^2 - 1)^2\, dx \) over piecewise \( C^1 \) functions with \( y(0) = y(2) = 0 \). Which of the following are true?
(a) The minimum value of \( J \) is 0. (b) The \( C^1 \) extremal \( y \equiv 0 \) is a minimiser. (c) \( y = 1 - \lvert x - 1 \rvert \) is a minimiser. (d) The minimiser is unique.
Answer: (a) and (c)
- (a) True. \( J \ge 0 \), and (c) attains 0.
- (b) False. \( J[0] = \int_0^2 1\, dx = 2 > 0 \).
- (c) True. Slopes are \( \pm 1 \), so \( F = 0 \) everywhere; the corner at \( x = 1 \) satisfies Weierstrass–Erdmann (\( F_{y'} = 0 \) and \( F - y'F_{y'} = 0 \) on both sides).
- (d) False. \( \lvert x - 1 \rvert - 1 \) and any saw-tooth with slopes \( \pm 1 \) also give 0.
Practice Q7. Let \( J[y] = \int_0^1 (y'^2 + 2yy' + 4y^2)\, dx \) with \( y(0) = 0 \), \( y(1) = 1 \). Which of the following are true?
(a) The Euler–Lagrange equation is \( y'' = 4y \). (b) The extremal is \( y = \frac{\sinh 2x}{\sinh 2} \). (c) Deleting the term \( 2yy' \) from the integrand does not change the extremal. (d) Because \( F_{yy'} \ne 0 \), the convexity test fails and \( J \) has no minimum.
Answer: (a), (b) and (c)
- (a) True. \( F_y = 2y' + 8y \), \( \frac{d}{dx}F_{y'} = \frac{d}{dx}(2y' + 2y) = 2y'' + 2y' \). E–L: \( 8y - 2y'' = 0 \).
- (b) True. \( y = c_1\sinh 2x + c_2\cosh 2x \) with the boundary values.
- (c) True. \( 2yy' = (y^2)' \) integrates to \( y(1)^2 - y(0)^2 = 1 \) for every admissible \( y \): a constant shift of \( J \).
- (d) False. Convexity needs the Hessian \( \begin{pmatrix} F_{yy} & F_{yy'} \\ F_{yy'} & F_{y'y'} \end{pmatrix} = \begin{pmatrix} 8 & 2 \\ 2 & 2 \end{pmatrix} \) to be positive semi-definite, not \( F_{yy'} = 0 \). Here the determinant is \( 12 > 0 \) and the trace is positive, so \( F \) is strictly convex and the extremal is the unique global minimum.
Quick Revision Table
| Idea | Fact |
|---|---|
| Second variation | \( \delta^2J = \int (Ph'^2 + Qh^2) \), \( P = F_{y'y'} \), \( Q = F_{yy} - \frac{d}{dx}F_{yy'} \) |
| Legendre | min \( \Rightarrow F_{y'y'} \ge 0 \); max \( \Rightarrow \le 0 \) |
| Jacobi equation | \( -(Pu')' + Qu = 0 \), \( u(a) = 0 \), \( u'(a) = 1 \) |
| Conjugate point | next zero of that \( u \); for \( y'^2 - k^2y^2 \) it is \( \frac{\pi}{k} \) |
| Weak min (sufficient) | E–L + \( F_{y'y'} > 0 \) + no conjugate point in \( (a, b] \) |
| Weierstrass \( E \) | \( F(w) - F(p) - (w - p)F_{y'}(p) \) |
| Strong min | add field + \( E \ge 0 \) for all \( w \) |
| \( \int y'^3 \) | weak, not strong (\( E = (w - p)^2(w + 2p) \)) |
| Convex \( F \) in \( (y, y') \) | extremal = global min |
| Corners | \( F_{y'} \), \( F - y'F_{y'} \) continuous |
| \( \int_{-1}^1 x^2y'^2 \) | inf 0, not attained |
Key Takeaways
What’s Next
- Variational Methods for Boundary Value Problems: the convexity theorem is what makes the Ritz method converge to the true solution.
- Extensions of the Euler–Lagrange Equation: constrained problems, where the multiplier plays the role of an eigenvalue.
- Functionals, First Variation and the Euler–Lagrange Equation.
- Calculus of Variations: Quick Revision.