A power-series recurrence can reveal special solutions, but it may leave their geometry obscure. Legendre polynomials provide a complementary construction: differentiating one explicit polynomial produces solutions, their orthogonality, and their normalization. They also show why an equation in Sturm–Liouville form need not be a regular Sturm–Liouville problem at its endpoints.
We study the real equation
The coefficient vanishes at both endpoints. Our conclusions about the polynomial solutions and their expansion will be proved directly; we will not assume the regular-endpoint spectral theorem.
Rodrigues' construction
Definition 2.6.1
(Legendre polynomials (Rodrigues' formula))
For each nonnegative integer , the Legendre polynomial of degree is
The factor is a normalization; it is chosen so that .
The first examples are
Theorem 2.6.2
(Polynomial eigenfunctions and their normalization)
has degree , satisfies and , and solves (1) with . Conversely, if a nonzero polynomial solves (1), its eigenvalue is for its degree , and it is a constant multiple of .
Proof
(Differentiate a polynomial identity)
Put . Its leading coefficient is one, so has degree and leading coefficient
The identity holds for all . Differentiate it times and use the product rule:
Dividing by and changing signs gives . Parity follows because is even and is differentiated times. For the endpoint value, write : after derivatives, at only the term differentiating all factors of survives, giving .
Conversely, if is a nonzero polynomial of degree , the coefficient of in (1) is times its leading coefficient. Thus . Subtract a suitable multiple of to cancel the leading term. A nonzero remainder of degree would have to satisfy , impossible. Hence the remainder is zero.
Orthogonality, norms and zeros
Theorem 2.6.3
(Orthogonality and squared norms)
For every polynomial of degree below ,
In particular the Legendre polynomials are mutually orthogonal, and
Proof
(Endpoint zeros remove the integration-by-parts terms)
Integrate the Rodrigues formula by parts times. Every boundary term vanishes: has a zero of order at each endpoint, so its derivatives of order below vanish there. The remaining integral contains , proving orthogonality.
For the norm, take instead. Its th derivative is . Hence
Set . Integrating the derivative of gives for , with . Therefore , and substitution yields .
Concept lensStructural
The same polynomials from Gram–Schmidt
Apply Gram–Schmidt in to . The th output is a polynomial of exact degree that is orthogonal to every polynomial of degree below . The theorem just proved shows that has the same two properties, and these properties determine a polynomial up to a scalar factor. Indeed, if has degree , leading coefficient , and is orthogonal to all polynomials of degree below , then has degree below and is orthogonal to all such polynomials, in particular to itself. Hence and .
So the differential-equation description and the inner-product description pick out the same line of polynomials in each degree; only the normalization differs. Gram–Schmidt with unit norms produces , whereas the Rodrigues factor makes .
Theorem 2.6.4
(Zero locations)
For , has exactly distinct real zeros in . All are simple.
Proof
(Orthogonality forces sign changes)
Let be the distinct zeros in at which changes sign. If , choose , with when . Across each sign-changing zero both and change sign; at any other zero of , its sign does not change. Thus the continuous polynomial has one sign on the interval and is not identically zero. Its integral cannot be zero, contradicting orthogonality because .
Hence there are at least sign-changing zeros. A degree- polynomial has at most zeros counted with multiplicity, so there are exactly , all simple and all in the open interval.
Proof X-ray
Where orthogonality is spent in the zero count
The proof uses a single fact from the orthogonality theorem: integrates to zero against every polynomial of degree below . Everything else manufactures one such polynomial against which the integral is visibly nonzero.
- The key move is to count only the zeros at which changes sign. A zero of even multiplicity leaves the sign unchanged; putting it into would make change sign there and destroy the constant-sign argument.
- The assumption is used exactly once, to make smaller than , so that orthogonality applies to .
- A strictly nonzero integral needs three things: is continuous, it has one sign, and, being a nonzero polynomial, it vanishes at only finitely many points.
- The closing step is algebra, not analysis: sign-changing zeros already exhaust the zeros, counted with multiplicity, that a polynomial of degree can have. No multiple zero and no zero outside remains possible.
No spectral theorem enters, which matters here because the endpoints are singular.
A three-term recurrence and an expansion
Proposition 2.6.5
(Three-term recurrence)
For ,
Proof
(Expand in the Legendre basis)
To see why only three terms appear, expand in the basis . For , orthogonality gives because has degree at most . The coefficient of also vanishes by parity. Comparing leading coefficients gives the coefficient of as . Evaluating at , where every , gives the coefficient of as . Rearranging proves (2). For , the separate identity is .
Worked example 2.6.6
(Generate from the recurrence and check it)
Take in (2):
Hence . Three independent checks agree with the theorems above. The polynomial is even, as requires. Its value at one is . Its leading coefficient equals . The recurrence therefore produces each new polynomial from the previous two without any differentiation.
Theorem 2.6.7
(Legendre expansion in )
For every , define
Then in as , and
This is an statement; it does not assert pointwise or uniform convergence or permit termwise differentiation.
Proof
(Density comes from polynomials)
Because has exact degree and nonzero leading coefficient, form a basis for all polynomials of degree at most . Polynomials are dense in : continuous functions are dense in on the finite interval, and Weierstrass approximation makes polynomials uniformly dense in the continuous functions. The normalized functions are therefore a complete orthonormal system. The Hilbert-space projection theorem gives the convergence and Parseval identity displayed above.
Worked example 2.6.8
(Expand a polynomial without an integral table)
From ,
The norm identity gives
A finite expansion is an exact identity. For general data the infinite expansion has the weaker norm-convergence meaning stated in the theorem.
The next example is the opposite extreme: a function with a jump, for which the expansion is genuinely infinite.
Worked example 2.6.9
(The sign function: an infinite expansion)
Let on . It is odd and each with even is even, so for every even . For the first two odd coefficients,
The partial sum through degree three is
Parseval's identity checks the bookkeeping. Here , while the two computed terms contribute
The remaining odd coefficients must supply exactly . The function is discontinuous, so no finite sum reproduces it, yet the theorem still applies. That theorem says nothing about the value of the series at the jump or at the endpoints.
Counterexample mode
The vanishing coefficient changes the endpoint problem
At , integrating (1) gives , so on
The polynomial solution is bounded, but the independent solution diverges at both endpoints. The coefficient vanishes at , so the regular Sturm–Liouville hypotheses requiring on the closed interval do not apply. In fact : its logarithmic endpoint singularities have integrable squares. Thus membership alone does not exclude it; boundedness excludes this particular solution. The direct polynomial and density arguments above remain valid, but a general endpoint classification and a claim that no other eigenvalues occur require additional analysis.
Common mistakes
Common mistake
(Orthogonal is not orthonormal)
Since , the coefficient of is . Using alone gives wrong coefficients already for : , whereas the correct coefficient of is .
Common mistake
(Dirichlet conditions at the singular endpoints)
Treating (1) as a regular problem on and imposing excludes every Legendre polynomial, because and . The endpoint condition satisfied by the polynomial solutions is boundedness near ; at it is exactly this condition that rules out the solution found above.
Exercises
Quick check
(Use Rodrigues' formula to compute and , and check in these cases.)
Solution · Answer
Differentiation gives and ; both equal one at .
Quick check
(Use the recurrence to express in the Legendre basis. Compute from the norm formula.)
Solution · Answer
Since , the recurrence gives and hence . Orthogonality gives .
Quick check
(Show that no polynomial of degree three can solve (1) with .)
Solution · Answer
A degree-three polynomial solution requires by its leading coefficient. Since , none exists.
Quick check
(Show that and solve (1) for , and prove their linear independence using initial data at .)
Solution · Answer
Both functions solve the equation: the derivative of is , so . Their initial-value vectors at zero are and , whose determinant is one. They are linearly independent; the equation is regular near zero although its endpoints are singular.