Numerov's method¶

Solve a smooth scalar second-order equation without an unreduced first-derivative term on a uniform grid using the fixed three-point 1:10:1 rule and a suitable closure.

Version: v1 · 2026-10-17 · History
Domain-specific #: 14233
Domain group: Formal Sciences
Origin domain: Mathematics
Subdomain: Numerical Ordinary Differential Equations → Mathematics
Aliases: Numerov method, Noumerov method

Core Idea¶

Numerov's method is a uniform-grid numerical procedure for a sufficiently smooth scalar second-order equation without an unreduced first-derivative term. Its classical three-point relation connects neighboring solution values to right-hand-side values with fixed weights \(1:10:1\). Together with suitable starting or boundary data and a solution rule, that relation becomes a method for computing a discrete solution. Caruso and Oguri present its homogeneous linear Schrödinger specialization; Bernardes uses the same weighted structure for a prescribed-source Poisson problem.^[1]^[2]

For \(y''=f(x,y)\), write \(x_i=x_0+ih\), \(y_i\approx y(x_i)\), and \(f_i=f(x_i,y_i)\). The smooth exact solution satisfies

\[ y(x_i+h)-2y(x_i)+y(x_i-h) =\frac{h^2}{12}\left[f(x_i+h,y(x_i+h)) +10f(x_i,y(x_i))+f(x_i-h,y(x_i-h))\right]+O(h^6). \]

The numerical equations omit the residual and solve for the grid values \(y_i\). Bernardes's convention is \(\phi''=-\rho\), so his weighted-source term has a minus sign.^[2] For general nonlinear \(f(x,y)\), the next-node term depends on the unknown next value: the relation is implicit, not an automatic explicit marching rule. A specified solver and closure are still needed.

Structural Signature¶

Sig role-phrases:

R1 — Second-order target without an unreduced first derivative. The target equation identifies the unknown scalar function, independent variable and right-hand side. Caruso's classical equation class excludes an unreduced \(y'\) term; his hydrogen treatment is a separately discussed generalization.^[1]
R2 — Uniform grid and step. Equally spaced nodes make the symmetric formula and its fixed coefficients meaningful. A variable-step formula needs a separate derivation.^[2]
R3 — Three neighboring solution values. The left side is the second difference of the solution at \(i-1,i,i+1\). Three source samples alone do not supply these unknown solution values.^[2]
R4 — Fixed weighted right-hand side. The same three nodes contribute right-hand-side values in the ratio \(1:10:1\), divided by twelve and multiplied by \(h^2\). Arbitrary stencil coefficients do not preserve this named specialization.^[1]^[2]
R5 — Closure, solver and solution output. Starting or boundary information must close the discrete problem, and an actual solution rule must produce the requested values or eigenvalue estimates. Two-sided matching and direct endpoint closure are different ways to fill this role.^[1]^[2]
R6 — Accuracy and validation account. This quality-bearing role states what error is assessed and what evidence supports the result. The source cases compare outputs with analytic solutions.^[1]^[2] Such a benchmark is not an additional defining operation that every execution must perform; omitting it leaves accuracy less established rather than changing the stencil's identity.

R1–R5 constitute the completed numerical procedure. R6 constrains what can responsibly be concluded about its output. A formula without a target and closure is only an ingredient.

What It Is Not¶

Numerov is not every three-point finite-difference formula. The fixed right-hand-side weighting is a distinguishing condition. Nor is it the differential equation being solved: the same equation can be approached by another method.

It is not universally explicit, a synonym for shooting, or a guarantee of a particular global accuracy. It does not automatically include variable-step, matrix, log-derivative or first-derivative extensions. Caruso explicitly distinguishes his hydrogen generalization from the classical no-first-derivative form.^[1]

Cowell's method is not an established alias for this bounded identity. Noumerov's original paper concerns special coordinates and extrapolation of celestial perturbations.^[3] The adjacent Cowell history passage belongs to a separately bylined Jackson note, not to Noumerov's article.^[4] Neither historical passage alone establishes equivalence to the modern scalar \(1:10:1\) procedure.

Scope of Application¶

The classical scope here is a smooth scalar second-order ordinary differential equation, a uniform grid and a suitable closure. The two documented settings are a quantum harmonic-oscillator eigenvalue problem and a one-dimensional Poisson Dirichlet problem. Their physical quantities differ, while the weighted numerical relation persists.^[1]^[2]

For a nonlinear right-hand side, the formal relation remains implicit; these examples do not supply a general nonlinear solver. Applicability outside the stated smoothness, equation and grid conditions requires additional justification.

Clarity¶

Three distinctions prevent common overclaims. First, solution output is different from validation: an estimated eigenvalue or potential array is the result, while comparison with an analytic answer assesses it. Second, an unscaled local residual is different from global solution error. An \(O(h^6)\) residual in the three-point equation does not, by itself, prove sixth-order or fourth-order global accuracy, stability or convergence.

Third, “linear” can refer to different objects. Bernardes reorganizes his direct summation so execution time grows linearly with grid dimension; his contrasted “non-linear” implementation has quadratic runtime.^[2] That description is not evidence that the underlying Poisson equation is nonlinear in its unknown.

Manages Complexity¶

The repeated three-node relation turns differential structure into local algebra. Instead of treating every grid point as a fresh problem, one reuses a fixed coefficient pattern. The remaining global work becomes visible in the closure: where to start, which boundary values are known, what parameter is unknown and how the solution is selected.

This separation prevents implementation details from becoming false definitions. Shooting is one closure strategy; a direct boundary solve is another. The weighting identifies the specialist rule, while closure determines how that rule becomes a usable computation.

Abstract Reasoning¶

Substituting \(f=-k^2y\) into the weighted relation and collecting terms gives

\[ \left(1+\frac{h^2k_{i+1}^2}{12}\right)y_{i+1} =2\left(1-\frac{5h^2k_i^2}{12}\right)y_i -\left(1+\frac{h^2k_{i-1}^2}{12}\right)y_{i-1}. \]

This is Caruso's linear homogeneous form.^[1] An algebraic next-value calculation requires the coefficient of \(y_{i+1}\) to be nonzero. That is a solvability condition, not an extra universal stage of the method.

Substituting \(f=-\rho(x)\) instead leaves prescribed source samples. The same local weighting survives, but two endpoint potentials can determine the boundary solution without an energy search.^[2] The useful inference is to separate what the stencil supplies from what the problem's closure must supply.

Knowledge Transfer¶

Transfer between these two equation settings is literal numerical reuse: retain the grid, three-value relation and weights, then derive the appropriate closure. One does not transfer the oscillator's match point into a Poisson problem merely because both use Numerov.

More broadly, recognizing a method requires checking its equation class, representation, solution rule and adequacy account. A visual resemblance between weighted triples in another domain is insufficient. The name does not carry a convergence guarantee into a new equation, boundary treatment or implementation.

Examples¶

Canonical: harmonic-oscillator eigenvalues¶

Caruso and Oguri apply the linear homogeneous rule to the one-dimensional harmonic oscillator. They propagate from both sides, match the wavefunctions and their derivatives, and adjust the trial energy. Their Table 1 compares six numerical eigenvalues with analytic oscillator values.^[1]

Mapped back: R1 is the no-\(y'\) Schrödinger target; R2 is the uniform integration step; R3 is the three neighboring wavefunction values; R4 is the weighted \(-k^2\psi\) contribution; R5 is the two-sided propagation, matching and energy adjustment producing eigenvalue/wavefunction estimates; R6 is the analytic comparison. The matching tolerance participates in that particular solution rule; the table is validation, not a universal runtime requirement.

This is a worked numerical eigenvalue example, not an empirical measurement or proof that every energy search terminates. The paper's separate hydrogen variant is not a second positive classical case.

Applied: Gaussian-source Poisson potential¶

Bernardes solves \(\phi''=-\rho\) with prescribed endpoint values. His worked test uses \(\rho(x)=e^{-x^2/4}\) on \([-10,10]\), with \(\phi(-10)=1\) and \(\phi(10)=2\). Equations (12)–(13) close the boundary problem; equation (16) reorganizes its direct solution through partial sums. Equation (18) and Figure 1 supply the analytic comparator and error/time comparisons.^[2]

Mapped back: R1 is the potential equation; R2 is the equal-spacing grid; R3 is the three neighboring potential values; R4 is the weighted negative source; R5 is the endpoint closure and direct computation producing a potential array; R6 is comparison with the analytic potential and reported error/runtime data. No unknown energy or two-sided shooting is required.

Bernardes describes this construction as sixth order and its reorganized execution as linear in grid dimension.^[2] Those are source-specific claims about that construction, not blanket properties of every Numerov use.

Structural Tensions¶

T1: Grid refinement vs total computational work. On a fixed interval, smaller \(h\) means more nodes. Refinement seeks a better-resolved solution, while a full-grid computation must produce more values. Bernardes's partial-sum organization reduces execution growth from quadratic to linear in grid dimension, and his Figure 1 compares cost and error across resolutions.^[2] It does not make additional outputs free.

A coarse grid reduces work but may fail the intended accuracy requirement. A finer grid increases work and still needs an appropriate global-error and stability assessment; the local residual exponent alone cannot select the grid. Diagnostic: What is the coarsest tested grid meeting the declared solution-error norm and tolerance for this closure, and what measured improvement does further refinement buy?

The distinction between local residual and global solution error is a boundary, not another opposed objective.

Structural–Framed Character¶

Numerov is predominantly structural within numerical analysis: its equation class, grid relations and coefficients determine membership. Evaluative weight enters the chosen tolerance and computational budget, but calling a result “accurate enough” does not establish that it used the Numerov rule. Human-practice dependence enters implementation, stopping decisions and validation; the mathematical relation itself can be identified independently of the institution using it.

Institutional origin and historical naming are not identity tests. Attribution does not establish modern formula equivalence or authorize the Cowell alias. Vocabulary travel is literal between wavefunction and potential computations because the mathematical roles remain, while a metaphorical use of “weighted neighbors” supplies too little structure. Import versus recognition asks whether the new problem actually has the same equation/grid/weight/closure package rather than merely receiving the name.

The portable skeleton is an effective computational procedure coupled to a surrogate-and-error question. These connect to Algorithm and Approximation, but a particular use still has to state its procedure and adequacy conditions. Its character: structural in its fixed numerical relation, domain-framed by smooth second-order equations and grid closure, with purpose-dependent validation criteria.

Structural Core vs. Domain Accent¶

The core contains the classical equation class, uniform grid, three neighboring solution values, fixed weighted right-hand side and sufficient closure with a solution rule. Having a closure is essential; choosing shooting rather than direct Dirichlet algebra is an accent. Wavefunctions, potentials, interval endpoints, programming languages and analytic benchmarks are also variable.

Reducing the entry to “a finite-difference method” would discard its fixed specialist relation. Treating that relation as an entirely separate cross-domain Prime would overstate its reach: both positive examples retain the same numerical-analysis carrier. Algorithm supplies the broader effective-procedure idea, while Approximation supplies the controlled-surrogate questions. Neither broader idea uniquely determines Numerov's weights, and neither makes its numerical guarantees automatic.

This entry is a kind of Numerical Method.

Algorithm is a real procedural bridge. At fixed grid size, Bernardes's direct construction has declared source/grid/endpoint inputs, finite partial-sum passes, a grid-valued output and a stated execution-cost account.^[2] A bare recurrence lacks that full procedure. Caruso's stopping test likewise must not be mistaken for proof that every possible eigenvalue search succeeds.

Approximation supplies the target, tractable surrogate, error and tolerance questions. The differential solution is the target and the computed grid solution is the surrogate. A complete application must still supply its error regime and acceptable tolerance; a local residual or an error plot alone does not establish that complete contract.

Numerical Method supplies the broader genus: both cases have a stated target, finite grid, solution rule, and assessable quality and cost conditions. Finite Difference Method is a closer family comparison because it organizes grid stencils, closure, consistency, stability, and convergence. Numerov fixes the 1:10:1 weighting. Its local truncation result and case-specific performance checks do not establish stability, convergence, or error-tolerance guarantees for every closure.

Relationships to Other Abstractions¶

Current abstraction Numerov's method Domain-specific

Parents (1) — more general patterns this builds on

Numerov's method is a kind of Numerical Method Domain-specific

A completed classical Numerov computation is a numerical method distinguished by its smooth scalar second-order target, uniform grid, and fixed three-point 1:10:1 rule.

Hierarchy path (1) — routes to 1 parentless root

Numerov's method → Numerical Method

Neighborhood in Abstraction Space¶

Numerov's method sits in a sparse region of the domain-specific corpus (81^st percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2721 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-17

Not to Be Confused With¶

Finite Difference: a difference operator or ingredient does not include the complete target-plus-closure procedure.
Recurrence Relation: a relation between indexed values is not, by itself, a numerical solution method.
Differential Equation or Schrödinger Equation: these identify problems, not this particular numerical procedure.
Cowell: historical association does not establish coextensiveness with the fixed classical scalar rule.
A qualified Numerov variant: a first-derivative, variable-step or matrix treatment needs its own assumptions and derivation.
An order label: a local stencil residual, a global solution error and a runtime growth rate answer different questions.

References¶

[1] Francisco Caruso and Vitor Oguri (2014). “Numerov numerical method applied to the Schrödinger equation”. arXiv:1403.7092v1, 17 March 2014; Portuguese original with an English title and abstract. Zero-based PDF pp. 2–3, §2 equations (1), (2), (7); pp. 4–6, matching procedure; p. 7, §3 and Table 1; pp. 7–10, §4 normal-form discussion and hydrogen generalization through equation (19). Full original author paper inspected. Its broad Euler/Runge–Kutta comparison is not adopted as a universal impossibility. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i

[2] Esmerindo Bernardes (2007). “A direct Numerov sixth order numerical scheme to accurately solve the unidimensional Poisson equation with Dirichlet boundary conditions”. arXiv:0712.1706v1. Zero-based PDF p. 0, equations (1)–(2); p. 1, equations (3)–(7), (11); p. 2, equations (12)–(16); pp. 3–4, equations (17)–(20) and Figure 1. Full original author paper inspected. The header dates the version 11 December 2007; the title page separately displays 4 September 2021. The discrepancy is retained, not resolved. Displayed equation (7) is the weighted recurrence; equation (11) is a first-derivative formula despite later cross-references calling it a recurrence. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n

[3] B. V. Noumerov (1924). “A Method of Extrapolation of Perturbations”. Monthly Notices of the Royal Astronomical Society 84, pp. 592–601. NASA ADS combined facsimile, zero-based PDF pp. 0–9; historical coordinate/perturbation discussion especially printed pp. 592–595. Full original article text inspected with noisy OCR; not used as a symbol-level proof of the modern scalar formula. registry ↩

[4] J. Jackson (1924). “Note on the Numerical Integration of d²x/dt² = f(x,t)”. Monthly Notices of the Royal Astronomical Society 84, opening printed p. 602 / combined PDF zero-based p. 10. Only the separately headed opening page and its Cowell context are cited; this is not full-work access or a proof of method equivalence. registry ↩