How statement and proof provenance work
The first chip identifies the source of the statement or construction; the second identifies the source of its local proof or verification.
- Literature-sourced: the exact statement appears in a cited source; only wording and notation differ.
- AI-adapted: a semantically identical restatement of literature-sourced material, modulo indexing, notation, and boundary cases adopted by the library.
- AI-generated: a genuinely novel statement formulated by AI, with no source for the claim itself.
These labels describe origin, not correctness: citations and verification chips remain separate evidence.
Restarted and block eigensolvers — Examples
1 · Prerequisites
- Algebraic Closure, Embeddings, and Separability
- Algebraic Extensions, Extension Degree, and Finite Fields
- Binary Operations, Monoids, Groups and Subgroups
- Composition Series, the Jordan–Hölder Theorem and Solvable Groups
- Computed Schur decompositions and nonnormal spectra
- Computed singular values and rank revealing factorizations
- Congruences, the Integers Modulo n and the Chinese Remainder Theorem
- Construction of the Natural Numbers
- Construction of the Real Numbers via Cauchy Sequences
- Construction of the Real Numbers via Dedekind Cuts
- Continuity, IVT, EVT, and Uniform Continuity
- Cosets, Index and Lagrange's Theorem
- Countability and Uncountability
- Cyclic Groups and Direct Products
- Determinants of Matrices over a Commutative Ring
- Diagonalisation and the Minimal Polynomial
- Direct Matrix Factorisations: LU, Cholesky and QR
- Divisibility, Euclidean Domains, Principal Ideal Domains and Unique Factorisation
- Divisibility, Greatest Common Divisors and Bézout's Identity
- Dual Spaces, Bilinear and Quadratic Forms, and Sylvester's Law of Inertia
- Eigenvalue Iterations and the QR Algorithm
- Eigenvalues, Eigenvectors and the Characteristic Polynomial
- Finite Counting, Factorials and Binomial Coefficients
- Finite-precision direct factorizations and least squares
- Foundations of the Real Numbers for Analysis
- Generalized eigenvalue pencils and QZ methods
- Group Actions, Orbits, Stabilisers and Cayley's Theorem
- Group Homomorphisms and the Isomorphism Theorems
- Hilbert Space Geometry and Riesz Representation
- Ideals, Quotient Rings and the Isomorphism Theorems for Rings
- Inner Product Spaces, Gram-Schmidt, Projections and Adjoints
- Krylov Subspaces, Arnoldi and GMRES
- Limits of Real Functions
- Linear Independence, Bases and Dimension
- Linear Transformations, Rank-Nullity and Quotient Spaces
- Matrices, the Matrix of a Linear Map, and Change of Basis
- Matrix Norms, Condition Numbers and Numerical Stability
- Monotone Sequences, Bolzano-Weierstrass, and Cauchy Completeness
- Normal Subgroups and Quotient Groups
- Order, Zorn's Lemma, and the Axiom of Choice
- Polynomial Rings, the Division Algorithm and Roots
- Primes, Euclid's Lemma and the Fundamental Theorem of Arithmetic
- Relations, Functions, and Quotients
- Restarted and block eigensolvers
- Rings, Subrings, Integral Domains and Fields
- Roots, Rational Powers, and Classical Inequalities
- Sequences and Limits
- Series: Convergence and the Nonnegative Tests
- Simple Field Extensions and the Construction of the Complex Numbers
- Splitting Fields
- Suprema and Infima
- Sylow's Theorems, p-Groups and Nilpotent Groups
- Symmetric Groups, Cycle Decomposition and the Sign Homomorphism
- The Determinant of a Linear Operator, Cofactors and Cramer's Rule
- The Fundamental Theorem of Algebra
- The Fundamental Theorem of Finite Abelian Groups
- The Galois Correspondence
- The Spectral Theorem, Positive Operators and Singular Value Decomposition
- The ZFC Axioms and the Basic Set Constructions
- Topology of ℝ
- Triangularisation, Generalised Eigenspaces and Jordan Canonical Form
- Vector Spaces, Linear Subspaces, Span and Direct Sums
2 · Summary
These examples compute an interior vibration mode with shift-invert, show a block trial space resolving two close modes with one exact Rayleigh-Ritz update, and exhibit a restart shift that geometrically suppresses a wanted eigenvector component. The finite models illustrate the assumptions and failure boundaries; they do not assert general convergence.
3 · Logical flowchart
4 · Definitions, theorems and proofs
None yet.
5 · Examples, counterexamples and false statements
A shift near the wanted eigenvalue suppresses its component under repeated implicit restarts
Example
Let , start from , and seek the largest eigenvalue . Use an implicit restart from dimension to with the fixed shift rule on each cycle, residual tolerance , and restart budget at least when considering cycles. This shift rule is allowed by the procedure; it is not claimed to be a good selection rule.
Supply the following finite-return projected solver for this example. For a Hermitian matrix , if , return success with the ordered pairs and . If , set
and return success with the ordered pairs and . Return unavailable on other shapes or non-Hermitian inputs. Each displayed vector is nonzero, direct multiplication verifies its eigenpair equation, and the two factors reproduce the quadratic characteristic polynomial. This supplied finite-dimensional example provider is not a general eigensolver or a new restart method.
Verification
Given: The exact diagonal matrix, starting vector, target, and shift above.
[F1] Exact implicit Arnoldi restart applies its shift polynomial and preserves an Arnoldi residual identity: for , the exact implicit restart retains the one-dimensional filtered Krylov space generated by .
[F2] Eigenvalues, eigenvectors, eigenspaces , and the spectrum of an endomorphism: an eigenvector is a nonzero vector satisfying .
Every projected in this example is Hermitian because , so the supplied two-dimensional solver returns valid complete projected data in finite exact work. By [F1], . The squared component weights are proportional to , so the Rayleigh quotient of the restarted vector is , whereas the initial quotient is . The wanted component-to- component ratio falls from 1 to 1/19.
Repeating the same exact restart multiplies the component by and the component by at every cycle. Hence after cycles their magnitude ratio is , which tends to zero. This shift choice suppresses the wanted eigendirection and gives no general convergence claim.
For each finite , the restart vector is proportional to , with all coordinates nonzero because none of vanishes. If a Ritz residual were zero, its lifted vector would satisfy ; the diagonal equations and distinct force to be proportional to some . No lies in : representing one would require for the other two distinct eigenvalues, hence . Also are independent because their coordinate determinant is the product of the three nonzero coordinates and the Vandermonde determinant for . Thus no pair is accepted at tolerance zero and the space is not invariant before each scheduled restart.
In a separate boundary run with shift , the filter has and , so the wanted component is removed exactly in one filtering step. The filtered vector is still nonzero, so this is not filtered breakdown; it is a shift-selection failure.
A two-vector LOBPCG block resolves a clustered pair in one finite Rayleigh-Ritz update
Example
Consider the mass-normalized vibration matrix with and preconditioner . The two wanted lowest modes are close, while the next mode is separated. Run the width- algorithm with residual tolerance and budget . Start with the orthonormal block
Verification
Given: The matrix, metric, preconditioner, and two starting vectors above.
[F1] Finite-budget block LOBPCG for a Hermitian-definite pencil with explicit rank-loss status: LOBPCG forms , , uses the trial span of , and selects its q smallest Ritz values after Rayleigh-Ritz.
[F2] A successful LOBPCG Rayleigh-Ritz update cannot increase the selected Ritz values: on a successful update retaining , each selected Ritz value cannot increase.
[F3] A principal-angle bound encloses the largest selected generalized Ritz value across a spectral gap: if the first q generalized eigenvalues are separated and is the maximum principal angle, then .
The vectors are orthonormal and have disjoint coordinate supports. Their initial Rayleigh quotients are and . Each has two nonzero coordinates associated with distinct diagonal entries of , so ; also , and lies in the same two-coordinate plane as . Thus are independent in that plane, and the disjoint planes make rank four.
Since , and the first trial space is all of . Its Rayleigh-Ritz values are exactly , and the selected block is . The update is successful because the trial rank is four; the returned residuals are zero, so the algorithm returns converged. In particular, and , as required by [F2].
For , every unit vector in has squared norm in , so . The strict gap holds, and [F3] gives , attained by the computed . The block captures both clustered modes; a one-dimensional trial space cannot contain this two-dimensional invariant subspace.
Shift-invert isolates an interior vibration mode and records solve error
Example
Use the exact shift-invert branch with start , Arnoldi dimension and residual tolerance . Use the mass-normalized three-mode vibration matrix and , and seek its interior eigenvalue with shift . The shifted matrix is invertible because . In a separate harmonic-branch invocation, use the same and , trial basis , QZ sweep budget , and QZ tolerance .
Verification
Given: , , , the exact target vector , and the separate harmonic invocation inputs , , and .
[F1] Shift-invert and harmonic Ritz extraction with solve and pencil status: in the exact shift-invert branch, a Ritz value maps to .
[F2] Shift-invert and harmonic Ritz extraction with solve and pencil status: the harmonic branch uses the projected pencil .
[F3] Shift-invert and harmonic Ritz extraction with solve and pencil status: a finite QZ diagonal pair maps to .
[F4] Shift-invert maps eigenvalues exactly and transfers an inexact solve residual to the original eigenpair: , with and .
Here and . Thus the interior mode 4 maps to , whose modulus exceeds the other transformed eigenvalue moduli and ; . In the separate harmonic invocation, , so and . The QZ input is this regular one-by-one pencil, already upper triangular with a nonzero diagonal pair, so the QZ procedure's computed condition holds and its status is qz-computed; . The scalar equation gives and the same value .
For an inexact solve take and . Then and , while the mapped value is . Its original residual is , which equals by [F4].
Sources
- Yousef Saad, Numerical Methods for Large Eigenvalue Problems, second edition, §7.3.1
- Peter Arbenz, Lecture Notes on Solving Large Scale Eigenvalue Problems, §11.2
- Peter Arbenz, Lecture Notes on Solving Large Scale Eigenvalue Problems, §13.6
- Andrew Knyazev, Recent implementations, applications, and extensions of LOBPCG, §1
- Yousef Saad, Numerical Methods for Large Eigenvalue Problems, second edition, §8.1
- Peter Arbenz, Lecture Notes on Solving Large Scale Eigenvalue Problems, §§11.4 and 12.6