1 Definition and Basic Properties
1.1 Inner product conventions
Bi-orthogonality is defined relative to an inner product (or sesquilinear pairing) on a vector space. Let the pairing be written as ⟨·,·⟩, linear in one argument and conjugate-linear in the other, according to the chosen convention. In finite-dimensional settings this is typically the standard complex inner product, but the notion extends to general Hilbert or Banach space frameworks where an appropriate pairing is available.
1.2 Right and left vector sets
A bi-orthogonal system consists of two families of vectors (or functions) usually described as a right set and a left set. The “right” vectors are arranged as {u₁, …, uₙ}, while the “left” vectors are arranged as {v₁, …, vₙ}. The naming reflects how they arise naturally for non-Hermitian operators: right eigenvectors are associated with action by the operator, while left eigenvectors are associated with the adjoint action or, more generally, with eigenvectors of a related operator.
1.3 Bi-orthogonality condition
The defining requirement is that each left vector pairs with exactly one matching right vector, producing a Kronecker delta: \[ \langle v_i, u_j\rangle = \delta_{ij}. \] Here δᵢⱼ equals 1 when i=j and 0 otherwise. This “cross” orthogonality differs from standard orthogonality, which pairs vectors within a single set.
1.4 Normalization and scaling freedom
The condition fixes relative scaling only up to reciprocal factors. If a pair (uᵢ, vᵢ) satisfies ⟨vᵢ, uᵢ⟩ = 1, then replacing uᵢ by αuᵢ and vᵢ by α^{-1}vᵢ (with α nonzero) preserves the bi-orthogonality relations. More generally, one can allow ⟨vᵢ, uⱼ⟩ = cᵢδᵢⱼ with nonzero constants cᵢ; rescaling can then convert these to the unit normalization used above.
1.5 Relation to ordinary orthogonality
If the left and right sets coincide, bi-orthogonality reduces to ordinary orthonormality: ⟨uᵢ, uⱼ⟩ = δᵢⱼ. Thus, bi-orthogonality can be seen as a generalization of orthonormal systems suited to contexts where a single set cannot simultaneously diagonalize or represent operator structure due to non-Hermitian behavior.
2 Construction in Finite-Dimensional Vector Spaces
2.1 Bi-orthogonal bases
In a finite-dimensional inner product space of dimension n, a bi-orthogonal system of n pairs typically forms a basis in each family. If {u₁, …, uₙ} is linearly independent, one can seek vectors {v₁, …, vₙ} such that ⟨v_i, u_j⟩ = δᵢⱼ for all i,j. When such a choice exists, both sets serve as “dual” bases under the given pairing.
2.2 Dual bases viewpoint
Bi-orthogonality is closely tied to dual spaces. Given a basis {u_j}, the dual basis {f_j} in the dual space is defined by f_j(u_k)=δⱼk. Under an inner product, each functional f_j can be represented as an inner product with some vector v_j: f_j(x)=⟨v_j, x⟩ (Riesz representation in Hilbert spaces). In that representation, the condition ⟨v_i, u_j⟩ = δᵢⱼ appears naturally.
2.3 Matrix formulation using Gram matrices
Let U be the matrix whose columns are the right vectors u_j, and let V be the matrix whose columns are the left vectors v_i. For complex spaces with the convention that ⟨x,y⟩ = x* y (conjugate-transpose), the bi-orthogonality condition becomes a matrix equation: \[ V^* U = I. \] A common equivalent formulation uses the Gram matrix G associated with the right vectors: G_{ij}=⟨u_i,u_j⟩. When left vectors are chosen via the dual basis representation, one obtains relations of the form \[ V = U (U^* U)^{-1} \] (up to convention-dependent conjugations). This shows that existence and construction reduce to invertibility of the relevant matrix.
2.4 Existence and uniqueness criteria
If the right vectors {u_j} form a basis (hence U is invertible), then the bi-orthogonal left vectors are uniquely determined by the requirement ⟨v_i, u_j⟩ = δᵢⱼ, provided the pairing is nondegenerate. In coordinates, this corresponds to solving a linear system for v_i. If the right family is not a basis (is linearly dependent), the Kronecker delta constraints are generally inconsistent or define only partial information.
2.5 Degeneracy and choice of compatible bases
In the presence of degeneracies, multiple bi-orthogonal pairs can be constructed. If an operator or constraint only determines the span of certain vectors, freedom remains to choose different bases within that span while maintaining bi-orthogonality with an appropriately modified dual set. Practically, one often selects a convenient normalization or imposes additional structure (such as smooth dependence on parameters) to fix the choice.
3 Bi-orthogonality for Linear Operators
3.1 Left and right eigenvectors
For a linear operator A on a finite-dimensional space, right eigenvectors satisfy \[ A u_i = \lambda_i u_i, \] while left eigenvectors are associated with the adjoint operator A^* in the Hilbert setting, typically in the form \[ A^* v_i = \overline{\lambda_i}\, v_i. \] With these definitions, bi-orthogonality between left and right eigenvectors can be arranged when eigenvalues and eigenspaces are handled appropriately, especially when A is not normal.
3.2 Eigenvalue problems and eigenvector pairing
When eigenvalues are simple (non-degenerate), left and right eigenvectors can often be chosen so that \[ \langle v_i, u_j\rangle = 0 \quad \text{for } i\neq j, \qquad \langle v_i, u_i\rangle \neq 0. \] Rescaling then yields the unit bi-orthogonality condition. This pairing is central to converting operator actions into diagonal-like expressions even when A does not admit an orthonormal eigenbasis.
3.3 Non-Hermitian operators and why bi-orthogonality appears
A non-Hermitian operator generally lacks a complete orthonormal set of eigenvectors; normality is the property that would guarantee such behavior. Without orthonormal eigenvectors, expansions become awkward if one attempts to use the standard inner product directly. Bi-orthogonality resolves this by introducing two compatible families—one capturing the action of A and the other capturing the action of a related adjoint or dual operator—so that coefficients in expansions can still be extracted cleanly.
3.4 Diagonalizable versus non-diagonalizable cases
If A is diagonalizable, there exist bases of right eigenvectors spanning the space, and a corresponding dual set of left eigenvectors can be chosen to satisfy bi-orthogonality. In contrast, if A is not diagonalizable (has Jordan blocks), eigenvectors alone may not span the space. In such cases, bi-orthogonal constructions are extended using generalized eigenvectors and associated dual chains, preserving a generalized delta pairing among these larger families.
3.5 Spectral projectors from bi-orthogonal data
When A is diagonalizable with eigenpairs {u_i, v_i}, one can define rank-one spectral projectors: \[ P_i = u_i \, v_i^*, \] again up to convention. These projectors satisfy \[ P_i P_j = \delta_{ij} P_i, \qquad \sum_i P_i = I, \qquad A = \sum_i \lambda_i P_i. \] The formulas rely precisely on the bi-orthogonality condition, since it ensures the correct cancellation when composing projectors.
4 Completeness, Resolution of the Identity, and Expansions
4.1 Completeness of bi-orthogonal systems
A bi-orthogonal system is complete if the right vectors span the space (or the relevant subspace), equivalently if the corresponding left vectors span the dual space in a compatible way. In finite dimensions, if {u_i} is a basis and ⟨v_i,u_j⟩=δᵢⱼ holds, then completeness follows immediately. Completeness is crucial for expansions, since it guarantees that every vector can be expressed in terms of the right family.
4.2 Resolution of the identity
Given complete bi-orthogonal sets, the identity operator can be written as \[ I = \sum_i u_i \, v_i^*, \] with convergence understood appropriately in infinite-dimensional settings. This “resolution of the identity” plays the same conceptual role as in orthonormal Fourier-type expansions, but with left vectors replacing the standard conjugate transpose of the right vectors when the operator is non-Hermitian.
4.3 Coordinate expansions using dual sets
For any vector x, completeness yields an expansion \[ x = \sum_i u_i \, c_i, \] where coefficients c_i are determined by pairing x with the left vectors: \[ c_i = \langle v_i, x\rangle. \] This follows from applying the resolution of the identity to x and using ⟨v_i, u_j⟩=δᵢⱼ to isolate the matching term.
4.4 Coefficient extraction formulas
The coefficient formula is often summarized as: \[ x = \sum_i u_i \langle v_i, x\rangle. \] Similarly, for an operator B, one may express its action in the bi-orthogonal basis via matrix elements \[ B_{ij} = \langle v_i, B u_j\rangle, \] which provides a consistent framework for transforming between operator and coordinate representations.
4.5 Practical computation in examples
In computations, bi-orthogonal pairs arise by first obtaining eigenvectors of A and of the adjoint (or the left eigenvectors via equivalent transposed eigenproblems). One then rescales each pair so that ⟨v_i,u_i⟩=1 and verifies cross-orthogonality numerically. In settings with nearly degenerate eigenvalues, small numerical errors can spoil the exact delta structure, so normalization procedures and conditioning checks are commonly used.
5 Algebraic Connections and Alternative Formulations
5.1 Relation to dual spaces and linear functionals
Bi-orthogonality can be described purely in algebraic terms without referencing adjoints: it is the relationship between a basis {u_j} of a vector space and a dual basis {f_i} of its dual space, with f_i(u_j)=δᵢⱼ. The inner product is then one mechanism for identifying dual functionals with vectors, turning linear functional pairing into an inner product pairing.
5.2 Bra–ket notation (dual vectors)
| In physics-oriented notation, one writes right vectors as kets | u_i⟩ and left vectors as bras ⟨v_i | , emphasizing that bras represent dual objects acting on kets through the pairing: |
|---|
\[
| \langle v_i | u_j \rangle = \delta_{ij}. |
|---|
\]
| This notation aligns with the operator projector form P_i = | u_i⟩⟨v_i | and can simplify expression of expansions and spectral decompositions. |
|---|
5.3 Similarity transforms and invariance under change of basis
Bi-orthogonality is stable under appropriate changes of basis that preserve the pairing structure. If one transforms right vectors by an invertible matrix and transforms left vectors by the inverse adjoint (or inverse transpose under real conventions), the relation ⟨v_i,u_j⟩=δᵢⱼ remains true. As a result, operator representations built from bi-orthogonal data behave consistently under similarity transforms, even when the underlying operator is not normal.
5.4 Connection to generalized inverses and conditioning
Because bi-orthogonal constructions effectively implement left/right duality, they relate to generalized inverses such as the Moore–Penrose pseudoinverse in special circumstances. In non-Hermitian problems, stable computation depends strongly on how well-conditioned the bi-orthogonal bases are: if left and right eigenvectors become nearly linearly dependent in an appropriate sense, coefficients extracted from ⟨v_i,x⟩ can become large and sensitive to perturbations.
5.5 Orthogonality in indefinite or weighted inner products
The notion can also be extended when the pairing is modified by a weight or uses an indefinite inner product. For example, with a positive-definite weight operator W defining ⟨x,y⟩_W = ⟨x,Wy⟩, bi-orthogonality is defined with respect to ⟨·,·⟩_W. More general indefinite settings require care because nondegeneracy may replace positivity, but the cross-orthogonality condition still provides a structured way to build dual families and derive expansion formulas.