Projection geometry and relaxed quasi-orthogonality for inf-sup stable Galerkin methods

• For indefinite, saddle-point, and nonsymmetric problems, exact Pythagoras is generally unavailable, motivating structural substitutes for cumulative Galerkin-increment control.Source: §1 · p. 1
• The paper directly analyzes uniformly bounded compatible Galerkin projections, avoiding stiffness matrices, wavelet bases, and LU-factorizations.Source: §1 · pp. 1–2
• The projection theorem provides explicit finite-window power-type square-function estimates for detail vectors, using an elementary basis-free argument applicable to higher-dimensional detail spaces.Source: §1 · pp. 2–3
• Uniform inf-sup stability with continuity constant C and discrete inf-sup constant γ gives projection bound K ≤ C/γ, making the relaxed exponent and prefactor uniform across targets and nested sequences sharing these constants.Source: §1 · p. 2
• For self-adjoint indefinite problems, positive and negative detail splittings yield a finite-window sign operator, replacing the generally unavailable uniformly bounded infinite splitting.Source: §1 · p. 2
• A fixed self-adjoint involution and uniformly stable nested Galerkin sequence can violate full quasi-orthogonality, with finite-support targets having full-tail ratios growing at least as N^α.Source: §1 · p. 2

• A compatible projection chain has commuting projections with increasing ranges; its detail projections are Δj = Pj − Pj−1 and detail spaces are Ej = ran Δj.Source: §2 · p. 3
• The uniform projection bound K controls the angle between an initial detail block and the consecutive block that follows it.Source: §2.1 · p. 3
• When K = 1, every projection is orthogonal and consecutive detail blocks are orthogonal, so the angle parameter vanishes.Source: §2.1 · p. 3
• Dyadic splitting of consecutive index intervals converts the angle estimate into an upper finite-window square-function estimate for arbitrary vectors yj ∈ Ej.Source: §2.2 · p. 4
• Applying the argument to the adjoint chain and using duality yields a two-sided square-function estimate.Source: §2.3 · pp. 4–5
• The two-sided estimate controls independent detail-space block multipliers and supports the finite-window sign splitting used later.Source: §2.3 · p. 5

2.4. A finite block-LU consequence

• The square-function result also recovers finite block-LU growth for matrices with uniformly controlled leading principal block sections and explicit exponent determined by C/γ.Source: §2.4 · pp. 5–6

3.1. Nested Petrov–Galerkin projections.

• The Galerkin projections form a compatible projection chain with uniform bound K = C_a/γ under nested spaces and uniform discrete inf-sup stability.Equal dimensions ensure existence and uniqueness, while compatibility follows from nested trial and test spaces and Galerkin uniqueness.Source: §3.1 · p. 8

3.2. Quantitative relaxed quasi-orthogonality.

• Relaxed GQO requires C_G(N) = o(N), whereas full GQO requires the stronger condition C_G(∞) < ∞.The paper distinguishes finite-window sublinear growth from uniform control over the full infinite tail.Source: §3.2 · p. 9
• For every starting level and window length N, the squared Galerkin increments are bounded by the initial error times C_qo(K)N^σ_K with σ_K < 1.The estimate applies to u_ℓ = G_ℓu and uses the detail projections and their adjoints in a two-sided square-function argument.Source: §3.2 · p. 9
• At the orthogonal endpoint K = 1, Galerkin increments are mutually orthogonal and the sharp finite-window GQO constant is at most 1.The general explicit constant is therefore not sharp when the projection chain is orthogonal.Source: §3.2 · p. 10
• The exponent and prefactor depend explicitly only on the common projection bound, so the guarantee is uniform over targets and adaptive paths with bounded C_a/γ.The constants are monotone in K, and the explicit prefactor is C_qo(K) = 4K^2(1 + ϑ_K).Source: §3.2 · p. 10

4. Consequences for adaptive methods

• Under estimator reduction, reliability, and quasi-monotonicity, the finite-window estimate yields linear estimator convergence with constants depending only on structural bounds.Proposition 4.1 assumes the listed estimator properties and produces C_lin ≥ 1 and 0 < q_lin < 1.Source: §4 · p. 10
• With the additional estimator, marking, and refinement hypotheses, adaptive algorithms converge with the optimal algebraic rates permitted by their approximation class.The optimality multiplier depends only on continuity, inf-sup, estimator, marking, and refinement constants, not on the target solution.Source: §4 · p. 11

4.1. Taylor–Hood discretizations of the Stokes problem.

• For Taylor–Hood discretizations of Stokes, uniform discrete inf-sup stability lets the projection theorem provide relaxed quasi-orthogonality for every nested admissible sequence.The required adaptive estimator and refinement properties then recover the rate-optimality conclusion.Source: §4.1 · p. 11

4. Consequences for adaptive methods

• The relaxed approach avoids wavelet-type Riesz bases, Jaffard-class decay, matrix representations, and LU-factorization.Source: §4.1 · p. 11
• The same projection-geometric argument applies to nonsymmetric Johnson–Nédélec coupling because its nested conforming spaces satisfy a uniform discrete inf-sup condition.This gives an alternative route to rate optimality without hierarchical coordinates or LU-factor estimates.Source: §4.2 · p. 11

5. Symmetric indefinite problems and finite-window sign geometry

• Each local detail compression B_j is self-adjoint and invertible, enabling the sign operator R_j = sgn(B_j) as an orthogonal involution.The local spectral splitting separates positive and negative detail modes before finite-window assembly.Source: §5 · p. 12
• The finite-window sign operator J_I combines local sign operators across consecutive detail spaces and satisfies a sublinear norm estimate when the window contains n levels.The construction controls the assembled operator through the projection square-function theorem rather than requiring an infinite decomposition.Source: §5 · pp. 12–13
• The detail spaces need not be mutually orthogonal, so J_I need not be self-adjoint in the ambient Hilbert inner product.Its algebraic directness remains sufficient to define the finite-window operator.Source: §5 · p. 12
• Finite-window sign separation recovers relaxed GQO, provided local spectral signs are assembled only over finite intervals with quantitatively controlled block-sign norms.The unresolved infinite positive–negative complementability issue is avoided by finite-window assembly.Source: §5 · pp. 11, 13
• Uniform inf-sup stability alone guarantees only the sublinear finite-window bound, while a stronger full-GQO bound can fail for a fixed self-adjoint involution and fixed stable Galerkin sequence.Source: §5 · p. 13

• Full GQO fails for a fixed self-adjoint involution and fixed nested uniformly stable Galerkin hierarchy.The construction keeps the operator and hierarchy fixed while varying the target in the finite-support family.Source: §6 · p. 13, §6.3 · p. 16, §6.4 · p. 16

6.1. A weighted trigonometric Schauder basis.

• The weighted trigonometric system forms a Schauder basis of Hα, ensuring strong convergence of the associated partial-sum projections.Uniform boundedness of partial sums and convergence on a dense span yield the basis property and Πm → I strongly.Source: §6.1 · p. 14

• The involution J exchanges the two Hilbert-space components and is self-adjoint and unitary, while its action yields the stated alternating orthogonality relations.These relations make the restricted indefinite form nondegenerate on the Galerkin spaces.Source: §6.1 · p. 14, §6.2 · p. 14
• The Galerkin projections are uniformly stable because the continuous inf-sup constant is 1 and the discrete stability constants remain uniformly positive.The hierarchy is therefore a valid uniformly inf-sup stable Galerkin sequence despite failure of full GQO.Source: §6.2 · p. 14
• The exponent’s dependence on the projection bound is essential: as α approaches 1, Kα diverges and the discrete inf-sup constants can approach zero.Thus nearly linear tail growth does not contradict a uniform sublinear exponent under a common projection bound.Source: §6.3 · p. 16

7.1. Uniform stability versus full GQO.

• Uniform inf-sup stability yields sublinear finite-window GQO, whereas full GQO requires additional hierarchical information in general.Similarity to an orthogonal projection chain is given as one stronger sufficient condition for full GQO.Source: §7.1 · p. 17
• The counterexample cannot be similar to an orthogonal projection chain through any boundedly invertible operator.Such similarity would provide the stronger hierarchical structure excluded by the constructed example.Source: §7.1 · p. 17

7.2. Sharp exponents and the LU connection.

• The sharp dependence of the universal exponent on the projection bound remains open.The weighted Fourier example supplies lower bounds at its own bound Kα, but explicit lower bounds at a prescribed K require additional quantitative control.Source: §7.2 · p. 17
• The projection argument bypasses LU factorization for relaxed GQO, but does not replace locality and decay analysis needed for full GQO.Earlier full-GQO proofs use hierarchical Riesz bases and off-diagonal decay, while relaxed estimates need only sublinear finite-factor growth.Source: §7.2 · p. 17

7.3. Further settings.

• Mixed finite element exterior calculus is directly covered when product spaces are nested, harmonic spaces are trivial, and discrete inf-sup constants are uniform.Nontrivial harmonic forms introduce moving discrete harmonic spaces and require additional analysis.Source: §7.3 · p. 17
• Sublinear cumulative nonorthogonality suffices for the standard adaptive conclusions of linear convergence and rate optimality.Full GQO remains useful for uniform infinite-tail estimates but is not required for these adaptive results.Source: §7.3 · p. 17