Sharp Ascent–Descent Spectral Stability under Strong Resolvent Convergence

Basic Information

• Paper ID: 2511.20971
• Title: Sharp Ascent–Descent Spectral Stability under Strong Resolvent Convergence
• Author: Marwa Ennaceur
• Classification: math.NA (Numerical Analysis), cs.NA, math.FA (Functional Analysis), math.SP (Spectral Theory)
• Submission Date: Submitted to arXiv on November 26, 2025
• Paper Link: https://arxiv.org/abs/2511.20971

Abstract

This paper establishes sharp stability results for the ascent and descent spectra of non-self-adjoint operators under the framework of strong resolvent convergence (SRS). The core quantitative hypothesis is the reduced minimum modulus $\gamma(T-\lambda)>0$, which ensures closed range and facilitates the transmission of the Kaashoek-Taylor criterion through gap convergence of operator graphs. At the essential level, B-Fredholm theory extends stability to powers $(T-\lambda)^m$, provided all intermediate powers satisfy $\gamma((T-\lambda)^j)>0$ ($1\le j\le m$). The author introduces a computable finite element diagnostic quantity $\gamma_h = \sigma_{\min}(M^{-1/2}(A_h-\lambda M)M^{-1/2})$ as a practical surrogate for $\gamma(T-\lambda)$, maintaining consistent positivity even in convection-dominated regimes with stabilized schemes (such as SUPG). Numerical experiments confirm that $\liminf_{h\to0}\gamma_h>0$ is a necessary and sufficient condition for spectral stability, while Volterra-type counterexamples demonstrate that the closed range condition is indispensable for powers.

Research Background and Motivation

1. Core Problem

The paper investigates the stability of fine spectral invariants (ascent and descent) for partial differential operators (particularly non-self-adjoint and singularly perturbed operators) during finite element discretization. Specifically:

• Ascent index $\text{asc}(S)$: the minimal order at which the kernel chain $\ker S \subset \ker S^2 \subset \cdots$ stabilizes
• Descent index $\text{dsc}(S)$: the minimal order at which the range chain $\text{Ran}(S) \supset \text{Ran}(S^2) \supset \cdots$ stabilizes

These invariants are extremely fragile in finite dimensions: the ascent index of a nilpotent matrix $S$ may be 3, but an arbitrarily small perturbation $S+\varepsilon I$ can collapse it to 0.

2. Problem Significance

Classical Example (Paradox in Introduction):

• The continuous one-dimensional transport operator $Lu=u'$ ($u(0)=0$) on $L^2(0,1)$ has closed range and $\text{asc}(L)=1$
• After discretization with second-order central differences, the matrix $A_h$ has eigenvalues clustering near zero (suggesting spectral convergence), but $\text{asc}(A_h)=\infty$ (catastrophic collapse)
• First-order upwind scheme correctly preserves $\text{asc}(A_h)=1$ and maintains uniformly closed discrete range

Root Cause: The central difference scheme has $\gamma_h \sim Ch \to 0$, while the upwind scheme satisfies $\gamma_h \ge c > 0$.

3. Limitations of Existing Methods

• Norm resolvent convergence: fails for rough coefficients ($L^\infty$ potentials) or singular limits (e.g., $\varepsilon \to 0$ in convection-diffusion)
• Mosco convergence: applies only to sectorial operators, cannot handle pure convection problems
• Lack of computable diagnostic quantities to predict success/failure of discretization

4. Research Motivation

To provide a unified theoretical framework:

• Work within strong resolvent convergence (SRS), a natural and broadly applicable setting
• Connect continuous operator theory to discrete numerical analysis through the quantitative condition $\gamma>0$
• Introduce practical diagnostic quantity $\gamma_h$ to guide the choice of stabilized schemes

Core Contributions

1. Sharp Stability Theorem (Theorem 3.2): Under SRS, proves persistence and closedness of ascent/descent spectra with necessary and sufficient condition $\gamma(S)>0$ or $\limsup_n \gamma(S_n)>0$
2. Power Propagation Mechanism (Lemma 3.4): Reveals that graph convergence $G(S_n^m) \xrightarrow{\text{gap}} G(S^m)$ requires all intermediate powers $\gamma(S^j)>0$ ($1\le j\le m$)
3. Computable Diagnostic Quantity: $\gamma_h = \sigma_{\min}(M^{-1/2}(A_h-\lambda M)M^{-1/2})$ serves as a discrete substitute for the continuous condition $\gamma(T-\lambda)>0$, efficiently computable via standard linear algebra tools
4. Three Stability Mechanisms (Theorems 3.15-3.16):
   • Sectorial case: Mosco convergence $\Rightarrow$ SRS $\Rightarrow$ $\gamma_h \ge \gamma(T-\lambda)>0$
   • SUPG stabilization: establishes uniform inf-sup condition $\Rightarrow$ $\gamma_h^{\text{stab}} \ge c>0$ (even as $\varepsilon \to 0$)
   • Pure convection limit: proposes Conjecture 3.17 with numerical evidence
5. Counterexamples and Necessity: Volterra operator (Appendix A) and central difference failure (Proposition 4.12) prove that the closed range condition cannot be relaxed
6. Adaptive Mesh Refinement Algorithm (Algorithm 2): Element marking strategy based on local diagnostic quantity $\gamma_h(K)$

Methodology Details

Task Definition

Given a continuous operator $T$ and its finite element approximation sequence $T_h$, determine when discrete ascent/descent spectra converge to continuous spectra: $\sigma_{\text{asc}}(T_h) \to \sigma_{\text{asc}}(T), \quad \sigma_{\text{dsc}}(T_h) \to \sigma_{\text{dsc}}(T)$

Input:

• Weak form $a(\cdot,\cdot)$ of continuous operator $T$
• Finite element space $V_h \subset H^1_0(\Omega)$
• Spectral parameter $\lambda \in \mathbb{C}$

Output:

• Stability diagnostic $\gamma_h$
• Verification of ascent/descent index preservation

Core Theoretical Framework

1. Reduced Minimum Modulus and Closed Range

For a closed operator $S$, define: $\gamma(S) := \inf\{\|Sx\| : x \in D(S), \, \text{dist}(x, \ker S) = 1\}$

Fundamental Equivalence (Kato 21): $\gamma(S) > 0 \iff \text{Ran}(S) \text{ is closed}$

2. Kaashoek-Taylor Criterion

$\text{asc}(S) < \infty \iff \exists m : \text{Ran}(S^m) \cap \ker S = \{0\}$$\text{dsc}(S) < \infty \iff \exists m : \text{Ran}(S) + \ker(S^m) = H$

These subspace transversality conditions remain stable under gap convergence, provided $\gamma(S)>0$ ensures closed range.

3. Strong Resolvent Convergence and Graph Convergence

SRS Definition: For $\mu \in \rho(T) \cap \bigcap_n \rho(T_n)$, $(T_n - \mu)^{-1}x \to (T - \mu)^{-1}x \quad \forall x \in H$

Key Lemma (Lemma 3.3): Graph gap convergence $G(S_n) \xrightarrow{\text{gap}} G(S)$ (i.e., $\|P_{G(S_n)} - P_{G(S)}\| \to 0$) preserves:

• Gap convergence of kernels/ranges
• Persistence of nontrivial intersections $E \cap F \ne \{0\}$
• Closedness and codimension of sums $E+F$

4. Obstacles to Power Propagation

Lemma 3.4: If $G(S_n^m) \xrightarrow{\text{gap}} G(S^m)$, then necessarily $\gamma(S^j)>0$ for all $1\le j \le m$.

Proof Sketch: If $\gamma(S^k)=0$, the forward graph map $\hat{S}^{k-1}: G(S^{k-1}) \to G(S^k), \quad (x, S^{k-1}x) \mapsto (S^{k-1}x, S^kx)$ lacks a lower bound. Even if $G(S_n^{k-1}) \to G(S^{k-1})$, the image $\hat{S}^{k-1}(G(S_n^{k-1})) = G(S_n^k)$ cannot converge to $G(S^k)$.

Design of Discrete Diagnostic Quantity

Finite Element Framework

• Mass matrix: $M_{ij} = \int_\Omega \varphi_i \varphi_j$
• Stiffness/convection matrix: $A_{ij} = a(\varphi_j, \varphi_i)$
• M-inner product: $\langle u, v \rangle_M := u^* M v$

Discrete Reduced Minimum Modulus

$\gamma_h := \inf_{u \ne 0} \frac{\|(A_h - \lambda M)u\|_{M^{-1}}}{\|u\|_M} = \sigma_{\min}(M^{-1/2}(A_h - \lambda M)M^{-1/2})$

Key Properties:

• M-self-adjoint case: $\gamma_h = \text{dist}(\lambda, \sigma(A_h, M))$ (spectral gap)
• Non-normal case: $\gamma_h \ge \text{dist}(\lambda, W_M(A_h))$ (numerical range distance)

Convergence Guarantee (Proposition 3.14)

If $T$ is an m-sectorial operator and $a_h \xrightarrow{\text{Mosco}} a$, then $\liminf_{h \to 0} \gamma_h \ge \gamma(T - \lambda) > 0$

Technical Innovations

1. Unified Framework: Incorporates both self-adjoint (Schrödinger) and non-self-adjoint (convection-diffusion) operators into a single theory through $\gamma$ rather than spectral gaps
2. Sharpness of Quantitative Conditions:
   • Necessity: Volterra operator and central difference counterexamples
   • Sufficiency: Constructive proof of Theorem 3.2
3. Theoretical Explanation of Stabilization Schemes: Proves that SUPG's uniform inf-sup condition (Proposition B.1) is equivalent to $\gamma_h^{\text{stab}} \ge c>0$
4. Algorithmic Diagnostics:
   • Algorithm 1: Adaptive selection of power $m$
   • Algorithm 2: Mesh refinement based on local $\gamma_h(K)$
5. Computational Strategy (Remark 3.8): Avoids explicit computation of $(A_h - \lambda M)^m$ (causing fill-in and ill-conditioning), instead uses Krylov subspace methods to solve generalized eigenvalue problems

Experimental Setup

Datasets (Model Operators)

1. Schrödinger Operator (Self-Adjoint Case)

$H = -\frac{d^2}{dx^2} + V(x) \text{ on } L^2(0,1), \quad V \in L^\infty(0,1)$

• 1D: uniform mesh, P1 finite elements
• 2D: unit square $(0,1)^2$ and L-shaped domain $(-1,1)^2 \setminus [0,1]^2$
• Test parameters: $V \equiv 25$, $\lambda = -1$ or $\lambda = 25$

2. Convection-Diffusion Operator (Non-Self-Adjoint Case)

$Lu = -\varepsilon u'' + \beta u' + cu \text{ on } (0,1), \quad u(0)=u(1)=0$

• Parameter ranges: $\varepsilon \in [10^{-5}, 1]$, $\beta \in [8, 50]$, $c=0$
• Discretization:
  • Standard Galerkin (central differences)
  • SUPG stabilization: $\delta_K = \delta h_K / \|\beta\|_{L^\infty(K)}$, $\delta \in (0, 1/2]$

3. Pure Transport Operator (Limiting Case)

$L_0 u = u', \quad u(0)=0$

• Central differences vs. first-order upwind
• Verify $\gamma_h \to 0$ (central) vs. $\gamma_h \ge c>0$ (upwind)

Evaluation Metrics

1. Discrete Reduced Minimum Modulus: $\gamma_h = \sigma_{\min}(M^{-1/2}(A_h - \lambda M)M^{-1/2})$
2. Numerical Range Distance: $\text{dist}(\lambda, W_M(A_h))$
3. Condition Number: $\kappa(A_h - \lambda M)$
4. Ascent/Descent Indices: $\text{asc}(A_h)$, $\text{dsc}(A_h)$ (via rank test $\text{rank}(A_h^{m+1}) = \text{rank}(A_h^m)$)
5. Convergence Rate: Slope of $|\gamma_\infty - \gamma_h|$ on log-log plot

Implementation Details

• Mesh Refinement: $h = 2^{-k}$, $k=4,\ldots,9$ (1D); $N \times N$, $N=20,40,80$ (2D)
• Singular Value Computation: ARPACK inverse iteration, tolerance $10^{-8}$
• CPU Time (Table 5): ~31 seconds for $N=10^5$ nodes, $m=3$
• Cholesky Factorization: Preconditioning of mass matrix $M$
• Adaptive Refinement: Marking threshold $\theta = \tau \cdot \min_K \gamma_h(K)$

Experimental Results

Main Results

1. Laplacian Operator (Table 4, Figure 2)

• $H = -\partial_x^2$, $\lambda = 25 \in (\pi^2, 4\pi^2)$
• Discrete eigenvalues: $\zeta_1(h) \downarrow \pi^2 \approx 9.87$, $\zeta_2(h) \downarrow 4\pi^2 \approx 39.48$
• $\gamma_h = \min\{|25-\zeta_1(h)|, |\zeta_2(h)-25|\} \to 14.48$ (Rayleigh-Ritz monotonicity)
• Convergence: $\gamma_h=14.88$ at $h=2^{-4}$, $\gamma_h=15.13$ at $h=2^{-8}$

2. Convection-Diffusion (Table 6, Figure 3)

• $\varepsilon=0.02$, $\beta=8$, $c=0$, $\lambda=-1$
• $\gamma_h$ increases from 6.8 ($h=2^{-4}$) to 8.7 ($h=2^{-8}$)
• Numerical range distance: $\text{dist}(\lambda, W_M(A_h))$ increases from 6.5 to 8.5 (slightly below $\gamma_h$)

3. SUPG Stabilization Robustness (Table 3)

Key Finding: SUPG maintains $\gamma_h \approx 8$ as $\varepsilon \to 0$, while unstabilized scheme has $\gamma_h \to 0$.

4. High Convection Number Case (Figure 4b)

• $\beta=50$, $\varepsilon \in [10^{-3}, 1]$, $h=1/200$
• $\min_\varepsilon \gamma_h \approx 79.38 > 0$ (uniform lower bound)

Ablation Studies

Power Dependence (Table 2, Remark 3.6)

Upwind scheme for $Lu=u'$ ($\lambda=0$):

Explanation: Although $\gamma(A_h) \ge c>0$, we have $\gamma(A_h^m) \sim h^{m-1} \to 0$ (approximating Volterra integral operator), validating Lemma 3.4's necessity.

Ascent Index Verification (Table 11, Example 4.16)

Mixed boundary conditions $T=-d^2/dx^2$, $u(0)=u'(1)=0$, $\text{asc}(T)=2$:

• All $h \in \{2^{-5}, \ldots, 2^{-9}\}$: $\text{asc}(A_h)=2$ (correct)
• $\gamma_h \in [2.41, 2.48]$ (uniformly positive)

Case Studies

Central Difference Failure (Proposition 4.12, Table 10)

$L_0 u = u', \quad \gamma_h \sim 4h \to 0$

• Continuous: $\text{asc}(L_0)=1$, $0 \notin \sigma_{\text{asc}}(L_0)$
• Discrete: $\text{asc}(A_h)=\infty$, $0 \in \sigma_{\text{asc}}(A_h)$ (catastrophic mismatch)

Physical Interpretation: Central differences lack numerical dissipation, and high-frequency modes (wavelength $\sim h$) pollute kernel/range structure.

Upwind Stabilization

$\gamma_h \ge c > 0 \Rightarrow \text{asc}(A_h)=1$ (restores correct behavior)

Two-Dimensional Extension

L-Shaped Domain (Table 8, Figure 5-6)

• Reentrant corner singularity: solution regularity $u \in H^{1+\pi/3-\varepsilon}(\Omega)$
• $\gamma_h$ convergence: 2.31 ($h=2^{-5}$) $\to$ 2.58 ($h=2^{-9}$)
• Convergence rate: $|\gamma_\infty - \gamma_h| \sim O(h)$ (consistent with reduced regularity)
• Local variation: $\gamma_h(K) \in [2.31, 2.56]$ (only 11% difference)

Conclusion: Even with geometric singularity, $\gamma_h$ remains uniformly positive, validating framework robustness.

Related Work

1. Operator Spectral Theory

• Kaashoek 20, Taylor 27: Subspace characterization of ascent/descent
• Berkani 10,11: B-Fredholm theory and essential spectrum
• Kato 21: Perturbation theory and graph convergence

2. Numerical Analysis

• Babuška-Osborn 8: Finite element methods for eigenvalue problems (norm resolvent convergence)
• Brooks-Hughes 12: SUPG stabilization
• Ern-Guermond 14: Generalized Mosco convergence for nonconforming/stabilized schemes

3. Convection-Diffusion Problems

• Numerical Range Methods: Gustafson-Rao 16, Trefethen-Embree 28
• Inf-sup Conditions: Houston et al. 17 (DG methods)

Advantages of This Work

• First to establish sharp necessary and sufficient conditions for ascent/descent spectral stability under SRS
• Unified treatment of self-adjoint and non-self-adjoint operators
• Provides computable diagnostic quantity $\gamma_h$ (existing work mostly stops at abstract convergence)
• Explains spectral stability mechanisms of SUPG and other stabilization schemes

Conclusions and Discussion

Main Conclusions

1. Necessary and Sufficient Condition: $\liminf_{h \to 0} \gamma_h > 0 \iff$ ascent/descent spectra are stable
2. Power Propagation: Requires verification of all intermediate powers $\gamma((T-\lambda)^j)>0$ ($1\le j \le m$)
3. Role of Stabilization: SUPG ensures $\gamma_h^{\text{stab}} \ge c>0$ through uniform inf-sup condition, even as $\varepsilon \to 0$
4. Diagnostic Quantity $\gamma_h$: Efficiently computable (Krylov methods), predicts discretization success/failure

Limitations

1. Pure Convection Limit ($\varepsilon=0$):
   • Conjecture 3.17 lacks rigorous proof (requires convergence framework beyond Mosco theory)
   • Numerical evidence is compelling (Table 3), but theoretical gap exists
2. Mesh Assumptions:
   • Quasi-uniformity requirement
   • Local refinement strategies (Algorithm 2) for non-smooth domains need further verification
3. High-Dimensional Computation:
   • 3D with $N \sim 10^6$: Cholesky factorization cost $O(N^2)$
   • Requires algebraic multigrid (AMG) preconditioning
4. Non-Sectorial Operators:
   • Example 4.15 (Helmholtz) is only preliminary exploration
   • SRS guarantees for general non-sectorial operators unclear

Future Directions

1. Theoretical Refinement:
   • Prove Conjecture 3.17 (graph convergence for pure convection)
   • Extend to block-structured operators (Stokes, Maxwell)
2. Algorithm Optimization:
   • Parallelize $\gamma_h^{(m)}$ computation
   • Rigorous error estimates for adaptive refinement
3. Application Extensions:
   • Semi-discrete stability for time-dependent problems
   • Local linearization for nonlinear operators
4. Software Implementation:
   • Integration into FEniCS/deal.II spectral analysis modules

In-Depth Evaluation

Strengths

1. Theoretical Innovation

• Sharpness: Necessary and sufficient condition $\gamma>0$ cannot be relaxed (Volterra counterexample)
• Universality: Self-adjoint/non-self-adjoint, sectorial/non-sectorial operators in unified framework
• Depth: Reveals $\gamma$ as "steady-state CFL condition"

2. Practical Value

• Operability: $\gamma_h$ computation requires only standard linear algebra (ARPACK)
• Predictiveness: Table 3 clearly distinguishes successful (SUPG) vs. failed (central difference) schemes
• Guidance: Algorithm 2 provides adaptive mesh refinement strategy

3. Experimental Sufficiency

• Multi-dimensional verification: 1D/2D, smooth/singular domains, $\varepsilon \in [10^{-8}, 1]$
• Quantitative comparisons: $\gamma_h$ vs. numerical range distance vs. condition number (Table 9)
• Complete counterexamples: central differences, Volterra, upwind power decay

4. Clarity of Exposition

• Logical chain (Figure 1): numerical failure $\to$ diagnostic quantity $\to$ subspace criterion $\to$ convergence mechanism
• Rich tables: Table 1 (continuous-discrete correspondence), Table 2 (power decay)
• Detailed appendices: Appendix A (necessity), B (SUPG proof), C (numerical stability)

Weaknesses

1. Theoretical Gaps

• Pure Convection Case: Conjecture 3.17 unproven, limiting theoretical completeness
• Non-Mosco Convergence: SRS guarantees for non-sectorial operators depend on specific operators (Example 4.15 is only special case)

2. Computational Cost

• High-Dimensional Bottleneck: $\gamma_h$ computation for 3D large-scale problems ($N>10^6$) insufficiently discussed
• Power Computation: Algorithm 1's stopping criterion $|\gamma_h^{(m+1)}/\gamma_h^{(m)} - 1| < \varepsilon_{\text{tol}}$ may fail for ill-conditioned problems

3. Experimental Limitations

• Limited 2D Results: Only Laplacian reported (Table 7-8), lacking detailed 2D convection-diffusion data
• L-Shaped Domain: Small local $\gamma_h(K)$ variation (11%) doesn't demonstrate significant adaptive refinement advantage

4. Application Scope

• Linearity Assumption: Nonlinear PDEs require local linearization, stability transfer unclear
• Time Dependence: Semi-discrete spectral stability not addressed

Impact

Contribution to Field

• Numerical Analysis: First computable criterion for ascent/descent spectral stability
• Operator Theory: Connects B-Fredholm theory to discrete approximation
• PDE Computation: Explains spectral mechanisms of SUPG and stabilization schemes

Practical Value

• Engineering Applications: Guidance for format selection in convection-dominated problems (e.g., CFD)
• Software Development: $\gamma_h$ as standard diagnostic tool for finite element libraries

Reproducibility

• Code Availability: No open-source implementation provided (but algorithms well-described)
• Data Reproducibility: Tables 4-11 have complete parameters for independent verification

Applicable Scenarios

Best Suited For

1. Non-Self-Adjoint PDEs: Convection-diffusion, Helmholtz, magnetohydrodynamics
2. Singular Perturbation: Boundary layer problems with $\varepsilon \ll 1$
3. Scheme Evaluation: Comparing different stabilization methods (SUPG, DG, upwind)

Requires Caution

1. Pure Convection ($\varepsilon=0$): Theory not fully covered
2. Nonlinear Problems: Requires additional theoretical support
3. Ultra-Large Scale ($N>10^7$): Computational cost may be prohibitive

Not Applicable

1. Time Evolution: Requires extension to semi-discretization
2. Strong Nonlinearity: E.g., Navier-Stokes turbulence
3. Stochastic PDEs: Uncertainty propagation not considered

Key References

1. Kato 21: Perturbation Theory for Linear Operators —— Classical textbook on perturbation theory and graph convergence
2. Kaashoek 20, Taylor 27: Subspace characterization of ascent/descent
3. Berkani 10,11: B-Fredholm theory
4. Brooks-Hughes 12: Original SUPG stabilization paper
5. Ern-Guermond 14: Modern reference for finite element theory
6. Babuška-Osborn 8: Authoritative survey on numerical methods for eigenvalue problems

Summary

This paper establishes the first sharp necessary and sufficient condition for ascent/descent spectral stability of non-self-adjoint operators under strong resolvent convergence, connecting abstract operator theory to practical numerical analysis through the computable diagnostic quantity $\gamma_h$. The theory is rigorous (necessity/sufficiency, counterexamples, B-Fredholm extension), experiments comprehensive (1D/2D, multiple parameters, stabilization schemes), and exposition clear. Main contributions:

1. Reveals $\gamma>0$ as quantitative threshold for spectral stability
2. Explains success mechanisms of SUPG and related methods
3. Provides practical algorithmic tools

Limitations include theoretical gaps for pure convection and high-dimensional computational costs. Overall, this is a high-quality work at the intersection of numerical analysis and operator theory, with significant implications for spectral analysis of PDE discretizations.