Solvability Complexity Index (SCI)

• SCI is a classification invariant that quantifies the minimum number of nested limits needed to compute a target quantity, such as an operator’s spectrum.
• It stratifies computational problems into levels (SCI 0, SCI 1, SCI 2, etc.), delineating precise thresholds for algorithmic convergence and error control.
• The framework drives algorithm design in spectral theory and infinite-dimensional analysis, revealing intrinsic computability boundaries and informing impossibility results.

The Solvability Complexity Index (SCI) is a classification invariant for computational problems that measures the minimal number of nested limiting processes required to algorithmically compute a desired quantity, such as the spectrum of an operator, to a prescribed accuracy. Originating in the context of computational spectral theory and infinite-dimensional analysis, the SCI hierarchy rigorously delineates boundaries between problems solvable by finite, one-limit, double-limit, and more iterated limiting algorithms, and underpins both algorithm design and impossibility results for broad classes of mathematical and computational tasks.

1. Formal Definition of the Solvability Complexity Index

A computational problem in the SCI framework is a quadruple

$\{\Xi,\;\Omega,\;(\mathcal{M},d),\;\Lambda\}$

where $\Omega$ is the domain of inputs (e.g., operators, transformations), $\Lambda$ is a set of admissible evaluation maps (oracle queries), $(\mathcal{M},d)$ is the target metric space (e.g., closed subsets of $\mathbb{C}$ with the Hausdorff or Attouch–Wets metric), and $\Xi:\Omega\to\mathcal{M}$ is the problem function.

A tower of algorithms of height $k$ is a family

$\Gamma_{n_k,\dots,n_1}:\Omega\to\mathcal{M}$

such that

$$\Xi(\omega) = \lim_{n_k\to\infty} \cdots \lim_{n_1\to\infty} \Gamma_{n_k,\dots,n_1}(\omega)$$

for all $\omega\in\Omega$, with each $\Gamma_{n_k,\dots,n_1}$ being a general algorithm depending on finitely many oracle evaluations. The SCI of $\Xi$, denoted $\mathrm{SCI}(\Xi)$ (or with subscript $A$ for arithmetic, $G$ for general, depending on computation model), is the minimal such $k$ if it exists, otherwise $\mathrm{SCI}(\Xi)=\infty$ (Ben-Artzi et al., 2015, Colbrook et al., 2019, Gazdag et al., 2022).

2. SCI Hierarchy and Interpretational Framework

The SCI hierarchy stratifies computational problems:

• SCI 0 (finite algorithm): The problem can be solved exactly, with no limiting process.
• SCI 1 (single-limit): There exists a sequence of algorithms $\Gamma_n$ such that, for all $\epsilon>0$, $d(\Gamma_n(\omega),\Xi(\omega))<\epsilon$ for sufficiently large $n$ (error-controlled uniform convergence).
• SCI 2 (double-limit): Two nested limits are required; i.e., no single-limit algorithm achieves convergence or error control uniformly, but $$\lim_{n_2\to\infty} \lim_{n_1\to\infty}\Gamma_{n_2,n_1}(\omega)=\Xi(\omega)$$.
• SCI $k$: $k$-fold nested limits are required.
• SCI $\infty$: The problem is not computable by any finite tower.

Canonical inclusion (strict): $$\Delta_0^\alpha \subsetneq \Delta_1^\alpha \subsetneq \Delta_2^\alpha \subsetneq \cdots,\quad \alpha\in\{A,G\}$$ where $\Delta_k^\alpha$ is the set of all computational problems with SCI $\leq k$ in model $\alpha$ (Ben-Artzi et al., 2015, Colbrook et al., 2019, Gazdag et al., 2022, Sorg, 17 Jan 2026). The SCI paradigm applies equally in Turing, arithmetic, and more general information-based models, with specific differentiations arising from algorithmic or information-theoretic lower bounds.

3. SCI in Infinite-Dimensional Spectral Theory

SCI has produced sharp classifications for operator spectral problems:

• Self-adjoint operators with convex essential spectrum have SCI $=1$ for spectrum computation, as one can use single-limit spectral approximations via the Galerkin method without spectral pollution (Rösler, 2019).
• General bounded operators: Computing the spectrum typically has SCI $=3$ (triple limit); self-adjoint or normal operators reduce to SCI $=2$; adding dispersion or resolvent growth bounds reduces this further to SCI $=1$ (Ben-Artzi et al., 2015, Colbrook et al., 2019).
• Spectral gap and decision problems: Tasks such as testing for the existence of a spectral gap or spectrum intersection with a compact set often require SCI $=2$ or higher (Colbrook et al., 2019).

For Koopman operators acting on $L^p(\mathcal{X},\omega)$ ($1<p<\infty$):

• The computation of the $\varepsilon$-approximate point spectrum $\sigma_{\text{app}}^\epsilon$ has SCI $=2$ for general continuous $F$, improving to SCI $=1$ for subclasses with a known modulus of continuity. Computation of the (true) approximate point spectrum $\sigma_{\text{app}}$ requires one further limit; e.g., SCI $=3$ ($=2+1$) for the general class, and SCI $=2$ for the known modulus case. Lower bounds are established via rotation/cycle counterexamples (Sorg, 19 Sep 2025).

SCI Classification Table for Koopman Operators (selected cases)

Category of F	$\mathrm{SCI}_G(\sigma_{\text{app}}^\epsilon)$	$\mathrm{SCI}_G(\sigma_{\text{app}})$
Continuous + modulus $(\Omega^\alpha)$	1	2
Measure preserving $(\Omega^m)$	2	3
General continuous $(\Omega)$	2	3
Both $(\Omega^{\alpha,m})$	1	2

(Sorg, 19 Sep 2025)

4. SCI, Limit Algorithms, and Error Control

Each SCI level has precise algorithmic implications:

• SCI 1 problems permit the design of uniform, error-controlled numerical schemes, often used in computer-assisted proofs and rigorous numerical analysis (Colbrook et al., 2019).
• SCI 2 (and higher) problems require nested approximation procedures. Convergence may occur “from below” ($\Sigma_k$), “from above” ($\Pi_k$), or in a “symmetric” ($\Delta_k$) fashion depending on monotonicity and error estimates.
• Algorithmic framework: For spectral approximation, the canonical approach is a (nested) finite-section method: discretizing on expanding subspaces, computing residuals or minimal singular values on a grid, updating via Hausdorff or Attouch–Wets convergence, and controlling for spectral pollution (Sorg, 19 Sep 2025, Rösler, 2019, Colbrook et al., 2019, Ben-Artzi et al., 2015).

5. Lower Bounds, Impossibility, and Endpoint Phenomena

The SCI theory also identifies sharp computational lower bounds:

• **Impossibility results:} For some problems (e.g., $L^\infty$ spectral computation for Koopman operators), no finite tower of algorithms (even information-based) suffices—$\mathrm{SCI}_G=\infty$—arising from the failure of separability and dense quadrature structure; reductions from descriptive set theory or Borel hierarchy demonstrate this (Sorg, 17 Jan 2026).
• **Generalised hardness of approximation (GHA):} A phase-transition phenomenon appears in underdetermined neural network training: below an accuracy threshold $\epsilon_0$, neither finite nor limiting algorithms suffice (even randomized). Above $2\epsilon_0$, error-controlled single-limit algorithms are possible. Thus, the SCI hierarchy exposes information-theoretic phase transitions in approximation, unrelated to classical P vs. NP intractability (Gazdag et al., 2022).
• **Prototype decision problems and Borel complexity:} There exist canonical problems with arbitrary finite SCI levels (e.g., $\Xi_m$ with SCI $=m$) constructed via nested quantification over matrix entries, providing reusable hardness reductions for future SCI lower bounds (Sorg, 17 Jan 2026).

6. Methodological Impact and Algorithm Design

SCI provides a blueprint for algorithm design and diagnosis:

• It systematizes the construction of adaptive, finite-evaluation algorithms tailored to sit at the minimal necessary height given structural input properties (e.g., symmetry, regularity, moduli, invariants).
• The SCI framework reveals precisely why standard discretizations fail—e.g., finite-section methods produce “spectral pollution” for problems at SCI $=2$ or higher, as they implicitly assume SCI $=1$ structure (Colbrook et al., 2019, Sorg, 19 Sep 2025).
• In classes where SCI bounds are sharp, the approach delivers convergence guarantees and avoids “invisible spectrum” or undecidable events.
• The index is model-independent, making it suitable for both classical (Turing, arithmetic) and general oracle-access computational paradigms, and aligns with logical classification via the Borel and Weihrauch hierarchies (Gazdag et al., 2022, Sorg, 17 Jan 2026).

7. Connections, Extensions, and Ongoing Research

• **Relation to computational complexity theory:} SCI formalizes a “hierarchy of unsolvability” for analytic and infinite-dimensional problems, complementing the discrete theory. GHA illustrates that information-based phase transitions can be sharper and more universal than Turing-classical barriers (Gazdag et al., 2022).
• **Type-2 and Weihrauch theory:} In the arithmetic model, SCI height corresponds to Weihrauch reducibility to iterated limits ($\lim^{(n)}$), but the general (information-based) SCI can be strictly higher due to lack of continuity or representation constraints (Sorg, 17 Jan 2026).
• **Applications beyond spectra:} The SCI methodology has already impacted polynomial root finding (e.g., Smale’s problems, McMullen and Doyle–McMullen towers), inverse problems, neural network training in AI optimization, and the computational theory of PDEs (Ben-Artzi et al., 2015, Colbrook et al., 2019, Gazdag et al., 2022).

The Solvability Complexity Index thus constitutes a central framework for the metatheory of computation in analysis and operator theory, allowing definitive, structure-sensitive statements about the computability and inherent limitations of fundamental problems in mathematics and computational science.