Catalytic Logspace: Complexity & Applications

• Catalytic logspace is a computational model featuring an O(log n) clean tape and a poly(n) catalytic tape that must be exactly restored.
• It offers increased computational power, enabling efficient deterministic algorithms for problems like bipartite matching and linear matroid intersection.
• Innovative techniques such as reversible compression and code-based methods connect this model to circuit simulation and broader complexity classes.

A catalytic logspace computational model consists of a deterministic Turing machine equipped with two distinct work tapes: an ordinary "clean" work tape with length $O(\log n)$, and a much larger "catalytic" tape of length $\mathrm{poly}(n)$, the contents of which $(\tau)$ must be restored exactly upon completion of the computation. This strict restoration mandate distinguishes catalytic logspace ($CL$) from conventional space-bounded complexity classes; the catalytic tape initially contains an arbitrary bitstring and can be used freely during the computation but must be preserved. Catalytic logspace machines possess provable computational power exceeding that of traditional logspace machines ($L$), yielding efficient solutions for problems such as bipartite matching and linear matroid intersection that are not known to be solvable in $L$.

1. Formal Model and Fundamental Properties

A catalytic Turing machine $M$ on input $x\in\{0,1\}^n$ takes a clean work tape of length $s(n)=O(\log n)$ and a catalytic tape of length $c(n)=\mathrm{poly}(n)$, filled with an adversarial string $\tau \in \{0,1\}^{c(n)}$. The machine must return the catalytic tape to its original configuration $\tau$ upon halting. The class $\mathrm{CSPACE}[s(n),c(n)]$ consists of all languages decided by catalytic machines with parameters $s(n),c(n)$; catalytic logspace is defined as

$$CL = \bigcup_{d\geq 1} \mathrm{CSPACE}[d \log n, n^d].$$

The restoration requirement is essential: computation must be correct for every initial tape, and no information leakage or "scrubbing" is permitted.

2. Lossy Catalytic Logspace

Gupta et al. established the variant lossy catalytic logspace $LCL[e]$, where the restoration requirement is relaxed to allow up to $e$ errors (Hamming distance $\leq e$) on the catalytic tape. The core characterization is

$$LCSPACE[s,c,e] = CSPACE[\Theta(s + e \log c), \Theta(c)],$$

i.e., $e$ errors on a catalytic tape of length $c$ are computationally equivalent to upgrading the clean workspace by $e \log c$ bits. The equivalence is bidirectional, and both BCH encoding (for simulating errors via extra workspace) and a combinatorial "chessboard mapping" (workspace via Hamming errors) underpin the conversions (Folkertsma et al., 2024). Consequently, catalysis is robust to $O(1)$ errors but collapses for superconstant $e$, as otherwise one would have an unexpected derandomization:

$SPACE[e \log n] \subseteq ZPP.$

This essentially fixes the boundary of lossy catalytic computation.

3. Collapse Results: CL vs. Nondeterminism and Randomness

A sequence of reductions (Cook et al.'s compress-or-compute) establishes that catalytic logspace, nondeterministic catalytic logspace, and randomized catalytic logspace coincide:

$CL = CNL = CPrL.$

This is achieved via reversible Euler tours on configuration graphs that efficiently compress parts of the catalytic tape or formulate small reachability instances. The main technical theorem is

$$CBSPACE[s, c] \subseteq CSPACE[O(s), O(c)]^{BSPACE[O(s)]},$$

where $$B \in \{N, \mathrm{coN}, \mathrm{BP}, \mathrm{Pr}\}$$. Consequently, catalytic analogues of Savitch's theorem ($NSPACE[s] \subseteq SPACE[s^2]$) and Immerman–Szelepcsényi ($NL = coNL$) hold with minimal overhead in the catalytic regime (Koucký et al., 11 Apr 2025). Randomness and nondeterminism add no power and all collapses are unconditional for polynomial-size catalyst and logarithmic free space.

4. Algorithmic Applications: Matching and Matroid Intersection

Catalytic logspace supports algorithms for combinatorial optimization problems that are inaccessible to conventional logspace. Agarwala and Mertz have shown that bipartite matching lies in $CLP$, the polynomial-time variant of catalytic logspace (machines halting in poly-time for every catalytic tape) (Agarwala et al., 14 Apr 2025). Their isolation-based method, combined with catalytic "compress-or-random" techniques, yields a deterministic, $O(\log n)$-space, $O(n^{O(1)})$-time algorithm for maximum matching, providing the first polynomial-time logspace algorithm for this central problem. Linear matroid intersection, a strict generalization of matching, has also been placed in $CLP$ via reduction and combinatorial exchange graph techniques (Agarwala et al., 8 Sep 2025). All catalytic routines guarantee restoration of the tape via compression/decompression schemes.

5. Beyond Logspace: Register Programs and Circuit Simulations

Catalytic machines are functionally equivalent to "clean" register programs over rings. The register program formalism enables simulation of uniform threshold circuits ($TC^1$) within $CL$, as demonstrated by Buhrman et al. (Alekseev et al., 24 Apr 2025). New advances extend simulation to $SAC^2$ (depth $\log^2 n$ circuits) with slight superpolynomial catalytic space, and matrix powering—pushing the frontier toward full containment of $NC^2$ in $CL$. These results are grounded in algebraic polynomial representations, recursive register composition, and layer merging for circuit depth compression. Matrix operations (product and inversion) over $\mathrm{poly}(n)$-size finite fields are implementable in-place with catalytic assistance (Cook et al., 13 Oct 2025).

6. Almost-Catalytic Computation and Code-Theoretic Methods

A relaxation called almost-catalytic computation allows restoration only when the initial catalytic tape satisfies a predicate (membership in a fixed set $A$). It is shown that $$\mathrm{ACL}(A) \cap \mathrm{ACL}(\bar{A}) \subseteq ZPP$$, and with specialized code-based constructions one can place broad classes such as $\mathrm{DSPACE}(n^k)$ and $\mathrm{DSPACE}(\log^k n)$ in almost-catalytic space for certain $A$—notably those defined via linear codes with specified random projection and subcube partition complexity (Bisoyi et al., 2024). However, covering radius lower bounds constrain further progress without new combinatorial or nonlinear ideas.

7. Structural Barriers and Complexity Relationships

Catalytic logspace is strictly more powerful than $NL$ and $TC^1$ but not known to contain $\mathrm{P}$ or $\mathrm{NC}$. Oracle separation and cryptographic barriers show that $\mathrm{CL} \subseteq \mathrm{P}$ would require non-relativizing techniques; there is an oracle $O$ for which $\mathrm{CL}^O = \mathrm{EXP}^O$ (Cook et al., 13 Oct 2025). Hard search problems such as lossy coding are in $\mathrm{searchCL}$ but not known to be in $\mathrm{P}$, further illustrating the nontriviality of the model.

Catalytic Logspace Class	Tape Restoration	Example Problem in Class
$CL$	Exact	Bipartite Matching (Agarwala et al., 14 Apr 2025)
$LCL[e]$	$\leq e$ errors	(Equivalent to $CL$ for $e=O(1)$)
$CLP$	Exact & Poly-time	Linear Matroid Intersection (Agarwala et al., 8 Sep 2025)
$\mathrm{ACL}(A)$	On predicate	Code-based simulation (Bisoyi et al., 2024)

8. Open Problems and Directions

Major open questions include whether $NC^2 \subseteq CL$, sharper lower bounds for separation, the minimization of required catalytic space, and more efficient deterministic algorithms for key graph and circuit problems. The extension to quantum catalytic logspace ($QCL$) establishes that quantum catalytic machines run in polynomial time and can simulate classical catalytic logspace as well as threshold circuits (Buhrman et al., 19 Jun 2025).

Catalytic logspace exposes deep connections between space-bounded complexity, derandomization, and combinatorial optimization, with compressive techniques and algebraic representations enabling computation well beyond the ordinary logspace regime. Its full power and barriers to further inclusion remain central open problems in complexity theory.