We give fast, simple, and implementable catalytic logspace algorithms for two fundamental graph problems. First, a randomized catalytic algorithm for $s\to t$ connectivity running in $\widetilde{O}(nm)$ time, and a deterministic catalytic algorithm for the same running in $\widetilde{O}(n^3 m)$ time. The former algorithm is the first algorithmic use of randomization in $\mathsf{CL}$. The algorithm uses one register per vertex and repeatedly ``pushes'' values along the edges in the graph. Second, a deterministic catalytic algorithm for simulating random walks which in $\widetilde{O}( m T^2 / \varepsilon )$ time estimates the probability a $T$-step random walk ends at a given vertex within $\varepsilon$ additive error. The algorithm uses one register for each vertex and increments it at each visit to ensure repeated visits follow different outgoing edges. Prior catalytic algorithms for both problems did not have explicit runtime bounds beyond being polynomial in $n$.

Hypergraph 2-colorability is one of the classical NP-hard problems. Person and Schacht [SODA'09] designed a deterministic algorithm whose expected running time is polynomial over a uniformly chosen 2-colorable 3-uniform hypergraph. Lee, Molla, and Nagle recently extended this to k-uniform hypergraphs for all k ≥ 3. Both papers relied heavily on the regularity lemma, hence their analysis was involved and their running time hid tower-type constants. Our first result in this paper is a new simple and elementary deterministic 2-coloring algorithm that reproves the theorems of Person-Schacht and Lee-Molla-Nagle while avoiding the use of the regularity lemma. We also show how to turn our new algorithm into a randomized one with average expected running time of only O(n). Our second and main result gives what we consider to be the ultimate evidence of just how easy it is to find a 2-coloring of an average 2-colorable hypergraph. We define a coloring oracle to be an algorithm which, given vertex v, assigns color red/blue to v while inspecting as few edges as possible, so that the answers to any sequence of queries to the oracle are consistent with a single legal 2-coloring of the input. Surprisingly, we show that there is a coloring oracle that, on average, can answer every vertex query in time O(1).

A catalytic machine is a space-bounded Turing machine with additional access to a second, much larger work tape, with the caveat that this tape is full, and its contents must be preserved by the computation. Catalytic machines were defined by Buhrman et al. (STOC 2014), who, alongside many follow-up works, exhibited the power of catalytic space (CSPACE) and, in particular, catalytic logspace machines (CL) beyond that of traditional space-bounded machines. Several variants of CL have been proposed, including nondeterministic and co-non-deterministic catalytic computation by Buhrman et al. (STACS 2016) and randomized catalytic computation by Datta et al. (CSR 2020). These and other works proposed several questions, such as catalytic analogues of the theorems of Savitch and Immerman and Szelepcsényi. Catalytic computation was recently derandomized by Cook et al. (STOC 2025), but only in certain parameter regimes. We settle almost all questions regarding randomized and nondeterministic catalytic computation by giving an optimal reduction from catalytic space with additional resources to the corresponding non-catalytic space classes. With regards to non-determinism, our main result is that CL = CNL and with regards to randomness we show CL = CPrL where CPrL denotes randomized catalytic logspace where the accepting probability can be arbitrarily close to 1/2. We also have a number of near-optimal partial results for non-deterministic and randomized catalytic computation with less catalytic space. We show catalytic versions of Savitch’s theorem, Immerman-Szelepscényi, and the derandomization results of Nisan and Saks and Zhou, all of which are unconditional and hold for all parameter settings. Our results build on the compress-or-compute framework of Cook et al. (STOC 2025). Despite proving broader and stronger results, our framework is simpler and more modular.

STOC ’25, Prague, Czechia

STOC ’25, Prague, Czechia

Abstract We obtain new catalytic algorithms for space-bounded derandomization. In the catalytic computation model introduced by (Buhrman, Cleve, Koucký, Loff, and Speelman STOC 2013), we are given a small worktape, and a larger catalytic tape that has an arbitrary initial configuration. We may edit this tape, but it must be exactly restored to its initial configuration at the completion of the computation. We prove that $$BPSPACE[S] \subseteq CSPACE[{S},{S^2}]$$ B P S P A C E [ S ] ⊆ C S P A C E [ S , S 2 ] where $$BPSPACE[S]$$ B P S P A C E [ S ] corresponds to randomized space S computation, and $$CSPACE[{S},{C}]$$ C S P A C E [ S , C ] corresponds to catalytic algorithms that use O(S) bits of workspace and O(C) bits of catalytic space. Previously, only $$BPSPACE[S]\subseteq CSPACE[{S},{2^{O(S)}}]$$ B P S P A C E [ S ] ⊆ C S P A C E [ S , 2 O ( S ) ] was known. In fact, we prove a general tradeoff, that for every $$\alpha \in [1,1.5]$$ α ∈ [ 1 , 1.5 ] , $$BPSPACE[S] \subseteq CSPACE[{S^{\alpha}},{S^{3-\alpha}}].$$ B P S P A C E [ S ] ⊆ C S P A C E [ S α , S 3 - α ] . We do not use the algebraic techniques of prior work on catalytic computation. Instead, we develop an algorithm that branches based on if the catalytic tape is conditionally random, and instantiate this primitive in a recursive framework. Our result gives

This paper revisits the study of two classical technical tools in theoretical computer science: Yao's trans-formation of distinguishers to next-bit predictors (FOCS 1982), and the “reconstruction paradigm” in pseudorandomness (e.g., as in Nisan and Wigderson, JCSS 1994). Recent works of Pyne, Raz, and Zhan (FOCS 2023) and Doron, Pyne, and Tell (STOC 2024) showed that both of these tools can be derandomized in the specific context of read-once branching programs (ROBPs), but left open the question of de randomizing them in more general settings. Our main contributions give appealing evidence that derandomization of the two tools is possible in general settings, show surprisingly strong consequences of such derandomization, and reveal several new settings where such derandomization is unconditionally possible for algorithms stronger than ROBPs (with useful consequences). Specifically: •We show that derandomizing these tools is equivalent to general derandomization. Specifically, we show that derandomizing distinguish - to- predict transformations is equivalent to prBPP=prP, and that derandomized reconstruction procedures (in a more general sense that we introduce) is equivalent to prBPP=prZPP. These statements hold even when scaled down to weak circuit classes and to algorithms that run in super-polynomial time. •Our main technical contributions are unconditional constructions of derandomized versions of Yao's transformation (or reductions of this task to other problems) for classes and for algorithms beyond ROBPs. Consequently, we deduce new results: A significant relaxation of the hypotheses required to derandomize the isolation lemma for logspace algorithms and deduce that NL=UL; and proofs that de-randomization necessitates targeted PRGs in catalytic logspace (unconditionally) and in logspace (conditionally). In addition, we introduce a natural subclass of prZPP that has been implicitly studied in recent works (Korten FOCS 2021, CCC 2022): The class of problems reducible to a problem called “Lossy Code”. We provide a structural characterization for this class in terms of derandomized reconstruction procedures, and show that this characterization is robust to several natural variations. Lastly, we present alternative proofs for classical results in the theory of pseudorandomness (such as two-sided derandomization reducing to one-sided), relying on the notion of deterministically transforming distinguishers to predictors as the main technical tool.

Random linear codes are a workhorse in coding theory, and are used to show the existence of codes with the best known or even near-optimal trade-offs in many noise models. However, they have little structure besides linearity, and are not amenable to tractable error-correction algorithms. In this work, we prove a general derandomization result applicable to random linear codes. Namely, in settings where the coding-theoretic property of interest is "local" (in the sense of forbidding certain bad configurations involving few vectors -- code distance and list-decodability being notable examples), one can replace random linear codes (RLCs) with a significantly derandomized variant with essentially no loss in parameters. Specifically, instead of randomly sampling coordinates of the (long) Hadamard code (which is an equivalent way to describe RLCs), one can randomly sample coordinates of any code with low bias. Over large alphabets, the low bias requirement can be weakened to just large distance. Furthermore, large distance suffices even with a small alphabet in order to match the current best known bounds for RLC list-decodability. In particular, by virtue of our result, all current (and future) achievability bounds for list-decodability of random linear codes extend automatically to random puncturings of any low-bias (or large alphabet) "mother" code. We also show that our punctured codes emulate the behavior of RLCs on stochastic channels, thus giving a derandomization of RLCs in the context of achieving Shannon capacity as well. Thus, we have a randomness-efficient way to sample codes achieving capacity in both worst-case and stochastic settings that can further inherit algebraic or other algorithmically useful structural properties of the mother code.

We provide compelling evidence for the potential of hardness-vs.-randomness approaches to make progress on the long-standing problem of derandomizing space-bounded computation. Our first contribution is a derandomization of bounded-space machines from hardness assumptions for classes of uniform deterministic algorithms, for which strong (but non-matching) lower bounds can be unconditionally proved. We prove one such result for showing that BPL=L “on average”, and another similar result for showing that BPSPACE[O(n)]=DSPACE[O(n)]. Next, we significantly improve the main results of prior works on hardness-vs.-randomness for logspace. As one of our results, we relax the assumptions needed for derandomization with minimal memory footprint (i.e., showing BPSPACE[S]⊆ DSPACE[c · S] for a small constant c), by completely eliminating a cryptographic assumption that was needed in prior work. A key contribution underlying all of our results is non-black-box use of the descriptions of space-bounded Turing machines, when proving hardness-to-randomness results. That is, the crucial point allowing us to prove our results is that we use properties that are specific to space-bounded machines.

Abstract The classic Impagliazzo–Nisan–Wigderson (INW) pseudorandom generator (PRG) (STOC ‘94) for space-bounded computation uses a seed of length $$O(\log n \cdot \log (nw/\varepsilon )+\log d)$$ O ( log n · log ( n w / ε ) + log d ) to fool ordered branching programs of length n, width w, and alphabet size d to within error $$\varepsilon $$ ε . A series of works have shown that the analysis of the INW generator can be improved for the class of permutation branching programs or the more general regular branching programs, improving the $$O(\log ^2 n)$$ O ( log 2 n ) dependence on the length n to $$O(\log n)$$ O ( log n ) or $${\tilde{O}}(\log n)$$ O ~ ( log n ) . However, when also considering the dependence on the other parameters, these analyses still fall short of the optimal PRG seed length $$O(\log (nwd/\varepsilon ))$$ O ( log ( n w d / ε ) ) . In this paper, we prove that any “spectral analysis” of the INW generator requires seed length $$\begin{aligned} \Omega \left( \log n\cdot \log \log \left( \min \{n,d\}\right) +\log n\cdot \log \left( w/\varepsilon \right) +\log d\right) \end{aligned}$$ Ω log n · log log min { n , d } + log n · log w / ε + log d to fool ordered permutation branching programs of length n, width w, and alphabet size d to within error $$\varepsilon $$ ε . By “spectral analysis” we mean an analysis of the INW generator that relies only on the spectral expansion of the graphs used to construct the generator; this encompasses all prior analyses of the INW generator. Our lower bound matches the upper bound of Braverman–Rao–Raz–Yehudayoff (FOCS 2010, SICOMP 2014) for regular branching programs of alphabet size $$d=2$$ d = 2 except for a gap between their $$O\left( \log n \cdot \log \log n\right) $$ O log n · log log n term and our $$\Omega \left( \log n \cdot \log \log \min \{n,d\}\right) $$ Ω log n · log log min { n , d } term. It also matches the upper bounds of Koucký–Nimbhorkar–Pudlák (STOC 2011), De (CCC 2011), and Steinke (ECCC 2012) for constant-width ( $$w=O(1)$$ w = O ( 1 ) </

For Boolean functions computed by read-once, depth-$D$ circuits with unbounded fan-in over the de Morgan basis, we present an explicit pseudorandom generator with seed length $\tilde{O}(\log^{D+1} n)$. The previous best seed length known for this model was $\tilde{O}(\log^{D+4} n)$, obtained by Trevisan and Xue (CCC `13) for all of $AC^0$ (not just read-once). Our work makes use of Fourier analytic techniques for pseudorandomness introduced by Reingold, Steinke, and Vadhan (RANDOM `13) to show that the generator of Gopalan et al. (FOCS `12) fools read-once $AC^0$. To this end, we prove a new Fourier growth bound for read-once circuits, namely that for every $F: \{0,1\}^n\to\{0,1\}$ computed by a read-once, depth-$D$ circuit, \begin{equation*}\sum_{s\subseteq[n], |s|=k}|\hat{F}[s]|\le O(\log^{D-1}n)^k,\end{equation*} where $\hat{F}$ denotes the Fourier transform of $F$ over $\mathbb{Z}^n_2$.

In this paper, we introduce a new, spectral notion of approximation between directed graphs, which we call singular value (SV) approximation. SV-approximation is stronger than previous notions of spectral approximation considered in the literature, including spectral approximation of Laplacians for undirected graphs [ST04], standard approximation for directed graphs [CKP+17], and unit-circle (UC) approximation for directed graphs [AKM+20]. Further, SV approximation enjoys several useful properties not possessed by previous notions of approximation, e.g., it is preserved under products of randomwalk matrices and bounded matrices. We provide a nearly linear-time algorithm for SV-sparsifying (and hence UC-sparsifying) Eulerian directed graphs, as well as $\ell$-step random walks on such graphs, for any $\ell \leq \operatorname{poly}(n)$. Combined with the Eulerian scaling algorithms of [CKK+18], given an arbitrary (not necessarily Eulerian) directed graph and a set S of vertices, we can approximate the stationary probability mass of the $\left(S, S^{c}\right)$ cut in an $\ell$-step random walk to within a multiplicative error of $1 / \operatorname{polylog}(n)$ and an additive error of $1 / \operatorname{poly}(n)$ in nearly linear time. As a starting point for these results, we provide a simple black-box reduction from SV-sparsifying Eulerian directed graphs to SV-sparsifying undirected graphs; such a directed-to-undirected reduction was not known for previous notions of spectral approximation.

Let $\mathcal{L}$ be a language that can be decided in linear space and let $\epsilon \gt 0$ be any constant. Let $\mathcal{A}$ be the exponential hardness assumption that for every n, membership in $\mathcal{L}$ for inputs of length n cannot be decided by circuits of size smaller than $2^{\epsilon n}$. We prove that for every function $f:\{0,1\}^{*} \rightarrow\{0,1\}$, computable by a randomized logspace algorithm R, there exists a deterministic logspace algorithm D (attempting to compute f), such that on every input x of length n, the algorithm D outputs one of the following:1)The correct value $f(x)$.2)The string: “I am unable to compute $f(x)$ because the hardness assumption $\mathcal{A}$ is false”, followed by a (provenly correct) circuit of size smaller than $2^{\epsilon n^{\prime}}$ for membership in $\mathcal{L}$ for inputs of length $n^{\prime}$, for some $n^{\prime}=\Theta(\log n)$; that is, a circuit that refutes $\mathcal{A}$. Moreover, D is explicitly constructed, given R.We note that previous works on the hardness-versus-randomness paradigm give derandomized algorithms that rely blindly on the hardness assumption. If the hardness assumption is false, the algorithms may output incorrect values, and thus a user cannot trust that an output given by the algorithm is correct. Instead, our algorithm D verifies the computation so that it never outputs an incorrect value. Thus, if D outputs a value for $f(x)$, that value is certified to be correct. Moreover, if D does not output a value for $f(x)$, it alerts that the hardness assumption was found to be false, and refutes the assumption.Our next result is a universal derandomizer for BPL (the class of problems solvable by bounded-error randomized logspace algorithms)1: We give a deterministic algorithm U that takes as an input a randomized logspace algorithm R and an input x and simulates the computation of R on x, deteriministically. Under the widely believed assumption $\mathbf{BPL}=\mathbf{L}$, the space used by U is at most $C_{R} \cdot \log n$ (where $C_{R}$ is a constant depending on R). Moreover, for every constant $c \geq 1$, if $\operatorname{BPL} \subseteq \operatorname{SPACE}\left[(\log (n))^{c}\right]$ then the space used by U is at most $C_{R} \cdot(\log (n))^{c}$.Finally, we prove that if optimal hitting sets for ordered branching programs exist then there is a deterministic logspace algorithm that, given a black-box access to an ordered branching program B of size n, estimates the probability that B accepts on a uniformly random input. This extends the result of (Cheng and Hoza CCC 2020), who proved that an optimal hitting set implies a white-box two-sided derandomization.1Our result is stated and proved for promise-BPL, but we ignore this difference in the abstract.

We give a deterministic white-box algorithm to estimate the expectation of a read-once branching program of length n and width w in space Õ(logn+√logn·logw). In particular, we obtain an almost optimal space Õ(logn) derandomization of programs up to width w=2√logn. Previously, the best known space complexity for this problem was O(min{logn· logw,log3/2n+√logn· logw}) via the classic algorithms of Savitch (JCSS 1970) and Saks and Zhou (JCSS 1999), which only achieve space Õ(logn) for w=polylog(n). We prove this result by showing that a variant of the Saks-Zhou algorithm developed by Cohen, Doron, and Sberlo (ECCC 2022) still works without executing one of the steps in the algorithm, the so-called random shift step. This allows us to extend their algorithm from computing the nth power of a w× w stochastic matrix to multiplying n distinct w× w stochastic matrices with no degradation in space consumption. In the regime where w≥ n, we also show that our approach can achieve parameters matching those of the original Saks-Zhou algorithm (with no loglog factors). Finally, we show that for w≤ 2√logn, an algorithm even simpler than our algorithm and that of Saks and Zhou achieves space O(log3/2 n).

We study query-to-communication lifting. The major open problem in this area is to prove a lifting theorem for gadgets of constant size. The recent paper [Paul Beame and Sajin Koroth, 2023] introduces semi-structured communication complexity, in which one of the players can only send parities of their input bits. They have shown that for any m ≥ 4 deterministic decision tree complexity of a function f can be lifted to the so called semi-structured communication complexity of f∘Ind_m, where Ind_m is the Indexing gadget. As our main contribution we extend these results to randomized setting. Our results also apply to a substantially larger set of gadgets. More specifically, we introduce a new complexity measure of gadgets, linear diversity. For all gadgets g with non-trivial linear diversity we show that randomized decision tree complexity of f lifts to randomized semi-structured communication complexity of f∘g. In particular, this gives tight lifting results for Indexing gadget Ind_m, Inner Product gadget IP_m for all m ≥ 2, and for Majority gadget MAJ_m for all m ≥ 4. We prove the same results for deterministic case. From our result it immediately follows that deterministic/randomized decision tree complexity lifts to deterministic/randomized parity decision tree complexity. For randomized case this is the first result of this type. For deterministic case, our result improves the bound in [Arkadev Chattopadhyay et al., 2023] for Inner Product gadget. To obtain our results we introduce a new secret sets approach to simulation of semi-structured communication protocols by decision trees. It allows us to simulate (restricted classes of) communication protocols on truly uniform distribution of inputs.

For a graph G on n vertices, naively sampling the position of a random walk of at time t requires work Ω(t). We desire local access algorithms supporting position_G(t) queries, which return the position of a random walk from some fixed start vertex s at time t, where the joint distribution of returned positions is 1/poly(n) close to those of a uniformly random walk in 𝓁₁ distance. We first give an algorithm for local access to random walks on a given undirected d-regular graph with Õ(1/(1-λ)√n) runtime per query, where λ is the second-largest eigenvalue of the random walk matrix of the graph in absolute value. Since random d-regular graphs G(n,d) are expanders with high probability, this gives an Õ(√n) algorithm for a graph drawn from G(n,d) whp, which improves on the naive method for small numbers of queries. We then prove that no algorithm with subconstant error given probe access to an input d-regular graph can have runtime better than Ω(√n/log(n)) per query in expectation when the input graph is drawn from G(n,d), obtaining a nearly matching lower bound. We further show an Ω(n^{1/4}) runtime per query lower bound even with an oblivious adversary (i.e. when the query sequence is fixed in advance). We then show that for families of graphs with additional group theoretic structure, dramatically better results can be achieved. We give local access to walks on small-degree abelian Cayley graphs, including cycles and hypercubes, with runtime polylog(n) per query. This also allows for efficient local access to walks on polylog degree expanders. We show that our techniques apply to graphs with high degree by extending or results to graphs constructed using the tensor product (giving fast local access to walks on degree n^ε graphs for any ε ∈ (0,1]) and Cartesian product.

A recent paper of Braverman, Cohen, and Garg (STOC 2018) introduced the concept of a weighted pseudorandom generator (WPRG), which amounts to a pseudorandom generator (PRG) whose outputs are accompanied with real coefficients that scale the acceptance probabilities of any potential distinguisher. They gave an explicit construction of WPRGs for ordered branching programs whose seed length has a better dependence on the error parameter ε than the classic PRG construction of Nisan (STOC 1990 and Combinatorica 1992). In this work, we give an explicit construction of WPRGs that achieve parameters that are impossible to achieve by a PRG. In particular, we construct a WPRG for ordered permutation branching programs of unbounded width with a single accept state that has seed length Õ(log^{3/2} n) for error parameter ε = 1/poly(n), where n is the input length. In contrast, recent work of Hoza et al. (ITCS 2021) shows that any PRG for this model requires seed length Ω(log² n) to achieve error ε = 1/poly(n). As a corollary, we obtain explicit WPRGs with seed length Õ(log^{3/2} n) and error ε = 1/poly(n) for ordered permutation branching programs of width w = poly(n) with an arbitrary number of accept states. Previously, seed length o(log² n) was only known when both the width and the reciprocal of the error are subpolynomial, i.e. w = n^{o(1)} and ε = 1/n^{o(1)} (Braverman, Rao, Raz, Yehudayoff, FOCS 2010 and SICOMP 2014). The starting point for our results are the recent space-efficient algorithms for estimating random-walk probabilities in directed graphs by Ahmadenijad, Kelner, Murtagh, Peebles, Sidford, and Vadhan (FOCS 2020), which are based on spectral graph theory and space-efficient Laplacian solvers. We interpret these algorithms as giving WPRGs with large seed length, which we then derandomize to obtain our results. We also note that this approach gives a simpler proof of the original result of Braverman, Cohen, and Garg, as independently discovered by Cohen, Doron, Renard, Sberlo, and Ta-Shma (these proceedings).

We prove that the Impagliazzo-Nisan-Wigderson [Impagliazzo et al., 1994] pseudorandom generator (PRG) fools ordered (read-once) permutation branching programs of unbounded width with a seed length of Õ(log d + log n ⋅ log(1/ε)), assuming the program has only one accepting vertex in the final layer. Here, n is the length of the program, d is the degree (equivalently, the alphabet size), and ε is the error of the PRG. In contrast, we show that a randomly chosen generator requires seed length Ω(n log d) to fool such unbounded-width programs. Thus, this is an unusual case where an explicit construction is "better than random." Except when the program’s width w is very small, this is an improvement over prior work. For example, when w = poly(n) and d = 2, the best prior PRG for permutation branching programs was simply Nisan’s PRG [Nisan, 1992], which fools general ordered branching programs with seed length O(log(wn/ε) log n). We prove a seed length lower bound of Ω̃(log d + log n ⋅ log(1/ε)) for fooling these unbounded-width programs, showing that our seed length is near-optimal. In fact, when ε ≤ 1/log n, our seed length is within a constant factor of optimal. Our analysis of the INW generator uses the connection between the PRG and the derandomized square of Rozenman and Vadhan [Rozenman and Vadhan, 2005] and the recent analysis of the latter in terms of unit-circle approximation by Ahmadinejad et al. [Ahmadinejad et al., 2020].