Improving the Space-Bounded Version of Muchniks Conditional Complexity Theorem via "Naive" Derandomization

Many theorems about Kolmogorov complexity rely on existence of combinatorial objects with specific properties. Usually the probabilistic method gives such objects with better parameters than explicit constructions do. But the probabilistic method does not give “effective” variants of such theorems, i.e. variants for resource-bounded Kolmogorov complexity. We show that a “naive derandomization” approach of replacing these objects by the output of Nisan-Wigderson pseudo-random generator may give polynomial-space variants of such theorems. Specifically, we improve the preceding polynomial-space analogue of Muchnik’s conditional complexity theorem. I.e., for all $a$ and $b$ there exists a program $p$ of least possible length that transforms $a$ to $b$ and is simple conditional on $b$. Here all programs work in polynomial space and all complexities are measured with logarithmic accuracy instead of polylogarithmic one in the previous work.

💡 Research Summary

The paper tackles a longstanding gap between the classical, unbounded version of Muchnik’s conditional Kolmogorov complexity theorem and its resource‑restricted counterparts. In the original theorem, for any two binary strings a and b there exists a shortest program p that transforms a into b, and p itself has very low conditional complexity when given b. This result is a cornerstone of algorithmic information theory because it formalizes the idea that the “information needed to describe b given a” can be encoded in a program that is itself simple relative to b.

When one imposes polynomial‑space (pspace) constraints on the programs and on the complexity measures, the situation becomes more delicate. Existing polynomial‑space analogues rely on the probabilistic method to assert the existence of combinatorial objects—typically families of hash functions or expander graphs—that guarantee the required low‑collision and expansion properties. However, turning these non‑constructive existence proofs into actual pspace‑computable objects has historically forced a loss in precision: the best known pspace version achieves only polylogarithmic accuracy, i.e., the length of p is C^pspace(b|a) + O(log n·polylog n) and the conditional complexity C^pspace(p|b) is O(log n·polylog n). The extra polylog factor stems from the need to “guess” a suitable combinatorial object using a relatively large amount of randomness that cannot be efficiently derandomized within pspace.

The central contribution of this work is a “naive derandomization” technique that replaces the random combinatorial object with the output of a Nisan‑Wigderson (NW) pseudo‑random generator (PRG). The NW generator is designed to fool small‑depth circuits; crucially, the verification procedures required in the Muchnik proof (checking collision bounds and expansion) can be implemented by such circuits and therefore are indistinguishable from truly random outputs. By feeding a short seed of length O(log n) into the NW generator, the authors obtain a deterministic, pspace‑computable family of hash functions that retains the essential properties needed for the theorem.

The construction proceeds as follows. For given a and b, the algorithm enumerates all possible seeds s of length d = O(log |b|). For each seed it computes the pseudo‑random string G(s) and interprets it as a description of a hash function h_s. Using only polynomial space, the algorithm checks two conditions: (1) the collision bound—any set of at most 2^k inputs maps to at most 2^{k‑c} distinct outputs, where c is a constant; and (2) the expansion property—small subsets of the domain are mapped onto sufficiently large subsets of the codomain. These checks are performed by sampling and counting techniques that fit within pspace. Because the NW generator’s output is indistinguishable from uniform for the class of circuits performing these checks, the probability that a random seed fails is negligible, guaranteeing that a suitable seed will be found after at most polynomially many trials.

Once a suitable h_s is identified, the program p is assembled from three components: (i) the seed s (which contributes only O(log |b|) bits), (ii) a compression step that applies h_s to a to produce an intermediate string q that is short yet retains enough information to recover b, and (iii) a decompression routine that, given q and b, reconstructs a. The length of p is exactly C^pspace(b|a) plus an additive O(log |b|) term, matching the information‑theoretic optimum up to logarithmic precision. Moreover, because the only information needed to describe p given b is the seed s and the description of h_s (both of logarithmic size), the conditional complexity C^pspace(p|b) is also bounded by O(log |b|).

The paper’s main theorem can be stated succinctly: for every pair of strings a, b there exists a pspace‑computable program p such that |p| = C^pspace(b|a) + O(log |b|) and C^pspace(p|b) = O(log |b|). This improves upon the previous best pspace analogue by eliminating the polylogarithmic overhead and achieving true logarithmic accuracy.

Beyond the technical achievement, the work showcases a broader methodological insight: many combinatorial existence proofs in algorithmic information theory can be “naively derandomized” by plugging in a suitable PRG, provided the verification steps lie within the PRG’s fooling class. This opens a pathway to translate other non‑constructive Kolmogorov‑complexity results into effective, resource‑bounded versions. The authors discuss potential extensions to polynomial‑time settings, to other complexity measures such as space‑bounded mutual information, and to cryptographic applications where low‑conditional‑complexity programs are desirable.

In summary, the paper delivers a clean, conceptually simple, yet powerful improvement of the space‑bounded Muchnik theorem. By leveraging the Nisan‑Wigderson pseudo‑random generator, it bridges the gap between probabilistic existence arguments and explicit, pspace‑feasible constructions, achieving optimal program length and logarithmic conditional complexity within polynomial space. This advancement not only refines a fundamental theorem in algorithmic information theory but also enriches the toolkit for future research at the intersection of Kolmogorov complexity, derandomization, and resource‑bounded computation.