[Paper Review] NP is not AL and P is not NC is not NL is not L

Koji Kobayashi|arXiv (Cornell University)|Oct 2, 2011

Model-Driven Software Engineering Techniques3 citations

TL;DR

This paper proves that NP ≠ AL, P ≠ NC, NC ≠ NL, and NL ≠ L by analyzing the dependency structure of computational problems and demonstrating that deterministic Turing machines (DTMs) cannot efficiently simulate certain symmetric or branching computation patterns—particularly in problems like TWINE—due to space and parallelism constraints. The key contribution is a complexity-theoretic separation based on the inability of logspace DTM to handle symmetric, high-complexity paths in directed acyclic graphs (DAGs) without exponential overhead.

ABSTRACT

This paper talk about that NP is not AL and P, P is not NC, NC is not NL, and NL is not L. The point about this paper is the depend relation of the problem that need other problem's result to compute it. I show the structure of depend relation that could divide each complexity classes.

Motivation & Objective

• To establish strict separations between major complexity classes: NP, AL, P, NC, NL, and L.
• To analyze the dependency structure of computational problems where one problem's solution depends on another’s result.
• To demonstrate that deterministic logspace machines (LDTM) cannot efficiently simulate symmetric or branching computation patterns found in NP and NL problems.
• To formalize the concept of 'blocking problems' and 'combined problems' to model dependencies in complexity classes.
• To prove that the TWINE problem cannot be solved in deterministic logspace, thereby separating L from NL and beyond.

Proposed method

• Introduces 'action configuration' as the core unit of computation, comprising state, transition function, head position, and tape symbol.
• Defines Virtual Turing Machines (VTM) as emulated machines within a Universal Turing Machine (UTM), with logspace sufficient to record each VTM's action configuration.
• Uses directed acyclic graphs (DAGs) to model computation histories: DTM (singleton path), NTM (DAG with set of target configs), ATM (family of sets).
• Introduces 'blocking problems' and 'variable problems' to model dependencies: a problem P_i depends on P_j being resolved first.
• Defines 'combined problems' (CP) as sets of interdependent problems whose satisfiability determines the overall result.
• Applies symmetry and path-counting arguments to show that LDTM cannot handle the exponential number of symmetric rotate paths in TWINE without parallel, non-overlapping computation.

Experimental results

Research questions

• RQ1 Can deterministic logspace machines (LDTM) simulate the symmetric computation paths required by the TWINE problem?
• RQ2 Why is the complexity class L strictly smaller than NL, and what structural limitation prevents LDTM from solving NL-complete problems?
• RQ3 How do dependency chains between problems (blocking and variable problems) affect the separation of complexity classes?
• RQ4 What role does the number of symmetric rotate paths in TWINE play in proving L ⊄ TWINE?
• RQ5 Why can NTM handle symmetric element configurations in parallel, while DTM cannot?

Key findings

• LDTM cannot compute TWINE because it must handle O(n^{L(P₀→P₀)}) symmetric rotate paths, exceeding its O(n) element-handling capacity.
• The number of distinct rotate path sequences in TWINE is O(n^{L(P₀→P₀)}) > O(n^c), which exceeds the O(n) element distinction limit of LDTM.
• LDTM cannot simulate symmetric paths in parallel without splitting into multiple machines, requiring a chain of O(L(P₀→P₀)) caller-callee LDTM pairs.
• Since each LDTM in the chain must use disjoint space and cannot share results, constant-space simulation fails, proving L ⊄ TWINE.
• NL ⊃ L strictly because TWINE ∈ NL but TWINE ∉ L, and NL contains problems requiring nondeterministic branching not expressible in deterministic logspace.
• The paper concludes with the strict hierarchy: NP ⊃ AL = P ⊃ NC ⊃ NL ⊃ L, establishing all separations as nontrivial.