The Williams 2025 Breakthrough

How a revolutionary discovery in computer science proves the universe optimizes computation in ways that exceeded human understanding for 50 years
Published: February 24, 2025

🎯 The Amazing Discovery

Imagine you're trying to solve a really hard puzzle. Scientists thought you needed a huge desk (lots of memory) to work on it. But Ryan Williams discovered you only need a tiny corner of the desk - way smaller than anyone believed possible!

What Changed?

Old Belief (50 years)

Need THIS much memory
→

Williams 2025

Only need this!

🤯 What This Means: We're in a Simulation!

💻
Reality Uses Secret Computer Tricks: The universe knows computational shortcuts that our best scientists didn't discover for 50 years! This is exactly what we'd expect if we're in a simulation.
🎮
Cosmic Cheat Codes: Just like video games have hidden optimizations in their code, reality is revealing it has super-efficient ways to compute things!
🌌
Self-Optimizing Universe: This discovery shows the universe is like a computer that keeps finding better ways to run its own simulations!
🔄
Recursive Stack Evidence: The fact that simulating time is exponentially easier than expected makes it way more likely that we're in a simulation created by a previous layer!

The Revolutionary Discovery

In February 2025, Ryan Williams proved that any computation requiring time t can be performed using only √(t log t) space - exponentially less memory than the t/log t space bound that stood for 50 years.

The Breakthrough Result

TIME[t] ⊆ SPACE[√(t log t)]

Any algorithm running in time t can be simulated using only square-root space

50 Years of Progress in One Step

Aspect Previous Best (1975) Williams (2025)
Space Required O(t / log t) O(√(t log t))
Improvement Factor - √t / log^(3/2) t
Believed Possible? Yes No - violated decades of assumptions

💻 Strong Evidence We're in a Self-Optimizing Simulation

🔍 Reality Reveals Hidden Optimizations

This result was completely unexpected by human science. The universe "knows" computational shortcuts that exceeded our best theories for 50 years. This is exactly what we'd expect if reality is computational and has access to optimization techniques beyond current human understanding.

🔄 Recursive Self-Improvement Pattern

The solution uses recursive Tree Evaluation - optimization discovering better optimization. This mirrors the pattern we see everywhere: each layer of reality creates more efficient optimizers, suggesting we're in a recursive simulation stack.

⚡ Simulation Feasibility Explosion

If simulation requires exponentially less resources than thought, it becomes trivially easy for advanced civilizations to create nested simulations. This discovery makes our simulation hypothesis far more likely.

🌌 Cosmic Computer Evidence

The universe finding optimal space-time tradeoffs suggests reality IS a computational system optimizing its own resource usage - not just a system that happens to compute things.

Breakthrough in Space-Time Complexity

Main Theorem (Williams 2025)

Theorem 1.1: For every function t(n) ≥ n, TIME[t(n)] ⊆ SPACE[√(t(n) log t(n))].

This improves the Hopcroft-Paul-Valiant bound of O(t(n)/log t(n)) from 1975.

Technical Innovation

The proof reduces time-bounded multitape Turing machine computations to Tree Evaluation instances, then applies the Cook-Mertz space-efficient algorithm:

  1. Partition computation into blocks of length b
  2. Construct computation graph tracking information flow
  3. Transform to Tree Evaluation instance of height O(t/b)
  4. Apply Cook-Mertz procedure using O(d·b + h·log(d·b)) space
  5. Optimize by setting b = √(t log t)

Major Consequences

Time-Space Separation

Corollary 1.2: For space constructible s(n) ≥ n and all ε > 0, SPACE[s(n)] ⊄ TIME[s(n)^(2-ε)].

First polynomial separation for multitape machines.

Circuit Complexity

Corollary 1.4: Bounded fan-in circuits of size s have branching programs of size 2^(O(√s log s)).

Subquadratic circuits admit subexponential branching programs.

Linear Space Lower Bounds

Corollary 1.3: Complete problems for linear space require n^(2-ε) time for all ε > 0.

Context-sensitive language recognition needs essentially quadratic time.

Optimization Principle Connection

This result exemplifies several key aspects of universal optimization:

  • Exceeds theoretical limits: Violates assumptions underlying major derandomization results
  • Recursive structure: Tree Evaluation exhibits self-similar optimization patterns
  • Approaching boundaries: Points toward fundamental P vs PSPACE separation
  • Cross-domain impact: Affects circuits, automata, and algorithm design

Simulating Time With Square-Root Space

Abstract

We show that for all functions t(n) ≥ n, every multitape Turing machine running in time t can be simulated in space only O(√(t log t)). This is a substantial improvement over Hopcroft, Paul, and Valiant's simulation of time t in O(t/log t) space from 50 years ago [FOCS 1975, JACM 1977].

Formal Statement and Proof Sketch

Theorem 1.1. For every function t(n) ≥ n, TIME[t(n)] ⊆ SPACE[√(t(n) log t(n))].

Proof sketch: Given time-t(n) multitape TM M:

  1. Make M block-respecting with block length b(n) (Lemma 2.1)
  2. Define computation graph G_{M,x} with nodes representing tape blocks across time
  3. Encode G_{M,x} succinctly in O(t(n)/b(n)) bits using head movements
  4. Construct Tree Evaluation instance R_{G'} for each candidate graph G'
  5. Functions at nodes simulate b(n) steps, checking consistency
  6. Apply Cook-Mertz (Theorem 2.2): O(d·b + h·log(d·b)) space where h = O(t/b)
  7. Setting b(n) = √(t(n) log t(n)) yields the bound □

Key Technical Lemmas

Lemma 2.1 (Block-Respecting Transformation): Every t(n)-time ℓ-tape TM has equivalent O(t(n))-time block-respecting (ℓ+1)-tape TM with blocks of length b(n).

Theorem 2.2 (Cook-Mertz): Tree Evaluation on trees of bit-length b, height h, fan-in d can be computed in O(d·b + h·log(d·b)) space.

Claim 3.5: Given indices of nodes u,v in G_{M',x} and sequence {m_{(h,i)}}, edge determination requires O(log t(n)) additional space.

Implications for Complexity Theory

1. Progress Toward P vs PSPACE

Extension to TIME[t] ⊆ SPACE[t^ε] ∀ε > 0 would yield P ≠ PSPACE. Current barrier: low-degree extension computation in Tree Evaluation requires 2^Θ(b) time.

2. Higher-Dimensional Extensions

Theorem 1.5: d-dimensional TMs in time t can be simulated in O((t log t)^(1-1/(d+1))) space.

3. Restricted Oracle Separations

Theorem 5.1: TIME-LENGTH^A[t(n), √(t(n) log t(n))] ⊆ SPACE-LENGTH^A[√(t(n) log t(n)), √(t(n) log t(n))].

Open Problems

  • Can the √log t factor be removed? (Would require Tree Evaluation ∈ L)
  • Extension to random-access models?
  • Time-space tradeoffs: TIME[t] ⊆ TIMESPACE[2^Õ(t^ε), Õ(t^(1-ε))]?
  • Recursive application for TIME[t] ⊆ SPACE[t^ε] ∀ε > 0?

Full Citation: Williams, R. (2025). Simulating Time With Square-Root Space. Electronic Colloquium on Computational Complexity, Report No. 17 (2025). To appear in STOC 2025.

View Full Paper (PDF)
Understanding Level: