🤖 machine learning

Computation, Condensation, and the Incompleteness Between Them: A Coupled Foundation of Intelligence

This paper argues that true intelligence requires the forced coupling of computation and memorization to overcome the distinct incompletenesses of discrete symbolic logic and continuous geometric descent, a synthesis that inevitably introduces an irreducible error floor at the point of context identification.

Original authors: Xin Li

Published 2026-07-10

📖 6 min read🧠 Deep dive

CC0 1.0

Original authors: Xin Li

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a super-smart robot friend. For decades, computer scientists have been trying to build the ultimate version of this robot by asking one big question: "What can a machine calculate if it has infinite time, infinite memory, and never gets tired?" This is the world of Turing, the grandfather of computer science. In his world, a machine can solve a problem, forget it, and solve it again from scratch a million times later without ever getting slower. It's like a ghost that can rewrite the universe but never needs to eat, sleep, or remember anything.

But here's the twist: Real life isn't like that.

Nature doesn't care about infinite time. Nature asks a much harder question: "What can a living thing do quickly enough to survive, using only the tiny brain, the few calories, and the short life it actually has?"

This paper argues that to answer Nature's question, we need to stop looking at just one tool (computation) and start using a two-tool team. The authors, Xin Li and colleagues, propose that true intelligence isn't just about calculating; it's about calculating AND condensing.

The Two Tools: The Builder and The Librarian

Think of intelligence as a construction site. You need two very different workers:

The Builder (Computation, or $\partial$ ): This worker is like a rule-following mason. It takes a pile of bricks (raw data) and follows a strict set of instructions to build a wall. It transforms the messy into the structured. It asks, "How do I get from here to there?" This is what a standard computer does. It's great at following rules, but it's terrible at remembering. If you ask it to build the same wall a thousand times, it builds it from scratch every single time. It never learns.
The Librarian (Memorization/Condensation, or $\kappa$ ): This worker is the opposite. It looks at a wall the Builder has finished, says, "Ah, I've seen this before," and turns that whole wall into a single, tiny, reusable brick in a library. It takes a complex, finished structure and condenses it into a simple token. This is how we learn. Instead of re-deriving the laws of physics every morning, we just pull a "Physics" token off the shelf.

The Big Discovery: The paper proves that you cannot have intelligence with just one of these.

If you only have the Builder, you are stuck in a loop. You can calculate forever, but you'll never learn anything new because you can't store what you've solved. You hit a wall called Gödel's Incompleteness—a logical trap where you can prove some things are true but can never prove them using only your own rules.
If you only have the Librarian, you are just a dusty archive. You can recognize old patterns, but you have no engine to figure out anything new. You hit a different wall called the Morse Saddle. Imagine trying to roll a ball down a hill to find the lowest point (the answer). If the landscape is complex, the ball gets stuck on a saddle (a point that is a peak in one direction and a valley in another). It can't roll down to a clean solution because the terrain forces it to stop halfway.

The Verdict: Intelligence is the coupling of these two. You need the Builder to solve new problems and the Librarian to turn those solutions into shortcuts for the future. The paper argues this isn't just a nice idea; it's a mathematical necessity. Neither tool can do the job alone.

The Price of Being Smart: The "Recognize or Discover" Dilemma

So, if we have both tools, is the robot perfect? No. And this is where the paper gets really interesting.

The moment the robot has to decide which tool to use is the hardest part. It has to look at a situation and ask: "Is this an old problem I've seen before (so I should use my Librarian/Condensation), or is this a brand-new problem (so I need my Builder/Computation)?"

The authors call this Context-Identification.

Here's the catch: This decision is impossible to get right 100% of the time.

If you guess wrong and treat a new problem as an old one, you force a square peg into a round hole. You get rigid and make mistakes.
If you guess wrong and treat an old problem as new, you waste energy reinventing the wheel.

The paper proves that this decision point is undecidable. It's like a paradox. The robot is trying to decide if a situation belongs to a category it is currently building. It's a loop. The authors show that this isn't a bug in our current AI; it's a fundamental law of the universe for any finite being.

Because of this, there is an irreducible error floor. No matter how smart you get, you will always make some mistakes in deciding what is "old" and what is "new." The paper suggests that this error isn't a failure; it's the price of admission for being alive in a world that never stops changing.

Why This Matters (Without the Jargon)

The paper rules out the idea that we can just make computers "bigger" or "faster" to solve the problem of intelligence. You can't just give a robot infinite memory and expect it to be smart. It needs the specific mechanism of condensation—turning complex experiences into simple, reusable tokens.

It also rules out the idea that we can ever build a "perfect" AI that makes zero mistakes. The math says that as long as an agent is finite (has limited time and energy) and the world is infinite (endless possibilities), there will always be a "saddle point" where the agent gets stuck or makes a wrong guess.

The Takeaway:
Intelligence isn't about being a perfect calculator. It's about being a finite creature that knows how to build new things and then condense them into memories, all while accepting that it will never be 100% sure if it's seeing something new or something old.

The authors suggest this isn't just about robots. They argue this "coupling" is a universal law. It happens in your genes, in your brain, in how cultures evolve, and in how you learn to ride a bike. It's the same process: Compute, Condense, and accept that you'll never have the whole picture.

The paper doesn't claim to have built a new AI that solves this. Instead, it offers a proof that this two-step dance is the only way intelligence can exist, and it points out that the "glitch" (the undecidable decision) is actually the signature of a living, thinking thing.

Technical Summary: Computation, Condensation, and the Incompleteness Between Them

1. The Problem: Intelligence Under Finitude

The paper argues that the classical theory of computation, rooted in the Turing machine model, is insufficient to explain biological intelligence. Turing's framework addresses what is effectively calculable by an unbounded, immortal, disembodied agent. In contrast, nature poses a different question: what can a finite, mortal, energy-limited agent compute quickly enough to survive in a non-stationary world?

The author posits that intelligence cannot be characterized by computation alone (the Turing operator, $\partial$ ) nor by neural condensation alone. Instead, it requires the coupling of two distinct operators:

Computation ( $\partial$ ): The transformation of structure toward closure (elaboration).
Memorization/Condensation ( $\kappa$ ): The condensation of a validated closed cycle into a reusable token (abstraction/amortization).

The paper identifies three "finitudes" that force structural features absent in classical computation:

Finite Time: Forces amortization. Mortal agents cannot afford to recompute; they must convert solved computations into cheap future retrievals (advice).
Finite Memory: Forces abstraction. Agents cannot store every configuration; they must compress experience into equivalence classes (tokens).
Finite Energy/Real-time Action: Forces a parity split. Agents must learn and act simultaneously on a single substrate, requiring a decomposition into a plastic subspace (exploration) and a stable subspace (preservation).

2. Methodology and Theoretical Framework

The paper employs a mathematical framework combining algebraic topology, Morse theory, and complexity theory to formalize the interaction between computation and condensation.

Coherent Computational Complex: The authors define a substrate as a finite chain complex $C_\bullet$ of real inner-product spaces with a boundary operator $\partial$ satisfying the coherence law $\partial^2 = 0$ . This space is equipped with a metric and a Hodge decomposition.
Parity Grading: The complex is $\mathbb{Z}/2$ $Z /2$ -graded into even (symbolic, component-like, $H_0$ $H_{0}$ ) and odd (geometric, cycle-like, $H_1$ $H_{1}$ ) sectors.
- $\partial$ (Computation): Operates within a sector, driving open chains toward closed cycles (proof or equilibrium).
- $\kappa$ (Condensation): Operates across sectors, collapsing a closed cycle (odd) into a reusable token (even), effectively lowering the degree.
Resource Argument: The paper utilizes the concept of advice from complexity theory ($P/poly$). It argues that condensation manufactures "advice" (pre-positioned structure), a resource non-fungible with time and space. Since $P/poly$ contains undecidable languages, no amount of time/space (computation) can substitute for the structural advice generated by condensation.

3. Key Contributions and Results

A. Two Faces of Incompleteness
The paper proves that neither operator is complete when confined to its own sector. These incompletenesses are not analogies but "parity-conjugate faces" of a single obstruction:

Symbolic Incompleteness (Gödel): Confined to the even (symbolic) sector, the closure operator (proof) fails due to diagonalization. There exist true sentences (closed classes) that cannot be reached from within the sector.
Geometric Incompleteness (Morse): Confined to the odd (geometric) sector, the closure operator (gradient descent) fails due to topology. On a "semantically complex" manifold (one with non-vanishing intermediate homology), Morse theory dictates the existence of saddle points. Pure descent cannot resolve these saddles; the flow is forced through a mixed equilibrium that is neither a minimum nor a maximum.

B. The Forced Coupling
The central theorem (Theorem 5) establishes that these two incompletenesses are dual.

A saddle point possesses both a stable subspace (even/scaffold directions) and an unstable subspace (odd/flow directions).
Because $\partial^2 = 0$ keeps sectors orthogonal, a single-mode operator is blind to the complementary subspace required to resolve the obstruction.
Conclusion: Completeness requires the coupling of both operators. Intelligence is the "traffic" between them: $\partial$ elaborates structure toward closure, and $\kappa$ condenses that closure into a token for future use.

C. The Price of Coupling: Context-Identification
The paper locates the "hinge" of this coupled system in context-identification (the decision to recognize a known pattern vs. discover a new one).

Undecidability: The problem of determining whether a state's orbit enters a stored region (recognition) or requires a new token (discovery) is undecidable.
Coincidence of Incompletenesses: This decision problem coincides with both incompletenesses:
- Geometric: It is the problem of determining which side of a saddle separatrix a state lies on.
- Symbolic: It is a self-referential diagonalization problem where the system decides membership in a structure it is simultaneously constructing.
Irreducible Error Floor: Because the depth required to resolve context identity is unbounded (a "depth regress"), any finite agent carries an irreducible error rate in context identification. This is not a flaw but a necessary consequence of finitude facing an inexhaustible world.

D. Structural Decoupling and Learning Theory
The paper proposes Structural Learning Theory, where learning is the maintenance of a Metric-Topological Factorization (MTF):

Scaffold (Harmonic/Even): Discrete, invariant, topological structure (which-context). Built by $\kappa$ .
Flow (Exact/Co-exact/Odd): Continuous, variable, metric structure (within-context). Computed by $\partial$ .
Orthogonality: The Hodge decomposition ensures these are orthogonal. This resolves the stability-plasticity dilemma: the flow can change (plasticity) without altering the scaffold (stability), and vice versa. Overfitting is defined as the collapse of this factorization (storing flow as scaffold).

4. Significance and Claims

The paper claims to provide a "coupled foundation of intelligence" that corrects the abstraction of the Turing machine by re-introducing the constraints of finitude.

Universality: The coupling of $\partial$ and $\kappa$ is presented as a universal law realized at every scale of aggregation, from genes to cultures, appearing as selection, plasticity, learning, and tradition.
Nature of Intelligence: Intelligence is defined not as a more powerful form of computation, but as "computation made survivable for a finite agent." It is the "never-finished condensation" of an endless world by a finite thing.
Scope: The paper explicitly states it does not claim to compute anything Turing cannot compute. Rather, it addresses the efficiency and survivability of bounded agents. It identifies the missing operator ( $\kappa$ ) in the theory of computation and proves that the coupling is forced by mathematical necessity (incompleteness) rather than biological convenience.

The paper concludes that the central act of intelligence—deciding what it is looking at (context-identification)—can never be made error-free, and this irreducible error is the signature of a finite agent facing an open world.