paper
active
paper:2022-08-21-prabros-act2022-slides-4223-pdf-890d16

2022-08-21_Prabros._ACT2022_slides_4223.pdf_890d16

Compositional SemanticsDenotational & semantic program designDenotational & semantic program designBell-stateMeasurement-Based Quantum Computationintertwining of grammar and meaningMBQCquantum cryptoquantum error correction

TL;DR

Bob Coecke's ACT 2022 presentation, delivered under the Quantinuum Compositional Intelligence programme, advances a principled taxonomy of compositionality that distinguishes at least four distinct regimes—Frege, Schrödinger, Whitehead, and Complete/LEGO compositionality—and argues that standard Frege compositionality (bottom-up meaning flow from parts to whole) is insufficient for quantum systems, natural language disambiguation, and AI reasoning alike. The presentation introduces the framework of Schrödinger compositional theory, defined as a process theory in which composition is non-trivial (the whole cannot be meaningfully decomposed back into parts) and all ingredients have clear ontological counterparts in reality—a condition quantum entanglement satisfies and that ZX-calculus (introduced in 2007) violates as a non-example of Whitehead compositionality. The broader toolchain instantiating these ideas includes DisCoCirc, DisCoPy, lambeq, BobCat, DisCoSpeed, and Q-Glove, spanning NLP, quantum hardware (photonics, MBQC, error correction), and biology. The programme spans collaborations across UCL/BBC, Oxford, Vienna, and Quantinuum hardware teams led by contributors including Dimitri, Konstantinos, Steve, and Giovanni. The work implies that building practically compositional quantum AI requires abandoning the assumption that sentence-level or circuit-level meaning decomposes cleanly into part meanings, and that a richer process-theoretic ontology—one that respects top-down context dependence alongside bottom-up structure—is the necessary foundation for both quantum computing and compositional NLP.

What to take away

  1. 1. Coecke distinguishes four named regimes of compositionality—Frege, Schrödinger, Whitehead, and Complete (LEGO)—where standard Frege compositionality assumes purely bottom-up meaning flow and is treated as the least expressive of the four.
  2. 2. A Schrödinger compositional theory is formally defined as a process theory in which composition is non-trivial (the whole cannot be meaningfully decomposed back into its parts) and all ingredients have clear ontological counterparts in reality.
  3. 3. ZX-calculus, introduced in 2007, is explicitly named as a non-example of Whitehead compositionality, meaning its diagrammatic ingredients lack direct ontological counterparts even though the calculus is formally complete.
  4. 4. The Quantinuum Compositional Intelligence programme fields at least seven named artefacts—DisCoCirc, DisCoPy, lambeq, BobCat, DisCoSpeed, Q-Glove, and the quantum-diagrammatic ZXW-theory—each targeting a distinct layer from NLP parsing to hardware circuit optimisation.
  5. 5. Frege's context principle (never ask for word meaning in isolation, only in the context of a sentence) introduces a top-down meaning flow that directly conflicts with classical Frege compositionality, and the presentation treats this tension as unresolved in current NLP models.
  6. 6. Quantum entanglement (illustrated via Bell states) is presented as the canonical case where Schrödinger non-decomposability holds: the joint state has no factored part-meanings, making it the motivating physical substrate for the new taxonomy.
  7. 7. The photonics implementation track (led by Bob and Giovanni, Vienna site) and the MBQC/error-correction/quantum-cryptography hardware tracks are identified as the two primary quantum-hardware application domains for compositional process theories.
  8. 8. An open hypothesis the presentation raises is whether a fully Whitehead-compositional framework—where every diagrammatic ingredient maps onto a real ontological entity—can be constructed for quantum AI tasks currently handled by ZX-based methods.
  9. 9. To replicate the process-theory methodology, a researcher should define a triple (S, P, D) of systems-as-wires, processes-as-boxes, and composition-as-wiring, then test whether the resulting diagrams satisfy non-trivial composition and ontological transparency as defined in the dodo-book formalism.
  10. 10. Lambeq (developed by the Oxford/Quantinuum team including Steve and Dimitri) serves as the primary NLP-to-quantum-circuit compilation tool in the stack, sitting between the linguistic DisCoCirc representation and hardware execution layers.

Peer brief — for seminar discussion

Presented at ACT 2022 by Bob Coecke under the Quantinuum Compositional Intelligence banner, this talk constructs a four-way taxonomy of compositionality—Frege, Schrödinger, Whitehead, and Complete/LEGO—and argues that each regime makes different and non-interchangeable demands on any AI or quantum system that aspires to reason compositionally. The vehicle for this taxonomy is what the presentation calls Schrödinger compositional theory, a species of process theory (formalised in the 'dodo-book' as a triple of systems-as-wires, processes-as-boxes, and composition-as-wirings) in which two conditions hold jointly: composition is non-trivial in the sense that a whole cannot be meaningfully decomposed back into its constituents, and every ingredient of the diagram has a clear ontological counterpart in reality. Quantum entanglement via Bell states is the canonical motivating example. The load-bearing finding is that these conditions are not simultaneously satisfied by widely used formalisms: ZX-calculus (2007) is named as a non-example of Whitehead compositionality precisely because its spiders and phases lack direct physical referents even though the calculus is formally complete. The practical instantiation spans DisCoCirc, DisCoPy, lambeq, BobCat, DisCoSpeed, and Q-Glove, with hardware tracks in photonics (Vienna, Bob and Giovanni) and error-correction/MBQC, and NLP tracks developed with UCL/BBC (Steve, Dimitri, Konstantinos). The implication drawn is that building genuinely compositional quantum AI requires moving beyond Frege's bottom-up meaning flow—acknowledging Frege's own context principle as a top-down corrective—and adopting a process-theoretic ontology that respects non-decomposability at the level of entangled or contextually disambiguated wholes. An alternative framing the talk could have used is categorial grammar or dependency-tree semantics, which also encode grammar-meaning interaction but do not impose the ontological-transparency requirement. A critical reader would push back on the ontological-transparency criterion itself: the claim that all diagrammatic ingredients must have 'clear meaningful ontological counterparts in reality' is asserted rather than operationalised, leaving it unclear how one would falsify or even consistently adjudicate whether a given process theory meets this standard—particularly in domains like NLP-bio where the mapping between circuit elements and biological or linguistic reality is contested. The open hypothesis is that a Whitehead-compositional replacement for ZX-calculus can be constructed that preserves completeness while grounding every generator physically, but no such construction is exhibited in the slides.

Methods (1)

Claims (6)

Questions (3)

Related work— refs + corpus + external arXiv

Cited / in-corpus / arXiv badges show which signals surfaced each row. Multi-source rows weighted higher.

Similar preprints — Semantic Scholar

Cross-corpus bridges (1)

same_concept_as · Nomic cosine

External markdown files that talk about the same concept as this entity.

  • alexander
    (QUANTUM) COMPOSITIONAL INTELLIGENCEpapers/extracted/2022-08-21_Prabros._ACT2022_slides_4223.pdf_890d16.md0.874