NOTE-299: A theory of concepts and their combinations II: A Hilbert space representation

Contribution

Part II turns Part I’s SCOP (LIT-340) into numbers. A concept gets a finite-dimensional complex Hilbert space, a state is a unit vector or density operator, and contexts and properties are orthogonal projectors (§2). A context changes the state by von Neumann–Lüders collapse, |x_q⟩ = P_e|x_p⟩/√⟨x_p|P_e|x_p⟩, with probability µ = ⟨x_p|P_e|x_p⟩ (eqs. 11–14).

• Fitting “pet”. It builds an explicit model of “pet” whose Born probabilities reproduce Part I’s exemplar frequencies (§3).
• Combination. It proposes the tensor product as the rule for combining concepts. Product states model “a pet and a fish”. An entangled state models “pet fish”, one creature that is both (§§4.3–4.4).
• Sentences. The same construction is extended to a three-word sentence, “The cat eats the food”, as an entangled state of “cat ⊗ eat ⊗ food” (§4.5). The authors suggest this could “solve” the bag-of-words problem of vector semantic models.

Key insight

The model encodes combination by correlation. A combined concept is a state of the product space in which a context acting on one factor also fixes the other. The entangled pet-fish state has the property that if “pet” collapses to goldfish, “fish” collapses to goldfish (eqs. 82–85). A product state does not: its “fish” factor is left untouched (eqs. 65–68). That is the paper’s reading of the difference between “pet fish” and “a pet and a fish”.

What does the numerical work is simpler. Basic contexts are basis vectors, contexts are sums of basis projectors, and the ground state is uniform. So every Born probability is a ratio of counts, n(E_i ∩ E_j)/n(E_j).

Assumptions

• Basic contexts (§3.1). The frequency ratings reflect a hidden set X of maximally specific “basic contexts”. Each is atomic and has exactly one eigenstate, and together they form an orthonormal basis B. There are 1,400 for “pet” and 408 for “fish”.
• Counts are chosen, not estimated. n(E_i) and n(X_ij) are “choose[n] … as in Table 2”, i.e. set by the modellers so that the ratios match the rounded frequencies. The dimension 1,400 is part of the same choice.
• Uniform ground state. |α_u|² = 1/1400 for every u (eq. 23), and the ground state is Σ_u |u⟩/√1400 (eq. 24).
• Every context is diagonal in B. P_e = Σ_{u∈E_e} |u⟩⟨u| (eqs. 25, 30, 36, 43, 60). This is what makes the model commutative.
• Shared basic contexts for pet fish (§4.4). The hypothesis is E^pet₆ = E^fish₃₀ = E^{pet,fish}: the basic contexts where the pet is a fish are the same ones where the fish is a pet. Both have 100 elements. The pet and fish spaces have different dimensions (1,400 and 408), so |u⟩⊗|u⟩ presupposes an identification of 100 basis vectors across the two spaces.
• Sentences (§4.5). E₄₇, the basic contexts of “The cat eats the food”, is shared by all three factor spaces.

Key results

• The Hilbert model is a SCOP (§2.5). Density operators with projector contexts and properties and the Lüders rule satisfy the SCOP axioms. This is asserted, with the construction given.
• Single-concept fits (§3.3). Examples: µ(cat | e1) = 75/303 = 0.25 (eq. 33); µ(cat | ground) = 168/1400 = 0.12 (eq. 35); µ(goldfish | e6) = 48/100 = 0.48 (eq. 45); µ(hedgehog | e6) = 0, because E₂₅ ∩ E₆ = ∅.
  • The authors add that the count table “corresponds more or less to a set theoretical model of the experimental data, such that the Hilbert space model can considered to be a quantization … of this set theoretical model”.
• Product states (§4.3). In p₆^pet ⊗ p₃₀^fish, goldfish has weight 0.48 (eqs. 69–71), and in the product of ground states it has 0.10 (eqs. 72–74). The difference is called the guppy effect in the compound. The authors then reject the product state as a model of “pet fish”: it describes “two ‘pet fish’ and not one”, since a collapse of “pet” to goldfish leaves “fish” unchanged.
• Entangled “pet fish” (§4.4, eq. 78). The state is |s⟩ = Σ_{u∈E^{pet,fish}} |u⟩⊗|u⟩/√100.
  • Its reduced states, Σ|u⟩⟨u|/100 on each side, “behave exactly like” p₆ and p₃₀ (eq. 80).
  • Under e₁₈ ⊗ 1 (“The pet is a goldfish”), both reduced states become the goldfish-and-pet-fish state (eqs. 82–85).
• Sentence (§4.5, eq. 89). The state is |s⟩ = Σ_{u∈E₄₇} |u⟩⊗|u⟩⊗|u⟩/√n(E₄₇). The context “The cat is Felix” on the cat factor collapses all three factors to states under “Felix and the cat eats the food” (eqs. 96–101).
• Memory (§4.6). A speculative reading: the ground state is an entangled state of “the compound of all concepts” held in memory. Context is “excitation”, forgetting is “decay”, and speech is compared to photon emission that restores entropy balance, with psychotherapy offered as an example. It is presented as “certainly still speculative”.

Claims

id	claim	strength	support
C1	A Hilbert-space model with projector contexts and collapse reproduces Part I’s exemplar frequencies for “pet”	moderate (construction), with defects	§3, Table 2. The counts are chosen to fit, not estimated. The parrot
C2	The same holds for “fish” under “The fish is a pet”	moderate, with defects	§4.2, Tables 3 and 5. The guppy entry is infeasible (39 > m(E₄₂) = 32)
C3	Property applicabilities are also reproduced	unsupported	Claimed in §5. No property projector is built anywhere
C4	”Pet fish” is an entangled state of pet ⊗ fish, and “a pet and a fish” is a product state	weak (modelling proposal)	§§4.3–4.4. Correct algebra, but the choice is justified only by the desired collapse behaviour. The entangled state’s statistics for every context in the model equal those of the separable, classically correlated mixture Σ
C5	This “provid[es] a natural solution to the pet fish problem” (abstract)	weak (overstated)	The guppy weights are inputs from a single-concept context rating, and no conjunction typicality is measured or predicted. The authors concede “this is not the real guppy effect” for the single-concept version (§5)
C6	The procedure extends to “an arbitrary number of combined concepts” and to sentences	weak (one schematic example)	§4.5. No data are involved, and no count or dimension is given for E₄₇
C7	Entanglement “accounts for the most meaningful combinations of concepts”	unsupported (assertion)	§1, §5
C8	The tensor-product construction can solve the bag-of-words problem of semantic vector models	unsupported (suggestion)	End of §4.5, citing Aerts & Czachor and Widdows
C9	Memory is a hugely entangled state of all concepts, and speech plays the role of photon emission in de-excitation	speculation (so labelled)	§4.6

Method

1. Take Part I’s contexts and exemplars.
2. Posit basic contexts and choose integer counts n(E_i) for each exemplar and n(X_ij) for each exemplar–context pair. The ratios must match the measured relative frequencies (rounded to 2 decimals), with totals of n = 1,400 for “pet” and m = 408 for “fish”.
3. Represent each basic context as an orthonormal basis vector and the ground state as the uniform superposition.
4. Represent each context as the projector onto its basic contexts.
5. Compute Born probabilities; these are count ratios.
6. For combination, take tensor products. Model “pet fish” by the maximally correlated vector on the shared basic contexts, and compute reduced density operators and post-collapse states.

Concepts

• Basic context / basic state — an atomic context with a single eigenstate, one per basis vector. The authors explicitly decline to say whether basic contexts are stored in memory (§3.1).
• Product state vs entangled (non-product) state — the standard definitions, illustrated with ℂ²⊗ℂ² (eqs. 48–54). Here product states mean independent combination (“and”), and entangled states mean one entity that instantiates both concepts.
• Reduced state — the partial trace, called here “exchanging one of the two products |u⟩⟨v| by the inproduct ⟨u|v⟩” (§4.4). The general construction is deferred to Jauch [13].
• Guppy effect — as used here, the rise of goldfish or guppy weight between a ground state and the state under a “pet is a fish” or “fish is a pet” context.

Connections

• Part I (LIT-340). Supplies the data and the SCOP vocabulary, and promises this model. Part I’s only mark of non-classicality was that the eigenstates of e ∨ f exceed those of e and of f (superposition). Part II realises it with superposed states. But every projector is in one commuting family, so the lattice of contexts actually used is Boolean.
• LIT-270 (Aerts et al. 2025, Rejected) and THEORY-013 / LIT-264. Part II’s non-classicality claims rest on entanglement, not on a Bell test. The later programme turned to CHSH tests on the same kind of concept combinations (“The Animal Acts”, Aerts & Sozzo 2011; LIT-264 §5; LIT-270). Contextuality-by-Default removes those violations once unequal marginals are accounted for. Part II is the conceptual ancestor.
  • No evidence of non-locality here. Part II’s own entangled state is, for every measurement in its model, statistically indistinguishable from a classical perfectly correlated mixture. So this paper provides no evidence of non-classical correlation even in principle. Its “entanglement” is a representational choice.
• DisCoCat (LIT-273) and Kartsaklis et al. (LIT-272). Part II proposes, in 2004, that a sentence’s meaning is a vector in the tensor product of its words’ spaces, and flags the bag-of-words problem. It predates Coecke–Sadrzadeh–Clark (2010) and anticipates the tensor-product framing. It has no grammar, no compact-closed reduction and no composition map. The sentence state is just posited from the shared basic contexts of the whole sentence, so it composes nothing from the parts. Its maximally correlated state Σ|u⟩⊗|u⟩⊗|u⟩ is exactly the image of the “copy” map δ(|u⟩) = |u⟩⊗|u⟩⊗|u⟩ applied to Σ|u⟩. This is the Frobenius comultiplication of the basis B (Coecke, Pavlovic & Vicary; see THEORY-017), which is the structure LIT-272 uses. The paper does not make that connection; it is my observation.
• Van Rijsbergen (LIT-262). Contemporary (2004) and independent: projectors as predicates on a document space. Part II cites Widdows [17], [18] rather than van Rijsbergen.

Bearing on the record

• The map’s role (§0, rows 1, 4; “concepts-as-projectors in Hilbert space and the guppy/pet-fish non-distributive combination”).
  • Concepts-as-projectors. Owned here, in the precise sense that contexts are projectors, states are vectors, and change is by collapse (§2). The map is right that this is prior art for “judgment = projector”.
  • “Non-distributive combination”. Not supported by this paper, and the map should drop “non-distributive” from its gloss of Aerts–Gabora 2005.
    • Commutative throughout. All contexts in every model here are sums of projectors diagonal in a single basis, so they commute and generate a Boolean algebra. The fitted “pet” and “fish” models are classical finite probability spaces written as vectors, as the authors’ own “quantization … of this set theoretical model” acknowledges.
    • Combination is by correlation. It is entanglement, not by any failure of distributivity. Within the model, the entanglement is observationally equivalent to classical correlation.
  • Pet fish. The guppy effect is imported from single-concept contextual ratings, not predicted for the conjunction.
• An important consequence for the map’s own §2 diagnosis. The map says the linear-representation / FCA lattice is “the single-context, Boolean shadow” of what quantum cognition studies. Aerts–Gabora’s own worked model is a single-basis, Boolean model: one orthonormal basis, all contexts diagonal in it. The genuinely non-commuting ingredient, which would be needed for order effects or non-distributivity, is never exercised in this paper.
  • Consequence for the map’s positioning. If the map wants a quantum-cognition source that uses incompatible (non-commuting) projectors on concept data, it has to cite later work (Aerts 2009, Busemeyer & Bruza 2012, LIT-316; neither read here). Aerts–Gabora 2005 should be cited for the proposal, not as a demonstration.
  • Why this helps the map. It sharpens the novelty claim: even the founding quantum-concept model sits in one Boolean context.
• THEORY-012 (KS contextuality = no global section). No bearing. There is no measurement cover and no gluing question. Everything is one context family.
• THEORY-013. Consistent with it. Nothing here survives as contextuality in the Contextuality-by-Default sense, and nothing claims to. The paper’s “contextual effects” are marginal shifts, and its “entanglement” produces only classically reproducible correlations for the contexts it uses.
• THEORY-017 / LIT-272 / LIT-273. A historical precursor for tensor-product sentence meaning, worth a line in any account of the DisCoCat lineage. It is not a technical source.
• ML practice. It carries nothing, and it does not belong in the Anthology.

Limitations

• Fitted counts are free parameters. The counts n(E_i), n(X_ij), n and m are chosen to match rounded data. With 1,400 free integers any finite table of conditional frequencies can be matched, subject to the subset constraints, which the table violates twice. The model has no predictive content for the data it fits.
• Only one commuting family. There are no incompatible contexts, no order effects and no interference terms. The Hilbert-space structure is idle.
• Properties are never modelled, despite the claim in §5.
• No combination data. No typicality of exemplars for the conjunction “pet fish” was collected. The “fish” experiment (§4.2) only rates “fish” under “The fish is a pet”, with the same subjects.
• Entanglement is posited. Why should a conjunction produce the maximally correlated state on shared basic contexts? The answer is the hypothesis E^pet₆ = E^fish₃₀. The sentence example has no numbers.
• Section 4.6 is speculation about memory, entropy and psychotherapy, with no argument connecting it to the model.

Open questions

• What would test the tensor-product account? A test needs incompatible contexts on at least one factor, so that entangled and classically correlated states make different predictions. The paper’s own construction never provides them.
• Can the pet-fish effect be predicted? That is, can the conjunction’s typicality be predicted from the constituents’ models without importing the “pet is a fish” rating? Only a measured conjunction-typicality data set, fitted out of sample, would settle it.
• Does the Frobenius-copy reading of the “pet fish” and sentence states extend to a compositional rule? The state would then be computed from word states and a grammatical map, as in LIT-272 and LIT-273, rather than posited from sentence-level basic contexts.

Corrections to the seeded skim

• Seeded from metadata; the text shows the summary is accurate as a description of what is proposed, but it overstates what is shown in two places. (i) “Contexts and properties as orthogonal projections”: the properties are never modelled. §2.3 gives the rule ν(p,a) = ⟨x_p|P_a|x_p⟩, but no property projector P_a is constructed for any of Part I’s 14 properties. §5 nonetheless says “the predictions about frequency values of exemplars and applicability values of properties of the model coincide with the values yielded by the experiment”. Only exemplar frequencies are reproduced (§3.3). (ii) “A solution to the pet-fish problem”. The model’s guppy and goldfish weights, 0.46 and 0.48, are copied from Part I’s rating of “pet” under the context “The pet is a fish” (e6). No typicality for the conjunction “pet fish” was ever measured, and the authors themselves write “Of course this is not the real guppy effect” (§5).
• Infeasible counts in the model. The exemplar sets E_i partition the 1,400 basic contexts: the sum of n(E_i) over the 14 exemplars is 1,400, and each context column sums correctly (303, 495, 500, 101, 200, 100). But X_ij = E_i ∩ E_j (eq. 21) must satisfy n(X_ij) ≤ n(E_i). Table 2 has n₁₇,₅ = 126 for parrot under e5, but n(E₁₇) = 98. For “fish”, Table 5 has m(X₄₂,₁) = 39 for guppy under “The fish is a pet”, but m(E₄₂) = 32. As specified, neither model exists. By my check the parrot row can be repaired by shrinking n(E5), which the data do not fix. So this is a bookkeeping error, not a proof of impossibility, but the claim that the model reproduces “exactly the weights” is false as printed.
• Dates. arXiv v1 is 26 Feb 2004. Crossref and the seed give Kybernetes 34(1/2):192–221, 2005 (pages not checked against the publisher). Under nucleation’s first-appearance convention, published: would be 2004-02-26. The seed’s value is left below for the owner to decide.
• Minor. Eq. (92)–(94) normalise the reduced states by n(E^cat₄₇) etc., while (89) uses n(E₄₇). Eqs. (99)–(101) give post-collapse reduced states without renormalising. Eq. (4)‘s second identity has ⟨y|z⟩ where ⟨x|z⟩ is meant. “Kybernetes, Summer 2004” on the PDF is the planned issue; it appeared in 2005.