Contribution
Two earlier lines compared contextual resources: the contextual fraction with a list of free operations it is monotone under (LIT-265), and Karvonen’s simulations (LIT-846). This paper joins them. It adds an adaptive free operation, conditional measurement, and an equational theory for the free operations. It builds a comonad of measurement protocols, so that adaptive simulations are co-Kleisli maps, and proves that a simulation exists exactly when a term in the free operations converts one model into the other.
Key insight
The operational notion (“use box d, with classical randomness and classical control, to behave like box e”) and the algebraic notion (“e is built from d by free operations”) pick out the same preorder. Adaptivity, which was the missing piece in Karvonen’s definition, is added by changing the scenario rather than the morphism: measure a protocol in MP(d) rather than a measurement in d.
Assumptions
- Finite scenarios ⟨X, Σ, O⟩ with Σ a simplicial complex (Definition 1), and no-signalling empirical models (Definition 3: compatible families; compatibility “generalizes … no-signalling”).
- Basic morphisms use simplicial maps, not relations (Definition 12); joint measurement of several source measurements is recovered through protocols.
- Randomness is shared and classical, supplied by an auxiliary non-contextual model c (Definition 18). The authors note that letting c range over a different class (e.g. quantum-realisable models) gives “relative simulatability”.
- Single-use boxes: a protocol is one round of possibly sequential measurements on one copy.
Key results
- Proposition 7. CF(z) = CF(u) = 0; CF(f*e) ≤ CF(e); CF(e/h) ≤ CF(e); CF(e +_λ e′) ≤ λCF(e) + (1 − λ)CF(e′); CF(e & e′) = max{CF(e), CF(e′)}; NCF(e ⊗ e′) = NCF(e)NCF(e′); CF(e[x?y]) = CF(e). Only the last is new here; the rest are cited to LIT-265.
- Proposition 8. A term without variables denotes a non-contextual model, and every non-contextual model is so denoted: a deterministic one as (f*u)/h, the rest by mixing.
- Propositions 10–11. The equations are sound up to isomorphism, and every term has a normal form: mixtures at the top, then translation and coarse-graining, then conditional measurements, then tensors of base cases. Completeness is open.
- Definitions 13–16, Theorem 17. Runs, protocols (prefix-closed, outcome-complete, deterministic in the next measurement), the scenario MP(X) whose compatibility is “can be interleaved without ever measuring an incompatible set”, the model MP(e), and the proof that MP is a comonoidal comonad (counit: a measurement as a one-step protocol; extension: run π(y) whenever Q calls y, reusing outcomes already obtained).
- Theorem 20. d simulates e (a deterministic map MP(d ⊗ c) → e, c non-contextual) iff e ≅ t[d/v] for some typed term t. Forward: c is a closed term and protocols are built by conditional measurements. Backward: by the normal form.
- Theorem 21. If d simulates e then NCF(d) ≤ NCF(e).
- Theorem 22 (no-cloning). e simulates e ⊗ e iff e is non-contextual. The proof first forces NCF(e) > 0 by pigeonhole on the finitely many protocol components, then rules out 0 < NCF(e) < 1 by Theorem 21 and NCF(e ⊗ e) = NCF(e)².
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Adaptive simulation with shared classical randomness coincides with conversion by free operations | strong (proof) | Theorem 20, via Propositions 8 and 11 |
| C2 | Measurement protocols form a comonoidal comonad on empirical models | strong (proof) | Theorem 17, §IV-C |
| C3 | Conditional measurement leaves the contextual fraction unchanged | strong (proof) | Proposition 7 |
| C4 | No contextual model can be cloned by classical simulation | strong (proof) | Theorem 22 |
| C5 | The equational theory is complete | not supported; open | §III-C |
Method
Category theory and term rewriting. Free operations are typing rules (Table II) whose closed terms are the free resources. A comonad on scenarios lifts deterministic simulations to adaptive ones, and the co-Kleisli category is where simulations live. The equivalence of the two views is proved through normal forms.
Concepts
- free operation: an operation on empirical models a classical agent can perform with no contextual resource beyond its arguments.
- conditional measurement x?y: measure x, and on outcome o measure y_o (a vertex in the link of x); its outcome is the pair.
- link lk_σ Σ: the faces still performable after σ has been measured.
- measurement protocol: a non-empty prefix-closed set of runs that answers every outcome and fixes the next measurement; a deterministic adaptive strategy, a “wiring”.
- simulation d ⇝ e: a deterministic morphism MP(d ⊗ c) → e with c non-contextual.
- isomorphism of models: a simplicial isomorphism and bijections of outcomes carrying one to the other (Definition 9).
Connections
It generalises Karvonen (LIT-846), whose morphisms had no preprocessing, and it adds conditional measurement to the free operations of Abramsky, Barbosa and Mansfield (LIT-265). It rests on Abramsky and Brandenburger (LIT-016) for the framework. Measurement protocols come from Acín, Fritz, Leverrier and Sainz, where they are “wirings”. The chapter by Barbosa, Karvonen and Mansfield (LIT-845, Remark 28) returns to non-adaptive procedures to characterise the maps they induce, and says why: its key lemma fails for adaptive protocols.
Bearing on the record
- CLAIM-100 (directed transport between scenarios with changing covers). Translation of measurements is along an arbitrary simplicial map between scenarios, and the conditional measurement extends a scenario’s cover with new measurements and faces. So the free operations already include cover-changing, stochastic, directed transport, and Theorems 20–21 say that every such transport is monotone for contextuality. Together with LIT-846 this answers the “global compatibility” half of A94’s question for no-signalling models. The “decision-relevant information” half is not addressed; nor is signalling data.
- The manuscript’s “repeated reconstruction” and chains. A composite of free operations is a free operation, and CF is monotone along each step, so along any chain of classical transports contextuality can only fall. This is a corollary I draw (Theorems 20–21 with composition in the co-Kleisli category); it bears on the manuscript’s interest in how compatibility evolves “during translation and repeated reconstruction” (A93), which CLAIM-100 quotes.
- Produces THEORY-174, with LIT-846 and LIT-845.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- No-signalling models only.
- Completeness of the equational theory is open.
- Simulations are qualitative: existence, not cost; there is no measure of how much resource a simulation uses beyond NCF monotonicity.
- Single-use boxes; sequential reuse with state is out of scope.
Open questions
- Completeness of equations (1)–(28), and whether Theorem 20 is a bijection between simulations and terms up to equality.
- Relative simulatability: grading MP by the auxiliary resource (for instance quantum-realisable c) and by run length.
- The link the outlook draws between simulations of possibilistic models and reductions between constraint-satisfaction problems.