Multiple Realisability: Concepts, Applications, and Significance

• Multiple realisability refers to the principle where a single higher-level entity, such as a function or specification, can be achieved through various physically, computationally, or structurally distinct means. This concept is widely applied in fields like philosophy of mind, quantum gravity, realizability semantics, and systems theory to highlight functional equivalence despite underlying differences.
• In timed assume-guarantee specification theory, multiple realisability involves interpreting specification models as collections of strategies. Strategies integrating various computations deliver the same functional outcome, demonstrating distinct implementations achieving the same function specification, like in elliptic automation in dynamical systems.
• Orthogonality-based semantics illustrate multiple realisability in concurrent computing, where formulas are sets of computational processes evolving through reduction sequences to model concurrent execution. This shows how multiplicitous pathways can yield the same elementary behavioural-functional outcomes upon divergent interactions and reduction procedures.

Multiple realisability is the principle that one higher-level object, function, specification, or behaviour can be instantiated by several physically, computationally, mathematically, or structurally different lower-level realisers. The term is used in several distinct research traditions rather than as one universally fixed formal doctrine. In philosophy of mind it concerns higher-level functional kinds; in quantum gravity it concerns the recovery of general relativity from different microstructures; in realizability semantics it concerns set-valued collections of witnesses or alternative interpretation schemes; in systems theory it concerns multiple strategies, controllers, executions, or protocol implementations; and in science and engineering it concerns distinct mechanisms or microstructures producing comparable effective behaviour. Across these uses, a central distinction is between multiplicity of realisers for one fixed target and multiplicity of interpretations, specifications, or target-level descriptions themselves.

1. Conceptual scope and principal distinctions

Multiple realisability is most sharply distinguished from uniqueness claims. A target may have several realisers even when each realiser is individually well-defined, and a specification may admit a set of behaviours without selecting one implementation. Conversely, a family of systems may exhibit similar performance without implementing exactly the same effective object. Claims of multiple realisability therefore require specifying the level at which equivalence is asserted.

Several levels recur across the literature:

1. Functional multiplicity: distinct lower-level structures implement the same higher-level function or input/output relation.
2. Mechanistic multiplicity: one specified process has different causal histories, execution orders, or token-level decompositions.
3. Structural multiplicity: different networks, reaction sets, geometries, or automata satisfy a common structural specification.
4. Semantic multiplicity: a proposition, formula, or contract denotes a collection of realisers rather than a unique witness.
5. Model-relative multiplicity: the same specification is realisable under one semantic or communication model but not another.
6. Interpretational multiplicity: one source logic or mathematical theory supports several realizability interpretations obtained by changing the treatment of atomic predicates or computational resources.

These senses must not be conflated. For example, a scalar performance metric can be shared by distinct systems without their full states, phases, causal structures, or microscopic dynamics being equivalent. Similarly, a formal semantics may contain many realisers without proving that a given formula has more than one syntactically distinct proof or implementation.

The term is not always the name of an established technical doctrine. “Realizability Models Separating Various Fan Theorems” constructs a Kripke model with relative realizability rather than a formal system named multiple realizability (Lubarsky et al., 2015). “Concurrent Realizability on Conjunctive Structures” develops an algebraic concurrent realizability model and explicitly treats its relation to multiple realizability as interpretive (Beffara et al., 2021). “Linear Realisability over nets: multiplicatives” similarly develops an orthogonality semantics for multiplicative linear logic, with multiple realizability arising from set-valued interpretations rather than constituting its primary problem (Ragot et al., 2024).

2. Realisability as a set of strategies and implementations

In timed assume–guarantee specification theory, realisability is defined as a restriction on the behaviours of timed specifications. A timed I/O automaton is

$P=(C,I,O,L,l^0,AT,Inv,coInv),$

where $C$ is a finite clock set, $I$ and $O$ are disjoint input and output alphabets, $L$ is a set of locations, and $AT$ is the action-transition relation. Its semantic model is a timed I/O transition system with special states $\bot$ and $\top$. The state $\bot$ represents incompatibility or assumption/guarantee error, while $\top$ represents violation of a system guarantee.

A strategy is a deterministic, tree-shaped timed I/O transition system in which every plain state accepts all environmental inputs but permits exactly one component move: either an output action or a maximal delay interval. A specification denotes a set $C$0 of such strategies. It is existentially realisable when

$C$1

A strategy $C$2 is realisable when it is free of $C$3 and semi-$C$4. This excludes explicit timestop and implicit time-locking, under the paper’s non-zenoness assumption. The realisable strategy set is informally

$C$5

The theory therefore does not identify realisability with the selection of one arbitrary controller. It generally retains all realisable strategies. Quotient can produce a canonical, most-general specification for a class of controllers, but it does not produce a unique implementation program. Normalisation removes behaviours from which incompatibility is unavoidable, while realisation removes behaviours from which system-guarantee violation is unavoidable. These are computed by dual local fixpoint procedures: $C$6-backpropagation for normalisation and $C$7-backpropagation for realisation.

The framework contains four principal operators. Parallel composition integrates components; conjunction combines viewpoints; disjunction preserves alternatives; and quotient synthesises the most-general controller completing a plant into a specification. In particular, disjunction can preserve strategies from either operand, while quotient denotes all realisable controllers satisfying the quotient condition. The congruence theorem states that realisable refinement equivalence is a congruence for $C$8, subject to alphabet composability. The theory therefore provides compositional manipulation of collections of realisable strategies rather than a cardinality theorem asserting that a specification has more than one implementation (Chilton et al., 2013).

Timed games make the multiplicity explicit. A game has a component, an environment, and a coin resolving simultaneous choices. The component and environment propose actions or maximal delays; actions take priority over delays, and smaller maximal delays take priority when both propose delays. Refinement can be characterised by inclusion of environment-winning strategies. The resulting semantics is set-valued, and closure operations include behaviours indistinguishable under realisable substitutive refinement. The paper does not define a predicate asserting $C$9, nor does it prove that every realisable specification has multiple implementations.

3. Realizability semantics and computational pluralism

In logic and computability theory, multiple realisability can concern either many realisers for one assertion or several realizability interpretations for one source logic. “Realizability Models Separating Various Fan Theorems” constructs a Kripke model whose computational resources are drawn from a countable ideal of Turing degrees rather than from one fixed oracle. Let

$I$0

and define

$I$1

The ideal is downward closed under Turing reducibility and closed under join, but has no greatest member. A Kripke model over $I$2 stages access to these oracles: at a node of length $I$3, only an appropriate initial segment of the sequence is computationally settled. The internal applicative structure consists of indices for functions computable using the internal generic real, and the resulting model is denoted $I$4.

This construction produces a node-dependent family of available computational resources. A realizer is an index in the applicative structure, but the computations available at a node depend on the oracle information settled there. Consequently, a proposition can be realized by computational information from different Turing degrees at different nodes. The model satisfies dependent choice at the root and is used to separate fan principles. With a Yu–Simpson sequence of reals, it validates W-D-FAN while refuting D-FAN, and it validates full FAN while refuting WKL in a different construction. The formal framework is called a Kripke model with relative realizability, but it provides a precise mathematical instance of computational pluralism.

A more general framework is given by “Uniform Realizability Interpretations.” It separates the treatment of atomic predicates from the uniform treatment of logical connectives and quantifiers. For a base interpretation, an atomic predicate $I$5 receives a relation

$I$6

and the uniform realizability clauses then define the interpretation of conjunction, implication, and quantification. In particular,

$I$7

and

$I$8

The quantified variable need not be encoded in the realizer. Computational content may instead enter through the interpretation of atomic predicates. Precise interpretations of $I$9 yield Kleene and modified realizability; uniform interpretations yield internal Herbrand-style quantifiers; approximate interpretations yield bounded or finite-set realizability; and state-dependent interpretations yield learning realizability. Classical realizability adds computational content to falsity and equality.

This framework distinguishes two forms of multiplicity. A fixed interpretation may assign many realisers to one formula, while different base interpretations may produce fundamentally different realizability schemes for the same source logic. The general soundness theorem states that if the non-logical axioms are realizable under a base interpretation, then every theorem of the source theory is realizable. The logical architecture is uniform, while witness production, approximation, standardness, state update, or control behaviour is determined by atomic interpretation (Berger et al., 4 Mar 2026).

4. Orthogonality, nets, and concurrent realisers

Orthogonality-based semantics makes multiplicity intrinsic by interpreting formulas as sets of computational objects. In concurrent realizability, the realisers are processes with global fusions,

$O$0

where $O$1 is a process in a non-replicating $O$2-calculus and $O$3 is an equivalence relation identifying channels. A pole $O$4 contains closed processes, and orthogonality is defined by successful interaction:

$O$5

For $O$6,

$O$7

and a behaviour is a biorthogonally closed set,

$O$8

A proposition is therefore a collection of process-with-fusion pairs. Parallel composition, name-passing, restriction, and global channel identification provide operationally distinct ways of realising the same behaviour. The semantics includes separated parallel composition, communicating parallel composition, tensor, par, and linear implication. Hom-sets

$O$9

may contain many realisers of one entailment, with composition given by

$L$0

The framework is restricted to a non-replicating fragment corresponding to multiplicative linear logic. The omission of weakening and contraction combinators reflects the linear resource discipline: processes cannot freely duplicate or discard resources. The paper establishes soundness of multiplicative linear logic in the resulting conjunctive parallel algebra and represents Honda–Yoshida combinators under regular poles (Beffara et al., 2021).

“Linear Realisability over nets: multiplicatives” gives a parallel construction for MLL and MLL*. Nets are finite labelled hypergraphs containing tensor, par, cut, and generalized-daimon links. A type is a biorthogonally closed set of nets, and a net realizes a formula or hypersequent when it belongs to the corresponding interpretation. Orthogonality is existential with respect to cut elimination:

$L$1

Because daimon–par reduction is nondeterministic, one successful reduction is sufficient; different reduction paths may lead to different normal forms. Thus multiplicity arises both from several nets belonging to one type and from several computational reductions of one interaction.

The paper proves adequacy for MLL* and MLL and completeness under suitable families of interpretation bases. Completeness classifies nets that realize a sequent uniformly across all appropriate interpretations; it does not assert a unique realizer. A formula interpretation may contain zero, one, finitely many, or infinitely many non-isomorphic nets. Multiple realizability is therefore a semantic consequence of set-valued orthogonality, not a separate cardinality theorem (Ragot et al., 2024).

5. Mechanistic and structural multiplicity in chemical pathways

Chemical pathway analysis distinguishes stoichiometric existence from executable realisability. A reaction network is represented as a directed hypergraph, and a pathway is specified by an integer-valued flow $L$2 assigning reaction multiplicities. Stoichiometric conservation alone does not ensure that the selected reactions can be fired from the available inputs. A flow is realisable when the Petri net constructed from the flow admits a firing sequence

$L$3

The Petri-net construction adds external places containing exactly the prescribed number of tokens for each reaction occurrence. Consequently, reachability of $L$4 enforces the selected reaction identities and multiplicities rather than merely asking whether some product is reachable in the unrestricted network.

One flow can have multiple executable realisations. Different firing sequences may use the same reactions with the same multiplicities, as in the sequences

$L$5

A realisability certificate is an occurrence-net process mapping individual places and transition occurrences back to the original Petri net. It records token histories and causal dependencies. A certificate has a topological ordering that yields an executable firing sequence, and one flow may admit many certificates. This is mechanistic multiple realisability: the stoichiometric pathway remains fixed while its reaction order and token-level causal structure vary.

The paper distinguishes this from alternative pathways. Different flows may use different reactions, multiplicities, or intermediates while satisfying the same transformation motif. Such alternatives constitute structural multiplicity, while multiple certificates of one flow constitute mechanistic multiplicity. Functional equivalence is treated operationally through a common input/output transformation rather than through a general semantic identity of biological function. Differences in regulation, kinetics, energetic cost, enzyme use, compartmentalisation, atom mapping, or physiological role are not eliminated by equal stoichiometric input/output behaviour.

Two extensions expose additional forms of collective realisability. A flow is scaled-realisable if some multiple $L$6 is realisable:

$L$7

A flow is borrow-realisable if additional molecules can be temporarily borrowed and returned at the end. Borrowing can function as a network catalyst that breaks a circular material dependency. In the pentose phosphate pathway example, a flow that is not scaled-realisable becomes borrow-realisable by borrowing one molecule of Glyald. These notions show that multiplicity may concern not only alternative executions but also alternative resource regimes (Andersen et al., 2023).

6. Physical realisation, communication models, and scientific significance

Physical realisation and effective behaviour

In elastodynamic cloak design, multiple realisability appears as the separation between an effective material field, an explicit microstructure, and the resulting wave response. Transformation elasticity supplies an ideal polar or Cosserat-type tensor that generally violates the minor symmetries of Cauchy elasticity. The paper instead optimises within the orthotropic Cauchy class using a coordinate-based neural field and a differentiable finite-element solver.

Several effective-field descriptions produce similar cloaking metrics: one homogeneous orthotropic material, two phase-specific materials, a $L$8 design, and cellwise fields with 46 or 100 active materials. These are alternative effective descriptions, not necessarily exact realisations of one identical tensor field. Manufacturability is imposed through a catalogue of approximately $L$9 homogenised binary $AT$0 solid/void cells and a Gaussian-mixture-model realisability manifold.

The microstructure map from binary geometries to effective descriptors is potentially many-to-one. Conditional diffusion generates multiple geometries conditioned on target effective parameters, while neural-field inverse design and nearest-neighbour selection provide alternative projection mechanisms. Diffusion-generated and neural-field-designed cells can be visibly different while approximating the same effective target. The paper therefore gives constructive evidence for multiple microstructural realisation, while not proving exact equality of homogenised tensors or enumerating the full equivalence class.

Multiplicity at the effective-property level does not entail equality of complete dynamic behaviour. Homogenised simulations recover approximately $AT$1 of the defect-free reference surface-displacement magnitude, whereas direct finite-element simulation of the fully resolved microstructure recovers approximately $AT$2. Moreover, the optimisation compares displacement magnitudes and can ignore phase differences. The reported effective cloak has approximately the same amplitude pattern as the ideal cloak but a phase lag of roughly $AT$3, or $AT$4. Thus equivalence must be indexed to a level: effective descriptor, homogenised response, scalar performance metric, or fully resolved wave field (Aznaurov et al., 4 Sep 2026).

Communication-model-relative realisability

In multiparty session types, a global type specifies global behaviour and projects canonically to communicating finite-state machines. The same projected system can have different executions under different communication models. A communication model is a set of allowed executions, and the principal models include bag-based, peer-to-peer FIFO, causally ordered, mailbox, bus, and synchronous communication.

A global type is realisable in a communication model when the projected implementation conforms to the global semantics and is deadlock-free. The execution language is model-relative:

$AT$5

Thus one global specification and one projection can have different realisability statuses under different communication models. One example is realisable under peer-to-peer communication but not mailbox communication because merging FIFO queues can introduce head-of-line blocking. Another is realisable under mailbox communication but not peer-to-peer communication because merged queues disambiguate a branch that separate channels leave inconsistent. Deadlock-free realisability is not monotone as communication becomes more or less constrained.

The paper sharply distinguishes this model-relative realisability from subtyping. Under communication models containing all synchronous executions, subtyping is characterised by inclusion of existential MSC languages and is independent of the communication model. Realisability, by contrast, depends on permitted linearisations, queue discipline, orphan messages, and deadlock behaviour. Decision procedures range from NLOGSPACE and PSPACE in restricted settings to EXPSPACE for regular causally closed asynchronous models with explicit complements (Giusto et al., 5 Dec 2025).

General relativity and functional multiple realisability

In philosophy of physics, multiple realisability is applied to the recovery of general relativity by quantum-gravity theories. The central claim is that recovery of the Einstein field equations does not by itself establish that a theory has identified the correct microscopic ontology. General relativity may be multiply realizable if similar macroscopic gravitational dynamics arise from different quantum structures.

Three mechanisms are examined. Induced gravity obtains an Einstein–Hilbert-type effective action from quantum-field fluctuations on a Lorentzian manifold. Thermodynamic gravity derives the Einstein field equations from entropy, temperature, heat flow, local equilibrium, causal horizons, and an area–entropy relation. Entanglement gravity derives them from entanglement entropy, area scaling, short-distance conformal behaviour, and entanglement equilibrium. These mechanisms differ in their underlying structures but can yield the same effective gravitational equations.

The argument is conditional. Induced gravity assumes a Lorentzian manifold in advance; thermodynamic derivations require nontrivial equilibrium and area-law assumptions; and entanglement derivations rely on assumptions concerning ultraviolet behaviour, area scaling, and local equilibrium. The paper therefore does not establish that every quantum-gravity microtheory recovers general relativity. Its conclusion is epistemic: recovery of general relativity is important evidence of classical adequacy, but multiple realisability reduces its discriminatory power among candidate microtheories. A theory can recover the Einstein field equations while being incorrect about strings, spin networks, causal sets, or other proposed microscopic entities (Jaksland, 2020).

Dynamical and topological multiplicity

In dynamical systems, a finite partial order can be realised as the Smale order of a structurally stable system in several ways. Every finite partial order is realisable as the Smale order of an $AT$6-stable diffeomorphism on a closed surface and as the Smale order of an Anosov flow on a closed $AT$7-manifold. For structurally stable surface diffeomorphisms with trivial attractors and repellers, realisability is characterised by a connectivity condition.

The same order can occur on different surfaces, including the sphere and torus, and the paper constructs infinitely many distinct conjugacy classes of stable diffeomorphisms for every order satisfying the connectivity condition. DA constructions, plug gluings, genus-increasing operations, and boundary-cycle choices can alter the global system without changing the Smale order. The order records the organisation of basic pieces but does not determine the ambient surface, conjugacy class, genus, or detailed attractor geometry (Iakovoglou, 2020).

These results provide a dynamical analogue of multiple realisability: one abstract relational structure admits many inequivalent geometric and dynamical realisations. They also show that realisability can be category-dependent. Orders failing the connectivity condition remain realisable in the $AT$8-stable surface and Anosov-flow categories but not in the category of structurally stable surface diffeomorphisms with only trivial attractors and repellers.

Multiple realisability therefore functions as a cross-disciplinary concept rather than a single formal property. Its rigorous use requires identifying the target object, the admissible realisers, the equivalence relation between outcomes, and the level at which multiplicity is asserted. In some settings it is established by explicit non-uniqueness results; in others it is represented by set-valued semantics, alternative execution structures, or model-relative changes in admissibility. The strongest general conclusion supported across these domains is not that every specification has many implementations, but that successful realisation alone does not generally determine a unique underlying mechanism.