Coherence and strictification for self-similarity

This paper studies questions of coherence and strictification related to self-similarity - the identity $S\cong S\otimes S$ in a (semi-)monoidal category. Based on Saavedra's theory of units, we first demonstrate that strict self-similarity cannot simultaneously occur with strict associativity -- i.e. no monoid may have a strictly associative (semi-)monoidal tensor, although many monoids have a semi-monoidal tensor associative up to isomorphism. We then give a simple coherence result for the arrows exhibiting self-similarity and use this to describe a `strictification procedure' that gives a semi-monoidal equivalence of categories relating strict and non-strict self-similarity, and hence monoid analogues of many categorical properties. Using this, we characterise a large class of diagrams (built from the canonical isomorphisms for the relevant tensors, together with the isomorphisms exhibiting the self-similarity) that are guaranteed to commute.


Introduction
An object S in a semi-monoidal category (C, ⊗) is self-similar when it satisfies the identity S ∼ = S ⊗ S. In a 1-categorical sense, self-similar objects are simply pseudo-idempotents and thus share many categorical properties with unit objects as characterised by Saavedra (see Section 2.1); they also provide particularly well-behaved examples of split idempotents (see Section 3.1). The most familiar non-unit example of a self-similar object is undoubtedly the natural numbers in the monoidal categories (Set, ×) and (Set, ), as illustrated by Hilbert's parable of the 'Grand Hotel' (see [Yanofsky 2013] for a good exposition in a general context). Topologically, self-similarity is clearly seen in the Cantor set & other fractals [Hines 1998, Leinster 2011; algebraically, it is very closely connected with Thompson's groups (see Section 5), the polycyclic monoids [Hines 1998, Lawson 1998, Hines 1999, and finds applications to algebraic models of tilings [Kellendonk, Lawson 2000].
In computer science, self-similarity plays a key role in categorical models of untyped systems such as the C-monoids of [Lambek, Scott 1986] (single-object Cartesian closed categories without unit objects modelling untyped lambda calculus -see [Hatcher, Scott 1986] for a survey). It is particularly heavily used in Girard's Geometry of Interaction program [Girard1988a, Girard 1988b where, as well as being a key feature of the 'dynamical algebra' it is used together with compact closure to construct monoids isomorphic to their own endomorphism monoid [Abramsky et al. 2002, Hines 1998, Hines 1999]. More recently, it has found applications in linguistic and grammatical models [Hines 2013c], categorical models of quantum mechanics , and -via the close connections with Thompson's groups -is relevant to cryptography and cryptanalysis [Shpilrain, Ushakov 2006, Hines 2014b, and homotopy idempotents (see Section 5).
The aim of this paper is to give coherence results and a strictification procedure for self-similarity, and so to relate the isomorphisms exhibiting selfsimilarity with canonical isomorphisms for the relevant (semi-)monoidal structures. The motivation for this is the observation (Theorem 4.2), based on Saavedra's theory of units (See Section 2.1), that strict self-similarity and strict associativity are mutually exclusive -either one or the other of these properties must be up to non-trivial isomorphism.

Categorical preliminaries
We refer to [MacLane 1998] for the theory of monoidal categories. We work with a slight generalisation that satisfies all the axioms for a monoidal category except for the existence of a unit object; following [Kock 2008], we refer to these as semi-monoidal. Definition 2.1. A semi-monoidal category is a category C with a functor ⊗ : C × C → C that is associative up to an object-indexed family of natural isomorphisms τ X,Y,Z : X ⊗ (Y ⊗ Z) → (X ⊗ Y ) ⊗ Z satisfying MacLane's pentagon condition ( τ W,X,Y ⊗ 1 Z ) τ W,X⊗Y,Z (1 W ⊗ τ X,Y,Z ) = τ W ⊗X,Y,Z τ W,X,Y ⊗Z A functor between semi-monoidal categories that (strictly) preserves the tensor is a (strict) semi-monoidal functor. We assume the obvious definition of semi-monoidal equivalence of categories.
(The above definition differs from that of [Kock 2008] in that we do not assume strict associativity -see Theorem 4.2 for the motivation for this). When we do consider unit objects, we will use Saavedra's characterisation, rather than MacLane & Kelly's original definition -see Section 2.1 below.
Remark 2.2. Any category C may be given a (degenerate) semi-monoidal tensor by fixing some object X ∈ Ob(C), and and defining A ⊗ B = X, f ⊗ g = 1 X for all objects A, B ∈ Ob(C) and arrows f, g ∈ Arr(C). In the theory of monoidal categories, standing assumptions such as (monoidal) well-pointedness are used to exclude such pathologies (see, for example, [Abramsky , Heunen 2012]); in the absence of a unit object we will instead use the assumption of 'faithful objects' given below.
. A consequence of monoidal well-pointedness is that for arbitrary A ∈ Ob(C), where A is not a strict retract of the unit object, the functors (A ⊗ ), ( ⊗ A) : C → C are faithful. Motivated by this, we say that an object A of a semi-monoidal category (C, ⊗) is faithful when the functors A ⊗ , ⊗ A : C → C are faithful.
Convention 2.4. Objects of semi-monoidal categories are faithful Semimonoidal categories in the literature commonly arise as semi-monoidal subcategories of monoidally well-pointed categories and have faithful objects. To avoid repeatedly emphasising this, we will instead indicate when an object of a semimonoidal category is not faithful.

Saavedra's theory of units
MacLane's original presentation of the theory of coherence for monoidal categories gave a single coherence condition for associativity (the Pentagon condition) and four coherence conditions for the units isomorphisms. Three of these four axioms were shown to be redundant in [Kelly 1964], leaving a single coherence condition expressing the relationship between the units isomorphisms, and associativity. In [Saavedra 1972] an alternative characterisation of units objects was given that required no additional coherence conditions (see [Kock 2008] for a comprehensive study of this, and [Joyal, Kock 2011] for extensions of this theory). We follow this approach; Definition 2.5 and Theorem 2.6 below are taken from [Kock 2008, Joyal, Kock 2011].
We also refer to [Kock 2008] for the following key results: Theorem 2.6.
1. Saavedra units are units in the sense of MacLane / Kelly, and thus a semi-monoidal category with a (Saavedra) unit is a monoidal category.

Self-similar objects and structures
The theory of self-similarity is precisely the theory of pseudo-idempotents in (semi-)monoidal categories. The difference in terminology arises for historical reasons -in particular, differing conventions in computer science, mathematics, and pure category theory.
Definition 3.1. An object S in a semi-monoidal category (C, ⊗) is called selfsimilar when S ∼ = S ⊗ S. Making the isomorphism exhibiting this self-similarity explicit, a self-similar structure is a tuple (S, ¡) consisting of an object S ∈ Ob(C), and an isomorphism ¡ : S ⊗ S → S called the code isomorphism. We denote its inverse by ¡ −1 = £ : S → S ⊗ S and refer to this as the decode isomorphism.
A strictly self-similar object is an object S such that (S, 1 S ) is a selfsimilar structure. The endomorphism monoid of a strictly self-similar object is thus itself a semi-monoidal category with a single object -i.e. it is a semimonoidal monoid.
Examples 3.2. Examples of non-strict self-similarity are discussed in Section 1. Strict examples include Thompson's group F (see [Brown 2004] for the semimonoidal tensor and associativity isomorphism of this group, and Section 5 for the explicit connection with strict self-similarity), and the group of bijections on the natural numbers, with the semi-monoidal tensor given by , σ(n) = n + 1 n even, n − 1 n odd.
An elementary arithmetic proof that the above data specifies a semi-monoidal monoid is given in [Hines 2013b]. More generally, it arises from a special case of a large class of representations of Girard's dynamical algebra (viewed as the closure of the two-generator polycyclic monoid [Nivat, Perrot 1970] under the natural partial order of an inverse semigroup) as partial functions, given in [Hines 1998, Lawson 1998]. Readers familiar with the Geometry of Interaction program will recognise the ( ) operation as Girard's model (up to Barr's l 2 : pInj → Hilb functor -see [Barr 1992, Heunen 2013) of the (identified) multiplicative conjunction & disjunction of [Girard1988a, Girard 1988b].

Self-similarity as idempotent splitting
Self-similar structures are a special case of idempotent splittings. We refer to [Freyd, Scedrov 1990, Lambek, Scott 1986, Selinger 2006, Street 1996 for the general theory and reprise some basic properties below: Definition 3.3. An idempotent e 2 = e ∈ C(A, A) splits when there exists some B ∈ Ob(C) together with arrows f ∈ C(A, B), g ∈ C(B, A) such that e = gf and f g = 1 B . We refer to the pair (f, g) as a splitting of the idempotent e 2 = e.
Remark 3.4. We may characterise self-similar structures in terms of splittings of identities: a self-similar structure (S, ¡) uniquely determines, and is uniquely determined by, an isomorphism ¡ such that (¡, £) is a splitting of 1 S and (£, ¡) is a splitting of 1 S⊗S .
This characterisation allows us to use standard results on idempotent splittings, such as their uniqueness up to unique isomorphism: Lemma 3.5. Given an idempotent e 2 = e ∈ C(A, A) together with splittings (f ∈ C(A, B), g ∈ C(B, A)) and (f ∈ C(A, B ), g ∈ C(B , A)), then there exists a unique isomorphism φ : B → B such that the following diagram commutes: Proof. This is a standard result of the theory of idempotent splittings. See, for example, [Selinger 2006].
Corollary 3.6. Self-similar structures at a given self-similar object are unique up to unique isomorphism Proof. The proof of this is somewhat simpler than the general case, as the splittings have both left and right inverses. Given a self-similar structure (S, ¡), and an isomorphism U : S → S, then (S, U ¡) is also a self-similar structure. Conversely, let (S, ¡ ) be a self-similar structure, and define U = ¡ £. Then (S, U ¡) = (S, ¡ ) and U = ¡ £ is the unique isomorphism satisfying this condition.
Remark 3.7. Uniqueness of self-similar structures Corollary 3.6 provides an illustration of the distinction between 'unique up to unique isomorphism', and 'actually unique'. If a self-similar structure at some object S ∈ Ob(C) is actually unique, then the only isomorphism from S to itself is the identity. Theorem 4.2 below then shows that S must be the unit object for the tensor given in Theorem 7.14.
The above characterisation of self-similar structures as idempotent splittings also gives an abstract categorical characterisation in terms of limits and colimits of diagrams: Proposition 3.8. Given a self-similar structure (S, ¡) of a semi-monoidal category (C, ⊗), then , and these two conditions characterise self-similar structures at S ∈ Ob(C).
Proof. A standard result on splitting idempotents is that a splitting of an idempotent

Strict self-similarity and strict associativity
Unit objects are special cases of self-similar objects -the distinction being that for an arbitrary self-similar object S, the functors S⊗ and ⊗S need not be fully faithful. We now describe how Saavedra's characterisation relates strictness for both associativity and self-similarity.  ( 1) and (1 ) are injective, and are isomorphisms precisely when the unique object of M is the unit.
Proof. Functoriality implies that 1 and 1 are monoid homomorphisms, and injectivity follows from the assumption (Convention 2.4) of faithful objects. These homomorphisms are isomorphisms precisely when they are fully faithful, in which case the unique object of M satisfies Saavedra's characterisation of a unit object.  M) , and thus by Lemma 4.1, the unique object of (M, ) ∼ = (η(M), ) is the unit object. (⇐) It is a standard result of monoidal categories that the endomorphism monoid of a unit object is an abelian monoid, and the tensor at this object coincides (up to isomorphism) with this strictly associative composition.

Remark 4.3. No simultaneous strictification
This paper is about strictification and coherence for self-similarity and its interaction with associativity. From Theorem 4.2, strictifying the associativity of a semi-monoidal monoid will result in non-strict self-similarity; conversely, strictifying self-similarity in a strictly associative setting will give a monoid with a non-strict semi-monoidal tensor (Proposition 7.18).
Examples 4.4. Finite and infinite matrices An illustrative example is given by infinitary matrix categories. Countable matrices over a 0-monoid R enriched with a suitable (partial, infinitary) summation 1 R form a category Mat R , with , and composition is given by the Cauchy product.
Assuming technical conditions on summation are satisfied, the subcategory of finite matrices has a strictly associative monoidal tensor, denoted ⊕ . On objects it is simply addition; given arrows, otherwise. Although this definition cannot be extended to infinite matrices, the endomorphism monoid Mat R (∞, ∞) can be given a non-strict semi-monoidal tensor (again, assuming technicalities on summation), such as x, y even, otherwise. This is the familiar 'interleaving' of infinite matrices, determined by the Cantor pairing c : N N → N given by c(n, i) = 2n + i (and used to great effect in modelling the structural rules of linear logic [Girard1988a, Girard 1988b). Any such isomorphism ¡ : N N → N will determine a (non-strict) semi-monoidal tensor on this monoid. However, by Theorem 4.2, no strict semi-monoidal tensor on Mat R (∞, ∞) may exist.

The group of canonical isomorphisms
In a semi-monoidal monoid, the isomorphisms canonical for associativity are closed under composition, tensor, and inverses, and thus form a group with a semi-monoidal tensor. As demonstrated in [Fiore, Leinster 2010], in the free case this group is the well-known Thompson group F (see [Cannon et al. 1996] for a non-categorical survey). An algebraic connection between this group and associativity laws is well-established (see [Brin 2004] for a more categorical perspective), and the tensor was given in [Brown 2004] -although not in categori-cal terms. An explicit connection with semi-monoidal monoids was observed in [Lawson 2007] where F is given in terms of canonical isomorphisms and singleobject analogues of projection / injection arrows for the tensor.
An interesting connection between Thomson's group F and the theory of idempotent splittings is given in [Brown, Geoghegan 1984], in the context of (connected, pointed) CW complexes. The unsplittable homotopy idempotents of a CW complex K are characterised by the fact that they give rise to a copy of F in π 1 (K). The categorical interpretation of this is unfortunately beyond the scope of this paper.

A simple coherence result for self-similarity
We now give a simple result that guarantees commutativity for a class of diagrams built inductively from the code / decode isomorphisms of a self-similar structure, and the relevant semi-monoidal tensor. This is modelled very closely indeed on MacLane's original presentation of his coherence theorem for associativity (briefly summarised in Section 6.1), in order to describe the interaction of self-similarity and associativity.

Coherence for associativity -the unitless setting
We first briefly reprise some basic definitions and results on coherence for associativity, taken from MacLane's original presentation [MacLane 1998], in the semi-monoidal setting. This is partly to fix notation and terminology, and partly to ensure that the absence of a unit object does not lead to any substantial difference in theory. We also restrict ourselves to the monogenic case, as this suffices to describe the interaction of associativity and self-similarity.
For more general, structural approaches to coherence, we refer to [Kelly 1974, Power 1989, Joyal, Street 1993] -these inspired the approach taken in Section 7 onwards.
Definition 6.1. The set T ree of free non-empty binary trees over some symbol x is is inductively defined by: x ∈ T ree, and for all u, v ∈ T ree then (uPv) ∈ T ree. The rank of a tree t ∈ T ree is the number of occurrences of the symbol x in t, or equivalently, the number of leaves of t.
We denote (the unitless version of ) MacLane's monogenic monoidal category by (W, P). This is defined by: Ob(W) = T ree and there exists a unique arrow (t ← s) ∈ W(s, t) iff rank(s) = rank (t). Composition is determined by uniqueness. Given p, q ∈ Ob(W), their semi-monoidal tensor is pPq; the tensor of arrows is again determined by uniqueness.
Remark 6.2. MacLane's definition [MacLane 1998] included the empty tree as a unit object, giving a monoidal, rather than semi-monoidal, category. Applying the common technique of adjoining a strict unit to a semi-monoidal category will recover MacLane's original definition, and MacLane's original theory in the exposition below. Definition 6.3. Given an object A of a semi-monoidal category (C, ⊗, τ , , ), MacLane's associativity substitution functor WSub A∈Ob(C) : (W, P) → (C, ⊗) is defined inductively below. When the context is clear, we elide the subscript on WSub . WSub(a)).
Remark 6.5. As W is posetal, all diagrams over W commute, so all (canonical) diagrams in C that are the image of a diagram in W are guaranteed to commute. Naturality and substitution are then used [MacLane 1998] to extend this to the general setting.

A preliminary coherence result for self-similarity
We now exhibit a class of diagrams based on identities, tensors, and the code / decode maps for a self-similar structure that are guaranteed to commute. This is based on a substitution functor from a posetal monoidal category that contains MacLane's (W, P) as a semi-monoidal subcategory.
Definition 6.6. The monogenic self-similar category (X , P) was defined in [Hines 1998] as follows: • Objects Ob(X ) = T ree • Arrows There exists unique (b ← a) ∈ X (a, b) for all a, b ∈ Ob(X ).
• Composition This is determined by uniqueness.
• Tensor Given u, v ∈ Ob(X ), their tensor is the binary tree uPv. The definition on arrows again follows from uniqueness.
• Unit object All objects e ∈ Ob(X ) are unit objects; the unique arrow (e ← ePe) is an isomorphism, and the functors (eP ), ( Pe) : X → X are fully faithful.
Remark 6.7. Abstractly, (X , P) may be characterised as the free monogenic indiscrete monoidal category. Thus, it is monoidally equivalent to the terminal monoidal category -in the semi-monoidal setting, it is more interesting.
(We again omit the subscript when the context is clear).
Remark 6.9. In stark contrast to MacLane's substitution functor, it is immediate that X Sub : (X, P) → (C, ⊗) is a strict semi-monoidal functor -no coherence conditions are needed to ensure functoriality.
We may now give a preliminary coherence result on self-similarity.
Lemma 6.10. Let (S, ¡) be a self-similar structure of a semi-monoidal category (C, ⊗), and let X Sub : (X, P) → (C, ⊗) be as above. Then every diagram over C of the form X Sub(D), for some diagram D over X , is guaranteed to commute.
Proof. As (X , P) is posetal, D commutes; by functoriality so does X Sub(D).
Remark 6.11. The diagrams predicted to commute by Lemma 6.10 are 'canonical for self-similarity', with arrows built from code / decode isomorphisms, identities, and the tensor. The more important question is about diagrams that are 'canonical for self-similarity & associativity' -when may these be guaranteed to commute? This follows as a special case of a more general result (Section 8).
Remark 6.12. Does WSub factor through X Sub? There is an immediate semi-monoidal embedding ι : (W, P) → (X , P).An obvious question is whether, or under what circumstances, the above substitution functors will factor through this embedding -i.e. when does the diagram of Figure 1 commute? Commutes?
It is immediate that this can only commute under very special conditions. Functoriality of WSub requires coherence conditions (i.e. MacLane's pentagon), whereas none are required for the functoriality of X Sub. Further, commutativity of this diagram would give a decomposition of canonical (for associativity) isomorphisms of (C, ⊗) into 'more primitive' operations built from {¡, £, ⊗}; in particular, τ S,S,S = (£ ⊗ 1 S )(1 S ⊗ ¡). A slight generalisation of Isbell's argument on the skeletal category of sets [MacLane 1998] would show that when (C, ⊗) admits projections, S is the terminal object. Instead of giving a direct proof of this, we will give a more general result in Corollary 7.22.

Strictification for self-similarity
We first describe a strictification procedure for self-similarity that gives a semimonoidal equivalence between a monoid and a monogenic category, then use this to give a coherence theorem that answers the question posed in Remark 6.11 in a more general setting. This strictification procedure generalises the 'untyping' construction of [Hines 1998, Hines 1999 (and indeed, corrects it in certain cases -see Remark 7.6).
Definition 7.1. Given a semi-monoidal category (C, ⊗, τ , , ) and arbitrary S ∈ Ob(C), the semi-monoidal category freely generated by S, denoted F S , is defined analogously to the usual monoidal definition. The assignment Inst : T ree → Ob(C) is defined inductively by Inst(x) = S ∈ Ob(C), and Inst(pPq) = Inst(p) ⊗ Inst(q) and based on this, objects and arrows are given by Ob(F S ) = T ree, and F S (u, v) = C(Inst(u), Inst(v)).
Composition is inherited in the natural way from C, as is the tensor: on objects this is simply the formal pairing P , and given arrows f ∈ F S (u, v), g ∈ F S (x, y) we have f Pg = f ⊗ g ∈ F S (uPx, vPy) = C(Inst(u) ⊗ Inst(x), Inst(v) ⊗ Inst(y)).
The assignment Inst : T ree → Ob(C) extends in a natural way to a strict semi-monoidal functor; using computer science terminology, we call this the instantiation functor Inst S : (F S , P) → (C, ⊗). It is as above on objects, and the identity on hom-sets, as F S (u, v) = C(Inst(u), Inst(v)). It is immediate that this epic strict semi-monoidal functor is a semi-monoidal equivalence of categories. When the object is clear from the context, we simply write Inst : The image of Inst S is the full semi-monoidal subcategory of C inductively generated by the object S, together with the tensor ⊗ . We refer to this as the semi-monoidal category generated by S within (C, ⊗), denoted (C S , ⊗).
Based on the above definitions, the following are immediate: Lemma 7.2. Let (S, ¡) be a self-similar structure of a semi-monoidal category (C, ⊗). Then the small semi-monoidal categories (W, P), (X , P) and (F S , P) have the same set of objects, and 1. The tuple (x, ¡) is a self-similar structure of (F S , P).
3. The following diagram commutes: Remark 7.3. The commutativity of the left hand triangle in the above diagram is well-established, and part of a standard approach to coherence for associativity and other properties. In particular, Joyal & Street phrased MacLane's theorem as an equivalence between the free monoidal category on a category and the free strict monoidal category on a category (see [Joyal, Street 1993] for details and extensions of this approach).

Functors from categories to monoids
The monic functor X Sub : (X, P) → (F S , P) specifies a distinguished wide semimonoidal subcategory of (F S , P); we use the following notation and terminology for its arrows: Definition 7.4. Given a self-similar structure (S, ¡) of a semi-monoidal category, we define an object-indexed family of arrows, the generalised code isomorphisms, by {¡ u = X Sub(x ← u) ∈ F S (u, x)} u∈Ob(F S ) . We refer to their inverses, {£ u = X Sub(u ← x) ∈ F S (x, u)} u∈Ob(F S ) as the generalised decode isomorphisms. Remark 7.5. As observed in Remark 7.10 below, the above object-indexed families of arrows are the components of a natural transformation. An alternative perspective is that (¡ u , £ u ) is a splitting of 1 u , and (£ u , ¡ u ) is a splitting of 1 x . The unique isomorphism X Sub(v ← u) = £ v ¡ u ∈ F S (u, v) is then the isomorphism exhibiting the uniqueness up to isomorphism of idempotent splittings described in Corollary 3.6. Remark 7.6. As (F S , P) is freely generated, we may give an inductive characterisation of the generalised code arrows by: and similarly for the generalised decode arrows. This is used as a definition in [Hines 1999], where it is (incorrectly) assumed that (C S , ⊗) ∼ = (F S , P) in every case -an assumption holds for the particular examples considered there, but not more generally.
These generalised code / decode arrows allow us to define a fully faithful functor from F S to the endomorphism monoid F S (x, x), considered as a singleobject category.
Definition 7.7. Let (S, ¡) be a self-similar object of a semi-monoidal category (C, ⊗). We denote the endomorphism monoid F S (x, x), considered as a singleobject category, by End(x), and define the generalised convolution functor are as in definition 7.4.
When the self-similar structure in question is apparent from the context, we will omit the subscript and write Φ : F S → End(x).
Proposition 7.8. The generalised convolution functor given above is indeed a fully faithful functor.
Proof. For all a ∈ Ob(F S ), Φ(1 a ) = 1 x . Given f ∈ F S (a, b), g ∈ F S (b, c), then , and thus Φ is full. Finally, given f, f ∈ F S (a, b), then A simple corollary is that the category freely generated by a self-similar object, and the endomorphism monoid of that object (considered a a singleobject category) are equivalent: Corollary 7.9. Let (S, ¡) be a self-similar object of a semi-monoidal category (C, ⊗). Then the categories F S and End(x) are equivalent.
Proof. Since Φ is fully faithful, it simply remains to prove that it is isomorphismdense. For arbitrary u ∈ Ob(F S ), the generalised code/decode arrows ¡ u ∈ F S (u, x) and In Definition 7.13 below, we give a semi-monoidal tensor on End(x) that makes the above equivalence a semi-monoidal equivalence of categories.
Remark 7.10. Corollary 7.9 guarantees the existence of suitable functors exhibiting this equivalence of categories; more explicitly, let us denote the obvious inclusion by ι : End(x) → F S . Then Φι = Id End(x) , and there is a natural transformation from ιΦ to Id F S whose components are the generalised decode isomorphisms of Definition 7.4: It is also almost immediate that a diagram over F S commutes iff its image under Φ commutes; we prove this explicitly in order to illustrate how this relies on uniqueness of generalised code / decode arrows: Corollary 7.11. Let (S, ¡) be a self-similar object of a semi-monoidal category (C, ⊗). Then a diagram D over F S commutes iff Φ(D) commutes.

Proof.
(⇒) This is a simple, well-known consequence of functoriality. (⇐) Let D be an arbitrary diagram over (F S , P). Up to the obvious inclusion ι : End(x) → F S , D and Φ(D) are diagrams in the same category; we treat their disjoint union D Φ ( D) as a single diagram. We then add edges to D Φ(D) by 9 9 x Φ(g) 9 9 £v y y linking each node n with its image using the unique generalised code / decode arrows. This is illustrated in Figure 2. Each additional polygon added to D Φ(D) commutes by definition of Φ. Thus the entire diagram commutes iff D commutes iff Φ(D) commutes.

Semi-monoidal tensors on monoids
We now exhibit a semi-monoidal tensor on the endomorphism monoid of a self-similar object such that the equivalence of Corollary 7.9 becomes a semi-monoidal equivalence.
Lemma 7.12. Let (S, ¡) be a self-similar structure of a semi-monoidal category and let (F S , P, t , , ) be the semi-monoidal category freely generated by S. Then, up to the inclusion End(x) → F S , Proof. Pv t u,v,w £ uP(vPw) , and by Remark 7.6, the following diagram commutes:

Given
By naturality of t , , , as required.
Based on the above lemma, we give a semi-monoidal tensor on the endomorphism monoid of a self-similar object.
Definition 7.13. Let (S, ¡) be a self-similar object of a semi-monoidal category. We define the semi-monoidal tensor induced by (S, ¡) to be the monoid homomorphism ¡ : End(x) × End(x) → End(x) given by When the self-similar structure is clear from the context, we elide the subscript, and write : Theorem 7.14. The operation defined above is a semi-monoidal tensor, and thus (End(x), ) is a semi-monoidal monoid.
Proof. This follows from Theorem 7.14 and Theorem 7.12.
At a given self-similar object S ∈ Ob(C), the semi-monoidal tensor is determined by the choice of isomorphism ¡ ∈ C(S ⊗ S, S); however, these are related by conjugation in the obvious way, and thus is unique up to unique isomorphism.
Proposition 7.16. Let (S, c) and (S, ¡) be self-similar structures at a given self-similar object. Then f ¡ g = ¡c −1 (f c g)c£ for all f, g ∈ End(x).
Proof. This follows by direct calculation on Definition 7.13; alternatively, and more structurally, it follows from the uniqueness up to unique isomorphism of idempotent splittings, and hence self-similar structures (Corollary 3.6).
Remark 7.17. The above Proposition does not imply that all semi-monoidal tensors on a given monoid are related by conjugation. As a counterexample, the monoid of functions on N has distinct semi-monoidal tensors, arising from the fact that it is a self-similar object in both (F un, ×) and (F un, ), that are clearly not related in this way (the relationship between the two is non-trivial and a key part of Girard's Geometry of Interaction program [Girard1988a, Girard 1988b, the details of which are beyond the scope of this paper).

From Theorem 4.2, (End(x),
) can only be strictly associative when x is the unit object for . When (S, ¡) is a self-similar structure of a strictly associative semi-monoidal category (e.g. the rings isomorphic to their matrix rings characterised in [Hines, Lawson 1998]), the associativity isomorphism for has the following neat form: Proposition 7.18. Let (S, ¡) be a self-similar object of a strictly associative semi-monoidal category (C, ⊗). Then the associativity isomorphism for (End(x), ) is given by α Proof. This follows by direct calculation on Part 2. of Lemma 7.12.

The strictly self-similar form of a monogenic category
The following is now immediate: Theorem 7.19. Let (S, ¡) be a self-similar structure of a semi-monoidal category (C, ⊗). Then (C S , ⊗), (F S , P) and (End(x), ) are semi-monoidally equivalent.
Proof. The semi-monoidal equivalence between (F S , P) and (C S , ⊗) is given by the semi-monoidal functor of Definition 7.1. From Corollary 7.15, the equivalence of categories between F S and End(x) gives a semi-monoidal equivalence between (F S , P) and (End(x), ).
Corollary 7.20. Every monogenic semi-monoidal category with a self-similar generating object is semi-monoidally equivalent to a semi-monoidal monoid. This justifies the description of (End(x), ) as the self-similarity strictification of (C S , ⊗).
Remark 7.21. A general principle is that 'categorical structures' are preserved by equivalences of categories. For example, if C S is closed, then so is End(x); this is used implicitly in [Lambek, Scott 1986] to construct single-object analogues of Cartesian closed categories, and in [Hines 1998, Hines 1999] to construct single-object analogues of compact closure. Similarly, when (C S , ⊗) admits projections / injections, (End(x), ) contains a copy of Girard's dynamical algebra [Hines 1998, Lawson 1998, Hines 1999] and under relatively light additional assumptions, admits a matrix calculus ]. In general, we may find single-object (i.e. monoid) analogues of a range of categorical properties.
We may now answer the question posed in Remark 6.12.
Corollary 7.22. The diagram of Figure 1 commutes precisely when the selfsimilar object in question is the unit object.
Proof. In the self-similarity strictification of (C S , ⊗), the self-similarity is exhibited by identity arrows. Commutativity of the diagram of Figure 1 implies that (End(x), ) has a strictly associative semi-monoidal tensor, so by Theorem 4.2, the unique object of End(x) is the unit object for . The equivalences of Theorem 7.19 then imply that S is the unit object for (C S , ⊗).
Self-similarity strictification also illustrates a close connection between the generalised convolution and instantiation functors; informally, generalised convolution is simply instantiation in an isomorphic category: Proposition 7.23. Let (S, ¡) be a self-similar object of a semi-monoidal category (C, ⊗), and denote by (F S , P) and (F x , Q) the semi-monoidal categories freely generated by S, and the unique object of (End(x), ), respectively. Then there exists a semi-monoidal isomorphism K : (F S , P) → (F x , Q) such that the following diagram of semi-monoidal categories commutes: Proof. We define the semi-monoidal functor K : (F S , P) → (F x , Q) as follows: • The small categories (F x , Q) and (F S , P) have the same underlying set of objects; we take K to be the identity on objects.
The inverse is immediate, as is the (strict) preservation of the semi-monoidal tensor. The commutativity of the above diagram follows by expanding out the definitions of Φ and Inst.

General coherence for self-similarity
We now consider coherence in the general case. Let us fix a a self-similar structure (S, ¡) of a semi-monoidal category (C, ⊗, τ , , ). We will abuse notation slightly; based on the monoid isomorphism End(x) ∼ = C(S, S), we treat the semi-monoidal tensor ¡ equally as an operation on C(S, S) = F S (x, x) and denote the (unique) associativity isomorphism for as α ∈ C(S, S). The question we address is the following: Given a diagram over C S with arrows built inductively from { ⊗ , τ , , , ¡ , , α , ( ) −1 }, when may it be guaranteed to commute?
We first fix some terminology.
Notation 8.1. Given a category C and a class Γ of operations and arrows of C, we say that a diagram is canonical for Γ when its edges are built inductively from members of Γ. For example, in a semi-monoidal category (C, ⊗ , τ , , ), a diagram canonical for { ⊗ , τ , , , ( ) −1 } is a diagram canonical for associativity, as usually understood.
The following demonstrates that a simple appeal to freeness is not sufficient: Proposition 8.2. Let (F S , P, t , , ) be the semi-monoidal category freely generated by S. Then, over F S 2. All diagrams canonical for { P , ¡ , , ( ) −1 } commute.
Proof. 1. is well-established; it follows from the monic-epic decomposition of MacLane's substitution functor described in Lemma 7.2 and (in the monoidal case) is commonly used [Joyal, Street 1993] to study coherence. 2. follows similarly from Lemma 7.2. For 3., the following diagram is canonical for selfsimilarity and associativity: Applying Φ : (F S , P) → (End(x), ) to this diagram gives the associativity isomorphism for (End(x), ) as α = 1 x , so by Theorem 4.2 the unique object of End(x) is the unit object for . Appealing to the semi-monoidal equivalences of Theorem 7.19 gives that S ∈ Ob(C S ) is the unit object for ⊗ .
We now introduce an equivalence relation on diagrams over F S that allows us to answer this question in the free case: Definition 8.3. As F S is a small category, we may treat a diagram D over F S , with underlying directed graph G = (V, E), as a pair of functions D V : V → Ob(F S ) and D E : E → Arr(F S ) satisfying D E (e) ∈ F S (D V (v), D V (w)) for all edges v e −→ w ∈ E We will omit the subscripts on D V and D E when the context is clear.
Given diagrams T, U with underlying graphs G = (V, E) and G = (V , E ) respectively, we say they are self-similarity equivalent, written T ∼ ¡£ U when there exists a graph isomorphism η : G → G such that, for all edges s e −→ t of G, the following diagram commutes: y y (An intuitive description is illustrated by example in Figure 3). This is an equivalence relation, since the object-indexed isomorphisms ¡ and £ specify a wide posetal subcategory of F S . We denote the corresponding equivalence classes by [ ] ¡£ . Proof. This is immediate from the definition, and a slight generalisation of the reasoning in the proof of Theorem 7.11.
The above theorem answers the question posed at the start of this section in the 'formal' setting (F S , P). To map this free setting to the concrete setting, we apply the Inst : (F S , P) → (C S , ⊗) functor, giving the following corollary: Corollary 8.6. Given a diagram E over C S canonical for {⊗, τ , , , ¡, , α, ( ) −1 }, then E is guaranteed to commute when there exists a diagram D over F S that is canonical for {P, t , , , ¡ , , α, ( ) −1 } satisfying 1. D is guaranteed to commute by Theorem 8.5 above.

Inst(D) = E.
The identification of generalised convolution as the instantiation functor of a semi-monoidally isomorphic category (Proposition 7.23) then translates the above into the following neat heuristic for characterising such diagrams: Corollary 8.7. Let E be a diagram over C S canonical for {⊗, τ , , , ¡, , α, ( ) −1 }. Let us form a new diagram E over C(S, S) by the following procedure: • Replace every object in E by S.
• Replace every occurrence of ⊗ by .
• Replace every occurrence of ¡ by 1 S .
Then E , which is canonical for { , α, ( ) −1 }, is guaranteed to commute by MacLane's coherence theorem for associativity iff E is guaranteed to commute by Corollary 8.6.