Diagrammatic Negative Information

Vincent Abbott Gioele Zardini

College of Engineering, Computing and Cybernetics Laboratory for Information and Decision Systems
Australian National University Massachusetts Institute of Technology
Canberra, ACT, Australia Cambridge, MA, USA
[email protected] [email protected]

The flow of information through a complex system can be readily understood with category theory.
arXiv:2404.03224v1 [math.CT] 4 Apr 2024

However, negative information (e.g., what is not possible) does not have an immediately evident cat-
egorical representation. The formalization of nategories using unconventional composition addresses
this issue, and lets imposed limitations on categories be considered. However, traditional nategories
abandon core categorical constructs and rely on extensive mathematical development. This creates a
divide between the consideration of positive and negative information composition. In this work, we
show that negative information can be considered in a natural categorical manner. This is aided by
functor string diagrams, a novel flexible diagrammatic approach that can intuitively show the opera-
tion of hom-functors and natural transformations in expressions. This insight reveals how to consider
the composition of negative information with foundational categorical constructs without relying on
enrichment. We present diagrammatic means to consider not only nategories, but preorders more
broadly. This paper introduces diagrammatic methods for the consideration of triangle inequalities
and co-designs DP/FeasBool , showing how important cases of negative information composition can
be categorically and diagrammatically approached. In particular, we develop systematic tools to rig-
orously consider imposed limitations on systems, advancing our mathematical understanding, and
present intuitive diagrams which motivate widespread adoption and usage for various applications.

1 Introduction
Category theory is the formal study of composition. Morphisms express how components can be con-
nected, and thus represent various “paths” which can be taken. Morphisms show both that objects can be
linked and, given the multitude morphisms possible, associate these links with information.
However, practical applications often need to consider negative (complementary) information (e.g.,
paths that are not possible). In trying to design or plan a connection between certain states, it is important
to consider minimality, and impossibilities from which to work. This concept of negative information is
central to widely used techniques such as Dijkstra’s algorithm or A* search [6].
However, negative information has a peculiar composition structure. A limitation on a → b carries
through to paths b → c and a → c, which does not match typical categorical structure. This follows from
paths b → c and a → c being composable into a → b routes, hence, a ban on the former imposes itself
on the latter. This is not in accordance with standard category theory which has a → b compose with
morphisms to a or from b.
Nategories, recently introduced in [4], offer a means to consider these effects by introducing ad-
ditional mathematical structure to categories, abandoning the general rule that morphism + morphism
→ morphism and adding “norphisms”, representing bans, which follow morphism + norphism → nor-
phism, representing carry through effects. Certain norphisms can be considered purely categorical using
an enriched construct on the mathematically advanced PN (positive-negative) category.
Diagrammatic category theory maps categorical properties onto diagrams, associating algebraic rules
of categories with the manipulation of diagrams [14]. This is central to monoidal string diagrams, whose
categorical structure corresponds exactly to diagrams with graphical isomorphism rules. However, tra-
ditional monoidal string diagrams are limited to expressing simple connections between objects and
morphisms, and are not able to readily represent functors and natural transformations, which requires
further constructs [10, 11, 13].

© V. Abbott, G. Zardini
Submitted to:
This work is licensed under the
ACT 2024
Creative Commons Attribution License.
2 Diagrammatic Negative Information

String diagrams developed by Marsden and Nakahira [10, 11, 13] represent categories with colored
regions, associating functors to the boundaries of regions and natural transformations to the intersection
of boundaries. This graphically encodes the behavior of functors and natural transformations through
similar intuitive isotopy rules. However, they are designed with to focus on algebraic constructs such as
monoids, Kan extensions, and adjunctions, but lack the flexibility to be used for or to give insight for
applied category theory.
Functor string diagrams are a novel approach which streamline string diagrams using solid principles
which ensure that complex diagrams are decipherable [3, 2]. They are used as the mathematical foun-
dations of neural circuit diagrams for machine learning [1]. However, they can also be used to consider
abstract algebraic ideas in an intuitive manner and, as we will soon show, the manipulation of natural
transformations and hom-functors they encourage are the key to revealing how foundational category
theory can encompass norphisms and negative information.

2 Background
2.1 Nategories
The current means of considering negative information in a category is with nategories. Nategories are
categories embedded with additional elements, called norphisms, with atypical composition rules.
Definition 1 (Nategory). A locally small nategory C is a locally small category with the following
additional structure. For each pair of objects a, b ∈ ObC , in addition to the set of morphisms HomC (a; b),
we also specify:
• A set of norphisms NomC (a; b).
• An incompatibility relation, which we write as a binary function
iab : NomC (a; b) × HomC (a; b) → 2. (1)
For all triples a, b, c, in addition to the morphism composition function
#abc : HomC (a; b) × HomC (b; c) → HomC (a; c), (2)
we require the existence of two inexact norphism composition functions
abc :HomC (a; b) × NomC (a; c) → NomC (b; c),
(3)
abc NomC (a; c) × HomC (b; c) → NomC (a; b),
:
and we ask that they satisfy two “equivariance” conditions:
ibc ( f n, g) ⇒ iac (n, f # g), (equiv-1)
iab (n g, f ) ⇒ iac (n, f # g). (equiv-2)
As can be seen, this construction extends typical categories, abandoning the universal notion of
composition in order to consider negative information. However, using a novel diagrammatic approach,
we can show these two forms of composition to be the natural interactions between hom-functors and
functions C(a, b) → 2 in the category Set.

2.2 Functor String Diagrams

To understand these constructs better, we use functor string diagrams, a novel diagrammatic method tai-
lored to expressing the behavior of functors and natural transformations within a category [3, 2]. We
show these in Figure 1. Diagrams are partitioned into vertical sections, each associated to an object
or morphism. Functors are represented by functor wires passing over relevant objects and morphisms,
which intuitively modify the underlying objects or morphisms while encoding the preservation of com-
position.
V. Abbott, G. Zardini 3

Figure 1: Functor string diagrams show categorical ex-

pressions by composing vertical sections, each of which
represents an object or a morphism.

Hom-functors are represented by a wire with a leftward arrow labeled with the relevant object. Nat-
ural transformations between hom-functors C(c, _) → C(a, _) correspond to morphisms f : a → c by the
Yoneda lemma. As shown in Figure 2, natural transformations are drawn on the functor wire, encoding
their ability to pass over underlying morphisms while maintaining equivalence.

Figure 2: For more complex expres-

sions, we develop equivalent expres-
sions which lets categorical algebra
be understood with graphical intu-
ition.

Consider that norphisms are defined by their interaction with iab : NomC (a; b) × HomC (a; b) → 2.
Currying iab , we see that a norphism n ∈ NomC (a; b) corresponds to a function in : HomC (a; b) → 2.
This lets us consider the curried form of norphisms, in : HomC (a; b) → 2, as morphisms C(a, b) → 2 in
the category Set.
This reveals two manners that morphisms can interact with C(a, b) → 2, as morphisms C(c, b) with
a C(a, _) functor applied, or natural transformations mapping C(c, _) → C(a, _) over an underlying b
object wire.

Figure 3: We can re-

express norphism and
their exact composition
with morphisms using
functor string diagrams.

The shape of inexact nategorical compositions are the same as the above. Inexact nategory com-
position yields abc : HomC (a; b) × NomC (a; c) → NomC (b; c) and abc : NomC (a; c) × HomC (b; c) →
NomC (a; b). To uniquely express these, we require reference to the constituent morphism (morphism in
4 Diagrammatic Negative Information

C) and norphism (corresponding to a function HomC (a; b) → 2). We do this by copying the shape of the
above exact compositions, but draw the norphism composition symbols on the connecting wires, indicat-
ing that we are not perofrming a typical composition. The equivariance condition, then, is implemented
by stating that these inexact compositions imply the exact compositions.
Figure 4: We use dotted thick
wires to indicate morphisms
produced from inexact nor-
phism composition. Note
that inexact composition im-
plies exact composition.

So far, we have developed diagrammatic tools to represent nategories. However, we have relied on
implications instead of pure equivalence of expressions, meaning the application of categorical algebra
is unclear. In the next section, we will provide tools to diagrammatically view preorders as transforms
which are pure equivalences. This perspective allows preorders to be understood as equivalences, avoid-
ing the complexities of 2-category algebra.

3 Diagrammatic Preorders
Preorders have second-order categorical structure. This complicates the mathematics, and makes ex-
pressions and algebra confusing to understand. This complexity obfuscates the insights preorders offer
and hampers their broader application despite their utility for understanding design and trade-offs in
engineered systems among other fields (see Section 5).
Here, we introduce a diagrammatic means of considering preorders and 2-categories more generally
with functor string diagrams, letting them be simultaneously be considered along with the hom-functors
and natural transformations we will frequently use. This makes the insights they offer easier to perceive.
This diagrammatic framework intuitively captures norphisms and co-designs DP/FeasBool in a straight-
forward manner. This both aids mathematical rigor by motivating the use of theoretical tools indicated
by diagrams and encourages adoption for applied cases given the clear manner in which diagrams enable
ideas to be expressed.

Preorders. Preorders X⩽ are a generalization of ⩽ relationships. A set X with internal preorders

can be expressed as a category X⩽ where objects are elements and morphisms are ⩽ relationships.
This encodes for the fact that elements exhibit x ⩽xx x, the identity relationship, and that x ⩽xy y and
y ⩽yz z implies there is a relationship x ⩽xz z. These correspond to morphisms X⩽ (x, x) and composition
X⩽ (x, y) × X⩽ (y, z) → X⩽ (x, z).

Preorders as Transforms. To maintain consistent types, we introduce an initial object 1 such that each
object x of X⩽ ∪ 1 has a unique identifying morphism xx1 : 1 → x. Then, ⩽xy relationships let us compose
y1y = xx1 ; ⩽xy , mapping from the unique identifier for x to the unique identifier for y.
V. Abbott, G. Zardini 5

Figure 5: A set with preorder structure can

be expressed as a category where morphisms
correspond to ⩽ relationships. This allows
for the transformation of objects into superior
ones. With a unique identifier from an initial
object 1, we can fix types and state that y1y =
xx1 ; ⩽xy .

Preorders of Hom-Sets. In our case, we desire preorders within hom-sets, C(a, b)⩽ ∪ 1. We associate
each hom-set C(a, b) with a pre-ordered set C(a, b)⩽ ∪ 1 where morphisms f , g, . . . ∈ C(a, b) are objects
f , g, . . . ∈ C(a, b)⩽ ∪ 1 with transforms to superior elements and with an identifying morphism f f1 : 1 → f
for each member.
In a functor string diagram, we typically express g : a → b with the symbol “g”. However, we can also
identify it by the identifying morphism g1g : 1 → g in the category C(a, b)⩽ ∪ 1. Then, using g1g = f f1 ; ⩽gf ,
we can substitute this identifier g1g with f f1 and the noted transformation ⩽gf .

Figure 6: Using a series of re-expressions, we can intuitively show superior morphisms as equivalent to
a transform of another.

Two-Categories. Consistent pre-orders on hom-sets in a category carry through. If f ⩽ g for f , g ∈

f ;k
C(a, b), and k ⩽ h for k, h ∈ C(b, c), then we have ⩽g;h in C(b, c)⩽ . This requires C to have mono-
tonic non-decreasing morphisms. Between adjacent pre-ordered hom-sets, we establish a bifunctor
⇑abc : C(a, b)⩽ ∪ 1 × C(b, c)⩽ ∪ 1 → C(a, c)⩽ ∪ 1 where ⇑ (1, 1) = 1, which implements composition
of morphisms, and which is associative with respect to subsequent compositions.
Therefore, a consistent pre-order on hom-sets of a monotonic non-decreasing category implements
the monotonic non-decreasing property that f ⩽ g → h; f ⩽ h; g categorically, by virtue of ⩽hh ⇑⩽gf yield-
ing a transform ⩽h; f
h;g in C(a, c)⩽ ∪ 1. This grants a two-category, and we show this in Figure 7.
Just as in algebraic manipulation of ⩽, we are often more interested in the fact that a pre-order exists
between expressions rather than the exact details of that preorder. So far, we have developed tools to
view preorders as transforms, and have integrated these tools into the vertical sections of functor string
diagrams. Now, we can confidently state lossy inequality expressions using diagrammatic categories.
When we state f ⩽ g for morphisms a → b, we mean that there exists a transform ⩽gf between those
morphisms in the preordered category C(a, b).
6 Diagrammatic Negative Information

Figure 7: We have composition be-

tween two-morphisms from adjacent
hom-sets.

Figure 8: We use ⩽ to state that a transform exists. This lets clearly and readily manipulate various
algebraic expressions.
3.1 Norphisms as Preorders
We can now get to describing norphisms as purely categorical constructs by encoding the equivariance
condition as a preorder as in Figure 9. Furthermore, this preorder presentation encodes for the additional
rule that bans described by norphisms are expansive, banning a set of morphisms and all others which
are superior. This encodes for f being infeasible implying that g is as well, should f ⩽ g, limiting the
search space of un-banned morphisms. Therefore, we not only present norphisms in a more streamlined
traditional categorical manner, but also present them more formally.
First, recognize that the set 2⩽ has a pre-order structure 0 ⩽ 1, which is akin to implication. Thus, we
can simply state the equivariance of inexact composition as a pre-order transform. Furthermore, as this
is an expression in a pre-ordered category with monotonic non-decreasing morphisms, the expansiveness
property of pre-orders is naturally implemented.

Figure 9: Using functor string dia-

grams and preorders, we give categor-
ical meaning to the equivariance con-
dition as shown in Figure 4.

Therefore, we have captured all the structure of categorical negative information diagrammatically.
V. Abbott, G. Zardini 7

4 Applications of Diagrammatic Norphisms

In addition to streamlining the theory of negative information by employing intuitive diagrams, dia-
grammatic category theory enables theorems to be intuitively seen enhancing our understanding during
applications. Here, we will cover cases from the original nategories paper, including negative informa-
tion stemming from the triangle inequality and co-designs, and show how diagrams can provide greater
insight while offering a platform for clear communication which motivates wider adoption.

4.1 The Triangle Inequality

Triangle inequalities abound in many cases, as knowledge of a set of processes to be completed al-
most necessarily indicates that some holistic approach may exist which performs them collectively. This
yields relationships of the form L( f ; g) ⩽ L( f ) + L(g) in many cases, such as the processing of proba-
bilistic grammars [9], or the construction of neural networks, as two layers trained separately will yield a
worse loss than training them together [7, 8]. Therefore, these compositional constructs are of particular
interest.
Furthermore, we often have concepts of minimum distance. For example, in path-finding we have
a Euclidean “as the crow flies” distance which is necessarily shorter than all possible paths. These
underestimates or admissible heuristics are of particular interest to A* search or Dijkstra’s algorithm and
can yield a generalized means to solve problems in a variety of fields. This generality of composition
indicates that category theory would be an appropriate method of analysis, however, these applications
rely on understanding negative information and how the inability to form a composition propagates
through composition rules. Hence, typical categorical approaches are insufficient and, instead, nategories
are necessary. We need a method of considering how a minimum on one hom-set propagates given
knowledge of possible paths in others.
In this section, we show how diagrammatic category theory can clearly dissect the algebra of tri-
angle inequalities and subsequently yield a form of norphism composition. The brevity of this section
highlights the utility of the tools we have developed, they allow for the fundamental positive and nega-
tive compositional structure of systems to be clearly seen, a significant advantage over current methods
and a substantial inroad towards considering this class of applied compositional problems in a universal,
categorical manner.
We start by taking the triangle inequality — L( f ; g) ⩽ L( f ) + L(g) for some generally defined L :
C(a, b) → R — and considering it diagrammatically as in Figure 10. This allows us to rearrange it so
that we get an expression over all elements f as morphisms of Set(1, C(a, b)). As this expression holds
for all elements, it fully identifies the function in Set which operates on.
We can take the negative of both sides to arrive at something with a similar form to norphism inexact
compositions shown in Figure 9. In Figure 11, we see that we can define inexact composition with a
domain object R rather than 2. Nonetheless, this expression is clearly defined and is analogous to the
categorical structure we have established so far. We see here an advantage of the general approach we
have used, allowing negative information to be generalized.
Within a consistent pre-ordered category with monotonic non-decreasing morphisms, we can safely
compose additional morphisms onto the above expressions, preserving the pre-order structure. A lower
bound µ on the distance L between objects a and c can be implemented by a norphism (−L); (⩾ −µ) :
C(a, c) → 2. We can inherit the inexact composition from Figure 11 as shown in Figure 12, deriving a
norphism in a category with a triangle inequality.
With this algebra, we clearly see how the ⩽ expression related to the triangle inequality in Figure 10
becomes associated to a norphism by composition with monotonic non-decreasing morphisms in Set⩽
as in Figure 12. The tools we have used so far are all purely categorical and largely straightforward.
Nonetheless, we are able to derive a norphism and the propagation of negative information concerning
some ban on morphisms.
8 Diagrammatic Negative Information

Figure 10: The triangle in-

equality can be rearranged
categorically, yielding a ≤
expression. This, of course,
can be viewed as a trans-
form within the category
Set(C(a, b), R)⩽ ∪ 1.

Figure 11: We can define inexact composition between a morphism g : b → b and C(a, c) → R as we
have a ⩽ relationship.
Figure 12: We can mod-
ify the inexact composition
in Figure 11 to give inexact
composition with codomain
object 2 by using the mono-
tonic non-decreasing mor-
phism ≤ −µ. This gives in-
exact composition mapping
to 2, yielding norphism com-
position.

This shows the applied utility of the tools we have developed, and their potential to be further devel-
oped to be applied to a host of domains. Furthermore, we see that negative information is–indeed–purely
categorical, emphasizing the usefulness of category theory even for considering what may seem to be
atypical forms of composition.

5 Diagrammatic Co-designs
Complex systems co-design is an exciting avenue for applied category theory [5, 15]. Interestingly, they
have powerful diagrammatic semantics. They describe the relationship between resources and processes,
encoding how functionality demands relate to provided resources. This can be applied to solve a wide
range of design optimization problems, with applications in autonomy, mobility, and automotive [16, 18,
19, 17, 12].
V. Abbott, G. Zardini 9

Co-designs are a domain where considering negative information is critical. As they relate to the
available resources and demanded functionalities of a system, we often want to utilize the knowledge
that some process is known to be impossible. A basic example is the knowledge that physical mass
or energy cannot be created. These limitations are norphisms — a restriction on the morphisms P →
Q. These limitations have flow on affects to other objects as, otherwise, the bans they impose can be
circumvented. This negative information can be considered with traditional nategories by establishing
special composition rules.
However, we can show that co-designs accept negative information in a far more natural, and more
categorical, manner. Indeed, by employing the insights of diagrams, we find that co-designs accept inter-
nal norphisms, meaning that all bans are represented as morphisms within the category itself. This insight
allows us to use traditional codesign solving tools to investigate and consider negative information.
Co-designs DP is a relations category where objects P are Boolean vector spaces indexed by the set P
and morphisms d are relations between vector spaces, employing typical relation contraction as composi-
tion. DP is distinct, however, in that the Boolean matrices which correspond to relations have a particular
pre-ordered structure as shown in Figure 13. Vectors, objects corresponding to output codomains, repre-
sent functionality requirements and have a non-decreasing structure. This represents some functionality
requirement being satisfied by superior resources. Covectors, objects corresponding to input domains,
represent resource availabilities and have a non-increasing structure. The contraction of vectors and cov-
ectors, therefore, answers whether some functionality requirement (or more) is found in the available
pool of resources (or less).

Figure 13: Just as in monoidal

string diagrams [14], we do not
draw the unit object 1, corre-
sponding to the Boolean vector
space with a single value. Func-
tionality requirements are mor-
phisms f : 1 → P and resource
availabilities are morphisms r :
P → 1. Contraction indicates
whether the requirements are sat-
isfied by the available resources.
This is also satisfied by superior
functionality being satisfied by
inferior resources.

Morphisms d : P → Q are design problems which process resources. Therefore, they map a function-
ality requirement in P, f : 1 → P, to a functionality requirement in Q, f ; d : 1 → Q. Similarly, they map
a pool of available resources in Q, r : Q → 1, to a pool of available resources in P, d; r : P → 1. This is
achieved by standard relation composition. These relations must, upon composition, produce valid non-
decreasing vectors and valid non-increasing covectors. Therefore, their corresponding Boolean matrices
op ⩽
must have a structure d : P⩽ × Q⩽ − → 2 such that contraction along either axis yields the appropriate
pre-ordered structure, as shown in Figure 14.
Relations can be understood as linear operations where addition is replaced by ∨ and multiplication
by ∧. We can understand the truth value of output indexes as whether a corresponding interaction be-
tween non-negative linear operations would yield non-negative values at those locations. From this, it
follow that all the rules of manipulating linear operations are present. Furthermore, as both ∨ and ∧ are
monotonic non-decreasing, relations are always monotonic non-decreasing. Therefore, DP, as a subcat-
egory of relations, has all monotonic non-decreasing morphisms. As inputs (functionality demanded)
increases, so does the output (available resources required).
10 Diagrammatic Negative Information

Figure 14: Morphisms in DP are co-designs, relations

between Boolean vector spaces. They are ensured to pro-
vide valid functionality vectors and availability covectors
by having a monotonic non-decreasing structure across
the codomain axis and a monotonic non-increasing struc-
ture across the domain axis.

Norphisms in Co-design. The nature of norphisms in co-designs states impossibility results, e.g., that
some functionality demand cannot be achieved by some pool of available resources. For instance, we
can state that f : 1 → Q cannot be provided by ϕ : P → 1 (the objects 1 correspond to Boolean vector
spaces 21 ) with any choice of design problem d : P → Q.
Generally, for a locally small category C⩽ with a separating object 1 (meaning morphisms x → _
are uniquely identified by the composition with the set of morphisms 1 → x), we have a faithful hom-
functor into set C⩽ (1, _) : C⩽ → Set⩽ . The preorder on 1 → y can be used to generate a predicate
(⩾ ϕ) : C⩽ (1, y) → 2 for some ϕ : 1 → y. We can use this to test the statement “is ψ; m ⩾ ϕ?”. With
the algebra of hom-functors [2], we rearrange this expression to get a generic “performance norphism”
C⩽ (x, y) → 2 as in Figure 15.

Figure 15: For a locally small cat-

egory C⩽ with a separating object,
we can construct a performance nor-
phism which bans all morphisms x →
y which return ϕ or greater for some
input ψ : 1 → x.

Therefore, we see that norphisms of the type we desire which place some limit on performance can
be naturally constructed from hom-functors and natural transformations. Co-designs are relations and,
therefore, have linear structure as their objects are fundamentally vector spaces and their composition
is fundamentally matrix multiplication with all positive values, albeit ignoring the magnitude of those
values. Therefore, they are closed categories with access to internal hom-functors, transpositions, and
monoidal products.
V. Abbott, G. Zardini 11

However, we are required to fix types to ensure that the correct pre-order structure is present within
the functionality demand vector spaces and resource availability provision covector spaces. This is
achieved by having the monoidal unit ηP : 1 → Pop ⊗ P, which provides transposes, switch P to a dual
object Pop [14]. By transposing and swapping the order of P⩽ , we get well-defined transposed design
problems as shown in Figure 16.

Figure 16: Within linear categories, we have internal hom-functors provided by outer ∨-products. Trans-
posed design problems on product axes offer natural transformations.

Using (op)monoidal products for internal hom-functors and transposes for natural transformations,
we can construct the performance norphism within DP as show in Figure 17. The separating object in
DP is 1, the vector space consisting of one Boolean value. The ψ : 1 → x morphism is f : 1 → P while
the ϕ : 1 → y morphism against which we test is r : Q → 1. The performance norphism C⩽ (x, y) → 2
corresponds to a design problem Pop ⊗ Q → 1. Therefore, the performance norphism asks if m : P →
Q (transposed into 1 → Pop ⊗ Q) is able to achieve the functionality demand of f from the available
resources of r. Furthermore, as relations are monotonic non-decreasing, morphisms m : P → Q which
provide more functionality than f from r are also banned.
Figure 17: Relations offer all the tools
to include a performance norphism inter-
nally.

Finally, as in Figure 18, we can take a ∨-sum of different performance norphisms n = ∨i ( f [i]op ⊗ r[i])
to get a relation Pop ⊗ Q → 1 which returns true if, for any i, the provided relation m : P → Q satisfies
the functionality demand f [i] with the resources r[i]. This process lets us construct any design problem
Pop ⊗ Q → 1 as a combination of bans and to view these as norphisms.
The norphisms we have constructed exhibit propagation of negative information as this property
follows from the axioms of categories. As displayed in Figure 19, a norphism on P → Q is a design
problem n : Pop ⊗ Q → 1 and provides a hook for composition with design problems e : P → R or
g : R → Q to form norphisms on R → Q and P → R respectively. These represent the fact that a ban
12 Diagrammatic Negative Information

Figure 18: Any design problem Pop ⊗ Q → 1

can be viewed as a norphism banning design
problems P → Q which can provide function-
ality f [i] from available resources r[i] for any
constituent i.

on P → Q and the existence of a morphism such as e : P → R implies a ban on R → Q must be placed,

otherwise composition with e would allow the ban to be circumvented. Norphisms composition in the
case of co-designs, then, is exact.
Figure 19: Norphisms on design problems
P → Q are themselves design problems
Pop ⊗ Q → 1. They can be composed with
design problems to form other design prob-
lems of the form _op ⊗ _ → 1, which cor-
respond to other design problems, showing
how negative information has a natural cate-
gorical structure.

Norphism Schemas. This framework can also be used to implement physical resource constraints,
wherein no composition of design problems can exist to expand a physical resource pool. Hence, a
resource pool ϕ : P → 1 can never provide the functionality to recreate more than itself, ϕ + : 1 → P.
This can be directly achieved by the transposed negation converting the resource pool into a functionality
pool, which we see in Figure 20. As there is no overlap between a pool and its transposed negation, this
does not ban the identity. Instead, it bans all design problems which offer more than the identity between
a resource and itself. We take a sum over all physically restricted components of the resource pool to
implement a resource limitation norphism.
V. Abbott, G. Zardini 13

Figure 20: We obtain a function-

ality vectors which corresponds
to just more than what an avail-
ability covector provides by a
transposed negation. As a per-
formance norphism, this imple-
ments a physical resource being
unable to be increased by any de-
sign problem.

6 Conclusion
Our diagrammatic investigation of co-designs and negative information reveals how basic categorical
constructs explain a number of complex properties. Co-designs are shown to ultimately be relations with
a fundamental linear structure, thereby inheriting graphical manipulation by monoidal string diagrams
which provides intuitive insight into complex algebra. Furthermore, as we have previously shown that
norphisms arise from predicates over hom-sets, the closedness of the codesign category allows negative
information to be considered internally. This leads to further natural properties. In the context of co-
designs, then, this investigation reveals that category theory is just as capable of considering positive as
negative information.

References
[1] Vincent Abbott (2023): Neural Circuit Diagrams: Robust Diagrams for the Communication, Implementation,
and Analysis of Deep Learning Architectures. Accepted to Transactions on Machine Learning Research.
Available at https://openreview.net/forum?id=RyZB4qXEgt.
[2] Vincent Abbott (2023): Robust Diagrams for Deep Learning Architectures: Applications and Theory. Hon-
ours Thesis, The Australian National University, Canberra.
[3] Vincent Abbott & Gioele Zardini (2024): Functor String Diagrams: A Novel Approach to Flexible Dia-
grams for Applied Category Theory. Available at https://arxiv.org/abs/2404.00249. Submitted to
7th Annual International Applied Category Theory Conference (ACT).
[4] Andrea Censi, Emilio Frazzoli, Jonathan Lorand & Gioele Zardini (2022): Categorification of Negative
Information using Enrichment. In: 5th Annual International Applied Category Theory Conference (ACT),
Glasgow, UK. Available at https://arxiv.org/abs/2207.13589.
[5] Andrea Censi, Jonathan Lorand & Gioele Zardini (2024): Applied Category Theory for Engineering. Avail-
able at https://bit.ly/3qQNrdR. Work-in-progress book.
[6] Daniel Delling, Peter Sanders, Dominik Schultes & Dorothea Wagner (2009): Engineering Route Planning
Algorithms. In Jürgen Lerner, Dorothea Wagner & Katharina A. Zweig, editors: Algorithmics of Large
14 Diagrammatic Negative Information

and Complex Networks: Design, Analysis, and Simulation, Springer, Berlin, Heidelberg, pp. 117–139,
doi:10.1007/978-3-642-02094-0_7. Available at https://doi.org/10.1007/978-3-642-02094-0_7.
[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren & Jian Sun (2015): Deep Residual Learning for Image Recog-
nition. CoRR abs/1512.03385. Available at http://arxiv.org/abs/1512.03385. ArXiv: 1512.03385.
[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren & Jian Sun (2016): Identity Mappings in Deep Resid-
ual Networks, doi:10.48550/arXiv.1603.05027. Available at http://arxiv.org/abs/1603.05027.
ArXiv:1603.05027 [cs].
[9] Dan Klein & Christopher D. Manning (2003): A parsing: fast exact Viterbi parse selection. In: Proceedings
of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on
Human Language Technology - Volume 1, NAACL ’03, Association for Computational Linguistics, USA,
pp. 40–47, doi:10.3115/1073445.1073461. Available at https://dl.acm.org/doi/10.3115/1073445.
1073461.
[10] Dan Marsden (2014): Category Theory Using String Diagrams. CoRR abs/1401.7220. Available at http:
//arxiv.org/abs/1401.7220. ArXiv: 1401.7220.
[11] Kenji Nakahira (2023): Diagrammatic category theory, doi:10.48550/arXiv.2307.08891. Available at http:
//arxiv.org/abs/2307.08891. ArXiv:2307.08891 [math].
[12] Marc-Philippe Neumann, Gioele Zardini, Alberto Cerofolini & Christopher H. Onder (2024): On
the Co-Design of Components and Racing Strategies in Formula 1. Available at https://www.
research-collection.ethz.ch/handle/20.500.11850/655029.
[13] Robin Piedeleu & Fabio Zanasi (2023): An Introduction to String Diagrams for Computer Scientists,
doi:10.48550/arXiv.2305.08768. Available at http://arxiv.org/abs/2305.08768. ArXiv:2305.08768
[cs].
[14] P. Selinger (2011): A Survey of Graphical Languages for Monoidal Categories. In Bob Coecke, edi-
tor: New Structures for Physics, Lecture Notes in Physics, Springer, Berlin, Heidelberg, pp. 289–355,
doi:10.1007/978-3-642-12821-9-4. Available at https://doi.org/10.1007/978-3-642-12821-9-4.
[15] Gioele Zardini (2023): Co-Design of Complex Systems: From Autonomy to Future Mobility Systems. Ph.D.
thesis, ETH Zurich, Switzerland.
[16] Gioele Zardini, Andrea Censi & Emilio Frazzoli (2021): Co-design of autonomous systems: From hard-
ware selection to control synthesis. In: 2021 European Control Conference (ECC), Rotterdam, Nether-
lands, pp. 682–689, doi:10.23919/ECC54610.2021.9654960. Available at https://ieeexplore.ieee.
org/document/9654960.
[17] Gioele Zardini, Nicolas Lanzetti, Andrea Censi, Emilio Frazzoli & Marco Pavone (2023): Co-Design to
Enable User-Friendly Tools to Assess the Impact of Future Mobility Solutions. IEEE Transactions on Network
Science and Engineering 10(2), pp. 827–844, doi:10.1109/TNSE.2022.3223912. Available at https://
ieeexplore.ieee.org/abstract/document/9963724.
[18] Gioele Zardini, Dejan Milojevic, Andrea Censi & Emilio Frazzoli (2021): Co-design of embodied intel-
ligence: a structured approach. In: IEEE/RSJ International Conference on Intelligent Robots and Sys-
tems (IROS), Prague, Czech Republic, pp. 7536–7543, doi:10.1109/IROS51168.2021.9636513. Available at
https://ieeexplore.ieee.org/document/9636513.
[19] Gioele Zardini, Zelio Suter, Andrea Censi & Emilio Frazzoli (2022): Task-driven Modular Co-design of
Vehicle Control Systems. In: 2022 61th IEEE Conference on Decision and Control (CDC), IEEE. Available
at https://ieeexplore.ieee.org/abstract/document/9993107.