ZX-calculus: Graphical Quantum Reasoning

• ZX-calculus is a graphical language that represents quantum processes using generators like Z- and X-spiders, Hadamard nodes, and wires with precise semantics.
• The formalism replaces traditional matrix algebra with local rewrite rules and normalization techniques, simplifying proofs in quantum mechanics and computation.
• Extensions of ZX-calculus enable its application to qudit systems, mixed states, and full qubit quantum mechanics while ensuring soundness, universality, and completeness.

Searching arXiv for recent and foundational ZX-calculus papers to ground the article. The ZX-calculus is a graphical language for reasoning about quantum mechanics and quantum computation in which quantum processes are represented by diagrams built from generators such as green $Z$-spiders, red $X$-spiders, Hadamard nodes, and wires, with composition represented by connecting outputs to inputs and with equality established by local rewrite rules rather than by matrix algebra (Backens, 2014). In the qubit setting, its diagrams denote linear maps between tensor powers of a 2-dimensional Hilbert space, and in later extensions the same diagrammatic methodology has been generalized to qudits, mixed-dimensional finite systems, mixed-state processes, and specialized fragments for classical or probabilistic computation (Wang, 2021). A central theme in the development of the formalism is completeness: the question of whether every equality of linear maps expressible in the intended semantics can also be derived purely diagrammatically. Across stabilizer, Clifford+T, full qubit, qudit, and finite-dimensional settings, the ZX-calculus has developed from a fragmentary but powerful heuristic language into a family of sound, universal, and, in several major senses, complete graphical calculi (Jeandel et al., 2019).

1. Definition and semantic interpretation

In its standard qubit form, a ZX-diagram is generated from green spiders $R_Z^{(n,m)}(\alpha)$, red spiders $R_X^{(n,m)}(\alpha)$, the Hadamard gate $H:1\to 1$, the empty diagram, identity wires, cups, and caps, with sequential and parallel composition corresponding respectively to ordinary composition and tensor product of linear maps (Jeandel et al., 2019). The diagrams are interpreted as matrices over $\mathbb{C}$, and a diagram with $n$ inputs and $m$ outputs denotes a linear map $\mathbb{C}^{2^n}\to\mathbb{C}^{2^m}$ (Backens et al., 2016). The formalism is therefore not merely pictorial: it is an equational language with a standard semantics in finite-dimensional Hilbert spaces.

The basic generators encode complementary observable structures. A green spider with phase $\phi$ acts like

$X$0

while a red spider with phase $X$1 acts like

$X$2

with $X$3 (Backens, 2014). In equivalent notation, a one-input one-output green spider is the matrix $X$4, and a one-input one-output red spider is $X$5 (Wetering, 2020). The Hadamard node mediates between these two bases and supports the color-change principle $X$6 in diagrammatic form (Wetering, 2020).

The calculus is usually presented in a compact closed or dagger-compact graphical setting, with cups and caps satisfying yanking equations and with the principle that only connectivity matters: planar deformation does not change semantic content so long as connectivity is preserved (Jeandel et al., 2019). In the stabilizer setting, this topological principle can even be reconstructed from a weaker braided autonomous categorical setting rather than assumed outright, which isolates more precisely the structural content required for ZX reasoning (Backens et al., 2017).

Algebraically, the two spider families each form a commutative special dagger Frobenius algebra, and their interaction realizes bialgebraic and Hopf-like structure (Wetering, 2020). This is the origin of the characteristic rewrite rules—spider fusion, copying, bialgebra, Hopf, Hadamard color change, and phase-commutation—which make large classes of matrix equalities accessible by local graphical transformation rather than by direct symbolic calculation.

2. Core rewrite theory and structural principles

The rewrite system of the ZX-calculus consists of local equations that preserve the standard interpretation. Among the most basic are spider fusion, under which adjacent same-color spiders fuse and their phases add; identity removal, under which a phase-free two-legged spider is just a wire; Hadamard self-inverse and Hadamard color change; $X$7-copy rules; and the bialgebra and Hopf laws governing red-green interaction (Wetering, 2020). In the single-qubit Clifford+T setting, a particularly small rule set suffices for the normalization argument: $X$8, $X$9, $R_Z^{(n,m)}(\alpha)$0, $R_Z^{(n,m)}(\alpha)$1, and $R_Z^{(n,m)}(\alpha)$2, all sound up to global scalar factors (Backens, 2014).

Soundness means that every derivable equality preserves denotation: $R_Z^{(n,m)}(\alpha)$3 Completeness is the converse implication: $R_Z^{(n,m)}(\alpha)$4 This distinction is central throughout the literature (Backens et al., 2016). Universality, by contrast, asserts that every linear map in the intended semantic domain is representable by some diagram (Wang, 2022).

A significant line of work simplified the stabilizer axiomatics. A reduced rule set consisting of $R_Z^{(n,m)}(\alpha)$5, $R_Z^{(n,m)}(\alpha)$6, $R_Z^{(n,m)}(\alpha)$7, $R_Z^{(n,m)}(\alpha)$8, $R_Z^{(n,m)}(\alpha)$9, $R_X^{(n,m)}(\alpha)$0, $R_X^{(n,m)}(\alpha)$1, and $R_X^{(n,m)}(\alpha)$2, together with the topology meta rule, is sound and complete for the stabilizer ZX-calculus (Backens et al., 2016). In this streamlined presentation, color symmetry and upside-down symmetry are no longer primitive axioms; they become derivable consequences of Hadamard conjugation and topological structure. The same work also shows that an additional scalar symbol and one of the previously used scalar rules are unnecessary (Backens et al., 2016).

Minimality questions were later studied explicitly. In a simplified stabilizer rule set with 9 explicit rewrite rules plus the meta-rule “only connectivity matters,” most remaining rules were shown necessary: $R_X^{(n,m)}(\alpha)$3, $R_X^{(n,m)}(\alpha)$4, $R_X^{(n,m)}(\alpha)$5, $R_X^{(n,m)}(\alpha)$6, $R_X^{(n,m)}(\alpha)$7, $R_X^{(n,m)}(\alpha)$8, and $R_X^{(n,m)}(\alpha)$9 cannot be derived from the others, while the status of $H:1\to 1$0 and $H:1\to 1$1 remained open, with at least one of them necessary (Backens et al., 2017). The unresolved status of the bialgebra rule is notable because that rule axiomatizes complementarity, one of the conceptual cornerstones of the calculus (Backens et al., 2017). This suggests that the foundational algebraic content of ZX has both a highly compressed form and a nontrivial residual ambiguity about what is genuinely primitive.

3. Completeness results from stabilizer to full qubit quantum mechanics

The stabilizer fragment was the first major domain in which completeness was established. In that fragment, all phases are multiples of $H:1\to 1$2, and the calculus captures stabilizer quantum mechanics, including Clifford operations and computational-basis preparation and measurement (Backens, 2014). The stabilizer ZX-calculus is sound and complete, and this completeness survives substantial axiom simplification (Backens et al., 2016).

A decisive intermediate result is that the ZX-calculus is complete for the single-qubit Clifford+T group (Backens, 2014). Here $H:1\to 1$3, $H:1\to 1$4, and $H:1\to 1$5 generates the single-qubit Clifford+T group, which is approximately universal for single-qubit unitaries (Backens, 2014). The proof proceeds through a unique normal form inspired by Matsumoto and Amano. Every single-qubit operator consisting of phase shifts that are multiples of $H:1\to 1$6 and Hadamards is either a Clifford operator or has the form

$H:1\to 1$7

with $H:1\to 1$8, $H:1\to 1$9, and $\mathbb{C}$0, and this normal form is unique (Backens, 2014). A non-identity theorem then shows that no nontrivial normal-form diagram of that shape equals the identity, yielding completeness by normalization.

The next major step was completeness for Clifford+T more broadly, and then for the full qubit calculus. One route used back-and-forth translation with the ZW-calculus. For Clifford+T diagrams, a complete axiomatisation $\mathbb{C}$1 was proved using translations between ZX and $\mathbb{C}$2, which is complete for dyadic rationals (Jeandel et al., 2019). The same work extended completeness beyond the fixed Clifford+T fragment to diagrams linear in variables with Clifford+T constants, and then obtained completeness for arbitrary ZX diagrams by adding a single non-linear axiom $\mathbb{C}$3 and translating through $\mathbb{C}$4, complete for arbitrary complex matrices (Jeandel et al., 2019).

A related thesis-level synthesis gives a complete axiomatisation for overall pure qubit quantum mechanics via translation from the ZW-calculus and introduces a triangle generator $\mathbb{C}$5 and a $\mathbb{C}$6-box $\mathbb{C}$7 with standard interpretations

$\mathbb{C}$8

which simplify the completeness proof (Wang, 2022). The same source also states that by restricting the coefficient ring to

$\mathbb{C}$9

one directly obtains a complete axiomatisation for Clifford+T quantum mechanics, and it further reports completeness for 2-qubit Clifford+T circuits using just 9 rules via verification of 17 circuit relations (Wang, 2022).

A separate algebraic reformulation replaced trigonometric side conditions in universally complete axiomatisations by ring-theoretic operations on complex numbers (Wang, 2019). This is important because earlier universal systems included rules where output phases were computed using $n$0, $n$1, $n$2, or $n$3, which complicated constructive use. The algebraic axiomatisation instead enlarges the generating set with general green spiders, triangles, and related gadgets so that the complete theory can be formulated using only algebraic operations over $n$4 (Wang, 2019).

4. Normal forms, graph-like diagrams, and algorithmic rewriting

Normal forms are a recurring mechanism behind completeness and practical simplification. In the single-qubit Clifford+T setting, the unique form $n$5 is the critical structural theorem (Backens, 2014). In finite-dimensional qudit and mixed-dimensional settings, constructive normal forms are also central. For qudit vectors in dimension $n$6, a radix-$n$7 basis labeling

$n$8

supports a unique diagrammatic normal form built from row-addition and row-multiplication gadgets, and from there map-state duality yields a normal form for arbitrary matrices (Wang, 2021). In the qufinite setting, this produces a unique normal form for every finite matrix and hence universality for finite quantum theory (Wang, 2021).

Graph-like forms play a parallel role in practical ZX rewriting. In graph-like diagrams, all spiders are taken to be of one color, all inter-spider connections are Hadamard edges, parallel Hadamard edges and self-loops are absent, and each boundary wire touches a single spider (Ewen et al., 2024). Every ZX-diagram can be converted into such a form using graph-theoretic simplification methods, after which local complementation and pivoting become canonical rewrite primitives (Ewen et al., 2024). This graph-like representation is also the basis of efficient stabilizer simplification and circuit extraction.

For 2-qubit Clifford+T circuits, ZX reasoning can be much more concise than a direct circuit-equation approach. A small extension of the stabilizer ZX rules, together with rule $n$9 and one further rule $m$0, suffices to derive all equations between 2-qubit Clifford+T circuits (Coecke et al., 2018). The paper explicitly notes that the ZX rules are “much simpler than the complete set of Clifford+T circuit equations due to Selinger and Bian” because intermediate ZX steps are not constrained by unitarity or by remaining within the gate set at every stage (Coecke et al., 2018). This suggests a general methodological advantage of ZX over strictly circuit-level calculi: semantic flexibility during intermediate rewrites can simplify both human and automated derivations.

The same structural perspective appears in scalable and algorithmic settings. A scalable ZX formalism introduces thick wires $m$1, dividers, gatherers, and scaled generators $m$2, with a normal form

$m$3

which preserves the rewriting discipline of the underlying calculus while providing a compact high-level syntax for algorithm proofs (Carette et al., 2021). This framework has been used to give fully graphical proofs of Deutsch–Jozsa, Bernstein–Vazirani, Simon, and Grover algorithms (Carette et al., 2021).

5. Extensions beyond qubits and beyond pure-state quantum mechanics

The qubit ZX-calculus has been generalized in several orthogonal directions. One major development is the qudit and mixed-dimensional extension. The qufinite ZX-calculus introduces qudit ZX-calculi for arbitrary finite dimension $m$4, with generalized green and red spiders, unnormalized Fourier-like Hadamards, and a carefully chosen triangle node $m$5 and its inverse (Wang, 2021). Each wire may carry its own dimension label, and new structural generators—the dimension-splitter and dimension-binder—allow a wire of dimension $m$6 to be identified with a pair of wires of dimensions $m$7 and $m$8 (Wang, 2021). The resulting diagram category is no longer a PROP in the strict single-sort sense, but remains compact closed (Wang, 2021).

A later completeness result for finite-dimensional Hilbert spaces strengthens this perspective. The finite-dimensional ZX-calculus $m$9, using mixed-dimensional Z-spiders and qudit X-spiders as generators, is universally complete for finite-dimensional Hilbert spaces (Poór et al., 2024). The mixed-dimensional Z-spider is parameterized by coefficients $\mathbb{C}^{2^n}\to\mathbb{C}^{2^m}$0 and uses the minimum dimension among all legs,

$\mathbb{C}^{2^n}\to\mathbb{C}^{2^m}$1

with semantics

$\mathbb{C}^{2^n}\to\mathbb{C}^{2^m}$2

while the X-spider is defined by modular addition over the local dimension (Poór et al., 2024). Completeness is proved by direct translations to and from the finite-dimensional ZW-calculus, exactly paralleling the qubit strategy but now in a genuinely mixed-dimensional setting (Poór et al., 2024).

The qutrit case has its own specific complications. Qutrit spiders carry two phases, the Hadamard is not self-adjoint, and graph-like normalization is more constrained because plain cups and caps are absent and same-color spiders connected by mixed edge types do not always fuse (Townsend-Teague et al., 2021). Even so, efficient simplification strategies for the stabilizer fragment have been derived using qutrit analogues of local complementation and pivoting. The stabilizer spiders are classified into three phase families $\mathbb{C}^{2^n}\to\mathbb{C}^{2^m}$3, $\mathbb{C}^{2^n}\to\mathbb{C}^{2^m}$4, and $\mathbb{C}^{2^n}\to\mathbb{C}^{2^m}$5, with $\mathbb{C}^{2^n}\to\mathbb{C}^{2^m}$6- and $\mathbb{C}^{2^n}\to\mathbb{C}^{2^m}$7-spiders eliminated by local complementation and adjacent $\mathbb{C}^{2^n}\to\mathbb{C}^{2^m}$8-pairs by pivoting (Townsend-Teague et al., 2021). This constitutes a first non-trivial step toward systematic qutrit circuit simplification.

ZX has also been extended to mixed states and classical or probabilistic processes. The discard ZX-calculus supports completely positive maps, and a decohered fragment isolates the part that survives after applying the decoherence projector

$\mathbb{C}^{2^n}\to\mathbb{C}^{2^m}$9

This decohered ZX-calculus is universal and complete for positive real matrices with affine support, with a normal form combining affine-support structure and Fourier-like positive coefficients (Carette et al., 6 Aug 2025). In that setting, a decohered red spider can represent a Bernoulli distribution $\phi$0 or a noisy bit-flip channel $\phi$1, and the support restriction precisely excludes non-affine Boolean operations such as AND (Carette et al., 6 Aug 2025).

A different classical direction is the fragment generated by Z and X spiders together with not and and gates, sometimes denoted ZX&, which yields a complete graphical calculus for classical circuits and “qubit multirelations” via a two-way translation with an extension of the Toffoli-generated category TOF (Comfort, 2020). This is not a phase-rich quantum fragment but a classical completion in which Z-spiders correspond to copying and X-spiders to addition (Comfort, 2020).

6. Applications, variants, and neighboring calculi

The ZX-calculus is widely used as an intermediate representation for circuit simplification, verification, compiler transformations, measurement-based computation, and quantum error correction (Wetering, 2020). In practical optimization, graph-like diagrams and phase gadgets are especially important. A phase gadget represents a diagonal action of the form

$\phi$2

for semi-Boolean phase polynomials, and gadget fusion is a standard mechanism for T-count reduction (Wetering, 2020). This is one reason ZX has become closely tied to automated rewriting tools such as Quantomatic (Jeandel et al., 2019).

The formalism has also been extended analytically. An analytic extension realizes differentiation and definite integration entirely within ZX-calculus, using the triangle and especially the W spider to encode the sum structure required by the Leibniz rule (Wang et al., 2022). In that framework, any differentiable ZX diagram can be differentiated as a ZX diagram and rewritten as a single diagram rather than as a sum of diagrams, and the same machinery yields a diagrammatic derivation of the parameter-shift rule and a ZX-based analysis of barren plateaus in variational quantum circuits (Wang et al., 2022). For a concrete ansatz with Hamiltonian $\phi$3, the variance bound

$\phi$4

was obtained diagrammatically (Wang et al., 2022).

Recent work has also pushed ZX into code-theoretic and dynamical settings. One approach represents stabilizer and subsystem codes by ZX-encoding graphs and models code switching by local Cliffords, qubit permutations, and gauge fixing, so that closed loops in code space correspond to logical Clifford gates (Frei et al., 1 Jun 2026). Another extends odd-prime-dimensional stabilizer ZX with a delay generator $\phi$5, yielding a delayed stabilizer ZX-calculus for translation-invariant and infinite stabilizer processes such as convolutional codes and lattice models (Comfort et al., 4 Jul 2026). In that delayed setting, every diagram admits a unique reduced AP-form, and the calculus is sound, universal, and complete for a generating-tableau semantics over $\phi$6 (Comfort et al., 4 Jul 2026).

There are also neighboring graphical calculi motivated by limitations of standard spiders. The ZQ-calculus keeps the phase-free Z-spider for entanglement while moving arbitrary single-qubit rotations to quaternion-labeled nodes, because unit quaternions cannot be absorbed into a generalized spider law: “Every monoid over $\phi$7 is commutative,” so no spider whose phases are labeled by unit quaternions can realize the non-commutative group structure required (Miller-Bakewell, 2020). This clarifies a common misconception: ZX is universal and complete in several qubit senses, but that does not imply every alternative single-qubit parametrization fits naturally into the spider formalism.

A more distant generalization reinterprets ZX in terms of Hopf–Frobenius algebra on $\phi$8 for a compact gauge group $\phi$9, replacing ordinary wires by ribbons and obtaining a ribbon ZX calculus for two-dimensional Yang–Mills theory (Wong et al., 11 Jun 2026). There the X spider encodes group multiplication, the Z spider pointwise multiplication, and the antipode becomes a ribbon twist (Wong et al., 11 Jun 2026). This suggests that the algebraic core of ZX extends well beyond quantum circuits in the narrow sense.

7. Conceptual significance and current outlook

The central significance of the ZX-calculus is that it replaces matrix reasoning by a local, compositional, topology-sensitive equational theory while retaining rigorous semantics and, in major settings, completeness (Jeandel et al., 2019). For stabilizer quantum mechanics, this yields a compact exact proof language; for Clifford+T and full qubit quantum mechanics, it yields a complete diagrammatic account of approximately universal and universal qubit computation; and for finite-dimensional systems more generally, recent results show that a mixed-dimensional ZX formalism is complete for all finite-dimensional Hilbert spaces (Poór et al., 2024).

Several recurring ideas organize the subject. First, completeness is often achieved either by normal-form theorems internal to ZX, as in the single-qubit Clifford+T proof, or by translations to and from another complete calculus such as ZW (Backens, 2014). Second, graph-like forms, local complementation, pivoting, and their higher-dimensional analogues provide an operational bridge between categorical structure and practical algorithmic rewriting (Townsend-Teague et al., 2021). Third, extensions typically preserve the core spider-based intuition while modifying the generator set just enough to capture additional algebraic structure: triangles and $X$00-boxes for simpler complete axiomatisations, binders and splitters for mixed dimensions, W spiders for analytic operations, delay for translation invariance, and discard/decoherence for classical–quantum interaction (Wang, 2022).

Common misconceptions are therefore best addressed by separating levels of generality. The ZX-calculus is not a single fixed rule set complete for every possible semantic domain; rather, different fragments and extensions have distinct completeness theorems, normal forms, and generator sets (Jeandel et al., 2019). Nor is it merely a heuristic diagram notation: in the stabilizer, Clifford+T, full qubit, and finite-dimensional settings cited above, semantic equality implies derivability within the appropriate calculus (Backens et al., 2016). At the same time, the existence of specialized variants such as algebraic ZX, qufinite ZX, decohered ZX, delayed ZX, and neighboring systems such as ZQ indicates that no single presentation exhausts all useful diagrammatic structures.

A plausible implication is that ZX-calculus is best understood not as one immutable syntax but as a family of tightly related graphical proof systems organized around interacting Frobenius and Hopf-like structures, compact closure, and compositional semantics. On that reading, the history of ZX is not only the refinement of one calculus, but the emergence of a general methodology for turning algebraic and operational structure into sound and, increasingly often, complete graphical reasoning.