Cartan–Khaneja–Glaser Decomposition

• The Cartan–Khaneja–Glaser decomposition is a Lie-theoretic factorization that expresses any unitary as a product of compact factors and an Abelian exponential, underpinning quantum circuit synthesis.
• It employs a recursive construction in SU(2^n) to systematically decompose n-qubit unitaries into explicit gate sequences using CNOTs and one-qubit rotations.
• The method optimizes resource counts, unifies various synthesis techniques, and enables fixed-depth implementations for both two-qubit gates and large-scale Hamiltonian simulations.

The Cartan–Khaneja–Glaser decomposition, often written as a KAK or KHK decomposition, is a Lie-theoretic factorization of unitary operators that is specialized in quantum information to $U \in SU(2^n)$ and related dynamical Lie groups. Its core statement is that, after choosing a Cartan involution and the associated symmetric-space splitting of the Lie algebra into compact and noncompact parts, every unitary can be written as a product of two factors from the subgroup $K$ and one Abelian exponential factor $A$, typically $U = K_1 \exp(a) K_2$ with $a$ in a maximal Abelian subalgebra of the $-1$-eigenspace. In quantum circuit synthesis this factorization becomes constructive and recursive, yielding exact decompositions of arbitrary $n$-qubit unitaries; in Hamiltonian simulation it yields fixed-depth realizations in which the time dependence is confined to the central Abelian term (Rodríguez et al., 5 Sep 2025, Alsheikh et al., 5 Dec 2025).

1. Lie-theoretic formulation

For a connected compact semisimple Lie group $G$ with Lie algebra $\mathfrak g$, a Cartan involution $\theta : \mathfrak g \to \mathfrak g$ induces a decomposition

$K$0

Choosing a maximal Abelian subspace $K$1, the KAK theorem states that every $K$2 may be written as

$K$3

For simply connected $K$4, the factors are unique up to the action of the Weyl group on $K$5 (Ding et al., 11 May 2026).

In the quantum-control variant, the decomposition is applied not only to $K$6 as a full group but also to the dynamical Lie algebra of a Hamiltonian. Given

$K$7

one chooses $K$8 with $K$9, and then a Cartan subalgebra $A$0. The resulting KHK factorization gives, for time evolution,

$A$1

so that the circuit depth is independent of $A$2 because the time dependence appears only in the Abelian middle factor (Alsheikh et al., 5 Dec 2025).

Two notational conventions coexist in the literature: $A$3 in the symmetric-space formulation and $A$4 in the Khaneja–Glaser recursion. The underlying structure is the same: an involutive splitting of the Lie algebra, a maximal Abelian subalgebra in the $A$5-eigenspace, and a group-level $A$6 factorization.

2. Specialization to $A$7 and recursive Cartan subalgebras

A concrete Khaneja–Glaser specialization for

$A$8

chooses

$A$9

and

$U = K_1 \exp(a) K_2$0

with

$U = K_1 \exp(a) K_2$1

At the group level, with $U = K_1 \exp(a) K_2$2 and $U = K_1 \exp(a) K_2$3, every $U = K_1 \exp(a) K_2$4 admits

$U = K_1 \exp(a) K_2$5

A second Cartan splitting of $U = K_1 \exp(a) K_2$6 refines this to the four-group form

$U = K_1 \exp(a) K_2$7

(Mansky et al., 2022).

The Cartan subalgebras $U = K_1 \exp(a) K_2$8 and $U = K_1 \exp(a) K_2$9 are built recursively. The construction begins with

$a$0

then defines

$a$1

followed by

$a$2

and finally

$a$3

This recursion determines which multiqubit generators populate the central Abelian exponentials and which appear in the auxiliary $a$4-layers (Mansky et al., 2022).

A complementary formulation describes the same recursive structure through alternating involutions of types AIII, A, and finally AI. In that framework, the first step uses an AIII involution by conjugation with $a$5 on one qubit, the second uses a block-swap involution on the resulting block-diagonal algebra, and the final two-qubit reduction uses the AI involution in the magic basis. Wierichs et al. show that the Quantum Shannon, Block-ZXZ, and Khaneja–Glaser decompositions implement the same recursive Cartan decomposition (Wierichs et al., 24 Mar 2025).

3. Constructive unitary synthesis

The recursive circuit-synthesis algorithm follows the algebraic decomposition directly. At a high level, one first finds a first-level split

$a$6

then decomposes the $a$7 factors by a second Cartan splitting to obtain the four-group form, and finally recurses on each block in $a$8. In practice the recursion is terminated at $a$9, where one uses the known optimal three-CNOT factorizations of arbitrary two-qubit gates due to Vatan–Williams and Shende–Markov–Bullock (Mansky et al., 2022).

The gate-level realization is explicit. Every Cartan generator in $-1$0 is a sum of two proportional monomials,

$-1$1

and the analogous statement for $-1$2 uses $-1$3. These are grouped into two-parameter blocks

$-1$4

with

$-1$5

Each $-1$6 is implemented, up to global phase, by two CNOTs framing two one-qubit rotations on the target wire. Appending a tensor $-1$7 to the generator corresponds to stamping two extra CNOTs controlled on the $-1$8th qubit. Appending $-1$9 corresponds to two fermionic SWAPs between the $n$0th wire and the appropriate subalgebra wire; each fSWAP can be decomposed into at most four CNOTs or absorbed into relabeling. A lone diagonal generator $n$1 is handled by a “Diag” gadget with four CNOTs and one rotation (Mansky et al., 2022).

This explicit realization is significant because the Cartan factorization is not merely an existence theorem. The algebraic generators themselves are mapped to circuits built from CNOT, SWAP, and one-qubit rotations, and the recursion reaches the level of elementary gates without leaving the Cartan framework. Mansky et al. further emphasize that the construction is independent of the standard CNOT implementation and can be adapted to other cross-qubit circuit elements by changing the block gadget rather than re-deriving the decomposition (Mansky et al., 2022).

4. Complexity, parameter counts, and comparison with other decompositions

For the explicit Khaneja–Glaser construction of $n$2, the CNOT counts for the Abelian factors are

$n$3

Both satisfy the same closed form,

$n$4

and one full Cartan step uses one $n$5 plus two $n$6 factors, so

$n$7

After adding the recursively decomposed size-$n$8 blocks,

$n$9

with $G$0, giving

$G$1

Thus the asymptotic CNOT cost is $G$2 (Mansky et al., 2022).

In the same comparison, earlier synthesis schemes are listed as follows: Barenco et al. give $G$3; Knill gives $G$4; the Gray-code approach gives $G$5; the cosine-sine decomposition of Möttönen et al. gives $G$6; the optimized quantum Shannon decomposition of Shende et al. gives $G$7; and the theoretical lower bound is $G$8 (Mansky et al., 2022).

A distinct but related notion of optimality appears in the recursive-Cartan framework of Wierichs et al. There, the alternating AIII$G$9A recursion yields a parameter-optimal decomposition: at level $\mathfrak g$0, the chosen Cartan subalgebra has dimension $\mathfrak g$1, and summing over levels gives $\mathfrak g$2 Cartan angles, with the final AI step supplying the remaining three two-qubit parameters. Their analysis unifies several synthesis methods by identifying a common recursive CD rather than by asserting identical gate-count constants (Wierichs et al., 24 Mar 2025).

5. The two-qubit case, Cartan coordinates, and equivalence classes

For $\mathfrak g$3, the Cartan–Khaneja–Glaser decomposition becomes the standard local–nonlocal–local factorization of two-qubit gates. One sets

$\mathfrak g$4

$\mathfrak g$5

and chooses the maximal Abelian subspace

$\mathfrak g$6

Then every $\mathfrak g$7 admits

$\mathfrak g$8

(Ding et al., 11 May 2026).

A constructive procedure forms the symmetric matrix

$\mathfrak g$9

diagonalizes $\theta : \mathfrak g \to \mathfrak g$0 by a real orthonormal basis $\theta : \mathfrak g \to \mathfrak g$1,

$\theta : \mathfrak g \to \mathfrak g$2

with $\theta : \mathfrak g \to \mathfrak g$3 and $\theta : \mathfrak g \to \mathfrak g$4, and then solves

$\theta : \mathfrak g \to \mathfrak g$5

This yields the Cartan angles up to Weyl permutations (Ding et al., 11 May 2026).

A central clarification in recent work concerns equivalence classes. Two distinct notions are separated: double-coset equivalence,

$\theta : \mathfrak g \to \mathfrak g$6

and projective-local equivalence,

$\theta : \mathfrak g \to \mathfrak g$7

For double-coset equivalence, the fundamental domain is the tetrahedral cell

$\theta : \mathfrak g \to \mathfrak g$8

whereas the usual “Weyl chamber” used in quantum-information practice is recovered only for projective-local equivalence: $\theta : \mathfrak g \to \mathfrak g$9 This resolves a long-standing inconsistency in the literature on two-qubit local equivalence (Ding et al., 11 May 2026).

Common gates are placed in these coordinates as

$K$00

$K$01

(Ding et al., 11 May 2026).

6. Alternative algebraic constructions and recent extensions

One algebraic generalization is the Quotient Algebra Partition framework. In this approach, $K$02 is partitioned into Abelian subspaces arranged in conjugate pairs, with commutator closure governed by a binary quotient-algebra rule. For $K$03, Su et al. construct a quotient algebra of rank zero consisting of a Cartan subalgebra plus $K$04 conjugate pairs, and show that, by selecting one member of each pair, one obtains Cartan decompositions of type AI. In the fourth paper of the series they further state that every Cartan decomposition is obtainable from the quotient algebra partition of the highest rank, and that the universality of the quotient algebra partition extends to classical and exceptional Lie algebras (Su, 2019, Su et al., 2019).

Another line of work reformulates the decomposition using involutive automorphisms. Mora Rodríguez et al. define

$K$05

at the group level and the corresponding $K$06 on $K$07, yielding

$K$08

Their algorithm then computes

$K$09

rotates $K$10 into a maximal Abelian subspace by optimization over $K$11, applies a second involution $K$12 on the residual compact factor, and recurses. The stated aim is to overcome reliance on ill-defined matrix logarithms and convergence issues of truncated Baker–Campbell–Hausdorff series. Their implementation is benchmarked on random unitaries in $K$13 and $K$14, with the reported averages

$K$15

$K$16

with $K$17 for $K$18 and $K$19 for $K$20 (Rodríguez et al., 5 Sep 2025).

The Hamiltonian-synthesis extension RedCarD applies a reductive refinement of the KHK idea to a dynamical Lie algebra. After generating $K$21, finding $K$22, and choosing a Cartan subalgebra $K$23, it fragments $K$24 by commutation with the $K$25 into subspaces

$K$26

and then performs nested optimizations over each $K$27. The stepwise cost function is

$K$28

with sequential update

$K$29

The cost evaluation can be shifted to hardware by rewriting

$K$30

and then minimizing each coordinate by Rotosolve, using three evaluations in $K$31 because the dependence is a single-mode sinusoid of period $K$32. On the 4-site transverse-field Ising model, the paper reports a $K$33–$K$34 reduction in classical runtime compared to the standard KHK approach for up to 20 spins, and experimental demonstrations on several IBM devices and Quantinuum’s H1-1 quantum computer (Alsheikh et al., 5 Dec 2025).

Within the broader recursive-Cartan program, Wierichs et al. also report an application to fast-forwardable Hamiltonian time evolution: the transverse-field XY model on $K$35 qubits is compiled into $K$36 gates in 22 seconds on a laptop (Wierichs et al., 24 Mar 2025). A plausible implication is that the Cartan–Khaneja–Glaser framework has evolved from a decomposition theorem for low-dimensional gate synthesis into a family of compiler architectures spanning exact unitary factorization, fixed-depth analogues of simulation circuits, and numerically stable recursive implementations.