---
title: Generalized Zassenhaus Formulas
url: https://www.emergentmind.com/topics/generalized-zassenhaus-formulas
type: topic
---

# Generalized Zassenhaus Formulas

Generalized Zassenhaus formulas are extensions, reformulations, and specialized variants of the classical factorization
\[
e^{X+Y}=e^X e^Y \prod_{n=2}^{\infty}e^{C_n(X,Y)},
\]
in which each \(C_n(X,Y)\) is a homogeneous Lie polynomial of degree \(n\). In the literature, “generalized” does not denote a single construction. It can refer to explicit non-product descriptions of the two-operator formula, multivariable factorizations, palindromic symmetric products, finite-order infinitesimal identities in synthetic differential geometry, exact closed forms under restrictive commutator hypotheses, or rigorously controlled finite truncations for unitary problems [1204.0389] [1702.04681] [1903.03140] [1808.00250] [1304.4667] [2107.01204] [2607.07692].

## 1. Classical framework and the meanings of generalization

The classical Zassenhaus formula is usually presented as the factorization dual to the Baker–Campbell–Hausdorff formula. Whereas BCH combines \(e^X e^Y\) into a single exponential, Zassenhaus disentangles \(e^{X+Y}\) into an ordered product beginning with \(e^X e^Y\) and followed by commutator corrections \(e^{C_2}e^{C_3}\cdots\). In the standard two-variable setting,
\[
C_2(X,Y)=-\frac12[X,Y], \qquad
C_3(X,Y)=\frac13[Y,[X,Y]]+\frac16[X,[X,Y]],
\]
and higher \(C_n\) are homogeneous Lie polynomials [1204.0389].

The literature uses “generalized Zassenhaus formula” in several distinct senses. Some works generalize the algebraic input from two operators to \(n\) operators. Others preserve the two-operator setting but alter the organization of the factorization, for example by imposing symmetry, replacing the infinite product by an explicit word series, or collapsing the full hierarchy of exponents under special commutator constraints. Still others generalize the formula operationally, treating truncated Zassenhaus products as computational schemes with convergence domains, error bounds, or numerical embeddings in operator splitting and quantum simulation [1903.03140] [1808.00250] [1702.04681] [2607.07692] [1204.0380] [2501.13922].

| Direction | Canonical form | Representative paper |
|---|---|---|
| Explicit two-operator reformulation | infinite sum of ordered \(B_n\)-words times \(e^A\) | [1702.04681] |
| Multivariable factorization | \(e^{X_1+\cdots+X_n}=e^{X_1}\cdots e^{X_n}\prod_{k\ge2}e^{W_k}\) | [1903.03140] |
| Symmetric palindromic factorization | palindromic product with \(C_{2k}=0\) | [1808.00250] |
| Infinitesimal finite-order identities | exact SDG factorizations by nilpotent degree | [1304.4667] |
| Closed-form solvable cases | single correction exponential \(e^{g[\!,\!]}\) | [2107.01204] |
| Rigorous truncation theory | finite products with explicit norm-error bounds | [2607.07692] |

A recurring misconception is that “generalized” must mean “many-operator.” The record is broader. Kimura’s work remains strictly two-operator but replaces the standard product-exponent viewpoint by an explicit combinatorial series [1702.04681]. The operator-splitting literature uses the term for generalized numerical deployment rather than for a new algebraic identity [1204.0380]. Taken together, these works suggest that generalized Zassenhaus formulas are best understood as a family of structurally different disentangling schemes rather than as a single theorem.

## 2. Explicit two-operator word-series reformulations

A particularly important reformulation is Kimura’s explicit description of the Zassenhaus formula, which does not present
\[
e^{A+B}=e^A e^B e^{C_2}e^{C_3}\cdots
\]
in the usual ordered-product form. Instead, it derives an explicit infinite series of ordered products of iterated adjoint-action operators multiplying \(e^A\) on the right [1702.04681]. Writing
\[
L_A O=[A,O], \qquad B_n=(L_A)^{n-1}B,
\]
the central formula is
\[
e^{A+B}
=
\left\{
1+\sum_{p=1}^{\infty}\sum_{n_1,\dots,n_p=1}^{\infty}
\frac{1}{n_p(n_p+n_{p-1})\cdots(n_p+\cdots+n_1)}
\,B_{n_p}\cdots B_{n_1}
\right\}e^A.
\]
This is an infinite sum, not an infinite product.

The construction begins by expanding \((A+B)^n\) and moving all \(A\)'s to the right:
\[
(A+B)^n \equiv \sum_{m=0}^n \frac{n!}{m!(n-m)!} X_m A^{n-m},
\]
where the \(X_m\) are polynomials in \(B\), iterated commutators of \(A\) with \(B\), and their products. From
\[
e^{A+B}=\left(\sum_{m=0}^{\infty}\frac{1}{m!}X_m\right)e^A,
\]
Kimura derives an operator recursion
\[
X_{m+1}=L_A X_m + B X_m, \qquad X_0=1,\quad X_1=B,
\]
and then decomposes \(X_m=\sum_{p=1}^m X_{m,p}\) by degree in \(B\). The extremal terms satisfy
\[
X_{m,1}=B_m, \qquad X_{m,m}=B^m,
\]
and the general \(X_{m,p}\) are obtained by a recursive combinatorial formula that ultimately yields the compact coefficient in the displayed series.

The same paper gives a transposed version,
\[
e^{A+B}
=
e^A
\left\{
1+\sum_{p=1}^{\infty}\sum_{n_1,\dots,n_p=1}^{\infty}
\frac{(-1)^{(n_p+\cdots+n_1)-p}}
{n_p(n_p+n_{p-1})\cdots(n_p+\cdots+n_1)}
\,B_{n_1}\cdots B_{n_p}
\right\},
\]
with reversed word order and an additional sign. It also rewrites \(e^X e^Y\) in analogous non-logarithmic series forms, thereby producing BCH-type companion expansions for the product side rather than for the logarithm [1702.04681].

This reformulation is equivalent to the classical Zassenhaus expansion but does not solve for the individual Lie-polynomial exponents \(Z_n\) or \(C_n\). That distinction is essential. The series gives explicit coefficients for every ordered monomial \(B_{n_p}\cdots B_{n_1}\), but it does not directly extract the factors \(e^{Z_m}\). Kimura explicitly notes that although the \(e^B\) part is visible by restricting to the \(B_1=B\) terms, it is hard to extract \(e^{Z_m}\) for arbitrary \(m\). Its value for generalized Zassenhaus theory lies elsewhere: it provides an all-orders coefficient system in a basis of iterated-adjoint words, which is particularly well suited to truncation, nilpotent situations, and potential extensions where direct product-exponent extraction is cumbersome [1702.04681].

## 3. Multivariable, symmetric, and infinitesimal extensions

The most direct algebraic extension is the multivariable Zassenhaus formula
\[
e^{X_1+X_2+\cdots+X_n}
=
e^{X_1}e^{X_2}\cdots e^{X_n}\prod_{k=2}^{\infty}e^{W_k},
\]
where each \(W_k\) is a homogeneous Lie polynomial of degree \(k\) in \(X_1,\dots,X_n\) [1903.03140]. Introducing a parameter \(t\),
\[
e^{t(X_1+\cdots+X_n)}
=
e^{tX_1}\cdots e^{tX_n}e^{t^2W_2}e^{t^3W_3}\cdots,
\]
one defines residual products \(R_m(t)\) and logarithmic derivatives \(F_m(t)=R_m'(t)R_m(t)^{-1}\), exactly as in the two-variable recursive framework. The first corrections are explicit:
\[
W_2=\frac12\sum_{1\le i<j\le n}[X_j,X_i],
\]
and
\[
W_3=
\frac13\sum_{1\le i<j,\;k\le n}[[X_j,X_i],X_k]
+\frac16\sum_{1\le i<j\le n}[[X_j,X_i],X_i].
\]
The paper develops explicit formulas up to \(W_6\) and general recursive formulas for \(W_7,\dots,W_{10}\), together with a composition-based combinatorial formula for the foundational coefficients \(f_{1,k}\) [1903.03140].

A different direction is the symmetric Zassenhaus formula, which replaces the one-sided core \(e^X e^Y\) by a palindromic shell,
\[
e^{X+Y}
=
e^{X/2}e^{Y/2}
\exp(C_2)\exp(C_3)\cdots
\exp(C_k)\cdots
\exp(C_k)\cdots
\exp(C_3)\exp(C_2)
e^{Y/2}e^{X/2}.
\]
The decisive structural result is that symmetry forces all even-degree exponents to vanish:
\[
C_{2k}\equiv 0,\qquad k\ge 1.
\]
Hence the factorization reduces to odd corrections only,
\[
e^{X+Y}
=
e^{X/2}e^{Y/2}
\exp(C_3)\exp(C_5)\cdots
\exp(C_{2k+1})\cdots
\exp(C_5)\exp(C_3)
e^{Y/2}e^{X/2},
\]
with, for example,
\[
C_3=\frac{1}{48}[X,[X,Y]]+\frac{1}{24}[Y,[X,Y]].
\]
This palindromic organization is a genuine Zassenhaus-type generalization rather than a mere rewriting of Strang splitting, because the correction factors remain homogeneous Lie polynomials and are generated recursively from logarithmic derivatives of partially corrected products [1808.00250].

Synthetic differential geometry yields a third kind of extension. In that setting, one works in a regular Lie group with nilpotent infinitesimals \(d_i\in D\), so identities truncate exactly by nilpotent degree [1304.4667]. The second-order Zassenhaus identity is
\[
\exp\bigl((d_1+d_2)(X+Y)\bigr)
=
\exp((d_1+d_2)X)\exp((d_1+d_2)Y)
\exp\!\left(-\frac{(d_1+d_2)^2}{2}[X,Y]\right).
\]
The third-order identity becomes
\[
\exp\bigl((d_1+d_2+d_3)(X+Y)\bigr)
=
\exp(\Sigma d_i X)\exp(\Sigma d_i Y)
\exp\!\left(-\frac{(\Sigma d_i)^2}{2}[X,Y]\right)
\exp\!\left(\frac{(\Sigma d_i)^3}{12}[X+2Y,[X,Y]]\right),
\]
and the fourth-order formula adds the quartic correction
\[
\exp\!\left(
\frac{(\Sigma d_i)^4}{24}
\bigl(
-[X,[X,[X,Y]]]-3[X,[Y,[X,Y]]]-3[Y,[Y,[X,Y]]]
\bigr)
\right).
\]
These are exact finite identities in SDG because the infinitesimal parameter is nilpotent, not because the series has been analytically summed [1304.4667].

## 4. Closed forms and algebraic collapse mechanisms

A major specialization occurs when the commutator closes linearly:
\[
[\hat X,\hat Y]=u\hat X+v\hat Y+c\mathcal I.
\]
Under this hypothesis, all higher nested commutators are proportional to \([\hat X,\hat Y]\),
\[
[\hat X,[\hat X,\hat Y]]=v[\hat X,\hat Y], \qquad
[\hat Y,[\hat X,\hat Y]]=-u[\hat X,\hat Y],
\]
so the infinite Zassenhaus product collapses to a single correction exponential [2107.01204]. The right-sided closed form is
\[
e^{\hat X+\hat Y}=e^{\hat X}e^{\hat Y}e^{g_r(u,v)[\hat X,\hat Y]},
\]
with equivalent centered and left-sided versions
\[
e^{\hat X+\hat Y}=e^{\hat X}e^{g_c(u,v)[\hat X,\hat Y]}e^{\hat Y}
=
e^{g_\ell(u,v)[\hat X,\hat Y]}e^{\hat X}e^{\hat Y},
\]
and coefficient relation
\[
g_r(u,v)=e^{\,v}g_c(u,v)=g_\ell(v,u).
\]
For \(u\neq v\),
\[
g_r(u,v)=
\frac{u\bigl(e^{u-v}-e^{u}\bigr)+v\bigl(e^{u}-1\bigr)}{uv(u-v)}.
\]
The special cases
\[
g_r(u,u)=\frac{1+u-e^u}{u^2},\qquad
g_r(0,0)=-\frac12,
\]
recover, in particular, the Glauber formula for a central commutator [2107.01204].

A different collapse mechanism is the no-mixed adjoint property,
\[
[\hat Y,\operatorname{ad}_{\hat X}^{\,i}\hat Y]=0 \qquad \forall i\ge 0,
\]
or equivalently
\[
ad_{\hat Y}ad_{\hat X}^{\,i}\hat Y=0 \qquad \forall i\in\mathbb N.
\]
Under this condition, every classical Zassenhaus exponent simplifies to a pure one-sided adjoint chain,
\[
\hat C_n(\hat X,\hat Y)=\frac{(-1)^{n-1}}{n!}\,ad_{\hat X}^{\,n-1}\hat Y,\qquad n\ge2.
\]
Moreover,
\[
[ad_{\hat X}^{p}\hat Y,\;ad_{\hat X}^{q}\hat Y]=0 \qquad \forall p,q\in\mathbb N,
\]
so the full infinite product recombines into a single exponential [2510.24364]:
\[
\exp(\hat X+\hat Y)
=
\exp(\hat X)\prod_{n=1}^{\infty}
\exp\!\left(\frac{(-1)^{n-1}}{n!}ad_{\hat X}^{\,n-1}\hat Y\right)
=
\exp(\hat X)\exp\!\left(
\sum_{n=1}^{\infty}\frac{(-1)^{n-1}}{n!}ad_{\hat X}^{\,n-1}\hat Y
\right).
\]
Formally,
\[
\sum_{n=1}^{\infty}\frac{(-1)^{n-1}}{n!}ad_{\hat X}^{\,n-1}\hat Y
=
\frac{1-\exp(-ad_{\hat X})}{ad_{\hat X}}\hat Y
=
\int_0^1 e^{-t\hat X}\hat Y e^{t\hat X}\,dt.
\]
This produces an exact specialized Zassenhaus formula rather than a truncation or asymptotic approximation [2510.24364].

These two collapse mechanisms differ in algebraic content. The commutator ansatz \([\hat X,\hat Y]=u\hat X+v\hat Y+c\mathcal I\) forces all nested commutators into the one-dimensional span of \([\hat X,\hat Y]\), whereas the no-mixed adjoint property preserves an entire commuting chain \(ad_{\hat X}^k\hat Y\). The first yields a single correction proportional to \([\hat X,\hat Y]\); the second yields a whole one-sided adjoint series that nevertheless exponentiates cleanly because its terms commute [2107.01204] [2510.24364].

## 5. Recursive computation, convergence, and truncation error

Efficient computation of Zassenhaus exponents is itself a major branch of the subject. Casas, Murua, and Nadinic introduce the parameterized factorization
\[
e^{\lambda(X+Y)}=e^{\lambda X}e^{\lambda Y}e^{\lambda^2 C_2}e^{\lambda^3 C_3}\cdots
\]
and define
\[
R_1(\lambda)=e^{-\lambda Y}e^{-\lambda X}e^{\lambda(X+Y)}, \qquad
R_n(\lambda)=e^{-\lambda^n C_n}R_{n-1}(\lambda),
\]
together with logarithmic derivatives
\[
F_n(\lambda)=\left(\frac{d}{d\lambda}R_n(\lambda)\right)R_n(\lambda)^{-1}.
\]
The central extraction rule is
\[
C_{n+1}=\frac{1}{(n+1)!}F_n^{(n)}(0),
\]
and the practical recursion is written in terms of coefficients \(f_{n,k}\):
\[
f_{1,k}
=
\sum_{j=1}^k \frac{(-1)^k}{j!(k-j)!}\operatorname{ad}_Y^{k-j}\operatorname{ad}_X^jY,
\]
\[
f_{n,k}
=
\sum_{j=0}^{[k/n]-1}\frac{(-1)^j}{j!}\operatorname{ad}_{C_n}^j f_{n-1,k-nj},
\]
with
\[
C_n=\frac1n\,f_{[(n-1)/2],\,n-1}\qquad (n\ge5).
\]
The distinctive point is not just recursion but minimality: via Lazard elimination, the resulting formulas are proved to consist of independent commutators and to contain the minimum number of commutator terms [1204.0389].

The computational advantages reported for this recursion are concrete. In a Mathematica implementation, the exponents \(C_n\) were computed up to \(n=20\) in less than 20 seconds, with memory about 35 MB; \(C_{20}\) contains 48,528 terms, all independent; and \(C_{16}\) has 54,146 terms in a word formulation but only 3,711 terms in the new commutator formulation [1204.0389]. In Banach algebras, the same paper derives a numerically enlarged convergence region. In particular, it contains the whole region
\[
\|X\|+\|Y\|<1.054,
\]
extends asymmetrically beyond radial bounds, contains
\[
\{(x,2.9216):x<0.00292\},\qquad
\{(2.893,y):y<0.0145\},
\]
and includes all points \((x,0)\) and \((0,y)\) with arbitrarily large \(x\) or \(y\) [1204.0389].

The symmetric Zassenhaus formula has its own convergence theory. In a Banach algebra with submultiplicative norm, a simple sufficient condition obtained from recursively bounded coefficients is
\[
\lambda(\|X\|+\|Y\|)<1.3225,
\]
which improves the previously established standard Zassenhaus guarantee \(\|X\|+\|Y\|<1.054\). A sharper asymmetric analysis produces a numerical region whose union with its \(X\leftrightarrow Y\) reflection contains the point \((0.001,1.539)\), as well as all \((x,0)\) and \((0,y)\) with arbitrarily large \(x\) or \(y\). The same paper also reports that truncations of the symmetric formula often converge faster than truncations of the standard one [1808.00250].

A more recent development is the derivation of explicit truncation-error bounds for unitary problems. For skew-adjoint \(A\) and \(B\), the truncated product
\[
e^{t(A+B)}\approx e^{tA}e^{tB}e^{t^2C_2(A,B)}\cdots e^{t^qC_q(A,B)}
\]
is treated as a rigorously controlled approximation in operator norm. For \(q=2\),
\[
\left\|
e^{t(A+B)}-e^{tA}e^{tB}e^{t^2C_2(A,B)}
\right\|
\le
\frac{t^3}{3}\|[B,[A,B]]\|
+
\frac{t^3}{6}\|[A,[A,B]]\|,
\]
and at \(t=1\),
\[
\left\|
e^{A+B}-e^Ae^Be^{-\frac12[A,B]}
\right\|
\le
\frac16\|[A,[A,B]]\|+\frac13\|[B,[A,B]]\|.
\]
For \(q=3\),
\[
\begin{aligned}
&\left\|
e^{t(A+B)}-e^{tA}e^{tB}e^{t^2C_2(A,B)}e^{t^3C_3(A,B)}
\right\| \\
&\le
\frac{t^4}{8}\|[B,[B,[A,B]]]\|
+\frac{t^4}{8}\|[B,[A,[A,B]]]\|
+\frac{t^4}{24}\|[A,[A,[A,B]]]\|
+\frac35 t^5\|[C_2,C_3]\|.
\end{aligned}
\]
In general,
\[
\left\|
e^{t(A+B)}-e^{tA}e^{tB}e^{t^2C_2(A,B)}\cdots e^{t^qC_q(A,B)}
\right\|
\le
\sum_{j=q+1}^{2q-1} t^j D_j^{[q]}(A,B),
\]
where each \(D_j^{[q]}(A,B)\) is a linear combination of norms of nested commutators. This converts the formal Zassenhaus product into a finite-order approximation scheme with explicit operator-norm control in the skew-adjoint setting [2607.07692].

## 6. Numerical splitting and quantum simulation

One practical use of generalized Zassenhaus formulas is as higher-order startup maps for operator splitting. For the linear Cauchy problem
\[
\frac{\partial c(t)}{\partial t}=Ac(t)+Bc(t),\qquad c(0)=c_0,
\]
a truncated Zassenhaus operator is defined by
\[
E_{\mathrm{Zassen},1}(t)=e^{At}e^{Bt},\qquad
E_{\mathrm{Zassen},j}(t)=\exp(C_j t^j)\quad (j\ge2),
\]
and
\[
E_{\mathrm{Zassen},i}(t)=\prod_{j=1}^i E_{\mathrm{Zassen},j}(t).
\]
The paper states that \(E_{\mathrm{Zassen},i}(t)\) has accuracy \(O(t^i)\). Embedding this into iterative splitting yields
\[
c_{i+j}(t)=E_{\mathrm{iter},j}(t)E_{\mathrm{Zassen},i}(t)c_0,
\]
with local error
\[
\|c(t)-c_{i+j}(t)\|\le O(t^{i+j})c_0.
\]
This construction is used after spatial discretization of PDEs into sparse matrix systems in CFD and transport-reaction models, where the Zassenhaus factors serve as accurate initializers for cheaper iterative corrections. The reported numerical observation is that with \(4\)-\(5\) iterative steps one obtains more accurate results than with the expensive standard schemes [1204.0380].

The no-mixed-adjoint specialization has also been deployed in a structured unitary coupled cluster setting. There, the operator class generated by \(\hat A_{ij}\) and \(\hat B_i\) satisfies
\[
[\hat A_{ij},\hat A_{kl}]=0,\qquad
[\hat B_i,\hat B_k]=0,\qquad
[\hat A_{ij},\hat B_k]=\delta_{ik}\hat B_j+\delta_{jk}\hat B_i,
\]
which implies the no-mixed adjoint property for
\[
\hat X=\sum_{1\le i<j\le N}\mu_{ij}\hat A_{ij},\qquad
\hat Y=\sum_{i=1}^{N}\mu_i\hat B_i.
\]
The infinite Zassenhaus series then collapses to an exact finite disentangling. In the invertible-matrix case,
\[
\sum_{k=0}^{\infty}\hat C_{k+1}(\hat X,\hat Y)
=
\vec\mu\cdot(\mathbb I-\exp(-M))M^{-1}\cdot \vec{\hat B},
\]
and the paper derives an exact finite product formula with at most
\[
\frac{N(N+1)}{2}
\]
elementary Givens-gate factors. The paper states that this ansatz requires no Trotterization and is exact on a quantum computer with a finite number of Givens gates equal to the number of free parameters [2510.24364].

A different quantum direction is provided by stochastic Zassenhaus expansions. For a Hamiltonian \(H=A+B\),
\[
e^{-itH}=e^{-itA}e^{-itB}\prod_{k=2}^{\infty}e^{-it^k H_k},
\]
where each \(H_k\) is the Hermitian representative of the corresponding Zassenhaus Lie polynomial [2501.13922]. The method nests Zassenhaus expansions recursively and approximates selected higher-order exponentials stochastically. If
\[
H=\sum_k c_k P_k,
\]
with Pauli decomposition, then
\[
e^{-it^m H}
=
\sum_k p_k e^{-i\theta(t)P_k'}
+
\mathcal O(|H|_1^2 t^{2m}),
\]
where
\[
p_k=\frac{|c_k|}{|H|_1},\qquad
P_k'=\operatorname{sign}(c_k)P_k,\qquad
\theta(t)=\sec^{-1}\!\left(\sqrt{1+(t^m|H|_1)^2}\right).
\]
This yields the \(\mathrm{SZE}_{k,p}\) hierarchy, whose leading error is
\[
\mathcal O\!\left(\|H'_{p+1}\|t^{p+1}+|H'_{k+1}|_1^2 t^{2(k+1)}\right).
\]
For a 10-qubit transverse-field Ising model, the paper reports an 11th-order stochastic Zassenhaus expansion with 42x fewer CNOTs than the standard 10th-order product formula, together with regimes in which the simulation error is reduced by many orders of magnitude relative to leading alternatives [2501.13922].

These numerical and algorithmic developments do not alter the classical Lie-theoretic core, but they broaden its operational meaning. A plausible implication is that generalized Zassenhaus formulas now function simultaneously as algebraic factorizations, combinatorial word-series expansions, symmetry-adapted products, exact collapse formulas on special Lie classes, and finite-order simulation primitives for sparse PDE solvers and quantum circuits.

Source: https://www.emergentmind.com/topics/generalized-zassenhaus-formulas