---
title: 'Almost Inner Derivations: Definitions and Rigidity'
url: https://www.emergentmind.com/topics/almost-inner-derivations
type: topic
---

# Almost Inner Derivations: Definitions and Rigidity

Almost inner derivations are derivations that satisfy an innerness condition weaker than the existence of a single global commutator implementer. The precise meaning depends on context. For finite-dimensional Lie algebras, an almost inner derivation is a derivation $D$ such that $D(x)\in [x,\mathfrak g]$ for all $x$, equivalently $D(x)=[a_x,x]$ with $a_x$ depending on $x$ [1704.06159]. In associative and operator-algebra settings, the same phrase can refer to local innerness on pairs, as in inner $2$-local derivations, or to norm-limits of inner derivations, as in separable unital $C^*$-algebras [1303.6033] [2508.21726]. A recurring theme is rigidity: in many natural classes, the almost-inner condition collapses to genuine innerness, while in others the quotient by inner derivations is small, explicitly computable, or even infinite-dimensional [1308.6117] [1905.08145] [2507.09484].

## 1. Definitions and principal formulations

The foundational Lie-algebraic definition used by Gordon–Wilson and developed systematically in later work is pointwise innerness: for a finite-dimensional Lie algebra $\mathfrak g$ over a field of characteristic zero, a derivation $D\in \operatorname{Der}(\mathfrak g)$ is almost inner if
$$
D(x)\in [x,\mathfrak g]\quad\text{for all }x\in \mathfrak g.
$$
The space of such derivations is denoted $\operatorname{AID}(\mathfrak g)$, and one has
$$
\operatorname{Inn}(\mathfrak g)\subseteq \operatorname{AID}(\mathfrak g)\subseteq \operatorname{Der}(\mathfrak g),
$$
with the quotient $\operatorname{AID}(\mathfrak g)/\operatorname{Inn}(\mathfrak g)$ measuring outer almost inner derivations [1905.08145]. A closely related exposition emphasizes the equivalent formulation that for every $x$ there exists $a_x\in\mathfrak g$ such that $D(x)=[a_x,x]$ [1704.06159].

Several adjacent notions appear in the literature. For associative rings, an inner $2$-local derivation on a ring $S$ is a map $\Delta$ such that for every pair $x,y\in S$ there exists $A_{x,y}\in S$ with
$$
\Delta(x)=[A_{x,y},x],\qquad \Delta(y)=[A_{x,y},y].
$$
This is a local two-point version of innerness rather than a single global commutator identity [1303.6033]. For derivations from a von Neumann algebra $\mathcal M$ into a quasi-normed bimodule $E\subset LS(\mathcal M)$, “almost inner” can mean that the derivation is implemented by some $a\in LS(\mathcal M)$ before one proves that the implementer can in fact be chosen in $E$ itself [1308.6117]. In separable unital $C^*$-algebras, the phrase can denote membership in the norm-closure of inner derivations inside the Banach space of derivations; in that setting, almost inner derivations exist precisely when the subspace of inner derivations is not norm-closed [2508.21726].

A useful summary is that the phrase “almost inner” is not tied to a single universal definition. It ranges from pointwise commutator conditions, to local-pair innerness, to ambient-algebra implementability, to norm-approximation by inner derivations. What unifies these usages is that innerness is required only after weakening either the domain of comparison, the topology, or the location of the implementer.

| Context | Almost-inner condition | Typical outcome |
|---|---|---|
| Lie algebras | $D(x)\in [x,\mathfrak g]$ for all $x$ | Often $\operatorname{AID}=\operatorname{Inn}$, but not always [1704.06159] |
| Matrix rings | Inner $2$-locality on each pair $(x,y)$ | Collapse to a single inner derivation under commutativity hypotheses [1303.6033] |
| Von Neumann bimodules | Implementer first found in $LS(\mathcal M)$ | Refined to an implementer in the target module $E$ [1308.6117] |
| Unital $C^*$-algebras | Norm-limit of inner derivations | Equivalent to non-closedness of inner derivations [2508.21726] |

## 2. Associative, operator-algebraic, and $C^*$-algebraic rigidity

A strong rigidity phenomenon appears for matrix rings over commutative associative rings. If $R$ is a commutative associative ring with identity and $n>1$, every inner $2$-local derivation $\Delta:M_n(R)\to M_n(R)$ is an inner derivation:
$$
\Delta(X)=[A,X]\qquad \text{for all }X\in M_n(R)
$$
for a single $A\in M_n(R)$ [1303.6033]. The proof uses matrix units $e_{ij}$, corner identities such as
$$
e_{kk}a(i,j)e_{ij}=e_{kk}a(i,k)e_{ij},
$$
a decomposition of $\Delta(e_{ij})$, and a normalization by the cyclic banded element $x_0=\sum_{k=1}^{n-1}e_{k,k+1}$ to pin down diagonal differences of local implementers [1303.6033]. The same paper also proves that every derivation $\delta:R\to R$ extends entrywise to a derivation $\delta_n:M_n(R)\to M_n(R)$ by
$$
\delta_n((a_{ij}))=(\delta(a_{ij})),
$$
with $\delta_n(XY)=\delta_n(X)Y+X\delta_n(Y)$ [1303.6033]. This situates matrix rings as a setting where local innerness collapses to global innerness, but ordinary derivations may still contain both inner and entrywise components.

For von Neumann algebras, the rigidity is stronger. If $\mathcal M$ is a von Neumann algebra and $E$ is a quasi-normed $\mathcal M$-bimodule of locally measurable operators, then every derivation $\delta:\mathcal M\to E$ is inner: there exists $d\in E$ such that
$$
\delta(x)=[d,x]\qquad \text{for all }x\in\mathcal M,
$$
and moreover
$$
\|d\|_E\le 2C_E\|\delta\|_{M\to E},
$$
with the sharper bound $\|d\|_E\le \|\delta\|_{M\to E}$ in the self-adjoint or skew-adjoint cases [1308.6117]. The proof proceeds by automatic continuity, extension to $LS(\mathcal M)$, $t(\mathcal M)$-continuity there, innerness on $LS(\mathcal M)$, and then a commutator inequality forcing the implementer into $E$ rather than merely into $LS(\mathcal M)$ [1308.6117]. This means that “inner in a larger ambient algebra” is not a genuinely weaker notion under the hypotheses of that theorem.

A related semifinite result treats derivations $\delta:\mathcal M\to E(\tau)$, where $E(\tau)$ is a symmetric operator space of $\tau$-compact operators. Under mild assumptions—namely that $E(\tau)$ is the dual of a symmetric space $F(\tau)$ and either separability or reflexivity holds—every continuous derivation $\delta:N\to E(\tau)$ is inner with implementer in $E(\tau)$ and
$$
\|d\|_{E(\tau)}\le \|\delta\|
$$
[1204.4052]. In particular, every derivation $\delta:\mathcal M\to L_p(\mathcal M,\tau)$ is inner for $1<p<\infty$ [1204.4052]. Here again, almost-inner behavior in an ambient measurable-operator sense collapses to genuine innerness.

For separable unital $C^*$-algebras, the perspective changes from algebraic localness to topological closure. A derivation is “almost inner” when it belongs to the norm-closure of the inner derivations $\{\delta_x:x\in M(A)\}$ [2508.21726]. The main equivalence is that the subspace of inner derivations is norm-closed if and only if $\operatorname{Inn}(A)$ is norm-closed in $\operatorname{Aut}(A)$, and this is further equivalent to a descriptive-set-theoretic condition:
$$
H^1(A)\text{ is a Banach space}\quad\Longleftrightarrow\quad \operatorname{Inn}(A)\text{ is norm-closed}\quad\Longleftrightarrow\quad \operatorname{Inn}(A)\in \Pi^0_3
$$
[2508.21726]. Thus almost inner derivations, in this norm-closure sense, exist precisely when inner derivations fail to be norm-closed [2508.21726]. This suggests a conceptual bifurcation: in von Neumann and measurable-operator contexts almost innerness tends to collapse algebraically, whereas in $C^*$-algebras it can survive as a genuine topological closure phenomenon.

## 3. Lie algebras: pointwise innerness, rigidity, and counterexamples

The Lie-algebraic theory begins with the observation that $\operatorname{AID}(\mathfrak g)$ is itself a Lie subalgebra of $\operatorname{Der}(\mathfrak g)$, that $\operatorname{Inn}(\mathfrak g)\subseteq \operatorname{CAID}(\mathfrak g)\subseteq \operatorname{AID}(\mathfrak g)$, and that for nilpotent $\mathfrak g$, every almost inner derivation is a nilpotent endomorphism and $\operatorname{AID}(\mathfrak g)$ is a nilpotent Lie algebra [1704.06159]. The same work introduced the method of fixed basis vectors: for $D\in \operatorname{AID}(\mathfrak g)$, one chooses $\varphi_D$ with $D(x)=[x,\varphi_D(x)]$, and sufficient coordinate equalities imply that each basis vector is “fixed”; once every basis vector is fixed, $D$ must be inner [1704.06159]. This method yields $\operatorname{AID}(\mathfrak g)=\operatorname{Inn}(\mathfrak g)$ for broad classes, including $2$-step nilpotent Lie algebras determined by graphs, free $2$-step and $3$-step nilpotent Lie algebras, free metabelian nilpotent Lie algebras on two generators over an infinite field, almost abelian Lie algebras over $\mathbb C$, and triangular Lie algebras [1704.06159].

A later paper extends this rigidity. Over a field of characteristic zero, one has
$$
\operatorname{AID}(\mathfrak f_{r,c})=\operatorname{Inn}(\mathfrak f_{r,c})
$$
for every free $c$-step nilpotent Lie algebra $\mathfrak f_{r,c}$ [1905.08145]. If $\mathfrak g$ has a codimension-one abelian ideal, then again $\operatorname{AID}(\mathfrak g)=\operatorname{Inn}(\mathfrak g)$ [1905.08145]. More generally, if $\mathfrak g=\mathfrak a\rtimes \mathfrak s$ over an algebraically closed field of characteristic zero, with $\mathfrak a$ the abelian solvable radical and $\mathfrak s$ semisimple, then all almost inner derivations are inner [1905.08145]. These results identify several structurally rigid settings in which the pointwise condition $D(x)\in[x,\mathfrak g]$ already forces a global commutator.

Filiform nilpotent Lie algebras show that rigidity is not universal. The standard filiform algebra $L_n$ satisfies $\operatorname{AID}(L_n)=\operatorname{Inn}(L_n)$, but for the family $R_n$ one has
$$
\operatorname{AID}(R_n)=\operatorname{Inn}(R_n)\oplus \operatorname{span}\{E_{n,2}\},
$$
so there is exactly one outer almost inner derivation [1905.08145]. For $W_n$, the almost inner space is larger:
$$
\operatorname{AID}(W_n)=\operatorname{Inn}(W_n)\oplus \operatorname{span}\{t_1,t_2,t_3\}\qquad (n\ge 9),
$$
with explicit witnesses $p_{t_i}(x)$ satisfying $t_i(x)=[x,p_{t_i}(x)]$ [1905.08145]. Most strikingly, the characteristically nilpotent filiform family $\mathfrak f_n$ for $n\ge 13$ satisfies
$$
\operatorname{Der}(\mathfrak f_n)=\operatorname{AID}(\mathfrak f_n),
$$
and $\operatorname{AID}(\mathfrak f_n)/\operatorname{Inn}(\mathfrak f_n)$ has dimension $3$ [1905.08145]. This is presented there as the first known family of nilpotent Lie algebras with all derivations almost inner [1905.08145].

Low-dimensional and family-based examples show that the quotient $\operatorname{AID}(\mathfrak g)/\operatorname{Inn}(\mathfrak g)$ can be arbitrarily large. In dimension $5$, examples such as $\mathfrak g_{5,6}$ and $\mathfrak g_{5,3}$ have
$$
\operatorname{AID}(\mathfrak g)=\operatorname{Inn}(\mathfrak g)\oplus \langle E_{5,2}\rangle
$$
or
$$
\operatorname{AID}(\mathfrak g)=\operatorname{Inn}(\mathfrak g)\oplus \langle E_{5,3}\rangle
$$
[1704.06159]. More dramatically, there exist $2$-step nilpotent Lie algebras $\mathfrak g_n$ of dimension $4n+2$ such that
$$
\dim\big(\operatorname{AID}(\mathfrak g_n)/\operatorname{Inn}(\mathfrak g_n)\big)=n
$$
[1704.06159]. This shows that even within nilpotent Lie algebras, the failure of innerness can be systematically large rather than sporadic.

## 4. Matrix pencils, algorithms, and explicit determination of $\operatorname{AID}$

For $2$-step nilpotent Lie algebras of genus $2$, the theory becomes highly explicit through matrix pencils. Writing $\mathfrak g=U\oplus Z$ with $Z=[\mathfrak g,\mathfrak g]=\operatorname{span}\{z_1,z_2\}$ and $U$ abelian, the bracket is encoded by skew-symmetric matrices $A,B$ via
$$
[x_i,x_j]=a_{ij}y_1+b_{ij}y_2,
$$
and the associated pencil is
$$
P(\lambda,\mu)=\mu A+\lambda B.
$$
The canonical form of this skew pencil is described by minimal indices and elementary divisors, and these invariants determine the dimension of $\operatorname{AID}(\mathfrak g)$ [2004.10567]. Over an algebraically closed field of characteristic not $2$,
$$
\dim \operatorname{AID}(\mathfrak g)=\dim \operatorname{Inn}(\mathfrak g)
+\sum_{\varepsilon_j\neq 0}(\varepsilon_j-1)
+2\sum (e_j-1)
+2\sum (f_i-1),
$$
while over $\mathbb R$ one adds the contribution
$$
4\sum m_r
$$
from complex-conjugate quadratic blocks [2004.10567]. In particular, over $\mathbb R$, a single real complex-conjugate block gives $\operatorname{AID}(\mathfrak g)=C(\mathfrak g)$ blockwise [2004.10567].

A complementary computational approach starts directly from a structure-constant table. For a finite-dimensional Lie algebra $\mathfrak g$ with basis $\{e_1,\dots,e_n\}$ and structure constants $c_{ij}^k$, one first solves the linear derivation equations
$$
\sum_{m=1}^n c_{ij}^m d_m^k
=
\sum_{p=1}^n d_i^p c_{pj}^k
+
\sum_{q=1}^n d_j^q c_{iq}^k
$$
to obtain $\operatorname{Der}(\mathfrak g)$, then computes $\operatorname{Inn}(\mathfrak g)$ from the matrices of $\operatorname{ad}_{e_s}$ [2403.05905]. The key test for almost innerness uses the matrix $M(z)$ defined by
$$
m_{k,j}(z)=\sum_{i=1}^n z_i c_{ij}^k
$$
and the vector $v_D(z)$ corresponding to $D(b_z)$ for $b_z=\sum z_i e_i$. Then
$$
D\in \operatorname{AID}(\mathfrak g)\quad \Longleftrightarrow\quad
\operatorname{rank}(M(z))=\operatorname{rank}([M(z)\mid v_D(z)])
\text{ for all }z\in F^n
$$
[2403.05905]. Over an algebraically closed field this becomes an ideal-membership condition in the polynomial ring generated by minors of $M(z)$ and of the augmented matrix [2403.05905].

This computational framework also answers a structural question about the quotient $\operatorname{AID}(\mathfrak g)/\operatorname{Inn}(\mathfrak g)$. For a specific $15$-dimensional Lie algebra $g_3$ in characteristic $3$, one has
$$
\dim \operatorname{Der}(g_3)=45,\qquad \dim \operatorname{Inn}(g_3)=12,
$$
and
$$
\operatorname{AID}(g_3)=\operatorname{Inn}(g_3)\oplus U
$$
for a $21$-dimensional subalgebra $U$ [2403.05905]. Two explicit derivations $d_1,d_2\in U$ have nonzero commutator modulo inner derivations, proving that
$$
\operatorname{AID}(g_3)/\operatorname{Inn}(g_3)
$$
can be non-abelian [2403.05905]. This answers a question of Kunyavskii and Ostapenko in the affirmative.

## 5. Extensions to Leibniz algebras, Lie superalgebras, and generalized derivation theories

For Leibniz algebras, the right Leibniz identity changes the role of inner derivations: right multiplications
$$
R_x(y)=[y,x]
$$
are derivations and constitute the inner derivations in the convention of the cited work [2010.01904]. An almost inner derivation is then a derivation $D$ such that for every $x$ there exists $a_x$ with
$$
D(x)=[x,a_x]\in [x,L].
$$
In addition to $\operatorname{AID}(L)$, the paper introduces right central almost inner derivations and central almost inner derivations, denoted $\operatorname{RCAID}(L)$ and $\operatorname{CAID}(L)$, with
$$
\operatorname{Inner}(L)\subset \operatorname{CAID}(L)\subset \operatorname{RCAID}(L)\subset \operatorname{AID}(L)\subset \operatorname{Der}(L)
$$
[2010.01904]. For null-filiform Leibniz algebras, $\operatorname{AID}=\operatorname{Inner}$, but in several filiform families the quotient acquires a one-dimensional extension generated by the map $E_{n,2}$:
$$
\operatorname{AID}(L)=\operatorname{Inner}(L)\oplus \langle E_{n,2}\rangle
$$
under explicit parameter conditions [2010.01904]. A related paper on two-step nilpotent Leibniz algebras with one-dimensional commutator ideal proves that every almost inner derivation is inner, with exactly three exceptional families: the Heisenberg Leibniz algebras $\mathfrak l_{2n+1}^{J_{\pm 1}}$ and the Dieudonné algebras $\mathfrak d_n$ [2301.06645].

For Lie superalgebras, the super-analogue of pointwise innerness is defined by requiring $D(x)\in [x,L]$ for every homogeneous $x$ [2509.00689]. Although finite-dimensional simple Lie superalgebras may admit outer derivations, the main theorem states that for every finite-dimensional simple Lie superalgebra $L$ over $\mathbb C$,
$$
aDer(L)=iDer(L).
$$
The proof shows that the known outer derivations—Euler derivations, odd derivations for $\mathfrak{psq}(n)$, and the outer $\mathfrak{sl}(2)$ for $\mathfrak{psl}(2|2)$—are not almost inner [2509.00689]. At the same time, naturally occurring non-inner almost inner derivations do exist for quasireductive Lie superalgebras built from prehomogeneous vector spaces; there the Euler derivation can be almost inner whenever all orbits are conical, while perfectness of the even part prevents innerness [2509.00689]. This identifies a super-analogue of the gap between simple rigidity and non-semisimple flexibility.

A different generalization replaces the Leibniz rule by a $\delta$-twisted one. A $\delta$-derivation $D:L\to M$ satisfies
$$
D([x,y])=\delta(x\cdot D(y)-y\cdot D(x)).
$$
For semisimple finite-dimensional Lie algebras over an algebraically closed field of characteristic $0$, the classification is extremely rigid: $\operatorname{Der}_\delta(\mathfrak g,V)\neq 0$ only for $\delta=1$, $\delta=\frac12$, or $\delta=-2$, with the non-inner cases confined to scalar maps on adjoint modules for $\delta=\frac12$ and explicit $\mathfrak{sl}(2)$ exceptions [2211.06645]. This is not the same notion as almost inner derivation, but it provides a parallel instance where derivations are “almost” ordinary only in a sharply delimited sense.

## 6. Cohomological and affine phenomena

A cohomological analogue of almost innerness appears for triple derivations on von Neumann algebras. Every triple derivation $\delta:M\to M$ is inner as a triple derivation, but for triple derivations $\delta:M\to M_*$ the picture changes [1309.3526]. For a factor $M$, the quotient
$$
\mathrm{Der}_{\mathrm{tr}}(M,M_*)\big/\overline{\operatorname{Inn}_{\mathrm{tr}}(M,M_*)}
$$
has dimension $0$ if and only if $M$ is finite, and dimension $1$ if $M$ is properly infinite [1309.3526]. In a properly infinite factor, every triple derivation decomposes as
$$
\delta=\delta_{\mathrm{inn}}+\lambda\,\delta_0,\qquad \lambda\in \mathbb R,
$$
with $\delta_{\mathrm{inn}}$ in the norm-closure of inner triple derivations and $\delta_0=D_{i\omega}\circ *$ representing the unique obstruction [1309.3526]. This is an “almost inner” phenomenon in the precise sense that innerness fails only by a one-dimensional quotient.

Loop algebras and affinizations furnish another sharply classified setting. Let $\mathfrak L$ be a minimal $Q$-graded subalgebra of a semisimple Lie algebra over an algebraically closed field of characteristic $0$. Its derived algebra
$$
I=[\mathfrak L,\mathfrak L]
$$
is abelian, and every almost inner derivation of $\mathfrak L$ is inner:
$$
AID(\mathfrak L)=Inn(\mathfrak L)
$$
[2507.09484]. If $S=F[t,t^{-1}]$, then under $\dim H=\dim I$ the loop algebra satisfies
$$
Der(\mathfrak L\otimes S)=Der(\mathfrak L)\otimes S\oplus Cent(\mathfrak L)\otimes Der(S),
$$
and every almost inner derivation of $\mathfrak L\otimes S$ is inner [2507.09484]. The affinization
$$
\widehat{\mathfrak L}=\mathfrak L\otimes S\oplus FK
$$
behaves differently: after modding out by inner derivations, almost inner derivations are exactly those vanishing on $I\otimes S$ and $K$ and sending $H\otimes S$ into $FK$ [2507.09484]. The quotient
$$
AID(\widehat{\mathfrak L})/Inn(\widehat{\mathfrak L})
$$
is spanned by explicit derivations $D_{ij}$, indexed by $1\le i\le \ell$ and $j\in \mathbb Z$, and this infinite set is linearly independent [2507.09484]. Thus a loop algebra can be rigid while its central extension acquires an infinite-dimensional almost-inner quotient.

From a broader perspective, these examples show that “almost inner” phenomena organize themselves around cohomological obstructions. Sometimes the obstruction vanishes completely, as in simple Lie superalgebras, matrix rings over commutative rings, and derivations into quasi-normed bimodules [1303.6033] [1308.6117] [2509.00689]. Sometimes it is finite-dimensional and explicit, as in triple derivations into the predual of a factor or filiform Lie algebras [1309.3526] [1905.08145]. Sometimes it is genuinely infinite-dimensional, as for affinizations of minimal $Q$-graded subalgebras [2507.09484]. This suggests that the topic is best understood not as a single theorem but as a family of rigidity and deformation problems centered on how far pointwise, local, ambient, or approximate innerness can deviate from actual innerness.

Source: https://www.emergentmind.com/topics/almost-inner-derivations