1 Introduction

Let \({\mathcal {A}}\) and \({\mathcal {B}}\) be Banach algebras, and let \({\mathscr {L}}^n({\mathcal {A}},{\mathcal {B}})\) be the space of n-linear bounded operators from \({\mathcal {A}}^n\) to \({\mathcal {B}}\). In this note, we study an Ulam type stability problem for a non-associative version of the multiplicativity equation, namely,

$$\begin{aligned} T(xy)=\Psi (T(x),T(y))\qquad (x,y\in {\mathcal {A}}), \end{aligned}$$
(1.1)

where \(T\in {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\) and \(\Psi \in {\mathscr {L}}^2({\mathcal {B}},{\mathcal {B}})\). An expected stability effect should be quite different from the one for almost multiplicative maps. Indeed, \(\Psi \) being merely ‘almost’ (but not exactly) associative on the range of T means that (1.1) cannot be satisfied accurately, since the left-hand side yields an associative operation on (xy). Hence, the best accuracy with which Eq. (1.1) can be satisfied should be expressed in terms of the ‘defect’ of \(\Psi \) concerning its associativity.

Our main result says, roughly speaking, that for certain Banach algebras, any operator T satisfying (1.1) with sufficient accuracy can be approximated to a prescribed accuracy by an operator for which the obvious obstacle of non-associativity of \(\Psi \) is actually the only obstacle in the way of satisfying (1.1).

To be more precise, for any linear map \(T:{\mathcal {A}}\rightarrow {\mathcal {B}}\) and any bilinear map \(\Psi :{\mathcal {B}}\times {\mathcal {B}}\rightarrow {\mathcal {B}}\), consider \(T^{\,\vee }:{\mathcal {A}}^2\rightarrow {\mathcal {B}}\) defined by

$$\begin{aligned} T^{\,\vee }(x,y)=T(xy)-\Psi (T(x),T(y)). \end{aligned}$$

We define the \(\Psi \)-multiplicative defect of T as \(\Vert T^{\,\vee }\Vert \), that is,

$$\begin{aligned} {\textsf{m}}\text{- }\text {def}_\Psi (T)=\sup \big \{\Vert T(xy)-\Psi (T(x),T(y))\Vert :x,y\in {\mathcal {A}},\, \Vert x\Vert , \Vert y\Vert \le 1\big \}. \end{aligned}$$

Similarly, we define the associative defect of \(\Psi \) by

$$\begin{aligned} {\textsf{a}}\text{- }\text {def}(\Psi )=\sup \big \{\Vert \Psi (u,\Psi (v,w))-\Psi (\Psi (u,v),w)\Vert :u,v,w\in {\mathcal {B}},\, \Vert u\Vert , \Vert v\Vert , \Vert w\Vert \le 1\big \}. \end{aligned}$$

Our main result then reads as follows.

Theorem 1

Let \({\mathcal {A}}\) be a unital amenable Banach algebra and \({\mathcal {B}}\) be a unital dual Banach algebra with an isometric predual. Then for arbitrary \(K,L\ge 1\) and \(\varepsilon ,\eta \in (0,1)\) there exists \(\delta >0\) such that the following holds true: If \(\Psi \in {\mathscr {L}}^2({\mathcal {B}},{\mathcal {B}})\) satisfies \(\Vert \Psi \Vert \le L\) and \(\Psi (1,u)=\Psi (u,1)=u\) for every \(u\in {\mathcal {B}}\), and \(T\in {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\) is a unital operator with \(\Vert T\Vert \le K\) and \({\textsf{m}}\text{- }\text {def}_\Psi (T)\le \delta \), then there exists a unital \(S\in {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\) such that

$$\begin{aligned} {\textsf{m}}\text{- }\text {def}_\Psi (S)<{\textsf{a}}\text{- }\text {def}(\Psi )^{1-\eta }\quad \text{ and } \quad \Vert T-S\Vert <\varepsilon . \end{aligned}$$

In order to put this result in a wider context, recall that the theory of almost multiplicative maps between Banach algebras was developed by B.E. Johnson in the series of papers [8,9,10,11], where he introduced the AMNM property. A pair \(({\mathcal {A}},{\mathcal {B}})\) of Banach algebras has this property, provided that for any \(K,\varepsilon >0\) there is \(\delta >0\) such that for every \(T\in {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\) satisfying \(\Vert T\Vert \le K\) and \(\Vert T(xy)-T(x)T(y)\Vert <\delta \Vert x\Vert \Vert y\Vert \) for all nonzero \(x,y\in {\mathcal {A}}\), there exists a multiplicative operator \(S\in {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\) with \(\Vert T-S\Vert <\varepsilon \). One of the brilliant ideas of Johnson was to define a ‘Banach-algebraic’ analogue of the Newton–Raphson approximation procedure which allowed him to show that if \({\mathcal {A}}\) is amenable and \({\mathcal {B}}\) is a dual Banach algebra, then \(({\mathcal {A}},{\mathcal {B}})\) has the AMNM property (see [11, Thm. 3.1]). His result was generalized in the PhD thesis of Horváth [7], and widely developed in the recent paper [4] by Choi, Horváth and Laustsen, where they exhibited a large collection of AMNM pairs of algebras of bounded operators on Banach spaces. Many other Ulam type stability problems on Banach algebras have been studied in recent years; see e.g. [1,2,3, 12, 13].

The proof of Theorem 1 closely follows the above-mentioned method of proof [11, Thm. 3.1] by using a Newton–Raphson-like algorithm defined in terms of an approximate diagonal. However, unlike in the case of Johnson’s method, unless \(\Psi \) is associative, our approximation process must terminate at a certain point.

2 Terminology and preparatory lemmas

Throughout this section, we fix unital Banach algebras \({\mathcal {A}}\) and \({\mathcal {B}}\), as well as operators \(T\in {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\) and \(\Psi \in {\mathscr {L}}^2({\mathcal {B}},{\mathcal {B}})\).

One of the main ideas of approximating almost multiplicative maps (see [4, 11]) is based on a modification of standard coboundary operators in the Hochschild cohomology theory, as defined in [14, Ch. 2], by introducing as a parameter a map \(\phi \in {\mathscr {L}}({\mathcal {A}},{\mathcal {B}})\) playing the role of ‘approximate action’ of \({\mathcal {A}}\) on \({\mathcal {B}}\). According to [4, Def. 5.2], for any \(n\in {\mathbb {N}}_0\), the n-coboundary operator \(\partial ^n_\phi :{\mathscr {L}}^n({\mathcal {A}},{\mathcal {B}})\rightarrow {\mathscr {L}}^{n+1}({\mathcal {A}},{\mathcal {B}})\) is defined by the formula

$$\begin{aligned} \begin{aligned} \partial ^n_\phi \psi (a_1,\ldots ,a_{n+1})&=\phi (a_1)\psi (a_2,\ldots ,a_{n+1}) +\sum _{j=1}^n (-1)^j \psi (a_1,\ldots ,a_ja_{j+1},\ldots ,a_{n+1})\\&\quad +(-1)^{n+1}\psi (a_1,\ldots ,a_n)\phi (a_{n+1}). \end{aligned} \end{aligned}$$

In fact, in the study of almost multiplicative maps only the 2-coboundary operator is relevant. For studying Eq. (1.1), we introduce the following modification of \(\partial _\phi ^2\). Define \(\delta _T^2:{\mathscr {L}}^2({\mathcal {A}},{\mathcal {B}})\rightarrow {\mathscr {L}}^3({\mathcal {A}},{\mathcal {B}})\) by the formula

$$\begin{aligned} \delta ^2_T\,\phi (x,y,z)=\Psi (T(x),\phi (y,z))-\phi (xy,z)+\phi (x,yz)-\Psi (\phi (x,y),T(z)). \end{aligned}$$

Observe that if \(\Psi \) is ‘almost’ associative, then the operator \(T^{\,\vee }\) ‘almost’ satisfies the relation \(T^{\,\vee }\in \textrm{ker}\,\delta _T^2\) in the following sense.

Lemma 2

For any \(T\in {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\) and \(\Psi \in {\mathscr {L}}^2({\mathcal {B}},{\mathcal {B}})\), we have

$$\begin{aligned} \Vert \delta _T^2\, T^{\,\vee }\Vert \le {\textsf{a}}\text{- }\text {def}(\Psi )\cdot \Vert T\Vert ^3. \end{aligned}$$

Proof

Note that

$$\begin{aligned} \begin{aligned}{}[\delta _T^2\, T^{\,\vee }](x,y,z)&= \Psi (T(x),T^{\,\vee }(y,z))+\Psi (T(xy),T(z))\\&\hspace{64pt}-\Psi (T(x),T(yz))-\Psi (T^{\,\vee }(x,y),T(z))\\&=\Psi [\Psi (T(x),T(y)),T(z)]-\Psi [T(x),\Psi (T(y),T(z))], \end{aligned} \end{aligned}$$

from which the result follows readily. \(\square \)

Lemma 3

The Fréchet derivative of the map

$$\begin{aligned} {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\ni S\mapsto {}S^{\,\vee }\in {\mathscr {L}}^2({\mathcal {A}},{\mathcal {B}}) \end{aligned}$$

at the point T is given by the formula

$$\begin{aligned}{}[{\textsf{D}}H](x,y)=H(xy)-\Psi (T(x),H(y))-\Psi (H(x),T(y))\qquad (H\in {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})) \end{aligned}$$

and, moreover,

$$\begin{aligned} \Vert (T+H)^{\,\vee }-T^{\,\vee }-{\textsf{D}}H\Vert \le \Vert \Psi \Vert \cdot \Vert H\Vert ^2\qquad (H\in {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})). \end{aligned}$$
(2.1)

Proof

It follows by a straightforward calculation:

$$\begin{aligned} \begin{aligned}{}[(T+H)^{\vee }-T^{\,\vee }](x,y)&=H(xy)-\Psi (T(x)+H(x),T(y)+H(y))+\Psi (T(x),T(y))\\&=H(xy)-\Psi (T(x),H(y))-\Psi (H(x),T(y))-r(H)(x,y), \end{aligned} \end{aligned}$$

where \(r(H)(x,y)=\Psi (H(x),H(y))\) and hence \(\Vert r(H)\Vert \le \Vert \Psi \Vert \cdot \Vert H\Vert ^2\). \(\square \)

In what follows, we use standard notation and facts concerning projective tensor products of Banach spaces, as described e.g. in [16, Ch. 2]. By \({\mathcal {A}}{\hat{\otimes }}{\mathcal {A}}\) we denote the completion of \({\mathcal {A}}\otimes {\mathcal {A}}\) in the projective norm. Recall that for every \(\Phi \in {\mathscr {L}}^2({\mathcal {A}},{\mathcal {B}})\) there exists a unique operator \(\Theta \in {\mathscr {L}}^1({\mathcal {A}}{\hat{\otimes }}{\mathcal {A}},{\mathcal {B}})\) such that \(\Theta (a\otimes b)=\Phi (a,b)\) for all \(a,b\in {\mathcal {A}}\) and, moreover, \(\Vert \Phi \Vert =\Vert \Theta \Vert \). From now on, we adopt the convention that for any \(\Phi \in {\mathscr {L}}^2({\mathcal {A}},{\mathcal {B}})\), we write \(\widetilde{\Phi }\) for the corresponding linear operator from \({\mathcal {A}}{\hat{\otimes }}{\mathcal {A}}\) to \({\mathcal {B}}\). Let also \(\pi _{\mathcal {A}}:{\mathcal {A}}{\hat{\otimes }}{\mathcal {A}}\rightarrow {\mathcal {A}}\) stand for the unique bounded linear map such that \(\pi _{\mathcal {A}}(a\otimes b)=ab\) for all \(a,b\in {\mathcal {A}}\).

Let us now recall some elementary facts concerning amenable Banach algebras; for a nice exposition of this theory, see [14, 15]. A net \((\Delta _\alpha )_{\alpha \in A}\subset {\mathcal {A}}{\hat{\otimes }}{\mathcal {A}}\) is called an approximate diagonal, provided that

$$\begin{aligned} \lim _\alpha (a\cdot \Delta _\alpha -\Delta _\alpha \cdot a)=0\quad \text{ and } \quad \lim _\alpha a\pi _{\mathcal {A}}(\Delta _\alpha )=a\quad \,\,(a\in {\mathcal {A}}). \end{aligned}$$

Note that in the case \({\mathcal {A}}\) is unital the latter condition means that \(\lim _\alpha \pi _{\mathcal {A}}(\Delta _\alpha )=1\) in norm. The approximate diagonal \((\Delta _\alpha )_{\alpha \in A}\) is called bounded if \(\sup _\alpha \Vert \Delta \Vert _\alpha <\infty \). Notice that, in our convention, \(\pi _{\mathcal {A}}=\widetilde{\Lambda }\), where \(\Lambda (a,b)=ab\). Therefore, \(\Vert \pi _{\mathcal {A}}\Vert =\Vert \Lambda \Vert =1\) and since \(\lim _\alpha \pi _{\mathcal {A}}(\Delta _\alpha )=1\), we must have \(\sup _\alpha \Vert \Delta _\alpha \Vert \ge 1\). A Banach algebra \({\mathcal {A}}\) is called amenable if there exists a bounded approximate diagonal \((\Delta _\alpha )_{\alpha \in A}\) for \({\mathcal {A}}\).

Concerning the codomain algebra in our stability problem, we need the following definition: A Banach algebra \({\mathcal {B}}\) is called a dual Banach algebra with isometric predual X, provided that X is a Banach space such that \({\mathcal {B}}\) and \(X^*\) are isometrically isomorphic and multiplication in \({\mathcal {B}}\) is separately \(\sigma ({\mathcal {B}},X)\)-continuous. In particular, \({\mathcal {B}}\) is equipped with the weak\(^*\) topology and the condition of the continuity of multiplication is equivalent to saying that \(i^*\circ \kappa (X)\) is a sub-\({\mathcal {B}}\)-bimodule of \({\mathcal {B}}^*\), where \(i:{\mathcal {B}}\rightarrow X^*\) is the said linear isometry and \(\kappa :X\rightarrow X^{**}\) is the canonical embedding (see [6, §2]).

It should be remarked that the original assumption made by Johnson was that the codomain algebra \({\mathcal {B}}\) is isomorphic, as a Banach \({\mathcal {B}}\)-bimodule, to the dual \(({\mathcal {B}}_*)^*\) of some Banach \({\mathcal {B}}\)-bimodule \({\mathcal {B}}_*\). Our choice follows the definition proposed in [4] which was influenced by [5, 6].

3 Proof of the main result

From now on, we fix Banach algebras \({\mathcal {A}}\) and \({\mathcal {B}}\) satisfying the assumptions of Theorem 1, as well as an isometric predual \({\mathcal {B}}_*\) of \({\mathcal {B}}\). Let \((\Delta _\alpha )_{\alpha \in A}\subset {\mathcal {A}}{\hat{\otimes }}{\mathcal {A}}\) be a bounded approximate diagonal for \({\mathcal {A}}\) with

$$\begin{aligned} \Delta _\alpha =\sum _j a_{\alpha ,j}\otimes b_{\alpha ,j}\qquad (\alpha \in A). \end{aligned}$$

All these are finite sums yet we do not need to indicate the sets over which we sum. The following observation is due to Johnson (see also [4, Lemma 5.11]); we include a short proof for the sake of completeness.

Lemma 4

For any \(R\in {\mathscr {L}}^2({\mathcal {A}},{\mathcal {B}})\), the limit

$$\begin{aligned} \lim _\alpha \sum _j R(a_{\alpha ,j},b_{\alpha ,j}) \end{aligned}$$

exists in the weak\(^*\) topology \(\sigma ({\mathcal {B}},{\mathcal {B}}_*)\) on \({\mathcal {B}}\), and it is bounded by \(\sup _\alpha \Vert \Delta _\alpha \Vert \Vert R\Vert \).

Proof

Since \(\sup _\alpha \Vert \Delta _\alpha \Vert <\infty \), using the Banach–Alaoglu theorem and passing to a subnet we may assume that \(\kappa _{{\mathcal {A}}{\hat{\otimes }}{\mathcal {A}}}(\Delta _\alpha )\) converges in the weak\(^*\) topology

to some \({\textsf{M}}\in ({\mathcal {A}}{\hat{\otimes }}{\mathcal {A}})^{**}\), and then \(\Vert {\textsf{M}}\Vert \le \sup _\alpha \Vert \Delta _\alpha \Vert \). Thus, by the weak\(^*\)-continuity of \(\kappa _{{\mathcal {B}}_*}^*\) and \((\widetilde{R})^{**}\), the net

$$\begin{aligned} \begin{aligned} \sum _j R(a_{\alpha ,j},b_{\alpha ,j})&=\sum _j \widetilde{R}(a_{\alpha ,j}\otimes b_{\alpha ,j})=\widetilde{R}(\Delta _\alpha )\\&=\kappa _{{\mathcal {B}}_*}^*(\widetilde{R})^{**}\kappa _{{\mathcal {A}}{\hat{\otimes }}{\mathcal {A}}}(\Delta _\alpha ) \end{aligned} \end{aligned}$$

converges in the weak\(^*\) topology to \(\kappa _{{\mathcal {B}}_*}^*(\widetilde{R})^{**}({\textsf{M}})\), and is bounded by \(\sup _\alpha \Vert \Delta _\alpha \Vert \Vert (\widetilde{R})^{**}\Vert =\sup _\alpha \Vert \Delta _\alpha \Vert \Vert R\Vert \). Note also that for any weak\(^*\) convergent net \((y_i)_{i\in I}\subset {\mathcal {B}}\) we have \(\Vert \!\lim _i y_i\Vert \le \limsup _i\Vert y_i\Vert \), hence the upper estimate by \(\sup _\alpha \Vert \Delta _\alpha \Vert \Vert R\Vert \) holds true for our weak\(^*\) limit. \(\square \)

Similarly as in Johnson’s method (see the proof of [11, Thm. 3.1]), we define an operator \({\textsf{J}}:{\mathscr {L}}^2({\mathcal {A}},{\mathcal {B}})\rightarrow {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\) by

$$\begin{aligned} {\textsf{J}}R(x)=\lim _\alpha \sum _j \Psi (T(a_{\alpha ,j}),R(b_{\alpha ,j},x)). \end{aligned}$$
(3.1)

This definition is correct in view of Lemma 4, and we have \(\Vert {\textsf{J}}R\Vert \le \sup _\alpha \Vert \Delta _\alpha \Vert \Vert T\Vert \Vert R\Vert \Vert \Psi \Vert \). We will use this observation several times in the sequel. Obviously, \({\textsf{J}}R\) depends on the operator T which we will not indicate explicitely, as it should be obvious from the context.

We shall show that formula (3.1) yields an approximate right inverse of the derivative \({\textsf{D}}\) given in Lemma 3 on the set of operators ‘almost’ belonging to \(\textrm{ker}\,\delta _T^2\). In order to simplify writing let us denote \(u\circ v=\Psi (u,v)\).

Proposition 5

Assume that T is unital and \(\Psi (1,u)=u\) for each \(u\in {\mathcal {B}}\). For every \(R\in {\mathscr {L}}^2({\mathcal {A}},{\mathcal {B}})\), we have

$$\begin{aligned} \begin{aligned} \Vert R+{\textsf{D}}{\textsf{J}}R\Vert \le \big (&2\!\cdot \!{\textsf{m}}\text{- }\text {def}_\Psi (T)\cdot \!\Vert \Psi \Vert \!\cdot \!\Vert R\Vert \\&+\Vert \psi \Vert \!\cdot \!\Vert T\Vert \!\cdot \!\Vert \delta _T^2 R\Vert +3\!\cdot \!{\textsf{a}}\text{- }\text {def}(\Psi )\!\cdot \!\Vert T\Vert ^2\!\cdot \!\Vert R\Vert \big )\!\cdot \!\sup _\alpha \Vert \Delta _\alpha \Vert . \end{aligned}\nonumber \\ \end{aligned}$$
(3.2)

Proof

First, observe that since \(\lim _\alpha \pi _{\mathcal {A}}(\Delta _\alpha )=1\), we have \(\lim _\alpha \sum _j a_{\alpha ,j}b_{\alpha ,j}=1\) in norm and hence \(\lim _\alpha \sum _j T(a_{\alpha ,j}b_{\alpha ,j})=1\). For any \(x,y\in {\mathcal {A}}\), using the identity \(\Psi (1,u)=u\), Lemma 3, and applying Lemma 4 to the bilinear maps \((a,b)\mapsto T(x)\circ (T(a)\circ R(b,y))\) and \((a,b)\mapsto (T(a)\circ R(b,x))\circ T(y)\), we thus obtain

$$\begin{aligned} \begin{aligned}{}[R+{\textsf{D}}{\textsf{J}}R](x,y)&=R(x,y)+[{\textsf{J}}R](xy)-T(x)\circ [{\textsf{J}}R](y)-[{\textsf{J}}R](x)\circ T(y)\\&=\lim _\alpha \sum _j\Big \{T(a_{\alpha ,j}b_{\alpha ,j})\circ R(x,y)+T(a_{\alpha ,j})\circ R(b_{\alpha ,j},xy)\\&\quad -T(x)\circ \big (T(a_{\alpha ,j})\circ R(b_{\alpha ,j},y)\big )\\&\quad -\big (T(a_{\alpha ,j})\circ R(b_{\alpha ,j},x)\big )\circ T(y)\Big \}. \end{aligned} \end{aligned}$$
(3.3)

Since

$$\begin{aligned} R(b_{\alpha ,j},xy)=\delta _T^2\,R(b_{\alpha ,j},x,y)-T(b_{\alpha ,j})\circ R(x,y)+R(b_{\alpha ,j}x,y)+R(b_{\alpha ,j},x)\circ T(y), \end{aligned}$$

we have

$$\begin{aligned} \begin{aligned}&\Big \Vert \lim _\alpha \sum _j\Big \{T( a_{\alpha ,j})\circ R(b_{\alpha ,j},xy)-T(a_{\alpha ,j})\circ R(b_{\alpha ,j}x,y)\\&\qquad +T(a_{\alpha ,j})\circ \big (T(b_{\alpha ,j})\circ R(x,y)\big )-T(a_{\alpha ,j})\circ \big (R(b_{\alpha ,j},x)\circ T(y)\big )\Big \}\Big \Vert \\&\quad \le \Vert \Psi \Vert \!\cdot \!\Vert T\Vert \!\cdot \!\Vert \delta _T^2 R\Vert \!\cdot \!\Vert x\Vert \Vert y\Vert \sup _\alpha \Vert \Delta _\alpha \Vert . \end{aligned} \end{aligned}$$

This estimate follows by considering the bilinear operator \(\Phi _1:{\mathcal {A}}\times {\mathcal {A}}\rightarrow {\mathcal {B}}\) given by

$$\begin{aligned} \Phi _1(a,b)=T(a)\circ \delta _T^2 R(b,x,y) \end{aligned}$$

for which \(\Vert \Phi _1\Vert \le \Vert \Psi \Vert \!\cdot \!\Vert T\Vert \!\cdot \!\Vert \delta _T^2 R\Vert \!\cdot \!\Vert x\Vert \Vert y\Vert \), thus

$$\begin{aligned}{} & {} \Big \Vert \lim _\alpha \sum _j T(a_{\alpha ,j})\circ \delta _T^2 R(b_{\alpha ,j},x,y)\Big \Vert \\{} & {} \quad \le \Vert \Psi \Vert \!\cdot \!\Vert T\Vert \!\cdot \!\Vert \delta _T^2 R\Vert \!\cdot \!\Vert x\Vert \Vert y\Vert \sup _\alpha \Vert \Delta _\alpha \Vert =:\!\eta _1(x,y). \end{aligned}$$

Coming back to (3.3) we see that \([R+{\textsf{D}}{\textsf{J}}R](x,y)\) is at distance at most \(\eta _1(x,y)\) from

$$\begin{aligned} \begin{aligned}&S(x,y){:}{=}\lim _\alpha \sum _j \Big \{T( a_{\alpha ,j}b_{\alpha ,j})\circ R(x,y)-T(a_{\alpha ,j})\circ \big (T(b_{\alpha ,j})\circ R(x,y)\big )\\&\qquad +T(a_{\alpha ,j})\circ \big (R(b_{\alpha ,j},x)\circ T(y)\big )-\big (T(a_{\alpha ,j})\circ R(b_{\alpha ,j},x)\big )\circ T(y)\\&\qquad +T(a_{\alpha ,j})\circ R(b_{\alpha ,j}x,y)-T(x)\circ \big (T(a_{\alpha ,j})\circ R(b_{\alpha ,j},y)\big )\Big \}. \end{aligned} \end{aligned}$$

Considering the bilinear operator \(\Phi _2:{\mathcal {A}}\times {\mathcal {A}}\rightarrow {\mathcal {B}}\) given by

$$\begin{aligned} \Phi _2(a,b)=T(a)\circ (T(b)\circ R(x,y))-(T(a)\circ T(b))\circ R(x,y), \end{aligned}$$

we have \(\Vert \Phi _2\Vert \le {\textsf{a}}\text{- }\text {def}(\Psi )\!\cdot \!\Vert T\Vert ^2\!\cdot \!\Vert R\Vert \!\cdot \!\Vert x\Vert \Vert y\Vert \). Hence, the difference in the first line of the definition of S(xy) can be estimated as follows:

$$\begin{aligned} \begin{aligned}&\Big \Vert \lim _\alpha \sum _j \Big \{ T(a_{\alpha ,j}b_{\alpha ,j})\circ R(x,y)-T(a_{\alpha ,j})\circ \big (T(b_{\alpha ,j})\circ R(x,y)\big )\Big \}\Big \Vert \\&\quad \le \Big \Vert \lim _\alpha \sum _j \Big \{\big (T(a_{\alpha ,j}b_{\alpha ,j})-T(a_{\alpha ,j})\circ T(b_{\alpha ,j})\big )\circ R(x,y)\Big \}\Big \Vert +\eta _2(x,y), \end{aligned} \end{aligned}$$
(3.4)

where

$$\begin{aligned} \eta _2(x,y){:}{=}{\textsf{a}}\text{- }\text {def}(\Psi )\!\cdot \!\Vert T\Vert ^2\!\cdot \!\Vert R\Vert \!\cdot \!\Vert x\Vert \Vert y\Vert \sup _\alpha \Vert \Delta _\alpha \Vert . \end{aligned}$$

Similarly, for the difference in the second line, we have

$$\begin{aligned}{} & {} \Big \Vert \lim _\alpha \sum _j \Big \{T(a_{\alpha ,j})\circ \big (R(b_{\alpha ,j},x)\circ T(y)\big )-\big (T(a_{\alpha ,j})\circ R(b_{\alpha ,j},x)\big )\circ T(y)\Big \}\Big \Vert \nonumber \\{} & {} \quad \le \eta _2(x,y). \end{aligned}$$
(3.5)

Finally, in order to estimate the difference in the third line of the definition of S(xy), consider the bilinear map \(\Phi _3:{\mathcal {A}}\times {\mathcal {A}}\rightarrow {\mathcal {B}}\) defined by \(\Phi _3(a,b)=T(a)\circ R(b,y)\). By the diagonal property, \(\lim _\alpha (x\cdot \Delta _\alpha -\Delta _\alpha \cdot x)=0\) in norm, therefore

$$\begin{aligned} 0=\lim _\alpha \widetilde{\Phi }_3(x\cdot \Delta _\alpha -\Delta _\alpha \cdot x)=\lim _\alpha \Big (\sum _j\Phi _3(xa_{\alpha ,j},b_{\alpha ,j})-\sum _j\Phi _3(a_{\alpha ,j},b_{\alpha ,j}x)\Big ) \end{aligned}$$

in norm, so also in the weak\(^*\) topology. By Lemma 4 applied to the bilinear maps \((a,b)\mapsto \Phi _3(xa,b)\) and \((a,b)\mapsto \Phi _3(a,bx)\), we infer that both \(\lim _\alpha \sum _j\Phi _3(xa_{\alpha ,j},b_{\alpha ,j})\) and \(\lim _\alpha \sum _j\Phi _3(a_{\alpha ,j},b_{\alpha ,j}x)\) exist in the weak\(^*\) sense, thus

$$\begin{aligned} \lim _\alpha \sum _j\Phi _3(xa_{\alpha ,j},b_{\alpha ,j})=\lim _\alpha \sum _j\Phi _3(a_{\alpha ,j},b_{\alpha ,j}x) \end{aligned}$$

in the weak\(^*\) sense. This means that the norm of the difference in the third line defining S(xy) equals

$$\begin{aligned} \begin{aligned}&\Big \Vert \lim _\alpha \sum _j \Big \{T(xa_{\alpha ,j})\circ R(b_{\alpha ,j},y)-T(x)\circ \big (T(a_{\alpha ,j})\circ R(b_{\alpha ,j},y)\big )\Big \}\Big \Vert \\&\quad \le \Big \Vert \lim _\alpha \sum _j \Big \{T(xa_{\alpha ,j})\circ R(b_{\alpha ,j},y)-\big (T(x)\circ T(a_{\alpha ,j})\big )\circ R(b_{\alpha ,j},y)\Big \}\Big \Vert \\&\qquad +\eta _2(x,y). \end{aligned} \end{aligned}$$
(3.6)

Combining (3.4), (3.5) and (3.6) we obtain

$$\begin{aligned} \begin{aligned} \Vert [R+{\textsf{D}}{\textsf{J}}R](x,y)\Vert&\le \Big \Vert \lim _\alpha \sum _j \Big \{\big (T(a_{\alpha ,j}b_{\alpha ,j})-T(a_{\alpha ,j})\circ T(b_{\alpha ,j})\big )\circ R(x,y)\Big \}\Big \Vert \\&\qquad +\Big \Vert \lim _\alpha \sum _j \Big \{T(xa_{\alpha ,j})\circ R(b_{\alpha ,j},y)-\big (T(x)\circ T(a_{\alpha ,j})\big )\circ R(b_{\alpha ,j},y)\Big \}\Big \Vert \\&\qquad +\eta _1(x,y)+3\eta _2(x,y)\\&\quad \le 2\!\cdot \!{\textsf{m}}\text{- }\text {def}_\Psi (T)\cdot \!\Vert \Psi \Vert \!\cdot \!\Vert R\Vert \!\cdot \!\Vert x\Vert \Vert y\Vert \sup _\alpha \Vert \Delta _\alpha \Vert +\eta _1(x,y)+3\eta _2(x,y), \end{aligned} \end{aligned}$$

which gives the desired estimate (3.2). \(\square \)

Corollary 6

Assume that T is unital and \(\Psi (1,u)=u\) for each \(u\in {\mathcal {B}}\). Let also \((\Delta _\alpha )_{\alpha \in A}\) be a bounded approximate diagonal in \({\mathcal {A}}{\hat{\otimes }}{\mathcal {A}}\) with \(M=\sup _\alpha \Vert \Delta _\alpha \Vert <\infty \). Then, the operator \({\textsf{J}}:{\mathscr {L}}^2({\mathcal {A}},{\mathcal {B}})\rightarrow {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\) defined by (3.1) satisfies

$$\begin{aligned} \begin{aligned}&\Vert T^{\,\vee }+{\textsf{D}}{\textsf{J}}\,T^{\,\vee }\Vert \le \big ( 2\!\cdot \!({\textsf{m}}\text{- }\text {def}_\Psi (T))^2\Vert \Psi \Vert \\&\quad +3\!\cdot \!{\textsf{a}}\text{- }\text {def}(\Psi )\!\cdot \!{\textsf{m}}\text{- }\text {def}_\Psi (T)\!\cdot \!\Vert T\Vert ^2+{\textsf{a}}\text{- }\text {def}(\Psi )\!\cdot \!\Vert \psi \Vert \!\cdot \!\Vert T\Vert ^4\big ) M. \end{aligned} \end{aligned}$$
(3.7)

Proof

We simply apply Proposition 5 to \(R=T^{\,\vee }\) and appeal to Lemma 2.

\(\square \)

Following Johnson’s idea, for any fixed bounded approximate diagonal \((\Delta _\alpha )_{\alpha \in A}\subset {\mathcal {A}}{\hat{\otimes }}{\mathcal {A}}\), we define the ‘improving operator’ \({\mathcal {F}}:{\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\rightarrow {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\) by

$$\begin{aligned} {\mathcal {F}}\,T=T+{\textsf{J}}T^{\,\vee }, \end{aligned}$$
(3.8)

where \({\textsf{J}}\) is given by formula (3.1).

Proposition 7

Under the same assumptions as in Corollary 6, there exist constants \(C_i=C_i(\Vert \Psi \Vert ,\Vert T\Vert ,M)\ge 0\), for \(i=1,2,3\), such that

$$\begin{aligned} \begin{aligned} {\textsf{m}}\text{- }\text {def}_\Psi ({\mathcal {F}}T)\le C_1(&{\textsf{m}}\text{- }\text {def}_\Psi (T))^2\\&+C_2\!\cdot \!{\textsf{a}}\text{- }\text {def}(\Psi )\!\cdot \!{\textsf{m}}\text{- }\text {def}_\Psi (T)+C_3\!\cdot \!{\textsf{a}}\text{- }\text {def}(\Psi ). \end{aligned} \end{aligned}$$

Moreover, these constant are given by the formulas

$$\begin{aligned} \left\{ \begin{array}{rcl} C_1(\Vert \Psi \Vert ,\Vert T\Vert ,M) &{} = &{} \Vert \Psi \Vert ^3\!\cdot \!\Vert T\Vert ^2\!\cdot \! M^2+2\Vert \Psi \Vert \!\cdot \! M\\ C_2(\Vert \Psi \Vert ,\Vert T\Vert ,M) &{} = &{} 3\Vert T\Vert ^2\!\cdot \! M\\ C_3(\Vert \Psi \Vert ,\Vert T\Vert ,M) &{} = &{} \Vert \Psi \Vert \!\cdot \!\Vert T\Vert ^4\!\cdot \! M. \end{array}\right. \end{aligned}$$
(3.9)

Proof

First, observe that considering the bilinear map \(\Phi :{\mathcal {A}}\times {\mathcal {A}}\rightarrow {\mathcal {B}}\) given by \(\Phi (a,b)=T(a)\circ T^{\,\vee }(b,x)\), for any fixed \(x\in {\mathcal {A}}\), by Lemma 4, we have

$$\begin{aligned} \Vert \Phi \Vert \le {\textsf{m}}\text{- }\text {def}_\Psi (T)\!\cdot \! \Vert \Psi \Vert \!\cdot \!\Vert T\Vert \Vert x\Vert , \end{aligned}$$

hence

$$\begin{aligned} \Vert {\textsf{J}}T^{\,\vee }(x)\Vert =\lim _\alpha \Vert \widetilde{\Phi }(\Delta _\alpha )\Vert \le {\textsf{m}}\text{- }\text {def}_\Psi (T)\!\cdot \! \Vert \Psi \Vert \!\cdot \!\Vert T\Vert \!\cdot \! M\Vert x\Vert . \end{aligned}$$
(3.10)

Using (3.10) in combination with Lemma 3, in particular estimate (2.1), as well as Corollary 6, we get

$$\begin{aligned} \begin{aligned} {\textsf{m}}\text{- }\text {def}_\Psi ({\mathcal {F}}T)&=\Vert (T+{\textsf{J}}T^{\,\vee })^{\,\vee }\Vert \\&\le \Vert (T+{\textsf{J}}T^{\,\vee })^{\,\vee }-T^{\,\vee }-{\textsf{D}}{\textsf{J}}T^{\,\vee }\Vert +\Vert T^{\,\vee }+{\textsf{D}}{\textsf{J}}T^{\,\vee }\Vert \\&\le \Vert \Psi \Vert \!\cdot \!\Vert {\textsf{J}}T^{\,\vee }\Vert ^2+\Vert T^{\,\vee }+{\textsf{D}}{\textsf{J}}T^{\,\vee }\Vert \\&\le C_1 ({\textsf{m}}\text{- }\text {def}_\Psi (T))^2+C_2\!\cdot \!{\textsf{a}}\text{- }\text {def}(\Psi )\!\cdot \!{\textsf{m}}\text{- }\text {def}_\Psi (T)+C_3\!\cdot \!{\textsf{a}}\text{- }\text {def}(\Psi ), \end{aligned} \end{aligned}$$

where \(C_1\), \(C_2\) and \(C_3\) are given as in (3.9). \(\square \)

The proof of our main result relies on an approximation procedure in which a given operator \(T_n\) is replaced by its ‘improved’ version \(T_n+{\mathcal {F}}T_n\). We iterate this process as long as the obtained multiplicative defects are roughly larger than the associative defect of \(\Psi \). This terminates at some point, unless we can continue and obtain an exact solution of equation (1.1).

Proof of Theorem 1

Fix any \(K,L\ge 1\), \(\varepsilon , \theta \in (0,1)\), pick a bounded approximate diagonal \((\Delta _\alpha )_{\alpha \in A}\) in \({\mathcal {A}}{\hat{\otimes }}{\mathcal {A}}\) with \(M=\sup _\alpha \Vert \Delta _\alpha \Vert <\infty \). Define \(\delta \in (0,1)\) by the formula

$$\begin{aligned} \delta =\Big [2\big (2LM+e^{4LM}(L^3M^2+3M)K^2+e^{8LM}LMK^4\big )\Big ]^{-1/\theta }\mathbf {\varepsilon }. \end{aligned}$$
(3.11)

Fix any unital operator \(T\in {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\) with \(\Vert T\Vert \le K\) and \({\textsf{m}}\text{- }\text {def}_\Psi (T)\le \delta \). Assume also that \(\Vert \Psi \Vert \le L\) and define \(\alpha ={\textsf{a}}\text{- }\text {def}(\Psi )\). Let \({\mathcal {F}}:{\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\rightarrow {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\) be the ‘improving operator’ defined by (3.8). We define a sequence \((T_n)_{n=0}^\infty \subset {\mathscr {L}}^1({\mathcal {A}},{\mathcal {B}})\) recursively by

$$\begin{aligned} T_0=T\quad \,\,\text{ and }\,\,\quad T_n={\mathcal {F}}T_{n-1}\qquad \text{ for } n\in {\mathbb {N}}. \end{aligned}$$

Suppose that \(\alpha \le ({\textsf{m}}\text{- }\text {def}_\Psi (T))^{1+\theta }\). We define \(N\in {\mathbb {N}}\cup \{\infty \}\) as the least natural number n for which

$$\begin{aligned} \alpha >({\textsf{m}}\text{- }\text {def}_\Psi (T_n))^{1+\theta }, \end{aligned}$$

provided that such an n exists, and we set \(N=\infty \) otherwise. Define \((\omega _n)_{n=0}^\infty \) recursively by \(\omega _0=0\) and \(\omega _n=1+(1+\theta )\omega _{n-1}\) for \(n\in {\mathbb {N}}\). For any \(n\in {\mathbb {N}}_0\), we also define

$$\begin{aligned} \delta _n=2^{-\omega _n}\delta \quad \text{ and } \quad \beta _n=\prod _{j=0}^{n-1}(1+LM\delta _j). \end{aligned}$$

Observe that since \(\sum _{j=0}^\infty \delta _j<\infty \), the product \(\prod _{j=0}^\infty (1+LM\delta _j)\) converges, hence we have \(\beta _n\nearrow B\) for some \(B<\infty \). In fact, since \(\omega _n\ge n\) for \(n\in {\mathbb {N}}_0\), we have

$$\begin{aligned} B\le \prod _{j=0}^\infty (1+2^{-j}LM\delta )\le \exp \Big \{\sum _{j=0}^\infty 2^{-j}LM\Big \}=\exp (2LM). \end{aligned}$$
(3.12)

Set \(K_n=\beta _n K\), so that we have \(K_n<BK\) for each \(n\in {\mathbb {N}}_0\).

\(\underline{{Claim.}}\) For each integer \(0\le n<N\), the following conditions hold true:

  1. (i)

    \(T_n\) is unital;

  2. (ii)

    \({\textsf{m}}\text{- }\text {def}_\Psi (T_n)\le \delta _n\);

  3. (iii)

    \(\Vert T_n\Vert \le K_n\).

We proceed by induction. In order to show (i) notice that the identity \(\Psi (u,1)=u\) implies \(T^{\,\vee }(b,1)=0\) for every \(b\in {\mathcal {B}}\). Hence, \({\textsf{J}} T^{\,\vee }(1)=0\) which gives \(T_1(1)={\mathcal {F}}T(1)=T(1)=1\) and, by induction, \(T_n(1)=1\) for every \(n\in {\mathbb {N}}\).

Assertions (ii) and (iii) are plainly true for \(n=0\) as \(\delta _0=\delta \) is not smaller than the defect of \(T=T_0\), and \(K_0=\beta _0K=K\). Fix \(n\in {\mathbb {N}}_0\) with \(n+1<N\) and suppose that estimates (ii) and (iii) hold true. Using Proposition 7, formulas (3.9), and the fact that \(\alpha \le ({\textsf{m}}\text{- }\text {def}_\Psi (T_n))^{1+\theta }\), we obtain

$$\begin{aligned} \begin{aligned} {\textsf{m}}\text{- }\text {def}_\Psi (T_{n+1})&={\textsf{m}}\text{- }\text {def}_\Psi ({\mathcal {F}}T_n)\\&\quad \le C_1(\Vert \Psi \Vert ,\Vert T_n\Vert ,M)\delta _n^2+C_2(\Vert \Psi \Vert ,\Vert T_n\Vert ,M)\alpha \delta _n\\&\quad +C_3(\Vert \Psi \Vert ,\Vert T_n\Vert ,M)\alpha \\&\quad \le \big (C_1(\Vert \Psi \Vert ,\Vert T_n\Vert ,M)+C_2(\Vert \Psi \Vert ,\Vert T_n\Vert ,M)\\&\quad +C_3(\Vert \Psi \Vert ,\Vert T_n\Vert ,M)\big )\delta _n^{1+\theta }\\&\quad \le \big (2LM+K_n^2L^3M^2+3K_n^2M+K_n^4LM\big )\delta _n^{1+\theta }\\&\quad \le \big (2LM+(L^3M^2+3M)B^2K^2+B^4K^4LM\big )\delta _n^{1+\theta }. \end{aligned} \end{aligned}$$

Therefore, appealing to the definition of \(\delta \), that is, formula (3.11), and inequality (3.12), we obtain

$$\begin{aligned} {\textsf{m}}\text{- }\text {def}_\Psi (T_{n+1})\le \frac{1}{2}\delta ^{-\theta }\delta _n^{1+\theta }=\frac{1}{2}\!\cdot \!2^{-(1+\theta )\omega _n}\delta =2^{-\omega _{n+1}}\delta =\delta _{n+1}. \end{aligned}$$

By inequality (3.10), we also obtain

$$\begin{aligned} \begin{aligned} \Vert T_{n+1}\Vert&=\Vert {\mathcal {F}}T_n\Vert \le \Vert T_n\Vert +\Vert {\textsf{J}}T_n^{\,\vee }\Vert \\&\le K_n+\delta _n K_nLM\\&=\beta _n(1+\delta _n LM)K=\beta _{n+1}K=K_{n+1}, \end{aligned} \end{aligned}$$

which completes the induction.

Now, suppose that the inequality \(({\textsf{m}}\text{- }\text {def}_\Psi (T_n))^{1+\theta }<\alpha \) is never satisfied, which means that our process does not terminate and we have \(N=\infty \). Then, since for each \(n\in {\mathbb {N}}\) we have

$$\begin{aligned} \Vert {\textsf{J}}T_n^{\,\vee }\Vert \le \delta _n K_nLM<\delta _n BKLM\le 2^{-n}\delta BKLM, \end{aligned}$$

the sequence \((T_n)_{n=1}^\infty \) is a Cauchy sequence, thus the norm limit \(S=\lim _{n\rightarrow \infty }T_n\) exists. In view of (ii), we have \(\lim _{n\rightarrow \infty }{\textsf{m}}\text{- }\text {def}_\Psi (T_n)=0\) which yields \({\textsf{m}}\text{- }\text {def}_\Psi (S)=0\) because for any \(x,y\in {\mathcal {A}}\) with \(\Vert x\Vert ,\Vert y\Vert \le 1\), we have

$$\begin{aligned} \Vert S(xy)-\Psi (S(x),S(y))\Vert= & {} \lim _{n\rightarrow \infty }\Vert T_n(xy)-\Psi (T_n(x),T_n(y))\Vert \\\le & {} \lim _{n\rightarrow \infty }{\textsf{m}}\text{- }\text {def}_\Psi (T_n)=0. \end{aligned}$$

Moreover, \(S-T=\sum _{n=0}^\infty {\textsf{J}}T_n^{\,\vee }\) and hence

$$\begin{aligned} \Vert S-T\Vert \le \sum _{n=0}^\infty 2^{-n}\delta BKLM\le 2\delta e^{2LM}KLM<\varepsilon , \end{aligned}$$

which means that in this case we have produced an exact solution of the equation \(S(xy)=S(x)\circ S(y)\) lying at distance smaller than \(\varepsilon \) from T.

If our process terminates, i.e. \(N<\infty \), then we have

$$\begin{aligned} {\textsf{m}}\text{- }\text {def}_\Psi (T_N)<\alpha ^{\frac{1}{1+\theta }} \end{aligned}$$

and, likewise in the previous case,

$$\begin{aligned} \Vert T_N-T\Vert \le \sum _{n=0}^{N-1}\Vert {\textsf{J}}T_n^{\,\vee }\Vert <\varepsilon . \end{aligned}$$

This completes the proof since for any \(\eta \in (0,1)\) we can pick \(\theta \in (0,1)\) so that \((1+\theta )^{-1}>1-\eta \) and then the parameter \(\delta \) defined by (3.11) does the job in the sense that whenever T is as above, then the approximation \(T_N\) satisfies \({\textsf{m}}\text{- }\text {def}_\Psi (T_N)<{\textsf{a}}\text{- }\text {def}(\Psi )^{1-\eta }\).

Remark

It may be of some interest to ask if the power function \(\alpha \mapsto \alpha ^{1-\eta }\) appearing in Theorem 1 is optimal and, in general, how close to \({\textsf{a}}\text{- }\text {def}(\Psi )\) one can get with the multiplicative defect \({\textsf{m}}\text{- }\text {def}_\Psi (S)\).