Generalized inverses of strictly monotone transformationsThanks:  The research has been supported by the EKÖP-25-0 University Research Scholarship Program of the Ministry for Culture and Innovation from the source of the National Research, Development and Innovation Fund, by the PhD Excellence Scholarship from the Count István Tisza Foundation for the University of Debrecen and by the University of Debrecen Program for Scientific Publication.

Abstract.

Let $C\subseteq\mathbb{R}^{n}$ be a convex set. The mapping $f:C\longrightarrow\mathbb{R}^{n}$ is strictly increasing, if $\langle f(x)-f(y),x-y\rangle>0$ for all distinct elements $x,y\in C$. Applying classical theorems of finite dimensional convex geometry and convex analysis, we show that the inverse functions has a unique extension $f^{(-1)}:\mbox{\rm conv}(f(C))\longrightarrow C$ such that $f^{(-1)}$ is monotone, continuous and it acts as a left-inverse of $f$. As an application we introduce the concept of vector-valued weighted quasi-arithmetic means and discuss their equality problem.

1. Introduction

It is an elementary fact that if $I\subseteq\mathbb{R}$ is an interval and $f:I\longrightarrow\mathbb{R}$ is a continuous, strictly monotone function, then $f(I)$ is again an interval, and the inverse function $f^{-1}:f(I)\longrightarrow I$ is also strictly monotone (in the same sense) and continuous. When $f$ is not continuous, the ordinary inverse $f^{-1}$ is still strictly monotone, but its domain of definition is not an interval. However, it is known that $f^{-1}$ can be extended to the convex hull of $f(I)$ such that this extension is still monotone, what is more, continuous. This result is proved explicitly in the paper [10] by Grünwald and Páles and it also appears in the textbook [9]. A famous particular case is the quantile function of a strictly increasing CDF of a random variable.

Proposition 1.

[10, Lemma 1] Let $I\subseteq\mathbb{R}$ be an interval and let $f:I\longrightarrow\mathbb{R}$ be a strictly monotone function. Then there exists a uniquely determined monotone function $f^{(-1)}:\mbox{\rm conv}\left(f(I)\right)\longrightarrow I$ such that

$$f^{(-1)}\left(f(x)\right)=x\hskip 28.45274pt\left(\,x\in I\,\right).$$

Moreover, $f^{(-1)}$ is continuous.

During the 60th International Symposium on Functional Equations Zs. Páles proposed an Open Problem (see [1, 4. Problem]) about the generalization of the previous observation to higher dimensional spaces. Before formulating the precise question we clarify some notions and notations. Throughout the whole paper $n$ denotes a positive integer. We consider the $n$ dimensional euclidean space $\mathbb{R}^{n}$ equipped with the standard inner product, the induced norm and topology. The inner product of $x,y\in\mathbb{R}^{n}$ will be denoted by $\langle x,y\rangle$. For an arbitrary nonempty set $S\subseteq\mathbb{R}^{n}$ its convex hull is denoted by $\mbox{\rm conv}(S)$. For a detailed introduction to fundamental concepts of convex geometry we recommend the monographs of Lay [15] and Rockafellar [22].

Definition 1.

Let $K\subseteq\mathbb{R}^{n}$ be a nonempty, convex set. We say that $f:K\longrightarrow\mathbb{R}^{n}$ is an increasing mapping, if

$$\left\langle\,f(x)-f(y)\,,\,x-y\,\right\rangle\geq 0$$

for all $x,y\in K$. Similarly, $f$ is called a decreasing mapping if

$$\left\langle\,f(x)-f(y)\,,\,x-y\,\right\rangle\leq 0$$

for all $x,y\in K$. Moreover, if the above inequalities are strict for all elements $x,y\in K$ such that $x\neq y$, then we call these strictly increasing/strictly decreasing mappings, respectively.

A notable example for increasing mappings is the gradient map of a convex function of $n$ variables. In the $n=1$ dimensional case the notion is equivalent to the standard (strict) monotonicity of a real function $f$ defined on an interval, as the inner product is the ordinary multiplication. Notice that multiplying an increasing mapping by $(-1)$ we obtain a decreasing one. Thus, in the sequel we mainly focus on (strictly) increasing mappings. It is easy to see that if $f$ is strictly increasing, then it has to be injective.

Consequently, the ordinary inverse $f^{-1}$ of a strictly increasing mapping always exists. The main question of our paper (based on the open problem by Páles) is the following: Does there exist an extension of $f^{-1}$ to the convex hull of the image, which acts as a left-inverse and remains monotone? Does such an extended inverse have some additional properties such as continuity? More precisely, we consider the problem below.

Problem 1.

[1, 4. Problem by Zs. Páles] Let $\emptyset\neq K\subseteq\mathbb{R}^{n}$ be a convex set and let $f:K\longrightarrow\mathbb{R}^{n}$ be a strictly increasing mapping. Does there exist an increasing mapping $f^{(-1)}:\mbox{\rm conv}\left(f(K)\right)\longrightarrow K$ such that $f^{(-1)}\left(f(x)\right)=x$ for all $x\in K$, moreover

$$\langle\,f^{(-1)}(x)-f^{(-1)}(y)\,,\,x-y\,\rangle\geq 0$$

for all $x,y\in\mbox{conv}\left(f(K)\right)$? Can $f^{(-1)}$ be continuous?

We give a positive answer to these questions assuming that the domain is closed and convex. We demonstrate that the additional topological property of the domain is crucial, which is in contrast to the $1$-dimensional setting. A function $f^{(-1)}$ fulfilling the desired properties is going to be called the extended monotone left-inverse of $f$. When it is not confusing, we simply refer to this function as the generalized inverse of $f$.

In Section 4, as an application, we introduce the concept of vector-valued weighted quasi-arithmetic means defined on a closed, convex subset of $\mathbb{R}^{n}$. This is the natural generalization of the classical notion: the $m$-variable quasi-arithmetic mean (briefly, QAM) generated by the strictly increasing, continuous function $\varphi:I\longrightarrow I$ is

$$\varphi^{-1}\left(\frac{\varphi(x_{1})+\dots+\varphi(x_{m})}{m}\right)\hskip 28.45274pt\left(x_{1}\,,\dots,x_{m}\in I\right),$$

where $I\subseteq\mathbb{R}$ is an interval. Different characterizations of the QAM are due to Kolmogorov [13], Nagumo [18], De Finetti [7], Aczél [2], and recently Burai–Kiss–Szokol [5, 6]. We discuss the equality problem of weighted vector-valued QAMs for a fixed number of variables. The solution in the real valued setting is elaborated in the classical monograph of Hardy–Littlewood–Pólya [11].

Recent results about the equality problem of several generalizations of QAMs are contained in [4, 10, 12, 17, 21, 20]. These articles concern real variables. For concepts of vector valued means we shall mention the conference paper [19] of Nielsen following a differential geometric approach and the paper of Leonetti [16].

2. Main results

Our first main objective is to guarantee the existence of a generalized left-inverse for strictly increasing mappings defined on closed, convex domains. In order to construct the generalized inverse, we begin with establishing necessary conditions for the existence.

Proposition 2.

Let $\emptyset\neq K\subseteq\mathbb{R}^{n}$ be a convex set, $f:K\longrightarrow\mathbb{R}^{n}$ be a strictly increasing mapping, and assume that $f^{(-1)}:\mbox{\rm conv}\left(f(K)\right)\longrightarrow K$ is an extended monotone left-inverse of $f$. Then

(1) $$f^{(-1)}(z)\in\bigcap_{x\in K}\left\{s\in K:\left\langle x-s\,,\,f(x)-z\right\rangle\geq 0\right\}$$

for all $z\in\mbox{\rm conv}\left(f(K)\right)$.

Proof.

Since $f^{(-1)}$ is an increasing mapping, the inequality

$$\langle f^{(-1)}(w)-f^{(-1)}(z),w-z\rangle\geq 0$$

holds for all vectors $w,z\in\mbox{\rm conv}\left(f(K)\right)$. In particular, if $w\in f(K)$, that is, there exists $x\in K$ such that $w=f(x)$, then $\langle f^{(-1)}(f(x))-f^{(-1)}(z),f(x)-z\rangle\geq 0.$ Due to the left-inverse property, this means that

$$\langle x-f^{(-1)}(z),f(x)-z\rangle\geq 0$$

for any $x\in K$ and $z\in\mbox{\rm conv}(f(K))$. Therefore $f^{(-1)}(z)$ is indeed contained in the intersection. ∎

Let us observe that in the intersection (1) each member is convex set. Indeed, every set is obtained as the intersection of the convex set $K$ and the solution set of the inner product inequality, which is a closed half-space by definition. A powerful classical result about the intersection of convex sets of $\mathbb{R}^{n}$ is Helly’s Theorem. The theorem has numerous variations, thus we explicitly formulate the most suitable one.

Theorem 1 (Helly’s Theorem).

Let $\mathcal{C}$ be a family of nonempty closed convex sets in $\mathbb{R}^{n}$ and suppose that $\mathcal{C}$ contains at least $n+1$ members. If there exists a finite subfamily $\mathcal{C}_{0}\subseteq\mathcal{B}$ such that $\bigcap\mathcal{C}_{0}$ is bounded, moreover every subfamily of $n+1$ sets from $\mathcal{C}$ has a nonempty intersection, then $\bigcap\mathcal{C}\neq\emptyset$.

A slightly stronger version can be found in [22, Corollary 21.3.2]. The following Lemma of pure linear algebra will be important to ensure that the assumptions of Helly’s Theorem hold.

Lemma 1.

Let $n,d\in\mathbb{N}$ and let us assume that for the vectors $x_{1}\,,\dots,x_{d}\in\mathbb{R}^{n}$ and $y_{1}\,,\dots,y_{d}\in\mathbb{R}^{n}$ the inequalities

(2) $$\left\langle x_{i}-x_{j}\,,y_{i}-y_{j}\right\rangle\geq 0\hskip 28.45274pt\left(i,j=1,\dots,d\right)$$

are fulfilled. Then, for any $z\in\mathbb{R}^{n}$, there exists $v\in\mbox{\rm conv}\left(x_{1}\,,\dots,x_{d}\right)$ such that the inequalities

(3) $$\left\langle x_{j}-v\,,y_{j}-z\right\rangle\geq 0\hskip 28.45274pt\left(j=1,\dots,d\right)$$

hold.

Proof.

Let us introduce the following two notations:

$$C:=\mbox{\rm conv}\left(x_{1}\,,\dots,x_{d}\right)\hskip 14.22636pt\mbox{ and }\hskip 14.22636pt\Delta_{d}:=\{a=\left(a_{1}\,,\dots\,a_{d}\right)\in\left[0,1\right]^{d}\,:\,a_{1}+\dots+a_{n}=1\}.$$

We have to show that there exists $v\in C$ such that

$$\min_{k=1,\dots,d}\left\langle x_{k}-v\,,y_{k}-z\right\rangle\geq 0.$$

However, this is equivalent to the assertion that

$$\max_{v\in C}\min_{k=1,\dots,d}\left\langle x_{k}-v\,,y_{k}-z\right\rangle\geq 0.$$

On the other hand, for any fixed $v\in C$, $\min_{k=1,\dots,d}\left\langle x_{k}-v\,,y_{k}-z\right\rangle$ is the minimum of $d$ real numbers, so it is the minimum of the convex hull of these numbers. That is,

$$\min_{k=1,\dots,d}\left\langle x_{k}-v\,,y_{k}-z\right\rangle=\min_{a\in\Delta_{d}}\sum_{k=1}^{d}a_{k}\left\langle x_{k}-v\,,y_{k}-z\right\rangle.$$

Consequently, we need to verify that

(4) $$\max_{v\in C}\min_{a\in\Delta_{d}}\sum_{k=1}^{d}a_{k}\left\langle x_{k}-v\,,y_{k}-z\right\rangle\geq 0.$$

Let us define the function $G:C\times\Delta_{d}\longrightarrow\mathbb{R}$ with the following formula:

$$G(v,a):=\sum_{k=1}^{d}a_{k}\left\langle x_{k}-v\,,y_{k}-z\right\rangle\hskip 28.45274pt\left(v\in C,a\in\Delta_{d}\right).$$

It is clear that $G$ is defined on the Cartesian product of two compact, convex subsets of $\mathbb{R}^{d}$, moreover $G$ is affine in its first variable and linear in its second variable. Therefore, Ky Fan’s Minimax Theorem [8] is applicable in (4), so we have to show

(5) $$\min_{a\in\Delta_{d}}\max_{v\in C}G(v,a)=\min_{a\in\Delta_{d}}\max_{v\in C}\sum_{k=1}^{d}a_{k}\left\langle x_{k}-v\,,y_{k}-z\right\rangle\geq 0.$$

For any fixed vector $a=\left(a_{1}\,,\dots a_{d}\right)\in\Delta_{d}$ the convex combination $v_{a}=a_{1}x_{1}+\dots+a_{d}x_{d}$ is contained in $C$. Now we may calculate

(6) $$G(v_{a},a) =\sum_{k=1}^{d}a_{k}\left\langle x_{k}-v_{a}\,,y_{k}-z\right\rangle=\sum_{k=1}^{d}a_{k}\left\langle x_{k}-\sum_{j=1}^{d}a_{j}x_{j}\,,y_{k}-z\right\rangle$$

(7) $$=\sum_{k=1}^{d}\sum_{j=1}^{d}a_{k}a_{j}\left\langle x_{k}-x_{j}\,,y_{k}\right\rangle-\sum_{k=1}^{d}\sum_{j=1}^{d}a_{k}a_{j}\left\langle x_{k}-x_{j}\,,z\right\rangle,$$

using the fact that $\sum_{j=1}^{d}a_{j}=1$. The second term is clearly zero, since

$$2\cdot\sum_{k=1}^{d}\sum_{j=1}^{d}a_{k}a_{j}\left\langle x_{k}-x_{j}\,,z\right\rangle=\sum_{k=1}^{d}\sum_{j=1}^{d}a_{k}a_{j}\bigl(\left\langle x_{k}-x_{j}\,,z\right\rangle+\left\langle x_{j}-x_{k}\,,z\right\rangle\bigr)=0.$$

For the investigation of the first term we shall utilize assumption (2), whence

$$0\leq\sum_{k=1}^{d}\sum_{j=1}^{d}a_{k}a_{j}\left\langle x_{k}-x_{j}\,,y_{k}-y_{j}\right\rangle$$

$$=\sum_{k=1}^{d}\sum_{j=1}^{d}a_{k}a_{j}\left\langle x_{k}-x_{j}\,,y_{k}\right\rangle-\sum_{k=1}^{d}\sum_{j=1}^{d}a_{k}a_{j}\left\langle x_{k}-x_{j}\,,y_{j}\right\rangle$$

$$=\sum_{k=1}^{d}\sum_{j=1}^{d}a_{k}a_{j}\left\langle x_{k}-x_{j}\,,y_{k}\right\rangle+\sum_{j=1}^{d}\sum_{k=1}^{d}a_{j}a_{k}\left\langle x_{j}-x_{k}\,,y_{j}\right\rangle=2\cdot\sum_{k=1}^{d}\sum_{j=1}^{d}a_{k}a_{j}\left\langle x_{k}-x_{j}\,,y_{k}\right\rangle,$$

so the first term in (7) is non-negative. Consequently, we have obtained that for any $a\in\Delta_{d}$ there exists $v_{a}\in C$ such that $G(v_{a},a)\geq 0$. Thus (5) is indeed fulfilled, which is equivalent to (4), and that completes the proof. ∎

Lemma 2.

Let $\emptyset\neq C\subseteq\mathbb{R}^{n}$ be a closed convex set and let $f:C\longrightarrow\mathbb{R}^{n}$ be a strictly increasing mapping. Suppose that, for any $z\in\mbox{\rm conv}\left(f(C)\right)$ and for any $x_{1}\,,\dots,x_{n}\,,x_{n+1}\in C$, we have

(8) $$\bigcap_{j=1}^{n+1}\{s\in C\,:\,\left\langle x_{j}-s\,,f(x_{j})-z\right\rangle\geq 0\}\neq\emptyset.$$

Then

$$\bigcap_{x\in C}\{s\in C\,:\,\left\langle x-s\,,f(x)-z\right\rangle\geq 0\}\neq\emptyset.$$

Proof.

The case when $C$ is a singleton is trivial, so from now on suppose that this is not the case. The idea is to use Helly’s Theorem. Let us consider the set

$$C_{x}:=\{s\in C\,:\,\left\langle x-s\,,f(x)-z\right\rangle\geq 0\}$$

for every $x\in C$. It is obvious that $x\in C_{x}\,$, moreover $C_{x}$ is the intersection of a half-space and $C$, therefore it is closed and convex. That is, $\mathcal{C}:=\{C_{x}\,:\,x\in C\}$ is an infinite family of nonempty closed convex subsets of $\mathbb{R}^{n}$. The assumption (8) ensures that every subfamily of $n+1$ sets from $\mathcal{C}$ has nonempty intersection. The last assumption of Helly’s Theorem which needs to be checked is the existence of a finite subfamily of $\mathcal{C}$ with a bounded intersection.

If $C$ itself is bounded, this is evident. Suppose the contrary. Then the recession cone of $C$ is non-trivial (see [22, Theorem 8.4]). That is,

$$R:=\{h\in\mathbb{R}^{n}\,:\,x+h\in C\mbox{ for all }x\in C\}\neq\{0\}.$$

It is well-known (see [22, Theorem 8.2]) that $R$ is a closed, convex cone, since $C$ is closed. We shall also consider

$$R_{0}:=\left\{\frac{h}{\|h\|}\,:\,h\in R\setminus\{0\}\right\},$$

which is the intersection of $R$ and the unit sphere $\mathcal{S}_{n-1}\,$, so $R_{0}$ is compact.

Claim 1.

For any $h\in R_{0}$ there exists $x_{h}\in C$ such that $\left\langle h\,,f(x_{h})-z\right\rangle>0$.

Proof of Claim 2.1.

Fix an arbitrary $h\in R_{0}$. In the first step we will show that there is a vector $w\in C$ fulfilling $\left\langle h\,,f(w)-z\right\rangle\geq 0$. Since $z\in\mbox{\rm conv}(f(C))$, there exist

$$\lambda_{1}\,,\dots,\lambda_{n}\,,\lambda_{n+1}\in\left[0,1\right]\hskip 8.53581pt\mbox{ and }\hskip 8.53581pty_{1}\,,\dots,y_{n}\,,y_{n+1}\in C$$

such that $z=\lambda_{1}f(y_{1})+\dots+\lambda_{n}f(y_{n})+\lambda_{n+1}f(y_{n+1})$ and $\sum_{i=1}^{n+1}\lambda_{i}=1$, according to Carathéodory’s Theorem. By calculating

$$0 =\left\langle h\,,0\right\rangle=\left\langle h\,,z-z\right\rangle=\left\langle h\,,\sum_{i=1}^{n+1}\lambda_{i}f(y_{i})-\sum_{i=1}^{n+1}\lambda_{i}z\right\rangle$$

$$=\sum_{i=1}^{n+1}\left\langle h\,,\lambda_{i}\left(f(y_{i})-z\right)\right\rangle=\sum_{i=1}^{n+1}\lambda_{i}\left\langle h\,,f(y_{i})-z\right\rangle,$$

we obtain that $\left\langle h\,,f(y_{i_{0}})-z\right\rangle\geq 0$ for at least one index $i_{0}\,$. Thus for $w:=y_{i_{0}}$ we indeed have $\left\langle h\,,f(w)-z\right\rangle\geq 0$. Since $h\in R$, we know that $x_{h}:=w+h\in C$ as well. Now

$$\left\langle h\,,f(x_{h})-z\right\rangle =\left\langle h\,,f(x_{h})-f(w)+f(w)-z\right\rangle=\left\langle h\,,f(x_{h})-f(w)\right\rangle$$

$$+\left\langle h\,,f(w)-z\right\rangle=\left\langle x_{h}-w\,,f(x_{h})-f(w)\right\rangle+\left\langle h\,,f(w)-z\right\rangle>0,$$

as the first term is positive (because $f$ is strictly increasing) while the second term is non-negative due to the previous step. $\boxtimes$

Now we are able to construct an open cover for $R_{0}\,$. For each $x\in C$ let us define the open half-space

$$H_{x}:=\left\{v\in\mathbb{R}^{n}\,:\,\left\langle v\,,f(x)-z\right\rangle>0.\right\}$$

According to Claim 1, every vector $h\in R_{0}$ is contained in the corresponding set $H_{x_{h}}\,$. Hence $R_{0}\subseteq\bigcap_{x\in C}H_{x}\,$. But $R_{0}$ is compact, so there exists a finite subcover, i. e. there exist $m\in\mathbb{N}$ and $u_{1}\,,\dots,u_{m}\in C$ such that

$$R_{0}\subseteq\bigcap_{k=1}^{m}H_{u_{k}}\,.$$

Claim 2.

$M:=\bigcap_{k=1}^{m}\{s\in C\,:\,\left\langle u_{k}-s\,,f(u_{k})-z\right\rangle\geq 0\}$ is a nonempty compact, convex set.

Proof of Claim 2.2.

The classical Helly’s Theorem for finite families of convex sets provides that $M\neq\emptyset$, using assumption (8). Closedness and convexity is obvious, so we only need to verify that $M$ is bounded. Let $Q$ denote its recession cone. Since $M\subseteq C$, we know that $Q\subseteq R$. Suppose that $Q$ is nontrivial, so there exists $0\neq q\in Q$ and hence

$$q_{0}:=\frac{q}{\|q\|}\in R_{0}\,.$$

But then there exists $\ell\in\{1,\dots,m\}$ such that $q_{0}\in H_{u_{\ell}}\,$. This means $\left\langle q_{0}\,,f(u_{\ell})-z\right\rangle>0$. If we pick an arbitrary $m\in M$, then $m+\mu\cdot q\in M$ for any $\mu>0$, because $q\in Q$. This implies

$$0\leq\left\langle u_{\ell}-\left(m+\mu q\right)\,,f(u_{\ell})-z\right\rangle=\left\langle u_{\ell}-m\,,f(u_{\ell})-z\right\rangle-\mu\cdot\left\langle q\,,f(u_{\ell})-z\right\rangle$$

for every $\mu>0$. However, we shall observe that, due to $\left\langle q_{0}\,,f(u_{\ell}-z)\right\rangle>0$, the right hand side tends to $-\infty$ as $\mu\to+\infty$. This contradiction means that $Q=\{0\}$, so $M$ has a trivial recession cone, hence it is bounded. $\boxtimes$

According to Claim 2, $\mathcal{C}$ indeed has a finite subfamily with bounded intersection, namely $\bigcap_{k=1}^{m}C_{u_{k}}=M$ is a compact set. Eventually we are able to apply Theorem 1 for $\mathcal{C}$ and obtain that $\bigcap\mathcal{C}\neq\emptyset$ which had to be verified. ∎

Remark 1.

Let us note that when $C$ is compact, then each set in the family $\mathcal{C}$ is bounded even if $z$ is not contained in $\mbox{\rm conv}(f(C))$. Thus if we suppose that $C$ is compact, convex, then the statement of Lemma 2 remains valid for arbitrary $z\in\mathbb{R}^{n}$, while the proof is just a direct application of Helly’s Theorem.

Combining the previous two Lemmas we are able to prove our main result about the existence and uniqueness of a generalized inverse.

Theorem 2.

Let $\emptyset\neq C\subseteq\mathbb{R}^{n}$ be a closed convex set and let $f:C\longrightarrow\mathbb{R}^{n}$ be a strictly increasing mapping. Then there exists a function $f^{(-1)}:\mbox{\rm conv}\left(f(C)\right)\longrightarrow C$ such that

• $f^{(-1)}\left(f(x)\right)=x$ for every $x\in C$,

• $\left\langle f^{(-1)}(u)-f^{(-1)}(v)\,,u-v\right\rangle\geq 0$ for all $u,v\in\mbox{\rm conv}\left(f(C)\right)$.

Moreover, $f^{(-1)}$ is uniquely determined.

Proof.

Let $z\in\mbox{\rm conv}(f(C))$ and $x_{1}\,,\dots,x_{n}\,,x_{n+1}\in C$ be arbitrary points. Using the notation $y_{j}:=f(x_{j})$ (for $j=1,\dots,n+1$), the strictly increasing property of $f$ yields

$$\left\langle x_{i}-x_{j}\,,y_{i}-y_{j}\right\rangle\geq 0\hskip 28.45274pt\left(i,j=1,\dots,d\right).$$

This means that the assumption (2) in Lemma 1 is fulfilled. Therefore there exists $v\in\mbox{\rm conv}\left(x_{1}\,,\dots,x_{n+1}\right)\subseteq C$ such that

$$\left\langle x_{j}-v\,,y_{j}-z\right\rangle\geq 0\hskip 28.45274pt\left(j=1,\dots,d\right).$$

That is,

$$\bigcap_{j=1}^{n+1}\{s\in C\,:\,\left\langle x_{j}-s\,,f(x_{j})-z\right\rangle\geq 0\}\neq\emptyset.$$

Now the assertion of Lemma 2 implies that

$$M_{z}:=\bigcap_{x\in C}\{s\in C\,:\,\left\langle x-s\,,f(x)-z\right\rangle\geq 0\}\neq\emptyset.$$

We are going to show that $M_{z}$ is actually a singleton. Suppose that $v,w\in M_{z}$ and consider the points $t_{1}:=\frac{v+3w}{4}$ and $t_{2}:=\frac{3v+w}{4}$. Now $t_{1}\,,t_{2}\in C$, moreover $t_{1}-t_{2}=\frac{w-v}{2}$. Since $v,w\in M_{z}$, the inequalities

$$\left\langle t_{1}-w\,,f(t_{1})-z\right\rangle\geq 0\hskip 8.53581pt\mbox{ and }\hskip 8.53581pt\left\langle t_{2}-v\,,f(t_{2})-z\right\rangle\geq 0$$

are fulfilled. But these are equivalent to

$$\left\langle\frac{v-w}{4}\,,f(t_{1})-z\right\rangle\geq 0\hskip 8.53581pt\mbox{ and }\hskip 8.53581pt\left\langle\frac{w-v}{4}\,,f(t_{2})-z\right\rangle\geq 0,$$

respectively. Multiplying by $2$ and adding them up we get

$$0\leq\left\langle\frac{w-v}{2}\,,f(t_{2})-f(t_{1})\right\rangle=\left\langle t_{1}-t_{2}\,,f(t_{2})-f(t_{1})\right\rangle.$$

Due to $f$ being strictly increasing, this can occur only when $t_{1}=t_{2}$ and, equivalently, $v=w$. Thus $M_{z}$ is indeed a singleton. The necessary condition in Proposition 2 states that if $f^{(-1)}$ exists then $f^{(-1)}(z)\in M_{z}$ must hold on its domain. This gives the unique definition of $f^{(-1)}:\mbox{\rm conv}(f(C))\longrightarrow C$, namely

$$\mbox{let }f^{(-1)}(z)\mbox{ be the unique element of }M_{z}\mbox{ for every }z\in\mbox{\rm conv}(f(C)).$$

Finally we have to show that this function is indeed a left inverse, and it is monotone increasing. For any $x,y\in C$ we have $\left\langle x-y\,,f(x)-f(y)\right\rangle\geq 0$, so $x\in M_{f(x)}$. Thus $f^{(-1)}(f(x))=x$ for all $x\in C$. On the other hand, consider any two vectors $u,v\in\mbox{\rm conv}(f(C))$ and introduce the notations

$$d:=\frac{f^{(-1)}(v)-f^{(-1)}(u)}{2}\hskip 8.53581pt\mbox{ and }\hskip 8.53581ptm:=\frac{f^{(-1)}(u)+f^{(-1)}(v)}{2}.$$

Due to $m\in C$ and using the definition of $f^{(-1)}$ we have

$$\left\langle m-f^{(-1)}(u)\,,f(m)-u\right\rangle\geq 0\hskip 8.53581pt\mbox{ and }\hskip 8.53581pt\left\langle m-f^{(-1)}(v)\,,f(m)-v\right\rangle\geq 0.$$

That is, $\left\langle d\,,f(m)-u\right\rangle\geq 0$ and $\left\langle-d\,,f(m)-v\right\rangle\geq 0$. After summation we get

$$0\leq\left\langle d\,,v-u\right\rangle\leq\left\langle 2d\,,v-u\right\rangle=\left\langle f^{(-1)}(v)-f^{(-1)}(u)\,,v-u\right\rangle.$$

Therefore $f^{(-1)}$ turns out to be increasing, thus the proof is complete. ∎

Remark 2.

We wish to emphasize that the proof provides the definition of the extended monotone left-inverse explicitly. For any $z\in\mbox{\rm conv}(f(C))$ we have

$$\{f^{(-1)}(z)\}=\bigcap_{x\in C}\{s\in C\,:\,\left\langle x-s\,,f(x)-z\right\rangle\geq 0\}.$$

One should also observe that the only occasion where we rely on the fact that $z$ is contained in $\mbox{\rm conv}(f(C))$ is when we apply Lemma 2. However, we have already mentioned in Remark 1 that if the domain is compact then Lemma 2 remains valid for any $z\in\mathbb{R}^{n}$. Thus if the domain of definition of a strictly increasing mapping is a compact convex set, then the generalized inverse has a unique extension to the whole space. We formulate this in the following Corollary.

Corollary 1.

Let $\emptyset\neq K\subseteq\mathbb{R}^{n}$ be a compact, convex set and let $f:K\longrightarrow\mathbb{R}^{n}$ be a strictly increasing function. Then there exists a function $f^{(-1)}:\mathbb{R}^{n}\longrightarrow K$ such that

• $f^{(-1)}\left(f(x)\right)=x$ for every $x\in K$,

• $\left\langle f^{(-1)}(u)-f^{(-1)}(v)\,,u-v\right\rangle\geq 0$ for all $u,v\in\mathbb{R}^{n}$.

Moreover, $f^{(-1)}$ is uniquely determined by the identity

$$\{f^{(-1)}(z)\}=\bigcap_{x\in K}\{s\in K\,:\,\left\langle x-s\,,f(x)-z\right\rangle\geq 0\}\hskip 28.45274pt\left(z\in\mathbb{R}^{n}\right).$$

3. Topological properties of $f^{(-1)}$ and the domain

In the one-dimensional case the only important property of the domain was convexity, the existence and uniqueness of the generalized inverse holds for open, closed and half-closed intervals. In the following example we demonstrate that in higher dimensions the closedness of the domain is a crucial condition. The idea is based on a counterexample presented by K. Okamura during the mentioned 60th ISFE meeting (see [1, 5. Remark]).

Example 1.

Consider the convex open half-disk $S=\{(x,y)\in\mathbb{R}^{2}:x>0\mbox{ and }x^{2}+y^{2}<1\}$ and the mapping $f:S\longrightarrow\mathbb{R}^{2}$ defined by $f(x,y)=\left(x^{2}-y^{2},2xy\right)$. Then $f$ is strictly increasing, but it has no extended monotone left-inverse .

Proof.

The fact that $f$ is strictly increasing is a matter of elementary calculations. An elegant reasoning is contained in [1, 5. Remark], applying an identification between $f$ and the complex square function $z\mapsto z^{2}$. Since

$$S=\left\{(r\cos\varphi,r\sin\varphi)\in\mathbb{R}^{2}\,:\,0<r<1,\mbox{ and }-\frac{\pi}{2}<\varphi<\frac{\pi}{2}\right\}$$

and $f\left(r\cos\varphi,r\sin\varphi\right)=\left(r^{2}\cos\left(2\varphi\right),r^{2}\sin\left(2\varphi\right)\right)$, it is clear that

$$f(S) =\left\{\left(r^{2}\cos\left(2\varphi\right),r^{2}\sin\left(2\varphi\right)\right)\,:\,0<r<1,-\frac{\pi}{2}<\varphi<\frac{\pi}{2}\right\}$$

$$=\left\{\left(R\cos\psi,R\sin\psi\right)\,:\,0<R<1,-\pi<\psi<\pi\right\}$$

$$=B\setminus\left\{(x,0)\,:\,x\leq 0\right\}.$$

Here $B=\{x\in\mathbb{R}^{2}\,:\,\|x\|<1\}$ is the open unit ball. Consequently, $\mbox{\rm conv}f(S)=B$. Suppose that there exists an extended monotone left-inverse $F:B\longrightarrow S$. In particular, there should exist $(a,b)\in S$ such that $F(0,0)=(a,b)$. Let us now investigate the necessary condition of Proposition 2 for $(0,0)\in\mbox{\rm conv}f(S)$. Then

$$0 \leq\left\langle\left(\frac{a}{2},0\right)-F(0,0)\,,f\left(\frac{a}{2},0\right)-(0,0)\right\rangle=\left\langle\left(\frac{a}{2},0\right)-(a,b)\,,\left(\frac{a^{2}}{4},0\right)-(0,0)\right\rangle$$

$$=\left\langle\left(-\frac{a}{2},-b\right)\,,\left(\frac{a^{2}}{4},0\right)\right\rangle=-\frac{a^{3}}{8}<0,$$

a contradiction. Thus $f$ cannot have an extended monotone left-inverse. ∎

In [1, 5. Remark] it was proved that for the function $f:S\longrightarrow\mathbb{R}^{2}$ in the previous example there is no continuous extension of $f^{-1}$ to $\mbox{\rm conv}f(S)$. This is an interesting difference compared to the one-dimensional setting, where the ordinary inverse always has a continuous and monotone extension to the convex hull of the image set.

However, it is possible to show that in the case of a closed, convex domain the generalized inverse function is continuous. Before we prove that, we formulate an easy geometric observation about euclidean spaces.

Proposition 3.

Let $(x_{k}):\mathbb{N}\longrightarrow\mathbb{R}^{n}$ be a sequence, $h\in\mathbb{R}^{n}$ be a vector and $\delta>0$ be a number such that $\|h\|=1$, $\|x_{k}\|>\delta$ for all $k\in\mathbb{N}$, moreover

$$\alpha_{k}:=\left\langle\frac{x_{k}}{\|x_{k}\|}\,,h\right\rangle\to 1\ \mbox{ as }\ k\to\infty.$$

Then, for any $d\in\left(0,\delta\right)$, we have

$$\lim_{k\to\infty}\left\langle\frac{x_{k}-dh}{\|x_{k}-dh\|}\,,h\right\rangle=1.$$

Proof.

Consider the function $g(t):=t-2d+\frac{d^{2}}{t}=\frac{1}{t}\left(t-d\right)^{2}$ defined for all $t>0$. Clearly $g$ is continuous, non-negative and $\lim_{t\to\infty}g(t)=+\infty$. Since $\delta>d$, there exists $M>0$ such that $g(t)>M$ for all $t\geq\delta$. In particular, $g\left(\|x_{k}\|\right)>M$ for every index $k\in\mathbb{N}$.

Since $\alpha_{k}\leq 1$ due to the CBS inequality, we have

$$\|x_{k}\|-2d\alpha_{k}+\frac{d^{2}}{\|x_{k}\|}\geq\|x_{k}\|-2d+\frac{d^{2}}{\|x_{k}\|}=g\left(\|x_{k}\|\right)>M\hskip 28.45274pt\left(k\in\mathbb{N}\right).$$

Now we shall calculate

$$\left(\frac{\|x_{k}\|-d}{\|x_{k}-dh\|}\right)^{2} =\frac{\|x_{k}\|^{2}-2d\|x_{k}\|+d^{2}}{\|x_{k}\|^{2}-2d\left\langle x_{k}\,,h\right\rangle+d^{2}}=1+2d\frac{\left\langle x_{k}\,,h\right\rangle-\|x_{k}\|}{\|x_{k}\|^{2}-2d\left\langle x_{k}\,,h\right\rangle+d^{2}}$$

$$=1+2d\frac{\alpha_{k}-1}{\|x_{k}\|-2d\alpha_{k}+\frac{d^{2}}{\|x_{k}\|}}\to 1\ \mbox{ as }k\to\infty,$$

because the numerator tends to $0$ while the denominator is bounded from below by $M>0$. Consequently, the positive sequence $\frac{\|x_{k}\|-d}{\|x_{k}-dh\|}$ also tends to $1$. Finally,

$$1 \geq\left\langle\frac{x_{k}-dh}{\|x_{k}-dh\|}\,,h\right\rangle=\frac{\|x_{k}\|}{\|x_{k}-dh\|}\cdot\left\langle\frac{x_{k}}{\|x_{k}\|}\,,h\right\rangle-\frac{d}{\|x_{k}-dh\|}=\frac{\|x_{k}\|-d}{\|x_{k}-dh\|}\cdot\alpha_{k}$$

$$+\frac{d}{\|x_{k}-dh\|}\left(\alpha_{k}-1\right)>\frac{\|x_{k}\|-d}{\|x_{k}-dh\|}\cdot\alpha_{k}+\frac{d}{\delta-d}\left(\alpha_{k}-1\right)\to 1\cdot 1+\frac{d}{\delta-d}\cdot 0=1.$$

Therefore $\left\langle\frac{x_{k}-dh}{\|x_{k}-dh\|}\,,h\right\rangle\to 1$ and the proof is complete. ∎

Theorem 3.

Let $\emptyset\neq C\subseteq\mathbb{R}^{n}$ be a closed, convex set and let $f:C\longrightarrow\mathbb{R}^{n}$ be a strictly increasing mapping. Then the unique extended monotone left-inverse function $f^{(-1)}:\mbox{\rm conv}\left(f(C)\right)\longrightarrow C$ is continuous.

Proof.

We have to show that if a sequence $\left(u_{k}\right):\mathbb{N}\longrightarrow\mbox{\rm conv}(f(C))$ converges to $v\in\mbox{\rm conv}(f(C))$ then $f^{(-1)}(u_{k})$ converges to $f^{(-1)}(v)$.

Suppose that, on the contrary, $\lim_{k\to\infty}f^{(-1)}(u_{k})\neq f(v)$. This means that there exists $\delta>0$ and a subsequence $\left(u_{\ell_{k}}\right)$ of $(u_{k})$ such that $\|f^{(-1)}\left(u_{\ell_{k}}\right)-f^{(-1)}\left(v\right)\|>\delta$. Without loss of generality we may assume that this sequence is the whole $\left(u_{k}\right)$. Let us observe that

$$\frac{f^{(-1)}\left(u_{k}\right)-f^{(-1)}\left(v\right)}{\|f^{(-1)}\left(u_{k}\right)-f^{(-1)}\left(v\right)\|}\in\mathcal{S}_{n-1}\ \mbox{ for all }k\in\mathbb{N}.$$

Since $\mathcal{S}_{n-1}$ is compact, the above sequence has a convergent subsequence. That is, there exist $h\in\mathbb{R}^{n}$ with $\|h\|=1$, and strictly increasing sequence of indices $(m_{k}):\mathbb{N}\longrightarrow\mathbb{N}$ such that if $x_{k}:=f^{(-1)}\left(u_{m_{k}}\right)-f^{(-1)}\left(v\right)$ for all $k\in\mathbb{N}$, then

$$\left\|\frac{x_{k}}{\|x_{k}\|}-h\right\|\to 0\ \mbox{ or, equivalently, }\ \left\langle\frac{x_{k}}{\|x_{k}\|}\,,h\right\rangle\to 1.$$

As $C$ is convex, $f^{(-1)}(v)+\frac{\delta}{\|x_{k}\|}x_{k}\in C$ for each $k\in\mathbb{N}$. But $C$ is closed as well, so $f^{(-1)}(v)+\delta h\in C$, whence

$$s:=f^{(-1)}(v)+\frac{\delta}{2}h\in C$$

also holds. Now let us apply Proposition 3 to the sequence $(x_{k})$, the vector $h$ and the constant $\delta$. The Proposition claims that, in particular, for $d=\frac{\delta}{2}$ the assertion

(9) $$1=\lim_{k\to\infty}\left\langle\frac{x_{k}-\frac{\delta}{2}h}{\|x_{k}-\frac{\delta}{2}h\|}\,,h\right\rangle=\lim_{k\to\infty}\left\langle\frac{f^{(-1)}\left(u_{m_{k}}\right)-s}{\|f^{(-1)}\left(u_{m_{k}}\right)-s\|}\,,h\right\rangle$$

is valid. In the next step let us fix another vector $t:=\frac{f^{(-1)}(v)+s}{2}\in C$. Then

$$\frac{\delta}{2}\left\langle h\,,f(s)-v\right\rangle =\left\langle s-f^{(-1)}(v)\,,f(s)-v\right\rangle=\left\langle s-t\,,f(s)-f(t)\right\rangle+\left\langle s-t\,,f(t)-v\right\rangle$$

$$+\left\langle t-f^{(-1)}(v)\,,f(s)-v\right\rangle=\left\langle s-t\,,f(s)-f(t)\right\rangle$$

$$+\left\langle t-f^{(-1)}(v)\,,f(t)-v\right\rangle+\frac{1}{2}\left\langle s-f^{(-1)}(v)\,,f(s)-v\right\rangle>0,$$

because the first term is positive as $f$ is strictly increasing, while the other two terms are non-negative, due to the defining properties of $f^{(-1)}$. Thus $\left\langle h\,,f(s)-v\right\rangle>0$, which implies that there exists $\varepsilon>0$ and $k_{0}\in\mathbb{N}$ such that

$$\left\langle h\,,f(s)-u_{m_{k}}\right\rangle>\varepsilon\ \mbox{ for all }k>k_{0}\,$$

since $u_{m_{k}}\to v$. Furthermore, from (9) we get that

$$\left|\left\langle f(s)-u_{m_{k}}\,,\frac{f^{(-1)}\left(u_{m_{k}}\right)-s}{\|f^{(-1)}\left(u_{m_{k}}\right)-s\|}-h\right\rangle\right|\leq\left\|f(s)-u_{m_{k}}\right\|\cdot\left\|\frac{f^{(-1)}\left(u_{m_{k}}\right)-s}{\|f^{(-1)}\left(u_{m_{k}}\right)-s\|}-h\right\|\to 0,$$

applying the CBS inequality and using that $\left(f(s)-u_{m_{k}}\right)$ is a bounded sequence. In particular, there exists an index $k_{1}>k_{0}$ such that

$$\left\langle f(s)-u_{m_{k}}\,,\frac{f^{(-1)}\left(u_{m_{k}}\right)-s}{\|f^{(-1)}\left(u_{m_{k}}\right)-s\|}-h\right\rangle>-\frac{\varepsilon}{2}\ \mbox{ for all }k>k_{1}.$$

Combining the previous results we obtain that, for all $k>k_{1}\,$, the inequality

$$\left\langle f(s)-u_{m_{k}}\,,\frac{f^{(-1)}\left(u_{m_{k}}\right)-s}{\|f^{(-1)}\left(u_{m_{k}}\right)-s\|}\right\rangle=\left\langle f(s)-u_{m_{k}}\,,\frac{f^{(-1)}\left(u_{m_{k}}\right)-s}{\|f^{(-1)}\left(u_{m_{k}}\right)-s\|}-h\right\rangle$$

$$+\left\langle f(s)-u_{m_{k}}\,,h\right\rangle>-\frac{\varepsilon}{2}+\varepsilon=\frac{\varepsilon}{2}>0$$

holds. But this implies

$$\left\langle f^{(-1)}\left(u_{m_{k}}\right)-s\,,u_{m_{k}}-f(s)\right\rangle<0\hskip 28.45274pt\left(k>k_{1}\right),$$

which contradicts the monotonicity of $f^{(-1)}$. Therefore our initial assumption about $f^{(-1)}(u_{k})$ not converging to $f^{(-1)}(v)$ was false, so $f^{(-1)}$ is continuous. ∎

4. Application for quasi-arithmetic means

In the final section of our paper we introduce the concept of vector valued quasi-arithmetic means.

Definition 2.

Let $\emptyset\neq C\subseteq\mathbb{R}^{n}$ be a closed, convex set and let $f:C\longrightarrow\mathbb{R}^{n}$ be a strictly monotone mapping. Moreover, let $m\in\mathbb{N}$ and $\lambda_{1}\,,\dots,\lambda_{m}>0$ be fixed numbers such that $m\geq 2$ and $\lambda_{1}+\dots+\lambda_{m}=1$.

Then the function $\mathcal{M}_{f,\lambda_{1},\dots,\lambda_{m}}^{[m]}:C^{m}\longrightarrow C$ defined by

$$\mathcal{M}_{f,\lambda_{1},\dots,\lambda_{m}}^{[m]}\left(x_{1}\,,\dots,x_{m}\right):=f^{(-1)}\bigl(\lambda_{1}f(x_{1})+\dots+\lambda_{m}f(x_{m})\bigr)\hskip 22.76219pt\left(x_{1}\,,\dots,x_{m}\in C\right)$$

is called the $m$-variable weighted quasi-arithmetic mean. The generator of the mean is $f$.

It is easy to see that for arbitrary weights $\mathcal{M}_{f,\lambda_{1},\dots,\lambda_{m}}^{[m]}=\mathcal{M}_{-f,\lambda_{1},\dots,\lambda_{m}}^{[m]}\,$, thus it is enough to consider the case of a strictly increasing generator. In [16] the idea of vector-valued QAMs appears, although it is a priori assumed that the generator has a convex image, which we find rather restrictive.

One of the first natural problems which arises immediately is the equality problem: for a fixed number of variables and fixed weights find those generators for which

(10) $$\mathcal{M}_{f,\lambda_{1},\dots,\lambda_{m}}^{[m]}\left(x_{1}\,,\dots,x_{m}\right)=\mathcal{M}_{g,\lambda_{1},\dots,\lambda_{m}}^{[m]}\left(x_{1}\,,\dots,x_{m}\right)\hskip 22.76219pt\left(x_{1}\,,\dots,x_{m}\in C\right).$$

In the scalar valued setting when the generators are supposed the be continuous, then the solution is a well-known classical result: $f=a\cdot g+b$ with some real numbers such that $a\neq 0$ (see [11, Theorem 83 and Section 3.7] or [21, Theorem A]).

In the case of non-continuous generators the situation is much more difficult. Even though it has been shown recently in [21, Corollary 4.2] that if the equality of two QAMs holds for all $m\in\mathbb{N}$, then the generators are affine transforms of each other, the same cannot be claimed in the case when we have a fixed number of variables $m$. What is more, a fresh result of Kiss and Pasteczka gives examples for three-variable QAMs which are equal but their generators are not affine transforms of each other (see [12, Example 1]).

The complete characterization of the solutions of the equality problem for fixed weights and a fixed number of variables is still open even in the real valued case. Therefore our aim here, instead of discussing the general equality problem, is to solve a particular case: for a fixed $m\geq 2$ when does the $m$-variable vector-valued weighted quasi-arithmetic mean coincide with the weighted arithmetic mean? The $m$-variable arithmetic mean (generated by the identity) with weights $\lambda_{1}\,,\dots,\lambda_{m}$ mean will be denoted by $\mathcal{A}_{\lambda_{1},\dots,\lambda_{m}}^{[m]}$.

Proposition 4.

Let $\emptyset\neq C\subseteq\mathbb{R}^{n}$ be a closed, convex set and let $f:C\longrightarrow\mathbb{R}^{n}$ be a strictly monotone mapping. Moreover, let $m\geq 2$ and $\lambda_{1}\,,\dots,\lambda_{m}>0$ such that $\lambda_{1}+\dots+\lambda_{m}=1$. Suppose that $\mathcal{M}_{f,\lambda_{1},\dots,\lambda_{m}}^{[m]}=\mathcal{A}_{\lambda_{1},\dots,\lambda_{m}}^{[m]}$. Then, for all $x\in C^{\circ}$,

$$y\in\mbox{\rm conv}(f(C))\setminus\{f(x)\}\ \mbox{ implies }\ f^{(-1)}(y)\neq x.$$

Proof.

Firstly, $\mathcal{M}_{f,\lambda_{1},\dots,\lambda_{m}}^{[m]}=\mathcal{A}_{\lambda_{1},\dots,\lambda_{m}}^{[m]}$ implies that, in particular,

(11) $$f^{(-1)} \left(\lambda_{1}f(x)+\left(1-\lambda_{1}\right)f(y)\right)=\mathcal{M}_{f,\lambda_{1},\dots,\lambda_{m}}^{[m]}\left(x,y,\dots,y\right)$$

$$=\mathcal{A}_{\lambda_{1},\dots,\lambda_{m}}^{[m]}\left(x,y,\dots,y\right)=\lambda_{1}x+\left(1-\lambda_{1}\right)y\hskip 42.67912pt\left(x,y\in C\right).$$

Suppose that, contrary to the statement, there exists $x\in C^{\circ}$ and $y\in\mbox{\rm conv}(f(C))$ such that $f^{(-1)}(y)=x$ but $y\neq f(x)$. Let us define the unit vector $h:=\frac{y-f(x)}{\|y-f(x)\|}$. Since $x\in C^{\circ}$, there exists $\alpha_{0}>0$ such that $x+\alpha h\in C$ for all $\alpha\in(0,\alpha_{0})$. Using the monotonicity of $f^{(-1)}$ we obtain

$$0\leq\left\langle f(x+\alpha h)-y\,,f^{(-1)}\left(f(x+\alpha h)\right)-f^{(-1)}(y)\right\rangle=\alpha\cdot\left\langle f(x+\alpha h)-y\,,h\right\rangle,$$

so $0\leq\left\langle f(x+\alpha h)-y\,,h\right\rangle$.

Now let us observe that, for all $\alpha\in(0,\alpha_{0})$, there uniquely exist $\mu_{\alpha}>0$ and $w_{\alpha}\in\mathbb{R}^{n}$ such that $\mu_{\alpha}\geq\|y-f(x)\|$ and $\left\langle w_{\alpha}\,,h\right\rangle=0$, moreover

(12) $$f(x+\alpha h)=f(x)+\mu_{\alpha}h+w_{\alpha}.$$

Indeed, the decomposition (12) is due to the Gram–Schmidt algorithm with

$$\mu_{\alpha}=\left\langle f(x+\alpha h)-f(x)\,,h\right\rangle\hskip 8.53581pt\mbox{ and }\hskip 8.53581ptw_{\alpha}=f(x+\alpha h)-f(x)-\mu_{\alpha}h,$$

while the previous step of the proof yields

$$0 \leq\left\langle f(x+\alpha h)-y\,,h\right\rangle=\left\langle f(x+\alpha h)-f(x)+f(x)-y\,,h\right\rangle=\left\langle f(x+\alpha h)-f(x)\,,h\right\rangle$$

$$+\left\langle f(x)-y\,,h\right\rangle=\mu_{\alpha}-\left\langle y-f(x)\,,\frac{y-f(x)}{\|y-f(x)\|}\right\rangle=\mu_{\alpha}-\|y-f(x)\|,$$

which is equivalent to $\mu_{\alpha}\geq\|y-f(x)\|$. Therefore

(13) $$\nu:=\inf\left\{\left\langle f(x+\alpha h)-f(x)\,,h\right\rangle\ :\ \alpha\in\left(0,\alpha_{0}\right)\right\}\geq\|y-f(x)\|.$$

In the next step let us choose $\beta\in\left(0,\alpha_{0}\right)$ such that for $z:=x+\beta h\in C$ we have

(14) $$\nu\leq\left\langle f(z)-f(x)\,,h\right\rangle<\frac{\nu}{\left(1-\lambda_{1}\right)}.$$

Moreover, let us fix $u:=\frac{1+\lambda_{1}}{2}x+\frac{1-\lambda_{1}}{2}z=x+\frac{1-\lambda_{1}}{2}\beta h\in C$. Then

$$0 \leq\left\langle f^{(-1)}\bigl(\lambda_{1}f(x)+(1-\lambda_{1})f(z)\bigr)-f^{(-1)}\bigl(f(u)\bigr)\,,\bigl(\lambda_{1}f(x)+(1-\lambda_{1})f(z)\bigr)-f(u)\right\rangle$$

$$=\left\langle\bigl(\lambda_{1}x+(1-\lambda_{1})z\bigr)-u\,,(1-\lambda_{1})f(z)-(1-\lambda_{1})f(x)+f(x)-f(u)\right\rangle$$

$$=\left\langle\frac{1-\lambda_{1}}{2}\beta\cdot h\,,\bigl((1-\lambda_{1})(f(z)-f(x))\bigr)-\bigl(f(u)-f(x)\bigr)\right\rangle$$

$$=\frac{1-\lambda_{1}}{2}\beta\cdot\Bigl((1-\lambda_{1})\left\langle f(z)-f(x)\,,h\right\rangle-\left\langle f(u)-f(x)\,,h\right\rangle\Bigr)$$

$$<\frac{1-\lambda_{1}}{2}\beta\cdot\Bigl((1-\lambda_{1})\frac{\nu}{1-\lambda_{1}}-\nu\Bigr)=0,$$

where we have used (11), (13) and (14) together with the monotonicity of $f^{(-1)}$. However, the obtained contradiction means that $f^{(-1)}(y)=x$ implies $y=f(x)$, so the proof is complete by contraposition. ∎

Theorem 4.

Let $m\geq 2$, and let $f:\mathbb{R}^{n}\longrightarrow\mathbb{R}^{n}$ be a strictly increasing mapping. Moreover let $m\geq 2$ and $\lambda_{1}\,,\dots,\lambda_{m}>0$ such that $\lambda_{1}+\dots+\lambda_{m}=1$. Then the following two statements are equivalent:

1. For all $x_{1}\,,\dots,x_{m}\in\mathbb{R}^{n}$ it holds that

   $$\mathcal{M}_{f,\lambda_{1},\dots,\lambda_{m}}^{[m]}\left(x_{1}\,,\dots,x_{m}\right)=\mathcal{A}_{\lambda_{1},\dots,\lambda_{m}}^{[m]}\left(x_{1}\,,\dots,x_{m}\right)$$

2. There exist a positive definite matrix $A\in\mathcal{M}_{n\times n}(\mathbb{R})$ and a vector $b\in\mathbb{R}^{n}$ such that

   $$f(t)=At+b\hskip 22.76219pt\left(t\in\mathbb{R}^{n}\right).$$

Proof.

We begin with the proof of (i) $\Longrightarrow$ (ii). From Proposition 4 we have that $f^{(-1)}$ is injective, so it coincides with the ordinary inverse. In particular, $\mbox{\rm conv}\left(f\left(\mathbb{R}^{n}\right)\right)=f\left(\mathbb{R}^{n}\right)$, i. e. $f\left(\mathbb{R}^{n}\right)$ is convex. Suppose that $t\in f\left(\mathbb{R}^{n}\right)$ is a boundary point of $f\left(\mathbb{R}^{n}\right)$. Then there exist a supporting hyperplane for $f\left(\mathbb{R}^{n}\right)$ at the point $t$ (cf. [22, Theorem 11.6]). This implies the existence of a vector $0\neq h\in\mathbb{R}^{n}$ such that

$$\left\langle f(x)-t\,,h\right\rangle\geq 0\hskip 28.45274pt\left(x\in\mathbb{R}^{n}\right).$$

However, $f$ is strictly increasing, so for $x_{0}:=f^{-1}(t)$ we have

$$0<\left\langle f(x_{0}-h)-f(x_{0})\,,(x_{0}-h)-x_{0}\right\rangle=\left\langle f(x_{0}-h)-t\,,-h\right\rangle,$$

which contradicts the choice of $h$. Thus $\partial\bigl(f\left(\mathbb{R}^{n}\right)\bigr)\cap f\left(\mathbb{R}^{n}\right)=\emptyset$, so $f(\mathbb{R}^{n})$ is open. Hence we may apply Brouwer’s invariance of domain theorem for the injective, continuous (see Theorem 3) map $f^{-1}:f\left(\mathbb{R}^{n}\right)\longrightarrow\mathbb{R}^{n}$ and conclude that its inverse $f:\mathbb{R}^{n}\to f(\mathbb{R}^{n})$ is also continuous.

Recalling (11) from the previous proof and applying $f$ to both sides, we have

(15) $$\lambda_{1}f(x)+(1-\lambda_{1})f(y)=f\left(\lambda_{1}x+(1-\lambda_{1})y\right)\hskip 28.45274pt\left(x,y\in\mathbb{R}^{n}\right).$$

Consequently, every coordinate function $f_{i}:\mathbb{R}^{n}\longrightarrow\mathbb{R}$ (for $i=1,\dots,n$) of $f$ is affine. That is, there exist $a_{i}\in\mathbb{R}^{n}$ and $b_{i}\in\mathbb{R}$ such that

$$f_{i}(x)=\left\langle a_{i}\,,x\right\rangle+b_{i}\hskip 28.45274pt\left(i=1,\dots,n\right)$$

for all $x\in\mathbb{R}^{n}$. The proof is similar to the solution process of Jensen’s equation (which corresponds to (15) with $\lambda_{1}=\frac{1}{2}$) discussed in [14, Section 13.2]. A slightly stronger statement can be found in [3, Theorem 3.1]. Summarizing the previous results, let $A\in\mathcal{M}_{n\times n}(\mathbb{R})$ be the matrix which contains the vector $a_{i}$ in its $i$-th row and let the $i$-th element of $b\in\mathbb{R}^{n}$ be $b_{i}$ (for $i=1,\dots,n$). Then

$$f(x)=\left(f_{1}(x),\dots,f_{n}(x)\right)=Ax+b\hskip 28.45274pt(x\in\mathbb{R}^{n}).$$

Furthermore, $A$ has to be positive definite, because, for every $0\neq x\in\mathbb{R}^{n}$, we have

$$\left\langle Ax\,,x\right\rangle=\left\langle(Ax+b)-(A\cdot 0+b)\,,x-0\right\rangle=\left\langle f(x)-f(0)\,,x-0\right\rangle>0.$$

Finally, for the converse direction (ii) $\Longrightarrow$ (i), let $A$ be positive definite, $b$ be arbitrary, and define $f(t)=At+b$ for all $t\in\mathbb{R}^{n}$. Then $f^{-1}(y)=A^{-1}(y-b)$ for all $y\in\mathbb{R}^{n}$, thus

$$\mathcal{M}_{f,\lambda_{1},\dots,\lambda_{m}}^{[m]} \left(x_{1}\,,\dots,x_{m}\right)=A^{-1}\bigl(\lambda_{1}\left(A(x_{1})+b\right)+\dots\lambda_{m}\left(A(x_{m})+b\right)-b\bigr)$$

$$=A^{-1}\bigl(A\left(\lambda_{1}x_{1}+\dots+\lambda_{m}x_{m}\right)\bigr)=\mathcal{A}_{\lambda_{1},\dots,\lambda_{m}}^{[m]}\left(x_{1}\,,\dots,x_{m}\right)$$

∎

5. Concluding remarks and open problems

In Theorem 4 the domain of the QAM is the whole space $\mathbb{R}^{n}$. Although this is quite restrictive, we demonstrate that $\mathbb{R}^{n}$ cannot be replaced by an arbitrary closed convex set.

Example 2.

Let $C:=\{(t,0)\in\mathbb{R}^{2}:t\in[0,1]\}$ and define $f:C\longrightarrow\mathbb{R}^{n}$ by

$$f\left(t,0\right):=\left(t,g(t)\right)\hskip 28.45274pt\left(\,t\in[0,1]\,\right),$$

where $g:[0,1]\longrightarrow\mathbb{R}$ is an arbitrary function. It is easy to check that $f$ is strictly increasing, moreover the generalized left inverse is

$$f^{(-1)}\left(t,s\right)=(t,0)\hskip 28.45274pt\bigl(\,(t,s)\in\mbox{\rm conv}(f(C))\,\bigr).$$

Clearly, $f$ is not necessarily an affine function. On the other hand, for arbitrary vectors $(x,0),(y,0)\in C$ we have

$$\mathcal{M}_{f,\frac{1}{2},\frac{1}{2}}^{[2]}\left((x,0),(y,0)\right)=f^{(-1)}\left(\frac{f(x,0)+f(y,0)}{2}\right)=f^{(-1)}\left(\frac{(x,g(x))+(y,g(y))}{2}\right)$$

$$=f^{(-1)}\left(\frac{x+y}{2},\frac{g(x)+g(y)}{2}\right)=\left(\frac{x+y}{2},0\right)=\frac{(x,0)+(y,0)}{2}=\mathcal{A}_{\frac{1}{2},\frac{1}{2}}^{[2]}\left((x,0),(y,0)\right).$$

On the other hand, this counterexample involves a pathological domain with empty interior.

Open Problem 1.

Is it possible to prove the analogue of Theorem 4 for a restricted domain, namely for any closed convex set $C\subset\mathbb{R}^{n}$ such that $C^{\circ}\neq\emptyset$?

Open Problem 2.

Is it possible to solve the equality problem of QAMs assuming that the generators are continuous? That is, how can we characterize the strictly increasing generators $f,g:C\longrightarrow\mathbb{R}^{n}$ for which (10) holds on an appropriate domain $C\subseteq\mathbb{R}^{n}$?

Open Problem 3.

We shall note that Definition 1 makes sense for an arbitrary inner product space instead of $\mathbb{R}^{n}$. Even though throughout the paper we use methods which are typically finite dimensional, it is natural to ask the following: Does there exist an extended monotone left-inverse for a strictly monotone mapping $f:C\longrightarrow H$, where $C$ is a closed, convex subset of an infinite dimensional Hilbert space $H$?