---
title: Normalized Duality Mapping
url: https://www.emergentmind.com/topics/normalized-duality-mapping
type: topic
---

# Normalized Duality Mapping

Searching arXiv for recent papers on normalized duality mappings and closely related duality-mapping work.
The normalized duality mapping is the duality mapping associated with the identity gauge $\varphi(r)=r$. In a real Banach space $X$ with dual $X^*$ and duality pairing $\langle x^*,x\rangle$, it is defined by
\[
Jx=\{x^*\in X^*:\langle x^*,x\rangle=\|x\|^2,\;\|x^*\|=\|x\|\}.
\]
It is a central object in nonlinear functional analysis because, in Hilbert spaces, $J$ is just the Riesz isomorphism, while in general Banach spaces it plays the role of a nonlinear “identity between $X$ and $X^*$” [2208.10689]. Recent work places the normalized duality mapping within a broader family of gauge-generated duality mappings $J_\varphi$, studies its quantitative monotonicity and its role in Yosida regularization [2208.10689], develops explicit formulas in matrix spaces equipped with Schatten norms [2009.07768], derives Gâteaux and partial Fréchet differentiability properties in $\ell^p$ [2606.15515], and analyzes its Mordukhovich coderivatives in $\ell^p$, $L^1$, and $C[0,1]$ [2407.21261].

## 1. Definition and normalization

Let $X$ be a real Banach space with dual $X^*$ and duality pairing $\langle x^*,x\rangle$. For a gauge function $\varphi:\mathbb{R}_+\to\mathbb{R}_+$ that is strictly increasing, continuous, satisfies $\varphi(0)=0$, and $\varphi(r)\to\infty$ as $r\to\infty$, the associated duality mapping is
\[
J_\varphi x=\Big\{x^*\in X^*:\langle x^*,x\rangle=\varphi(\|x\|)\|x\|,\;\|x^*\|=\varphi(\|x\|)\Big\},\quad x\in X.
\]
The normalized duality mapping is obtained by taking $\varphi(r)=r$, so that
\[
Jx=\{x^*\in X^*:\langle x^*,x\rangle=\|x\|^2,\;\|x^*\|=\|x\|\}
\]
[2208.10689].

A basic structural relation is
\[
J_\varphi x=\frac{\varphi(\|x\|)}{\|x\|}\,Jx,\qquad x\in X,\;x\neq 0,
\]
with the obvious interpretation at $x=0$ [2208.10689]. This shows that general duality mappings are radial rescalings of the normalized one. A plausible implication is that many properties proved first for $J$ can be transported to $J_\varphi$ once the geometry of $X$ is strong enough.

In finite-dimensional normed spaces, an equivalent normalization is often expressed through a dual pair $(u,v)$ satisfying both Hölder saturation and symmetric norm equality:
\[
\langle v,u\rangle=\|v\|_{X'}\|u\|_X,\qquad \|v\|_{X'}=\|u\|_X.
\]
Under this normalization one again has
\[
\langle u,v\rangle=\|u\|_X^2,\qquad \|v\|_{X'}=\|u\|_X
\]
[2009.07768]. The Schatten-space literature adopts this formulation explicitly, while the Banach-space papers use the standard $J$ notation.

## 2. Geometric structure and single-valuedness

The mapping $J$ is generally set-valued. Its single-valuedness and continuity are controlled by the geometry of $X$ and $X^*$. If $X^*$ is strictly convex, then $X$ is smooth, and the duality mapping $J_\varphi:X\to X^*$ is single-valued. If, in addition, both $X$ and $X^*$ are reflexive and strictly convex, then $J_\varphi$ is a bijection, and its inverse is the duality mapping on $X^*$ associated with $\varphi^{-1}$. If $X$ and $X^*$ are locally uniformly convex, then $J_\varphi$ is a homeomorphism between $X$ and $X^*$ [2208.10689].

The normalized duality mapping inherits these properties as the special case $\varphi(r)=r$. In the terminology recalled in the coderivative study, $J$ has the following standard structural properties: for each $x\in X$, $J(x)$ is nonempty, bounded, closed, and convex; $J(0)=\{0^*\}$; $J(ax)=aJ(x)$ for scalar $a\in\mathbb{R}$; and if $j(x)\in J(x)$ and $j(y)\in J(y)$, then
\[
(j(x)-j(y),x-y)\ge 0
\]
[2407.21261]. In smooth or uniformly smooth spaces, $J$ is continuous, and in uniformly smooth spaces it is uniformly continuous on bounded sets [2407.21261].

The Hilbert-space case is the canonical reference point. Under the Riesz identification, $J=\operatorname{Id}_H$ [2407.21261]. In $\ell^2$, the same statement appears as $J(x)=x$ [2606.15515]. This makes $J$ the Banach-space analogue of the gradient of the squared norm. The $\ell^p$ analysis states this explicitly: in Hilbert spaces, $J$ is the identity, i.e. $\nabla(\tfrac12\|x\|^2)=x$, whereas in $\ell^p$, $\nabla(\tfrac12\|x\|_p^2)$ is represented, up to identification with $\ell^q$, by $J(x)$ [2606.15515].

## 3. Quantitative monotonicity and the extension to general gauges

A principal structural result for general duality mappings is a local quantitative monotonicity inequality. If $X$ is a locally uniformly convex Banach space and $\varphi$ is any gauge function, then for every $R>0$ and every $x_0\in X$ there exists a nondecreasing function
\[
\psi=\psi(R,x_0):\mathbb{R}_+\to\mathbb{R}_+
\]
with $\psi(0)=0$ and $\psi(r)>0$ for $r>0$, such that
\[
\langle x^*-x_0^*,x-x_0\rangle\ge \psi(\|x-x_0\|)\,\|x-x_0\|
\]
for all $x$ with $\|x-x_0\|\le R$ and all $x^*\in J_\varphi x$, $x_0^*\in J_\varphi x_0$ [2208.10689].

For monotone operators, the basic inequality is only nonnegativity of the pairing. The inequality above gives more: it provides a positive lower bound depending on $\|x-x_0\|$ inside a fixed ball. The paper interprets this as a modulus of monotonicity for $J_\varphi$ [2208.10689]. In particular, it implies strict monotonicity: if $x\neq x_0$ and $\|x-x_0\|\le R$, then
\[
\langle x^*-x_0^*,x-x_0\rangle>0.
\]

A related global inequality is
\[
\langle J_\varphi u_1-J_\varphi u_0,\;u_1-u_0\rangle
\ge
\big(\varphi(\|u_1\|)-\varphi(\|u_0\|)\big)\big(\|u_1\|-\|u_0\|\big)\ge 0
\]
[2208.10689]. This connects the monotonicity of $J_\varphi$ to the gauge function itself.

When $\varphi(r)=r$, the result specializes exactly to the normalized duality mapping:
\[
\langle x^*-x_0^*,x-x_0\rangle\ge \psi(\|x-x_0\|)\,\|x-x_0\|
\]
for all $x$ in the ball and all $x^*\in Jx$, $x_0^*\in Jx_0$ [2208.10689]. The paper identifies this as precisely Theorem A of Kartsatos, itself extending an earlier result of Prüss from uniformly convex to locally uniformly convex Banach spaces. The novelty is not a new structural principle for $J$ itself, but the proof that the same principle holds for all $J_\varphi$ [2208.10689].

Under additional assumptions, one obtains a convergence consequence: if $X$ is smooth and locally uniformly convex and
\[
\langle J_\varphi x_n-J_\varphi x_0,\;x_n-x_0\rangle\to 0,
\]
then $x_n\to x_0$ in $X$ [2208.10689]. This is presented as one way of expressing that $J_\varphi$ is of type $(S_+)$.

## 4. Normalized duality mapping in $\ell^p$ and Schatten spaces

In $\ell^p$, with $1<p<\infty$ and $1/p+1/q=1$, the normalized duality mapping is single-valued and admits the coordinate formula
\[
J(0)=0,
\]
and for $x=(t_1,t_2,\dots)\in\ell^p\setminus\{0\}$,
\[
J(x)=\Big(\frac{|t_1|^{p-1}\operatorname{sign}(t_1)}{\|x\|_p^{\,p-2}},\,
\frac{|t_2|^{p-1}\operatorname{sign}(t_2)}{\|x\|_p^{\,p-2}},\,
\dots\Big),
\]
equivalently
\[
[J(x)]_n=\frac{|x_n|^{p-2}x_n}{\|x\|_p^{\,p-2}}
\]
[2606.15515]. This factor $\|x\|_p^{2-p}$ enforces the normalization
\[
\|J(x)\|_q=\|x\|_p,\qquad \langle J(x),x\rangle=\|x\|_p^2
\]
[2606.15515]. The adjoint mapping $J^*:\ell^q\to\ell^p$ has the analogous form with $p$ and $q$ interchanged, and the relation
\[
J^*\circ J=\operatorname{Id}_{\ell^p},\qquad J\circ J^*=\operatorname{Id}_{\ell^q}
\]
appears in the coderivative analysis [2407.21261].

The Schatten-space analogue is structurally parallel. For $V=\mathbb{R}^{m\times n}$ equipped with the trace inner product
\[
\langle A,B\rangle=\mathrm{Tr}(A^TB)
\]
and the Schatten–$p$ norm, Theorem 3.1 states that for $1<p<\infty$ the normalized duality mapping is single-valued and given by
\[
J_{S_p}(A)=U_r\,\operatorname{diag}(J_p(\boldsymbol{\sigma}))\,V_r^T,
\]
where $A=U_r\operatorname{diag}(\sigma_1,\dots,\sigma_r)V_r^T$ is the reduced SVD and $J_p$ is the vector duality mapping on the singular value vector $\boldsymbol{\sigma}$ [2009.07768]. The singular vectors are preserved, and the singular values are transformed by the scalar $\ell_p$-duality map. The normalization carries over exactly:
\[
\langle A,J_{S_p}(A)\rangle=\|A\|_{S_p}^2,\qquad \|J_{S_p}(A)\|_{S_q}=\|A\|_{S_p}
\]
[2009.07768].

For $p=1$, the Schatten duality mapping is set-valued because the dual norm $S_\infty$ lacks strict convexity. The paper introduces a rank-constrained sparse duality mapping
\[
J_{S_1,\operatorname{rank}}(A)=\|A\|_{S_1}\,U_rV_r^T,
\]
which is single-valued and Borel-measurable but not continuous [2009.07768]. This suggests that in non-strictly convex settings the normalized duality principle remains meaningful, but canonical single-valued selections may require an additional structural criterion.

## 5. Differentiability and coderivatives

The differentiability theory in $\ell^p$ shows a marked contrast between the ranges $2<p<\infty$ and $1<p<2$. For all $1<p<\infty$, the normalized duality mapping is Gâteaux differentiable at $0$, with
\[
J'(0)(v)=J(v)\quad\text{for all }v\in\ell^p\setminus\{0\}
\]
[2606.15515]. This already reflects the nonlinearity of $J$ at the origin.

For $2<p<\infty$, $J$ is Gâteaux differentiable at every $x\neq 0$. If $x=(t_n)\in\ell^p\setminus\{0\}$, $v=(s_n)\in\ell^p$, and $\lambda_n=0$ when $t_n=0$, otherwise $\lambda_n=s_n/t_n$, then
\[
J'(x)(v)=J(x)(p-1)D(\Lambda)-J(x)(p-2)\frac{\sum_{i=1}^{\infty}\lambda_i|t_i|^p}{\sum_{i=1}^{\infty}|t_i|^p}\,I
\]
[2606.15515]. Coordinatewise, for $n\in N(x)$,
\[
[J'(x)(v)]_n
=
(p-1)\lambda_n[J(x)]_n
-
(p-2)\Big(\frac{\sum_{i=1}^{\infty}\lambda_i|t_i|^p}{\sum_{i=1}^{\infty}|t_i|^p}\Big)[J(x)]_n.
\]
The paper emphasizes that the normalization introduces a rank-one correction depending on $\sum_i\lambda_i|x_i|^p$, unlike the simpler pure diagonal derivative of the non-normalized power map [2606.15515].

For $1<p<2$, directional differentiability is restricted. If $N(v)\not\subset N(x)$, then $J'(x,v)$ does not exist; if $N(v)\subset N(x)$ and $v$ is ratio bounded with respect to $x$, then the same derivative formula holds [2606.15515]. This is presented as a precise manifestation of the irregular behavior created by the exponent $p-2<0$.

The same paper proves a partial Fréchet differentiability statement: for $2<p<\infty$, if $x\in\ell^p\setminus\{0\}$ satisfies $1<|N(x)|<\infty$, then $J$ is Fréchet differentiable at $x$, and the Fréchet derivative coincides with the Gâteaux derivative formula [2606.15515]. The coderivative then agrees with the adjoint of the Fréchet derivative at such points.

A separate study of Mordukhovich derivatives in Banach spaces emphasizes a different feature: the coderivative of $J$ is often highly restrictive. In $\ell^p$, for all $x\in\ell^p$,
\[
D^*J(x)(0)=\{0\}
\]
[2407.21261]. If $(J(x),y)\neq 0$, then
\[
0\notin D^*J(x)(y),
\]
and for nonzero $x\in\ell^p$ and any $a>0$ with $a\neq 1$,
\[
aJ(x)\notin D^*J(x)(x)
\]
[2407.21261]. In $L^1(S)$ and $C[0,1]$, where $J$ is typically multi-valued, the coderivative is again described as extremely restrictive, frequently reducing to $\{0\}$ at natural embedded directions or excluding natural candidates such as scalar multiples of current dual values [2407.21261].

A plausible implication is that norm smoothness and coderivative richness are distinct phenomena: explicit first-order formulas for $J$ do not entail a large Mordukhovich derivative.

## 6. Resolvents, Yosida approximants, and topological degree

The normalized duality mapping is a standard device for building resolvents and Yosida approximants of maximal monotone operators in Banach spaces. For a maximal monotone operator $A:X\supset D(A)\to 2^{X^*}$, a gauge function $\varphi$, and $\lambda>0$, one considers
\[
0\in J_\varphi(x_\lambda-x)+\lambda A x_\lambda.
\]
Under the standing assumptions that $X$ is reflexive and $X$ and $X^*$ are strictly convex, this inclusion has a unique solution $x_\lambda\in D(A)$ [2208.10689]. The generalized resolvent and Yosida approximant are then defined by
\[
J_\lambda^\varphi x:=x_\lambda,\qquad
A_\lambda^\varphi x:=\frac1\lambda J_\varphi(x-J_\lambda^\varphi x)
\]
[2208.10689]. They satisfy
\[
A_\lambda^\varphi x\in A(J_\lambda^\varphi x),\qquad
x=J_\lambda^\varphi x+J_\varphi^{-1}(\lambda A_\lambda^\varphi x)
\]
[2208.10689].

In the normalized case $\varphi(r)=r$, this becomes
\[
x=J_\lambda x+\lambda J^{-1}(A_\lambda x),
\]
and one can rewrite
\[
A_\lambda=(A^{-1}+\lambda J^{-1})^{-1}x
\]
in the classical Banach-space form [2208.10689]. In Hilbert spaces, where $J$ is the Riesz map, the corresponding formulas reduce to the familiar resolvent $(I+\lambda A)^{-1}$ and Hilbert-space Yosida regularization [2208.10689].

The main continuity theorem in the general-gauge setting states that if $X$ is reflexive and both $X$ and $X^*$ are locally uniformly convex, then
\[
(\lambda,x)\mapsto A_\lambda^\varphi x,\qquad
(\lambda,x)\mapsto J_\lambda^\varphi x
\]
are continuous on $(0,\infty)\times X$ [2208.10689]. When $\varphi(r)=r$, this recovers earlier continuity results for the classical Yosida approximant and resolvent, but without relying on an explicit formula [2208.10689]. The normalized duality mapping thus appears as the prototype of a broader Yosida framework.

The topological implication discussed in the same work concerns the Browder degree. For operators of the form $A+F$, with $A$ maximal monotone and $F$ an $(S_+)$-mapping, the degree is defined using Yosida approximants:
\[
\deg_B(A+F,G,p):=\lim_{\lambda\to 0^+} d_{(S_+)}(A_\lambda+F,G,p).
\]
The paper shows that if one replaces the normalized duality mapping by any $J_\varphi$, the family $A_\lambda^\varphi$ forms a pseudomonotone homotopy as $\lambda$ varies, and the resulting degree coincides with the classical Browder degree [2208.10689]. Conceptually, this means that the normalized duality mapping is not topologically privileged, even though it remains the standard model.

## 7. Interpretations, special cases, and limitations

Several recurring themes organize the modern theory of normalized duality mappings.

First, $J$ is best understood as the prototype within the larger family $\{J_\varphi\}$. The formula
\[
J_\varphi x=\frac{\varphi(\|x\|)}{\|x\|}Jx
\]
shows that normalized duality is the base case from which gauge-adapted variants are obtained by radial rescaling [2208.10689]. This is particularly useful when one wants Yosida approximants and resolvents adapted to non-Hilbertian or non-power-type geometries [2208.10689].

Second, the distinction between strictly convex and non-strictly convex settings is decisive. For $1<p<\infty$, Schatten spaces are strictly convex, and $J_{S_p}$ is continuous and single-valued [2009.07768]. For $p=1$ or $p=\infty$, the duality mapping becomes set-valued [2009.07768]. Likewise, in general Banach-space terms, $X^*$ strictly convex implies single-valuedness of $J$, while nonsmooth spaces such as $L^1$ and $C[0,1]$ admit rich multivalued behavior [2407.21261].

Third, first-order regularity does not imply strong metric regularity. In $\ell^p$, explicit Gâteaux derivatives exist for $2<p<\infty$, and partial Fréchet differentiability is established under finite-support assumptions [2606.15515]. Yet the coderivative-based covering constant is zero in the cases analyzed:
\[
\hat{\alpha}(J,x,y)=0
\]
for $2<p<\infty$ and finite-support $x$ with $y=J(x)$ [2606.15515]. The paper describes this as showing that the normalized duality mapping has no positive covering modulus even where it is Fréchet differentiable.

Finally, several limitations remain explicit in the literature. The differentiability and covering-constant analysis in $\ell^p$ is carried out in real spaces, not complex ones [2606.15515]. For the normalized mapping, full Gâteaux differentiability is obtained for $2<p<\infty$, while for $1<p<2$ only restricted directional differentiability is proved [2606.15515]. Fréchet differentiability of normalized $J$ is proved only under the strong assumption $|N(x)|<\infty$, and the paper formulates as an open problem whether for $2<p<\infty$ the same Fréchet derivative formula holds at every nonzero $x$ [2606.15515].

Taken together, these results locate the normalized duality mapping at the intersection of Banach-space geometry, monotone operator theory, and variational analysis. It is the canonical normalized correspondence between a point and its norm-attaining supporting functionals; it is the prototype from which general duality mappings are generated; and it remains central both for explicit model calculations, such as those in $\ell^p$ and Schatten spaces, and for abstract constructions such as Yosida approximation and Browder degree [2208.10689; 2009.07768; 2606.15515; 2407.21261].

Source: https://www.emergentmind.com/topics/normalized-duality-mapping