betaFeedback

WorksMathematics

Hadamard inverse mapping theorem for certain continuously Gateaux differentiable mappings

Abstract

Hadamard inverse mapping theorem concerns global inversion of continuously differentiable mappings between Banach spaces. Here we consider a weaker notion of smoothness and work with continuously Gateaux differentiable mappings. We prove that a continuously Gateaux differentiable mapping with invertible derivatives such that the inverses are uniformly bounded, possesses a continuous right inverse on every compact set. We also provide sufficient conditions for global invertibility of certain continuously Gateaux differentiable mappings between Banach spaces.

Introduction

Invertibility of mappings is a cornerstone in Mathematical Analysis. Let XX and YY be Banach spaces and let f:X→Yf : X \to Y be a C1C^{1}-smooth mapping whose derivative f′(x)f'(x) is an isomorphism for every x∈Xx \in X. Then the classical Inverse Mapping Theorem implies that ff is locally invertible around each point. If, in addition, ff satisfies the condition

∥[f′(x)]−1∥≤M,∀x∈X,\left\|[f'(x)]^{-1}\right\| \le M,\qquad\forall x \in X,

then ff is globally C1C^{1} invertible on XX, and hence a global diffeomorphism of XX onto YY. In other words, there is a C1C^{1} function gg such that

g(f(x))=f(g(x))=x,∀x∈X.g(f(x)) = f(g(x)) = x,\qquad\forall x \in X.

This condition is due to Hadamard [10] for mappings between finite-dimensional spaces, and to Lévy [15] for mappings between Hilbert spaces and it is known as the Hadamard-Lévy condition. Detailed information about this and other conditions for global inversion of smooth mappings between Banach spaces can be found, for instance, in [17, 8]. The result has been extended in several directions over the years. Some authors weaken the C1C^{1} assumption on the mapping using a pseudo-Jacobian, Clarke subdifferential concept instead, see [18, 19, 9]; others consider set-valued maps [16, 2] or more general underlying spaces [6, 4]; a local surjection theorem with continuous inverse is obtained in [7], etc.

We are interested in the global inversion of mappings between Banach spaces under smoothness assumptions weaker than C1C^{1}. Mappings called continuously Gâteaux differentiable, written CG1C^{1}_{G}, seem to be good candidates. In contrast to the C1C^{1} case, a CG1C^{1}_{G} mapping may fail to be invertible on any neighborhood of a point at which its derivative is invertible, see Example 4.1. Hence, even local invertibility of a CG1C^{1}_{G} mapping f:X→Yf:X\to Y around a fixed point xx needs an additional requirement. One such condition is

lim sup⁡x′→x∥f′(x′)−f′(x)∥<∥[f′(x)]−1∥−1.\limsup_{x'\to x}\lVert f'(x')-f'(x)\rVert< \lVert[f'(x)]^{-1}\rVert^{-1}.

It appears in [14 Theorem 10] and shows that a certain controlled discontinuity of the Gâteaux derivatives allows local invertibility.

The paper is organized as follows. § 2 contains the necessary definitions and preliminaries. In § 3 we prove that a mapping f:X→Yf:X\to Y which is CG1C^{1}_{G} on XX, has invertible derivatives and satisfies the Hadamard-Lévy condition is surjective. § 4 is devoted to local invertibility of CG1C^{1}_{G} mappings. In § 5 we prove our main result on the global invertibility of CG1C^{1}_{G} mappings.

Preliminaries and notation

Let (X,∥⋅∥X)(X,\lVert\cdot\rVert_{X}) and (Y,∥⋅∥Y)(Y,\lVert\cdot\rVert_{Y}) be Banach spaces. In XX we denote the unit sphere by SXS_{X}, and the closed (resp. open) unit ball by BXB_{X} (resp. BX∘B_{X}^{\circ}). For a non-empty set A⊆XA\subseteq X we denote by C(A,Y)C(A,Y) the space of all norm-to-norm continuous mappings f:A→Yf:A\to Y. It is well-known that, equipped with the supremum norm

∥f∥A:=sup⁡{∥f(x)∥Y:x∈A},\lVert f\rVert_{A}:=\sup\{\lVert f(x)\rVert_{Y}:x\in A\},

the space (C(A,Y),∥⋅∥A)(C(A,Y),\|\cdot\|_A) is a Banach space provided that AA is compact.

A mapping f:X→Yf:X\to Y has a directional derivative at a point x∈Xx\in X in the direction h∈Xh\in X whenever the limit

f′(x,h):=lim⁡t→0f(x+th)−f(x)t(1)f'(x,h):=\lim_{t\to0}\frac{f(x+th)-f(x)}{t} \tag*{(1)}

exists. The mapping ff is said to be Gâteaux differentiable at xx whenever the directional derivatives f′(x,h)f'(x,h) exist for all h∈Xh\in X and are bounded linear operators, denoted by f′(x,h)=f′(x)hf'(x,h)=f'(x)h. The norm of the linear map f′(x):X→Yf'(x):X\to Y is the standard linear operator norm ∥f′(x)∥:=sup⁡{∥f′(x)h∥Y:∥h∥X≤1}\|f'(x)\|:=\sup\{\|f'(x)h\|_Y:\|h\|_X\le1\}.

A mapping f:X→Yf:X\to Y is called Fréchet differentiable at x∈Xx\in X if there is a bounded linear operator f′(x):X→Yf'(x):X\to Y such that

lim⁡∥h∥X→0f(x+h)−f(x)−f′(x)h∥h∥X=0,\lim_{\|h\|_X\to0}\frac{f(x+h)-f(x)-f'(x)h}{\|h\|_X}=0,

and it is called smooth, denoted f∈C1f\in C^1, if the mapping x→f′(x)x\to f'(x) is norm-to-norm continuous.

A mapping f:U→Yf:U\to Y is said to be continuously Gâteaux differentiable on an open set U⊆XU\subseteq X, or ff is CG1C^1_G on UU, if ff is continuous on UU, the directional derivatives f′(x,h)f'(x,h) exist for all x∈Ux\in U and h∈Xh\in X and f′(⋅,⋅):U×X→Yf'(\cdot,\cdot):U\times X\to Y is a jointly continuous mapping on the product U×XU\times X. As noted at [11 p.73] this notion does not agree with the usual definition of a C1C^1-smooth mapping in the Fréchet sense. If ff is CG1C^1_G on UU, then the derivative will of course be the same linear map, but the continuity requirement is weaker. It is well known, see e.g. [11 Lemma 3.2.3, Theorem 3.2.5], that if ff is CG1C^1_G on UU, then for any x∈Ux\in U, ff is Gâteaux differentiable at xx. It is worth mentioning that for any x0∈Ux_0\in U there exist δ>0\delta>0 and L>0L>0 such that ∥f′(x)(h)∥Y≤L∥h∥X\|f'(x)(h)\|_Y\le L\|h\|_X for all x∈{x0}+δBX∘x\in\{x_0\}+\delta B_X^\circ and h∈Xh\in X, so ff is locally Lipschitz on UU [11 Theorem 5.1.1].

If ff is CG1C^1_G on XX, then ∀x∈X, ∀h∈X\forall x\in X,\ \forall h\in X,

f(x+th)−f(x)t=∫01f′(x+uth)h du,(2)\frac{f(x+th)-f(x)}{t}=\int_0^1 f'(x+uth)h\,du, \tag*{(2)}

see [11 Lemma 3.2.4].

In the sequel we will make essential use of the following auxiliary result.

Lemma 2.1. Let f:X→Yf:X\to Y be CG1C^1_G on XX. Let K⊆XK\subseteq X be a non-empty compact set, E⊆XE\subseteq X be a finite dimensional subspace and let r>0r>0. Then there exists a function α:[−1,1]→R+\alpha:[-1,1]\to\mathbb{R}^+ such that α(t)→0\alpha(t)\to0 as ∣t∣→0|t|\to0, and

sup⁡{∥f(x+th)−f(x)−tf′(x)h∥Y:x∈K, h∈rBE}≤∣t∣α(t).\sup\{\|f(x+th)-f(x)-tf'(x)h\|_Y:x\in K,\ h\in rB_E\}\le|t|\alpha(t).

Proof. Let t∈Rt \in\mathbb{R} be such that ∣t∣≤1|t| \le1. Using (2), we have

∥f(x+th)−f(x)−tf′(x)h∥Y=∥t∫01[f′(x+uth)h−f′(x)h] du∥Y≤∣t∣∫01∥f′(x+uth)h−f′(x)h∥Y du.\begin{aligned} \left\|f(x+th)-f(x)-tf'(x)h\right\|_Y = \left\|t\int_0^1\left[f'(x+uth)h-f'(x)h\right]\,du\right\|_Y \\ &\le|t|\int_0^1\left\|f'(x+uth)h-f'(x)h\right\|_Y\,du. \end{aligned}

Set

α(t):=sup⁡{∫01∥f′(x+uth)h−f′(x)h∥Y du:(x,h)∈K×rBE}.\alpha(t):=\sup\left\{\int_0^1\left\|f'(x+uth)h-f'(x)h\right\|_Y\,du:(x,h)\in K\times rB_E\right\}.

Fix ε>0\varepsilon>0. By uniform continuity of the continuous mapping f′:X×X→Yf':X\times X\to Y on the compact (K+rBE)×rBE(K+rB_E)\times rB_E, there is δ>0\delta>0 such that ∀x1,x2∈K+rBE\forall x_1,x_2\in K+rB_E and ∀h1,h2∈rBE\forall h_1,h_2\in rB_E satisfying ∥x1−x2∥X+∥h1−h2∥X<δ\|x_1-x_2\|_X+\|h_1-h_2\|_X<\delta, it holds that

∥f′(x1)h1−f′(x2)h2∥Y<ε.\left\|f'(x_1)h_1-f'(x_2)h_2\right\|_Y<\varepsilon.

For ∣u∣≤1|u|\le1, ∣t∣<1|t|<1 and (x,h)∈K×rBE(x,h)\in K\times rB_E it follows that (x+uth,h)∈(K+rBE)×rBE(x+uth,h)\in(K+rB_E)\times rB_E. Then, ∀∣t∣<min⁡{1,δ/r}\forall|t|<\min\{1,\delta/r\} we get

∥f′(x+uth)h−f′(x)h∥Y<ε,\left\|f'(x+uth)h-f'(x)h\right\|_Y<\varepsilon,

which implies α(t)<ε\alpha(t)<\varepsilon. Therefore, α(t)→0\alpha(t)\to0 as ∣t∣→0|t|\to0. □

The next lemma is a precursor to the Ekeland Variational Principle, see [1 Chapter 5, Section 1]. See also the comments concerning the Basic Lemma on [12 p. 93]. It easily follows from the Ekeland Variational Principle itself, see e.g. [12 Basic Lemma], and is a partial case of Takahashi’s Nonconvex Minimization Theorem [20]. Here we state it in the form given in [13].

Lemma 2.2. [13 Lemma 3] Let (X,∥⋅∥X)(X,\|\cdot\|_X) be a Banach space and μ:X→R+∪{∞}\mu:X\to\mathbb{R}_+\cup\{\infty\} be a lower semicontinuous proper function such that for some r>0r>0

∀x:0<μ(x)<∞⇒∃y:μ(y)<μ(x)−r∥y−x∥X.\forall x: 0<\mu(x)<\infty\Rightarrow\exists y:\mu(y)<\mu(x)-r\|y-x\|_X.

Then, for each x∈dom⁡μx\in\operatorname{dom}\mu there is y∈Xy\in X such that

μ(y)=0andr∥y−x∥X≤μ(x).\mu(y)=0\quad\text{and}\quad r\|y-x\|_X\le\mu(x).

Surjectivity

We will first prove that any f:X→Yf:X\to Y which is CG1C^1_G on XX and possesses invertible Gâteaux derivatives satisfying the Hadamard-Lévy condition has a continuous right inverse on every compact set in YY. In Theorem 3.2 we will use this result to prove the surjectivity of ff.

Proposition 3.1. Let (X,∥⋅∥X)(X,\|\cdot\|_X) and (Y,∥⋅∥Y)(Y,\|\cdot\|_Y) be Banach spaces. Let the mapping f:X→Yf:X\to Y be CG1C^1_G on XX and possesses an invertible Gâteaux derivative at every point, that is ∀x∈X\forall x\in X there exists [f′(x)]−1:Y→X[f'(x)]^{-1}:Y\to X. Let the Hadamard-Lévy condition

∥[f′(x)]−1∥≤M,∀x∈X(3)\|[f'(x)]^{-1}\|\le M,\quad\forall x\in X \tag*{(3)}

hold.

Then, for every non-empty compact K⊆YK\subseteq Y one can find a continuous mapping g∈C(Y,X)g\in C(Y,X) which is a right inverse of ff on KK, that is

f(g(y))=y,∀y∈K.f(g(y))=y,\quad\forall y\in K.

Furthermore, if 0∈K0\in K and f(0)=0f(0)=0, then g(0)=0g(0)=0.

Proof. Let us fix a non-empty compact set K⊆YK\subseteq Y and a mapping g∈C(Y,X)g\in C(Y,X). If f(g(y))=yf(g(y))=y on KK, there is nothing to prove. If not, then there is some yˉ∈K\bar{y}\in K with f(g(yˉ))≠yˉf(g(\bar{y}))\ne\bar{y}. Then the mapping u:Y→Yu:Y\to Y defined by u(y):=y−f(g(y))u(y):=y-f(g(y)) is continuous on YY and such that u≢0u\not\equiv0. In the Banach space (C(K,X),∥⋅∥K)(C(K,X),\|\cdot\|_K) consider the function

μ(g):=∥u∥K=sup⁡y∈K∥u(y)∥Y.\mu(g):=\|u\|_K=\sup_{y\in K}\|u(y)\|_Y.

Obviously, μ:C(K,X)→R+\mu:C(K,X)\to\mathbb{R}^{+} is a lower semicontinuous function such that μ(g)>0\mu(g)>0. Roughly speaking, the function tells us how ”far” the mapping g∈C(K,X)g\in C(K,X) is from being a right inverse of ff on KK. Therefore, our task is to find a continuous mapping gˉ\bar{g} from KK to XX at which μ\mu vanishes. We will accomplish this in several steps.

Step 1. For any fixed y∈Yy\in Y we consider the minimization problem

(P)min⁡w∈X∥f′(g(y))w−u(y)∥Y.(P)\qquad\min_{w\in X}\|f'(g(y))w-u(y)\|_Y.

It is clear that a solution to (P)(P) is

w(y):=[f′(g(y))]−1u(y)(4)w(y):=[f'(g(y))]^{-1}u(y) \tag*{(4)}

at which (P)(P) attains its minimal value zero. Furthermore, as f′(x)f'(x) is injective, w(y)w(y) is the unique solution. In this way, one generates a single valued mapping w:Y→Xw:Y\to X, defined by (3.2). We should note that ww is not continuous in general. However, for any fixed y∈Yy\in Y the function

∥f′(g(⋅))w(y)−u(⋅)∥Y\|f'(g(\cdot))w(y)-u(\cdot)\|_Y

is continuous on YY and vanishes at yy. Now, take ε>0\varepsilon> 0 such that

ε<μ(g)/4.(5)\varepsilon< \mu(g)/4. \tag*{(5)}

For every fixed y∈Yy \in Y, by continuity of ∥f′(g(⋅))w(y)−u(⋅)∥Y\lVert f'(g(\cdot))w(y)-u(\cdot)\rVert_Y at yy, there exists δy>0\delta_y > 0 such that y~∈y+δyBY∘\widetilde{y} \in y+\delta_y B_Y^\circ we have

∥f′(g(y~))w(y)−u(y~)∥Y<ε/3.(6)\lVert f'(g(\widetilde{y}))w(y)-u(\widetilde{y})\rVert_Y < \varepsilon/3. \tag*{(6)}

For the compact set K\mathcal{K} we have an open cover

K⊆⋃y∈K({y}+δyBY∘).\mathcal{K} \subseteq\bigcup_{y\in\mathcal{K}} \left(\{y\}+\delta_y B_Y^\circ\right).

from which we can take a finite subcover

K⊆⋃i=1n({yi}+δyiBY∘)\mathcal{K} \subseteq\bigcup_{i=1}^{n} \left(\{y_i\}+\delta_{y_i}B_Y^\circ\right)

for some {y1,…,yn}⊆K\{y_1,\ldots,y_n\}\subseteq\mathcal{K}. Set wi:=w(yi)∈Xw_i:=w(y_i)\in X for i=1,…,ni=1,\ldots,n, and consider the finite dimensional subspace W:=span⁡{w1,…,wn}W:=\operatorname{span}\{w_1,\ldots,w_n\} in XX. Let ∣⋅∣W\lvert\cdot\rvert_W be a Euclidean norm on WW. Set L:=max⁡{∣wi∣W2,i=1,…,n}>0L:=\max\{\lvert w_i\rvert_W^2,i=1,\ldots,n\}>0. Note that if wi=0w_i=0 for all ii then (3.4) gives μ(g)=∥u∥K<ε/3\mu(g)=\lVert u\rVert_{\mathcal{K}}<\varepsilon/3 which contradicts (3.3).

Next, for any fixed y∈Yy\in Y, we define the function py:W→Rp_y:W\to\mathbb{R} by

py(w):=∥f′(g(y))w−u(y)∥Y+ε3L∣w∣W2,p_y(w):=\lVert f'(g(y))w-u(y)\rVert_Y+\frac{\varepsilon}{3L}\lvert w\rvert_W^2,

and consider the minimization problem

(Py)min⁡w∈Wpy(w).(P_y)\qquad\min_{w\in W}p_y(w).

Claim 1. For any fixed y∈Yy\in Y, the corresponding function py:W→Rp_y:W\to\mathbb{R} is coercive and strictly convex.

Proof. The function pyp_y is coercive, as

py(w)≥ε3L∣w∣W2→+∞ when ∣w∣W→∞,p_y(w)\geq\frac{\varepsilon}{3L}\lvert w\rvert_W^2\to+\infty\ \text{when }\lvert w\rvert_W\to\infty,

and pyp_y is strictly convex because the norm ∥⋅∥Y\lVert\cdot\rVert_Y is convex and the squared Euclidean norm ∣⋅∣W2\lvert\cdot\rvert_W^2 is strictly convex.

Therefore, for any y∈Yy \in Y, the function py:W→Rp_y: W \to\mathbb{R} attains its minimum on WW at a single point which is the unique solution to the problem (Py)(P_y). Let us denote this solution by z(y)z(y). Obviously, this defines a single-valued mapping z:Y→Wz: Y \to W.

Claim 2. The set z(K)z(\mathcal{K}) – the image of K\mathcal{K} under zz – is a bounded set in WW, and py(z(y))<2ε/3p_y(z(y)) < 2\varepsilon/3, ∀y∈K\forall y \in\mathcal{K}.

Proof. Since for any w∈Ww \in W,

∥f′(g(y))z(y)−u(y)∥Y+ε3L∣z(y)∣W2≤∥f′(g(y))w−u(y)∥Y+ε3L∣w∣W2.\left\|f'(g(y))z(y)-u(y)\right\|_Y+\frac{\varepsilon}{3L}|z(y)|_W^2\leq\left\|f'(g(y))w-u(y)\right\|_Y+\frac{\varepsilon}{3L}|w|_W^2.

taking w=0w=0 we get

ε3L∣z(y)∣W2≤∥u(y)∥Y.\frac{\varepsilon}{3L}|z(y)|_W^2\leq\|u(y)\|_Y.

Since the mapping u(⋅)u(\cdot) is continuous, it is bounded on the compact K\mathcal{K}, and the latter implies boundedness of zz on K\mathcal{K}.

Let y∈Ky\in\mathcal{K}. Then there exists yi∈Ky_i\in\mathcal{K} such that y∈yi+δyiBY∘y\in y_i+\delta_{y_i}B_Y^\circ. Recalling that wi=w(yi)∈Ww_i=w(y_i)\in W, using (3.4) and the definition of LL, we get

py(z(y))=min⁡w∈Wpy(w)=min⁡w∈W{∥f′(g(y))w−u(y)∥Y+ε3L∣w∣W2}≤∥f′(g(y))wi−u(y)∥Y+ε3L∣wi∣W2<ε3+ε3LL=2ε3.\begin{aligned} p_y(z(y))&=\min_{w\in W}p_y(w)=\min_{w\in W}\left\{\left\|f'(g(y))w-u(y)\right\|_Y+\frac{\varepsilon}{3L}|w|_W^2\right\}\\ &\leq\left\|f'(g(y))w_i-u(y)\right\|_Y+\frac{\varepsilon}{3L}|w_i|_W^2<\frac{\varepsilon}{3}+\frac{\varepsilon}{3L}L=\frac{2\varepsilon}{3}. \end{aligned}

Claim 3. The mapping z:Y→Wz: Y \to W is continuous on YY.

Proof. Let {ξn}n=1∞\{\xi_n\}_{n=1}^{\infty} be a convergent sequence of points in YY with limit point ξ0\xi_0. We have to show that z(ξn)→z(ξ0)z(\xi_n)\to z(\xi_0) as nn tends to infinity. Since for any w∈Ww\in W, pξn(z(ξn))≤pξn(w)p_{\xi_n}(z(\xi_n))\leq p_{\xi_n}(w), i.e.

∥f′(g(ξn))z(ξn)−u(ξn)∥Y+ε3L∣z(ξn)∣W2≤∥f′(g(ξn))w−u(ξn)∥Y+ε3L∣w∣W2,∀w∈W,(7)\left\|f'(g(\xi_n))z(\xi_n)-u(\xi_n)\right\|_Y+\frac{\varepsilon}{3L}|z(\xi_n)|_W^2 \leq\left\|f'(g(\xi_n))w-u(\xi_n)\right\|_Y+\frac{\varepsilon}{3L}|w|_W^2,\qquad\forall w\in W, \tag*{(7)}

taking w=0w=0, we get

ε3L∣z(ξn)∣W2≤∥u(ξn)∥Y\frac{\varepsilon}{3L}|z(\xi_n)|_W^2\leq\|u(\xi_n)\|_Y

which by continuity of u(⋅)u(\cdot) implies that {z(ξn)}n\{z(\xi_n)\}_n is a norm bounded sequence in the finite dimensional space WW. Hence, we can take a convergent subsequence, z(ξnj)→zˉz(\xi_{n_j})\to\bar z as j→∞j\to\infty. Fixing w∈Ww \in W and using the joint continuity of f′(⋅,⋅)f'(\cdot,\cdot), and continuity of g(⋅)g(\cdot) and u(⋅)u(\cdot), passing to the limit in (3.5) as j→∞j \to\infty we obtain that

∥f′(g(ξ0))zˉ−u(ξ0)∥Y+ε3L∣zˉ∣W2≤∥f′(g(ξ0))w−u(ξ0)∥Y+ε3L∣w∣W2,∀w∈W.\left\|f'(g(\xi_0))\bar{z}-u(\xi_0)\right\|_Y+\frac{\varepsilon}{3L}|\bar{z}|_W^2\leq\left\|f'(g(\xi_0))w-u(\xi_0)\right\|_Y+\frac{\varepsilon}{3L}|w|_W^2,\qquad\forall w\in W.

Therefore,

∥f′(g(ξ0))zˉ−u(ξ0)∥Y+ε3L∣zˉ∣W2≤inf⁡w∈W{∥f′(g(ξ0))w−u(ξ0)∥Y+ε3L∣w∣W2}.\left\|f'(g(\xi_0))\bar{z}-u(\xi_0)\right\|_Y+\frac{\varepsilon}{3L}|\bar{z}|_W^2\leq\inf_{w\in W}\left\{\left\|f'(g(\xi_0))w-u(\xi_0)\right\|_Y+\frac{\varepsilon}{3L}|w|_W^2\right\}.

The latter means that zˉ\bar{z} is a solution to (Pξ0)(P_{\xi_0}), and since the solution is unique, zˉ≡z(ξ0)\bar{z}\equiv z(\xi_0). As this is true for every choice of a convergent subsequence of {z(ξn)}n=1∞\{z(\xi_n)\}_{n=1}^{\infty}, it has a single accumulation point and the continuity of zz follows.

Step 2. For t∈Rt\in\mathbb{R} we will consider the mappings gt:Y→Xg_t:Y\to X defined by gt(y)=g(y)+tz(y)g_t(y)=g(y)+tz(y) for y∈Yy\in Y. By Claim 3 it is clear that for every t∈Rt\in\mathbb{R}, gt∈C(Y,X)g_t\in C(Y,X). We denote ut(y):=y−f(gt(y))u_t(y):=y-f(g_t(y)) and for any fixed y∈Ky\in\mathcal{K}, we consider the function φy:R→R+\varphi_y:\mathbb{R}\to\mathbb{R}^+ defined by

φy(t):=∥y−f(gt(y))∥Y.\varphi_y(t):=\left\|y-f(g_t(y))\right\|_Y.

Then, for the function μ\mu at gtg_t we have

μ(gt)=max⁡y∈K{φy(t)}=max⁡{∥ut(y)∥Y:y∈K}=∥ut∥K.\mu(g_t)=\max_{y\in\mathcal{K}}\{\varphi_y(t)\}=\max\{\left\|u_t(y)\right\|_Y:y\in\mathcal{K}\}=\left\|u_t\right\|_{\mathcal{K}}.

We apply Lemma 2.1 to ff which by assumption is CG1C_G^1 on XX, the compact set K:=g(K)K:=g(\mathcal{K}) in XX (as a continuous image of a compact set), the finite dimensional space E:=W⊆XE:=W\subseteq X, and r:=sup⁡{∥z(y)∥X:y∈K}r:=\sup\{\left\|z(y)\right\|_X:y\in\mathcal{K}\} (which is finite, as z(K)z(\mathcal{K}) is bounded by Claim 2) to obtain a function α(t)→0\alpha(t)\to0 as ∣t∣→0|t|\to0 such that for ∣t∣≤1|t|\leq1,

max⁡y∈K{∥f(gt(y))−f(g(y))−tf′(g(y))z(y)∥Y}≤∣t∣α(t).(8)\max_{y\in\mathcal{K}}\left\{\left\|f(g_t(y))-f(g(y))-tf'(g(y))z(y)\right\|_Y\right\}\leq|t|\alpha(t). \tag*{(8)}

Let us denote for simplicity q(y):=f′(g(y))z(y)q(y):=f'(g(y))z(y), thus q∈C(Y,Y)q\in C(Y,Y). For any fixed y∈Ky\in\mathcal{K}, by (3.6) we have that for t∈[0,1]t\in[0,1],

φy(t)=∥y−f(gt(y))∥Y=∥y−tf′(g(y))z(y)−f(g(y))−[f(gt(y))−f(g(y))−tf′(g(y))z(y)]∥Y≤∥y−tf′(g(y))z(y)−f(g(y))∥Y+∥[f(gt(y))−f(g(y))−tf′(g(y))z(y)]∥Y≤∥u(y)−tq(y)∥Y+tα(t).\begin{aligned} \varphi_y(t)&=\left\|y-f(g_t(y))\right\|_Y\\ &=\left\|y-tf'(g(y))z(y)-f(g(y))-\left[f(g_t(y))-f(g(y))-tf'(g(y))z(y)\right]\right\|_Y\\ &\leq\left\|y-tf'(g(y))z(y)-f(g(y))\right\|_Y+\left\|\left[f(g_t(y))-f(g(y))-tf'(g(y))z(y)\right]\right\|_Y\\ &\leq\left\|u(y)-tq(y)\right\|_Y+t\alpha(t). \end{aligned}

On the other hand,

∥u(y)−tq(y)∥Y=∥u(y)−tu(y)+t[u(y)−q(y)]∥Y=∥(1−t)u(y)+t[u(y)−q(y)]∥Y≤(1−t)∥u(y)∥Y+t∥u(y)−q(y)∥Y.\begin{aligned} \lVert u(y)-tq(y)\rVert_Y &= \lVert u(y)-tu(y)+t[u(y)-q(y)]\rVert_Y\\ &= \lVert(1-t)u(y)+t[u(y)-q(y)]\rVert_Y\\ &\le(1-t)\lVert u(y)\rVert_Y+t\lVert u(y)-q(y)\rVert_Y. \end{aligned}

Furthermore,

∥u(y)−q(y)∥Y=∥f′(g(y))z(y)+f(g(y))−y∥Y≤∥f′(g(y))z(y)+f(g(y))−y∥Y+ε3L∣z(y)∣W2=py(z(y))<2ε/3<ε,(9)\begin{aligned} \lVert u(y)-q(y)\rVert_Y &= \lVert f'(g(y))z(y)+f(g(y))-y\rVert_Y\\ &\le\lVert f'(g(y))z(y)+f(g(y))-y\rVert_Y+\frac{\varepsilon}{3L}\lvert z(y)\rvert_W^2 \\ &= p_y(z(y))<2\varepsilon/3<\varepsilon, \tag*{(9)} \end{aligned}

by Claim 2. Gathering all these yields

φy(t)≤(1−t)∥u(y)∥Y+tε+tα(t),∀y∈K, ∀t∈[0,1].\varphi_y(t)\le(1-t)\lVert u(y)\rVert_Y+t\varepsilon+t\alpha(t),\quad\forall y\in\mathcal{K},\ \forall t\in[0,1].

Since α(t)→0\alpha(t)\to0, as ∣t∣→0\lvert t\rvert\to0, α(t)<μ(g)/4\alpha(t)<\mu(g)/4 for sufficiently small t>0t>0. Then for such tt,

max⁡y∈K{φy(t)}≤(1−t)∥u∥K+tα(t)+tε,\max_{y\in\mathcal{K}}\{\varphi_y(t)\}\le(1-t)\lVert u\rVert_{\mathcal{K}}+t\alpha(t)+t\varepsilon,

which implies

μ(gt)≤(1−t)μ(g)+tα(t)+tε=μ(g)+t[α(t)+ε−μ(g)].\mu(g_t)\le(1-t)\mu(g)+t\alpha(t)+t\varepsilon=\mu(g)+t[\alpha(t)+\varepsilon-\mu(g)].

In addition, by the choice of ε\varepsilon, see (3.3), we have that α(t)+ε<μ(g)/2\alpha(t)+\varepsilon<\mu(g)/2, thus α(t)+ε−μ(g)<−μ(g)/2\alpha(t)+\varepsilon-\mu(g)<-\mu(g)/2 and finally, for sufficiently small t>0t>0,

μ(gt)≤μ(g)−tμ(g)/2.(10)\mu(g_t)\le\mu(g)-t\mu(g)/2. \tag*{(10)}

Since the condition (3.1) holds for ff we have ∥[f′(g(y))]−1∥≤M\lVert[f'(g(y))]^{-1}\rVert\le M for all y∈Yy\in Y. It remains to observe that, for any y∈Ky\in\mathcal{K}, using (3.7),

∥z(y)∥X=∥[f′(g(y))]−1q(y)∥X≤∥[f′(g(y))]−1∥∥q(y)∥Y≤M∥q(y)∥Y≤M∥q(y)+u(y)−u(y)∥Y≤M(∥q(y)−u(y)∥Y+∥u(y)∥Y)≤M(ε+∥u(y)∥Y).\begin{aligned} \lVert z(y)\rVert_X &=\lVert[f'(g(y))]^{-1}q(y)\rVert_X\\ &\le\lVert[f'(g(y))]^{-1}\rVert\lVert q(y)\rVert_Y\\ &\le M\lVert q(y)\rVert_Y\\ &\le M\lVert q(y)+u(y)-u(y)\rVert_Y\\ &\le M(\lVert q(y)-u(y)\rVert_Y+\lVert u(y)\rVert_Y)\\ &\le M(\varepsilon+\lVert u(y)\rVert_Y). \end{aligned}

Therefore, by the choice of ε\varepsilon, see (3.3),

∥z∥K=max⁡y∈K{∥z(y)∥X}≤M(ε+max⁡y∈K{∥u(y)∥Y})<M(μ(g)/4+μ(g))=(5M/4)μ(g).\begin{aligned} \lVert z\rVert_{\mathcal{K}} &=\max_{y\in\mathcal{K}}\{\lVert z(y)\rVert_X\}\\ &\le M(\varepsilon+\max_{y\in\mathcal{K}}\{\lVert u(y)\rVert_Y\})\\ &<M(\mu(g)/4+\mu(g))\\ &=(5M/4)\mu(g). \end{aligned}

This means that −μ(g)<−(4/(5M))∥z∥K-\mu(g)<-(4/(5M))\|z\|_{\mathcal K}, which substituted into (10) gives that for t>0t>0 small enough,

μ(g+tz)<μ(g)−(2/(5M))∥tz∥K.\mu(g+tz)<\mu(g)-(2/(5M))\|tz\|_{\mathcal K}.

We apply Lemma 2.2 to the Banach space (C(K,X),∥⋅∥K)(C(\mathcal K,X),\|\cdot\|_{\mathcal K}) and the function μ\mu with r=2/(5M)r=2/(5M) to conclude that μ\mu vanishes at some gˉ∈C(K,X)\bar{g}\in C(\mathcal K,X), hence f(gˉ(y))=yf(\bar{g}(y))=y, ∀y∈K\forall y\in\mathcal K. Finally, we continuously extend gˉ\bar{g} to the whole YY by applying the Dugundji Extension Theorem, see e.g. [5 Theorem 4.1].

It remains to be noted that if f(0)=0f(0)=0 and 0∈K0\in\mathcal K then one can easily modify the above by considering instead of the Banach space C(K,X)C(\mathcal K,X), the Banach space C0(K,X)C_0(\mathcal K,X) of all continuous g:K→Xg:\mathcal K\to X such that g(0)=0g(0)=0. It is clear that in this case u(0)=w(0)=0u(0)=w(0)=0, also z(0)=0z(0)=0, so gt(0)=0g_t(0)=0 and all the rest works in the same way.

It easily follows that a mapping satisfying the assumptions of the proposition above, is surjective.

Theorem 3.2. Let (X,∥⋅∥X)(X,\|\cdot\|_X) and (Y,∥⋅∥Y)(Y,\|\cdot\|_Y) be Banach spaces. Let the mapping f:X→Yf:X\to Y be CG1C_G^1 on XX, let it have invertible Gâteaux derivatives and let the Hadamard-Lévy condition (3.1) hold. Then ff is surjective.

Proof. Let y∈Yy\in Y be arbitrary and set K={y}\mathcal K=\{y\}. From Proposition 3.1 it follows that there is a continuous g∈C(Y,X)g\in C(Y,X) such that f(g(y))=yf(g(y))=y, so there is x=g(y)∈Xx=g(y)\in X such that f(x)=yf(x)=y. Therefore ff is surjective, i.e. f(X)=Yf(X)=Y.

Local invertibility

We will start by giving an example of a CG1C_G^1 mapping f:H→Hf:H\to H on the separable Hilbert space H=ℓ2(N)H=\ell_2(\mathbb{N}), such that f′(0)=Id⁡f'(0)=\operatorname{Id}, but ff is not invertible in any neighbourhood of 00. The example is elaborated on the basis of [3 Example 3].

Example 4.1. Let us consider the separable Hilbert space H=ℓ2(N)H=\ell_2(\mathbb{N}). There exists a mapping φ:H→H\varphi:H\to H which is CG1C_G^1 on HH such that

φ(0)=0,φ′(0)=0,(11)\varphi(0)=0,\qquad\varphi'(0)=0, \tag*{(11)}

and for λn:=2−n2\lambda_n:=2^{-n^2} it holds that

φ(λnen)=λnen,(12)\varphi(\lambda_n e_n)=\lambda_n e_n, \tag*{(12)}

where en=(0,…,1⏟n,0,…)e_n=(0,\ldots,\underbrace{1}_{n},0,\ldots), n∈Nn\in\mathbb{N} is the canonical basis of HH.

Then the mapping f(x):=x−φ(x)f(x):=x-\varphi(x) is CG1C^1_G on HH with f′(0)=Id⁡f'(0)=\operatorname{Id}, but

f(λnen)=f(0)=0,∀n∈N,f(\lambda_n e_n)=f(0)=0,\qquad\forall n\in\mathbb{N},

so ff is not injective in any neighbourhood of the origin.

Proof. Let ψ:R→R+\psi:\mathbb{R}\to\mathbb{R}^{+} be a C∞C^\infty-smooth bump function with ψ(0)=1=max⁡ψ\psi(0)=1=\max\psi, supp⁡ψ⊆[−12,12]\operatorname{supp}\psi\subseteq[-\frac{1}{2},\frac{1}{2}], i.e., ψ(t)=0\psi(t)=0, for all tt with ∣t∣≥1/2|t|\ge1/2.

For a,b>0a,b>0 and h∈Hh\in H with ∥h∥=1\|h\|=1 we set

φh,a,b(x):=ψ(∥Phx∥/a)ψ(∥Ph⊥x∥/b),\varphi_{h,a,b}(x):=\psi(\|P_hx\|/a)\psi(\|P_h^\perp x\|/b),

where PhP_h is the orthogonal projection onto Rh\mathbb{R}h, e.g. Penx=⟨x,en⟩enP_{e_n}x=\langle x,e_n\rangle e_n, and Ph⊥P_h^\perp is the orthogonal projection onto the hyperplane orthogonal to hh.

It is straightforward to check that the mapping

φ(x):=∑n=1∞λnφen,λn,2nλn(x−λnen)en\varphi(x):=\sum_{n=1}^{\infty}\lambda_n\varphi_{e_n,\lambda_n,2^n\lambda_n}(x-\lambda_ne_n)e_n

is CG1C^1_G on HH and satisfies (4.1) and (4.2).

From this example it is clear that for a CG1C^1_G mapping one needs to assume more than just invertibility of its Gâteaux derivative to ensure that it is locally invertible. The following local inversion result can be found in [14].

Theorem 4.2. [14 Theorem 10] Let (X,∥⋅∥X)(X,\|\cdot\|_X) and (Y,∥⋅∥Y)(Y,\|\cdot\|_Y) be Banach spaces and f:X→Yf:X\to Y be a continuous mapping which is Gâteaux differentiable. Suppose that for some x∈Xx\in X the derivative f′(x)f'(x) is invertible and

lim sup⁡x′→x∥f′(x′)−f′(x)∥<∥[f′(x)]−1∥−1.(13)\limsup_{x'\to x}\|f'(x')-f'(x)\|<\|[f'(x)]^{-1}\|^{-1}. \tag*{(13)}

Then there exist open sets U⊆XU\subseteq X and V⊆YV\subseteq Y, containing xx and f(x)f(x) respectively, such that f∣U:U→Vf|_U:U\to V is a bi-Lipschitz homeomorphism, and g:=[f∣U]−1g:=[f|_U]^{-1} is Gâteaux differentiable on VV with g′(y)=[f′(g(y))]−1g'(y)=[f'(g(y))]^{-1} for y∈Vy\in V. If in addition ff is CG1C^1_G on UU, then gg is CG1C^1_G on VV.

Main result

Here we present the proof of our main result along the lines of the proof of [13 Theorem 1].

Theorem 5.1. Let (X,∥⋅∥X)(X,\|\cdot\|_X) and (Y,∥⋅∥Y)(Y,\|\cdot\|_Y) be Banach spaces and f:X→Yf:X\to Y be a CG1C^1_G mapping on XX such that for any x∈Xx\in X its Gâteaux derivative f′(x)f'(x) is invertible and (13) holds. Let also the Hadamard-Lévy condition (3) hold. Then ff is CG1C^1_G invertible on XX, i.e. there exists g:Y→Xg:Y\to X which is CG1C^1_G on YY, such that

g(f(x))=f(g(x))=x,∀x∈X.g(f(x))=f(g(x))=x,\quad\forall x\in X.

Moreover, gg is globally Lipschitz with Lipschitz constant MM.

Proof. First we will show that ff is bijective.

As the conditions of Theorem 3.2 are satisfied, the surjectivity of ff follows. So, f(X)=Yf(X)=Y.

To get injectivity of ff, we exploit the idea of the proof of [13 Theorem 1]. Let us take a,b∈Xa,b\in X be such that f(a)=f(b)f(a)=f(b). By considering instead of ff the function x→f(a)−f(a−x)x\to f(a)-f(a-x) we can and do assume without loss of generality that b=0b=0 and f(0)=0f(0)=0. Then f(a)=0f(a)=0 and we need to show that a=0a=0. Set K:=f([0,a])\mathcal{K}:=f([0,a]). Since ff is continuous, K\mathcal{K} is a nonempty compact.

As the assumptions of Proposition 3.1 are fulfilled, there is a continuous mapping

g:Y→X,such that g(0)=0 and f(g(y))=y,∀y∈K.g:Y\to X,\quad\text{such that }g(0)=0\text{ and }f(g(y))=y,\quad\forall y\in\mathcal{K}.

Consider the set

I:={t∈[0,1]:g(f(ta))=ta}.I:=\{t\in[0,1]:g(f(ta))=ta\}.

Obviously, 0∈I0\in I, because g(0)=0g(0)=0. Due to the continuity of gg and ff the set II is closed and, therefore, compact. Let

tˉ:=max⁡{t:t∈I}.\bar{t}:=\max\{t:t\in I\}.

Assume that tˉ<1\bar{t}<1. By the local inversion result – Theorem 4.2 applied at tˉa\bar{t}a – there are open sets U∋tˉaU\ni\bar{t}a and V∋f(tˉa)V\ni f(\bar{t}a) such that for each y∈Vy\in V there is a unique x∈Ux\in U satisfying f(x)=yf(x)=y.

From the continuity of ff and gg there is β>0\beta>0 such that for all t∈(tˉ,tˉ+β)⊂(0,1)t\in(\bar{t},\bar{t}+\beta)\subset(0,1) we have ta∈Uta\in U, g(f(ta))∈Ug(f(ta))\in U, and f(ta)∈Vf(ta)\in V.

Hence, for t∈(tˉ,tˉ+β)t\in(\bar{t},\bar{t}+\beta), the point tata is the only solution of f(⋅)=f(ta)f(\cdot)=f(ta).

Since f(g(y))=yf(g(y))=y for all y∈Ky\in\mathcal{K} and since f(ta)∈Kf(ta)\in\mathcal{K}, we have that

f(g(f(ta)))=f(ta).f(g(f(ta)))=f(ta).

From the uniqueness of the solution to f(⋅)=f(ta)f(\cdot)=f(ta) one gets that g(f(ta))=tag(f(ta))=ta for all t∈(tˉ,tˉ+β)t\in(\bar{t},\bar{t}+\beta) which contradicts the definition of tˉ\bar{t}.

So, tˉ=1\bar{t}=1 meaning that g(f(a))=ag(f(a))=a. But f(a)=0f(a)=0. Since g(0)=0g(0)=0, it follows that a=0a=0.

We have proved that if f(a)=f(b)f(a)=f(b) then a=ba=b, so ff is injective. Till now we have established that ff is invertible and for any y∈Yy \in Y there is a unique x∈Xx \in X such that f(x)=yf(x)=y. Therefore, there is a mapping g:=f−1:Y→Xg:=f^{-1}:Y\to X. From Theorem 4.2 it follows that for any x∈Xx\in X there is an open set U∋xU\ni x and an open set V∋f(x)V\ni f(x) such that [f∣U]−1[f|_U]^{-1} is CG1C^1_G on VV. In other words, gg is CG1C^1_G on VV, and in particular gg is CG1C^1_G at yy. As yy was arbitrary, gg is CG1C^1_G on YY.

Finally, as g′(y)=[f′(g(y))]−1g'(y)=[f'(g(y))]^{-1} for all y∈Yy\in Y the condition (3) implies that ∥g′(y)∥≤M\lVert g'(y)\rVert\le M for all y∈Yy\in Y. The global Lipschitz continuity of the CG1C^1_G mapping gg easily follows from (2). □\square

Data Availability Statement. No datasets were generated or analyzed during the current study.

Funding. M. Hamamdjiev is supported by the Science Fund of Sofia University, grant No. 80-10-176/22.04.2026. N. Zlateva is supported by the Bulgarian National Science Fund, grant No. KP-06-H92/6 signed December 8, 2025.

Acknowledgements. The authors are extremely grateful to Dr. Milen Ivanov for his constant support and encouragement and especially for the help with Example 4.1.

References

  1. [1]J.-P. Aubin, I. Ekeland. Applied nonlinear analysis, New York, John Wiley & Sons, 1984.
  2. [2]J.-P. Aubin, H. Frankowska, On inverse function theorems for set-valued maps, J. Math. Pures et Appl. 66 (1987) 71–89.
  3. [3]V. Averbukh, O. Smolyanov, The various definitions of the derivative in linear topological spaces, Russian Math. Surveys 23(4) (1968) 67–113.DOI
  4. [4]R. Cibulka, M. Fabian, T. Roubal, An Inverse Mapping Theorem in Fréchet-Montel Spaces, Set-Valued Var. Anal. 28 (2020) 195–208.DOI
  5. [5]J. Dugundji, An extension of Tietze’s theorem, Pacific J. Math. 1 (1951), 353–367.DOI
  6. [6]I. Ekeland, An inverse function theorem in Fréchet spaces, Ann. Inst. Henri Poincaré (C) Anal. Non Lineaire 28(1) (2011) 91–105.DOI
  7. [7]I. Ekeland, E. Séré, A local surjection theorem with continuous inverse in Banach spaces, Rend. Lincei Mat. Appl. 35 (2024), 1–15.DOI
  8. [8]O. Gutù, On global inverse theorems, Topol. Methods in Nonlinear Anal. 49(2) (2017) 401–444.DOI
  9. [9]O. Gutù, J. A. Jaramillo, Surjection and inversion for locally Lipschitz maps between Banach spaces, J. Math. Anal. Appl. 478(2) (2019) 578–594.DOI
  10. [10]J. Hadamard, Sur les transformations ponctuelles, Bull. Soc. Math. France 34 (1906) 71–84.DOI
  11. [11]R. S. Hamilton, The inverse function theorem of Nash and Moser, Bull. Amer. Math. Soc. (N.S.) 7(1) (1982) 65–222.DOI
  12. [12]A. Ioffe, Variational Analysis of Regular Mappings: Theory and Applications, Springer Monogr. Math. Cham, Springer, 2017.
  13. [13]M. Ivanov, N. Zlateva. Hadamard inverse function theorem proved by variational analysis, Serdica Math. J. 49(1-3) (2023) 97–106.DOI
  14. [14]M. Ivanov, J. A. Jaramillo, S. Lajara, N. Zlateva, Continuous selections and invertibility of nonsmooth maps between Banach spaces, J. Optim. Theory Appl. 206 (2025), art. no. 41.DOI
  15. [15]P. Lévy, Sur les fonctions des lignes implicites, Bull. Soc. Math. France 48 (1920) 13–27.DOI
  16. [16]Z. Páles, Inverse and Implicit Function Theorems for Nonsmooth Maps in Banach Spaces, J. Math. Anal. Appl. 209(1) (1997) 202–220.DOI
  17. [17]R. Plastock, Homeomorphisms between Banach spaces, Trans. Amer. Math. Soc. 200 (1974) 169–183.DOI
  18. [18]B. H. Pourciau, Hadamard’s theorem for locally Lipschitzian maps, J. Math. Anal. Appl. 85(1) (1982) 279–285.DOI
  19. [19]B. H. Pourciau, Global invertibility of nonsmooth mappings, J. Math. Anal. Appl. 131(1) (1988) 170–179.DOI
  20. [20]W. Takahashi, Existence theorems generalizing fixed point theorems for multivalued mappings, In: M. A. Thera, J. B. Baillon (Eds.) Fixed point theory and applications, Pitman Research Notes in Mathematics Series 252 (1991), Longman Scientific & Technical, Harlow, 397–406.

Paper details

vlnc:W_y8xeqxqu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents