An Agmon–Allegretto–Piepenbrink principle for Schrödinger operators

We prove that each Borel function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V : \Omega \rightarrow [{-\infty }, +\infty ]$$\end{document}V:Ω→[-∞,+∞] defined on an open subset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Omega \subset {\mathbb R}^{N}$$\end{document}Ω⊂RN induces a decomposition \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Omega = S \cup \bigcup _{i} D_{i}$$\end{document}Ω=S∪⋃iDi such that every function in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W^{1,2}_{0}(\Omega ) \cap L^{2}(\Omega ; V^{+} \,\mathrm {d}x)$$\end{document}W01,2(Ω)∩L2(Ω;V+dx) is zero almost everywhere on S and existence of nonnegative supersolutions of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$-\Delta + V$$\end{document}-Δ+V on each component \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_{i}$$\end{document}Di yields nonnegativity of the associated quadratic form \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \int _{D_{i}} (|\nabla \xi |^2+V\xi ^2). $$\end{document}∫Di(|∇ξ|2+Vξ2).


Introduction
When V belongs to a Lebesgue space L p ( ) with p > N /2 or, more generally, to the Kato class K ( ), see Definition 2.1 below, the Agmon-Allegretto-Piepenbrink (AAP) principle provides a strong link between existence of nonnegative supersolutions and certain spectral properties of the Schrödinger operator − + V ; see e.g. [27]. A typical formulation of the AAP principle for (1.1) is the following Theorem 1.1 If V ∈ K ( ), then the Dirichlet problem (1.1) has a nontrivial nonnegative solution u ∈ W 1,2 0 ( ) for some nonnegative μ ∈ L 2 ( ) if and only if inequality (1.2) holds for every ξ ∈ W 1,2 0 ( ).
Such a statement is known to specialists but in Sect. 2 we present a proof for the reader's convenience. A key ingredient is the compact imbedding W 1,2 0 ( ) ⊂ L 2 ( ; V dx) that holds when V ∈ K ( ).
Various attempts have been made to further generalize the class of potentials V in the AAP principle; see for instance [13][14][15]21]. However, an equivalence as in Theorem 1.1 for a general Borel function V is hopelessly false. For example, taking = B 1 to be the unit ball centered at 0 and V (x) = 1/|x 1 | α with 1 ≤ α < 2, one has that (1.2) is trivially satisfied by positivity of V but there exists no nonnegative function v that satisfies − v + V v ≥ 0 in the sense of distributions in , other than the trivial one; see [17,Theorem 1.4] and Example 5.5 below.
In contrast, existence of nonnegative supersolutions typically yields nonnegativity of the quadratic form under weaker assumptions on V , as it has been proved in [15] for V ∈ L 1 ( ). As we shall see, this property is true even for general Borel potentials V .
One of the novelties of our work is to address this issue by adapting the concept of supersolution to the Schrödinger operator − + V . Our approach is based on the existence of a Borel set S ⊂ with the following properties: (a) every solution u of (1.1) satisfies u = 0 almost everywhere in S, (b) if moreover u is nonnegative, then μ ≤ 0 on S, regardless of μ on \ S.
The Borel set S that we consider is related solely to the positive part V + and has been introduced in [19] as the level set S := { ζ 1 = 0}, (1.3) where ζ 1 denotes the precise representative of the torsion function ζ 1 , which minimizes the energy functional One verifies that ζ 1 exists at every point of since ζ 1 can be written as the difference between a continuous and a bounded subharmonic function; see e.g. [18,Proposition 8.1]. This set S has been used in [19] to investigate the failure of the strong maximum principle for the Schrödinger operator − + V + and non-existence of solutions of the associated Dirichlet problem. That S defined by (1.3) satisfies properties (a) and (b) is justified in Sect. 5. An alternative way of identifying S without explicitly relying on the torsion function is also available using the Wiener criterion of Dal Maso and Mosco [6]; see [24]. In a classical situation where V + ∈ K ( ), the strong maximum principle holds and then one has S = ∅. However, for general potentials V + one should expect to have S = ∅. For example, if a ∈ and V (x) = 1/|x − a| 2 , then S = {a}. The set S may be larger than a finite union of singletons: Take a smooth open set ω and V (x) = 1/d(x, ∂ω) 2 , then S = ∂ω.
Choosing instead V (x) = 1/d(x, ω) 2 , one has S = ω. These two examples are physically related to the localized effect of particles in Quantum Mechanics that has been beautifully addressed in recent years by I. Díaz [9,10]. From [19], we know that \ S can be written as a disjoint union of its Sobolev-connected components, namely (1.4) where the index set I is finite or countably infinite. In the terminology introduced in [19, Sections 10 and 11], each set D i is Sobolev-open in the sense that there exists v i ∈ W 1,2 0 ( ) whose precise representative v i is defined everywhere in and D i = { v i > 0}. In addition, D i is Sobolev-connected meaning that it cannot be further written as a disjoint finite reunion of Sobolev-open subsets. When S is closed in with respect to the Euclidean topology, the sets D i coincide with the usual connected components of \ S.
Decomposition (1.4) propagates to every ξ ∈ W 1,2 0 ( ) ∩ L 2 ( ; V + dx) as we prove in Sect. 4 that ξ = 0 almost everywhere in S and, for every i ∈ I , ξχ D i ∈ W 1,2 0 ( ) ; see Proposition 4.3. Thus, ξ can be written as a sum of Sobolev functions, ξ = i∈I ξχ D i almost everywhere in . (1.5) In the spirit of the decompositions (1.4) and (1.5), we localize the AAP principle each component D i : We recall that u is a distributional solution of (1.1) whenever u ∈ W 1,1 0 ( )∩ L 1 ( ; V dx) and − u + V u = μ in the sense of distributions in . (1.7) We show in Sects. 4 and 5 that this equation can be localized in D i so that uχ D i satisfies for every Borel set A ⊂ and τ is a measure carried by S, that is |τ |( \ S) = 0. Under the assumptions of Theorem 1.2, we have that τ is nonnegative and we may apply the strong maximum principle for − + V + in D i to deduce that lim inf is the average integral on the ball B r (x). Theorem 1.2 thus provides one with a Poincaré inequality that gives the imbedding which can be seen as a compatibility condition between V + and V − to have existence of supersolutions in D i . When (1.6) is satisfied, we consider the vector space equipped with the seminorm We then denote by H i ( ) the completion of H i ( ).
In the spirit of the concept of weighted spectral gap by Pinchover and Tintarev [22], it is natural to investigate whether there is room for an improvement of (1.6) in the form We prove Theorem 1.3 in Sect. 8. The possibility that the measure in the statement of Theorem 1.2 satisfies μ(D i ) = 0 does not exclude the existence of another one with μ ≥ 0 in D i and μ(D i ) > 0, for which a nonnegative solution exists and then the stronger conclusion of Theorem 1.3 holds; see Example 7.7. This phenomenon is due to the fact that distributional solutions need not belong to L 2 ( ; V − dx), see Corollary 8.2, and is in sharp contrast to what happens with the operator − − λ 1 , where λ 1 is the first eigenvalue of the Laplacian. In such a case, one has S = ∅ and if u solves for some finite nonnegative measure μ, then using the first eigenfunction ϕ of − as test function one deduces that necessarily μ = 0 and u = αϕ with α ∈ R. Under the stronger assumption (1.8), any element ξ ∈ H i ( ) can be naturally associated to a function in L 2 (D i ; w i dx) and in particular is well-defined almost everywhere in D i . Moreover, using the direct method of the Calculus of Variations one easily deduces that, for every h ∈ L 2 (D i ; w i dx), there exists a unique minimizer θ i,h of the energy functional An interesting open problem is to find a weight w i such that the conclusion of Theorem 1.4 holds for every nonnegative h ∈ L 2 (D i ; w i dx).

Proof of Theorem 1.1
In this section we provide a proof of Theorem 1.1. We assume that N ≥ 3 ; the case N = 2 requires some adaptation as one needs to modify the definition of the Kato class using the logarithm in replacement of the potential 1/|z| N −2 . Let us start recalling the definition of the Kato class K ( ).

Definition 2.1 A Borel function
The proof of Theorem 1.1 is based on the following lemma: To our knowledge, the proof of this lemma goes back to [25, Chapter 6, Theorem 9.3]. We provide a more direct proof based on [29]. The estimate above implies that W 1,2 0 ( ) ⊂ L 2 ( ; V dx) and that the inclusion is compact. Indeed, by density of C ∞ c ( ) in W 1,2 0 ( ), the inequality holds with ϕ = ξ ∈ W 1,2 0 ( ). Moreover, if (ξ n ) n∈N is a bounded sequence in W 1,2 0 ( ), then by the Rellich-Kondrashov Compactness Theorem we may extract a subsequence such that ξ n j → u in L 2 ( ) for some u ∈ W 1,2 0 ( ). Using the inequality above for ξ n j − u we then have, for every > 0, Since is arbitrary, the convergence ξ n j → u in L 2 ( ; V dx) then follows.
Proof of Lemma 2. 2 We first show that, for every ϕ ∈ C ∞ c ( ), This inequality is proved in [25, Chapter 6, Theorem 8.8] using Bessel potentials, but here we follow a simpler idea from [29]: Since From both estimates and Fubini's theorem, we get Combining this estimate with (2.2), we obtain (2.1).
To conclude, we cover with a regular net of balls B δ (x i ) with i = 1, . . . , M, where the number of overlaps depends solely on the dimension N . Let ψ 1 , . . . , ψ M be a smooth partition of unity subordinated to this covering. Given ϕ ∈ C ∞ c ( ), we havê for some constant C 3 > 0 depending on N . By (2.1) applied to ϕψ i , we get We have 0 ≤ ψ i ≤ 1 and we may assume that |∇ψ i | ≤ C 4 /δ for every i. Combining (2.3) and (2.4), we getˆ where C 5 > 0 depends on N . Since V ∈ K ( ), we have η(δ) → 0 as δ → 0. Thus, given > 0, we may take δ > 0 small enough so that C 5 η(δ) ≤ .
Proof of Theorem 1.1 "⇐ " We prove the reverse implication, so we assume that Our goal is to prove that the minimization problem achieves its infimum for some nonnegative function u. Note that, thanks to (2.5), the infimum λ 1 is well-defined and nonnegative. Once one knows that the minimum exists, it is standard to show that Take a minimizing sequence (ξ n ) n∈N of (2.6). We show that (∇ξ n ) n∈N is bounded in L 2 ( ; R N ). To this end, for n large enough, we apply Lemma 2.2 with = 1/2 to get Recalling the normalization condition of ξ n , it follows that By boundedness of (ξ n ) n∈N in W 1,2 0 ( ) and compactness of the inclusion W 1,2 0 ( ) ⊂ L 2 ( ; V dx), we may extract a subsequence (ξ n j ) j∈N that converges weakly to some v in W 1,2 0 ( ) and also strongly in L 2 ( ) and L 2 ( ; V dx). We thus havê with´ v 2 = 1. We deduce that v is a minimizer of (2.6). Hence, the nonnegative function |v| is also a minimizer and we have the conclusion by taking u = |v| and μ = λ 1 u.
To prove the direct implication of Theorem 1.1, we show that existence of a positive supersolution yields a weighted Poincaré inequality. The following lemma is based on a typical variational argument that can be implemented even for a general Borel function V .
Then, for every ξ ∈ W 1,2 and η is nonnegative. Using η as a test function in (2.7), we get Since ∇u = 0 almost everywhere in {u = 0}, Note that, as → 0, As μ ∈ L 2 ( ), it then follows from the equation satisfied by u and the density of Thus, by Lemma 2.3, Since V ∈ L 1 ( ) and u is a nontrivial nonnegative function such that by the strong maximum principle, see [23,Proposition 22.2], we have u > 0 almost everywhere in and then (2.10) giveŝ

Duality solutions for nonnegative potentials
For each f ∈ L ∞ ( ), let ζ f be the unique minimizer of the energy functional Denoting by M( ) the vector space of finite Borel measures in , we recall the following notion from [16]: One verifies that the precise representative ζ f is defined at every point in , that is for Hence, the integral in the right-hand side of (3.3) is well-defined.
Every distributional solution is a duality solution, with the same datum. The converse however need not be true; see Example 5.5. A major advantage of dealing with duality solutions is that, in contrast with distributional solutions, they always exist for every ν ∈ M( ). For example, if g ∈ L ∞ ( ), then a combination of the Euler-Lagrange equations satisfied by ζ f and ζ g implies that Since ζ f = ζ f almost everywhere in by Lebesgue's differentiation theorem, we deduce that ζ g is the duality solution of (3.2) with datum ν = g dx. Moreover, from the estimate for every g ∈ L ∞ ( ) we then have We investigate in this section a counterpart of (3.5) that is valid for any duality solution of (3.2). We begin with the following general property: As we have pointed out, Proposition 3.2 is straightforward when u = ζ g with g ∈ L ∞ ( ). To prove the proposition in full generality, we use the property that, for every ν ∈ M( ), the duality solution u of (3.2) satisfies the representation formula where G x denotes the duality solution of Formula (3.6) is proved in [19,Lemma 13.2] when ν is nonnegative. For a general signed measure ν, one first applies (3.6) to the duality solutions of (3.2) with nonnegative data ν + and ν − . We recall that this step can be performed since duality solutions exist for any data in M( ). It then follows from (3.6) applied to ν + and ν − that G x d|ν| < +∞ for almost every x ∈ and, taking differences, we deduce (3.6) for ν = ν + − ν − .

Proof of Proposition 3.2
Since G x is a duality solution, using ζ 1 as test function we havê for every x ∈ . Then, by nonnegativity of G x we deduce that We now prove a sharper version of Proposition 3.2, where the Lebesgue measure is replaced by the W 1,2 capacity. More precisely, we recall that every duality solution u has a precise representative u which is defined quasi-everywhere, that is except in a set of W 1,2 capacity zero. This is a consequence of the fact that u is a finite measure in ; see [19,Proposition 4.1] and [23,Proposition 8.9]. We then have the following statement whose proof follows along the lines of [19, Proposition 13.3]: Let us denote by F the fundamental solution of − and focus our discussion on dimension N ≥ 3. The starting observation is that if ν ∈ M( ) is extended as 0 outside , then for any smooth bounded open set U ⊃ the distributional solution of satisfies |w| ≤ F * |ν| almost everywhere in U . Hence, if F * |ν| is bounded, then so is w and, by interpolation, we have w ∈ W 1,2 0 (U ). As a result, ν belongs to the dual space (W 1,2 0 (U )) and is diffuse with respect to the W 1,2 capacity, that is, for any Borel set A ⊂ with W 1,2 capacity zero, one has ν(A) = 0. The action of ν as an element in (W 1,2 0 (U )) is given by In dimension N = 2, the strategy above needs to be slightly adapted since F(x) → −∞ as |x| → ∞ and one cannot expect to have F * |ν| bounded in R 2 .
Proof of Proposition 3. 3 We first assume that the Newtonian potential F * |ν| is bounded. In this case, by comparison with F * |ν| we have u ∈ L ∞ ( ) and then u ∈ W 1,2 0 ( ) ∩ L 2 ( ; V + dx). Moreover, u satisfies the Euler-Lagrange equation Given a sequence of radial mollifiers (ρ n ) n∈N in C ∞ c (B 1 ), let u n be the duality solution of (3.2) with datum ρ n * ν, so that by uniqueness u n = ζ ρ n * ν . By (3.5), for every n ∈ N we then have We claim that To this end, we subtract the Euler-Lagrange equations satisfied by u n and u, and apply u n − u as test function. Using the nonnegativity of V + , we get For the first integral in the right-hand side of (3.11), we apply Fubini's theorem, Extending u n − u by 0 in R N \ , we have that u n − u ∈ W 1,2 (R N ). The sequence (ρ n * (u n − u)) n∈N is supported in a smooth bounded open set U and converges weakly to zero in W 1,2 0 (U ). From (3.8) and (3.12), we get which implies (3.10). Now passing to a subsequence (u n j ) j∈N , one deduces that In view of (3.9), the conclusion holds when F * |ν| is bounded.
In the case of a general measure ν ∈ M( ), it suffices to prove that the truncated function T 1 (u) satisfies the conclusion. Observe that T 1 (u) ∈ W 1,2 0 ( ) ∩ L 2 ( ; V + dx) and − T 1 (u) + V + T 1 (u) = ν in the sense of distributions in , for some diffuse measure ν ∈ M( ) ; see [3,4,7]. By a classical property in Potential Theory [12, Theorem 3.6.3] applied to ν + and ν − , this measure ν can be strongly approximated in M( ) by a sequence of measures (ν j ) j∈N such that |ν j | has bounded Newtonian potential, for which the conclusion holds from the first part of the proof.
Denote by v n the duality solution of (3.2) with datum ν n . By uniqueness of duality solutions, v n and T 1 (u) are also minimizers of their associated energy functionals and therefore satisfy an Euler-Lagrange equation. Subtracting those equations, we then havê . We now fix k > 0. Applying the identity above with test function ξ = T k (v n − T 1 (u)), by nonnegativity of V + we get Thus, as n → ∞,

Localization on D i
In this section, we prove that duality solutions of (3.2) and functions in W 1,2 0 ( ) ∩ L 2 ( ; V + dx) can be localized on each component D i . We rely on an orthogonality property of the Green's functions G x that is implicitly used in [19] to identify the components D i . It relies on (3.7) above and [19, Proposition 9.1], and is valid for each i ∈ I : Hence, uχ D i is the duality solution of (4.1).

Proof of Proposition 4.1 By the representation formula (3.6), the duality solution
As a consequence of Proposition 4.1, we deduce a localized strong maximum principle for duality solutions of (3.2). We first recall that the strong maximum principle holds for ζ g with nonnegative g ∈ L ∞ ( ) in the sense that, for each i ∈ I , we have that either ζ g > 0 on D i or ζ g ≡ 0 on D i ; see [19,Theorem 12.1]. Next, every duality solution u of (3.2) with nonnegative data ν can be bounded from below by some duality solution with nonnegative datum in L ∞ ( ). More precisely, by [19,Proposition 7.1], there exists a bounded nondecreasing continuous function Q : , where α > 1 and C > 0 is a constant that depends on α and .

Corollary 4.2 Let i ∈ I and let u be a nonnegative duality solution of
Proof Letū i = uχ D i . Then,ū i is a duality solution of (3.2) with nonnegative datum ν D i . By the comparison principle (4.5), Since´D i u > 0, the function Q(ū i ) is nonzero on D i . Then, by the strong maximum principle, we have ζ Q(ū i ) > 0 on D i . Thus, for every We now prove that decomposition (1.4) of \ S in terms of its components D i induces a natural splitting (1.5) of functions in W 1,2 0 ( ) ∩ L 2 ( ; V + dx).
Lemma 4.4 suffices for our purposes and is trivially satisfied when S is negligible for the Lebesgue measure. A small adaptation of the proof yields a sharper version in terms of the W 1,2 capacity: Proof of Proposition 4. 5 We assume by contradiction that S∩{ ξ = 0} has positive W 1,2 capacity. Then, there exist a compact set K ⊂ S ∩{ ξ = 0} with positive capacity and a nonnegative diffuse measure ν ∈ M( ) ∩ (W 1,2 0 ( )) supported in K with ν(K ) > 0 ; see [23, Proposition A.17]. We then have that´ ξ dν > 0 and the functional has a nontrivial nonnegative minimizer v, which is the duality solution of (3.2). In particular, since ν is supported in S,ˆ v =ˆ ζ 1 dν = 0.

Proof of Proposition 4.3
Since and, by Lemma 4.4, ξ = 0 almost everywhere in S, we are left to prove that ξχ D i ∈ W 1,2 0 ( ) for every i ∈ I .
Given ∈ C ∞ ( ; R N ), let v be the duality solution of (3.2) with datum ν = div dx. Then, by Proposition 4.1,v i := vχ D i is the duality solution of (3.2) with datum ν = (div )χ D i dx. By uniqueness, v andv i are also variational solutions for the same data. In particular, they both belong to W 1,2 0 ( ) ∩ L 2 ( ; V + dx) and satisfy the associated Euler-Lagrange equations.
We now observe that ∇v i = (∇v)χ D i almost everywhere in . (4.7) Indeed, this follows from the facts that ∇v i = 0 almost everywhere in {v i = 0} ⊃ \ D i and , using the Euler-Lagrange equation satisfied byv i and (4.7) we then havê where we use the norm Applying v as a test function in the Euler-Lagrange equation satisfied by v itself, we have We deduce from the Riesz Representation Theorem that ∇(ξ χ D i ) ∈ L 2 ( ; R N ). Since is smooth and (4.9) holds for every smooth that need not have compact support in , we conclude that ξχ D i ∈ W 1,2 0 ( ).

Duality solutions for signed potentials
We begin by extending the definition of duality solutions to the Dirichlet problem with a signed potential V : where ζ f is the minimizer of (3.1).
It is convenient to observe that, as the test functions ζ f depend on V + and ζ f = ζ f almost everywhere in , a duality solution of (5.1) is also a duality solution of where μ + V − u is regarded as the datum of the problem. In particular, the precise representative u is also well-defined quasi-everywhere in and we have the following direct consequences of Propositions 3.3 and 4.1, respectively: Corollary 5.2 If u is a duality solution of (5.1) with datum μ ∈ M( ), then u = 0 quasieverywhere in S.

Corollary 5.3
Let i ∈ I . If u is a duality solution of (5.1) with datum μ ∈ M( ), then uχ D i is a duality solution of (5.1) with datum μ D i .
We now rely upon [19,Proposition 4.1] for the operator − + V + to deduce that a duality solution is a distributional solution with possibly different datum:

Proof (Proof of Proposition 5.4)
Since u is a duality solution of (5.2) involving a finite measure in , we have u ∈ W 1,1 0 ( ). Take the duality solutions u 1 and u 2 of (3.2) with data (μ + V − u dx) + and (μ + V − u dx) − , respectively. By [19,Proposition 4.1], there exist nonnegative diffuse measures τ 1 and τ 2 such that τ 1 ( \ S) = τ 2 ( \ S) = 0 with in the sense of distributions in . From Proposition 3.2, one has u = 0 almost everywhere in S. Thus, Subtracting the equations satisfied by u 1 and u 2 , and using (5.4), we then get Even for a nonnegative potential V the measure τ which appears in Proposition 5.4 can be nonzero: In this case, S = B 1 ∩ {x 1 = 0} and then B 1 \ S has two connected components. Since ζ 1 is a duality solution with constant datum 1, we have for some finite measure τ supported in {x 1 = 0}. When α ≥ 2, we have τ = 0, but when 1 ≤ α < 2 the Hopf lemma holds in each component of B 1 \ S and implies that τ (S) > 0 ; see [17,Theorem 9.1].
Proof We recall that a distributional solution is also a duality solution for the same datum. By comparison with the conclusion of Proposition 5.4, we deduce that μ S = −τ in the sense of distributions in , and then equality holds as measures; see [23,Proposition 6.12].
We are left to show that τ ≥ 0. To this end, we first identify the diffuse part ( u) d of the measure u with respect to the W 1,2 capacity. Since V u = 0 almost everywhere in S and τ = −μ S is diffuse, we have On the other hand, as a consequence of Kato's inequality, by nonnegativity of u in we have

Approximation scheme
Existence of a unique duality solution of (3.2) entitles us to construct a sequence of duality solutions for a family of truncated problems associated to (5.1) with signed V . Proposition 6.1 Let μ be a nonnegative measurable function on and let (u n ) n∈N be the sequence defined by induction as u 0 = 0 and, for n ≥ 1, u n is the duality solution of Then, for every n ∈ N, we have u n ∈ W 1,2 0 ( ) ∩ L 2 ( ; V + dx) ∩ L ∞ ( ) and 0 ≤ u n ≤ u n+1 almost everywhere in .
Proof For every nonnegative f ∈ L ∞ ( ), we have ζ f ≥ 0 almost everywhere in . Thus, by nonnegativity of μ,ˆ Hence, u 1 ≥ 0 = u 0 almost everywhere in . We next assume that the conclusion holds for some n ∈ N. By nonnegativity of μ, T n+2 (μ) ≥ T n+1 (μ). Then, for every nonnegative By the induction assumption, the integrand in the right-hand side is nonnegative and we deduce that u n+2 − u n+1 ≥ 0 almost everywhere in . Finally, since T n (μ) ∈ L ∞ ( ), we have by induction that u n ∈ L ∞ ( ) for every n ∈ N. As u n ∈ L 1 ( ; V + dx), we then have u n ∈ L 2 ( ; V + dx). Finally, since u n is a finite measure in , we also have by interpolation that u n ∈ W 1,2 0 ( ).

Variational setting
Throughout the section, we assume that there exists a measurable function w i : D i → (0, +∞) such that (1.8) holds, whence Denoting by E the functional defined in H i ( ) by (1.9), the following property is standard in the Calculus of Variations:

the functional E has a unique minimizer θ i,h and any minimizing sequence of E is a Cauchy sequence in H i ( ).
Proof Uniqueness of the minimizer follows from the parallelogram identity satisfied by the norm · i which yields If u and v are both minimizers of E, then u − v 2 i ≤ 0. We now observe that E is bounded from below, whence inf E ∈ R. Taking a minimizing sequence (ξ j ) j∈N , it follows from (7.1) that, for every j, k ∈ N, Observe that the right-hand side converges to zero as j, k → ∞. Hence, (ξ j ) j∈N is a Cauchy sequence and converges in H i ( ) to the unique minimizer of E.
Since h is nonnegative and (1.8) holds, we have E(−ξ − j ) ≥ 0 and then Thus, (ξ + j ) j∈N is also a minimizing sequence. By Proposition 7.1, it converges to θ i,h in H i ( ) and then also in L 2 (D i ; w i dx). In particular, θ i,h ≥ 0 almost everywhere in D i .
As an alternative to the previous proof based on minimizing sequences, one may rely on the decomposition We rely on the following closure property which is a convenient tool to identify the weak limit of a sequence in H i ( ) :

Proof of Lemma 7.4
By the Banach-Saks property in a Hilbert space (which is a more comfortable and precise version of Mazur's lemma), we can take a subsequence (v n j ) j∈N that converges strongly to v in H i ( ) in the Cesàro sense, that is and then also in L 2 (D i ; w i dx). By pointwise convergence of (v n j ) j∈N to v in D i , we deduce that v = v almost everywhere in D i .
In particular, the sequences (min {u n , v n }) n∈N and (max {u n , v n }) n∈N are bounded in H i ( ). Hence, we may extract subsequences that converge weakly in H i ( ) to f and g, respectively.
As they converge almost everywhere to min {u, v} and max {u, v} in D i , the conclusion follows from Lemma 7.4 and the lower semicontinuity of the norm under weak convergence.
In contrast with the existence of a positive and negative parts in H i ( ), it is unclear whether every u ∈ H i ( ) admits a truncation T k (u) ∈ H i ( ) for every k > 0. Let us now show that the minimizer θ i,h is an upper bound for the approximating sequence constructed in Sect. 6.

Proposition 7.5
If h ∈ L 2 (D i ; w i dx) is nonnegative and if (u n ) n∈N is the sequence defined in Proposition 6.1 with μ = w i hχ D i , then, for every n ∈ N, we have Proof Since u 0 = 0, the case n = 0 is a consequence of Proposition 7.2. We now assume by induction that the inequality holds for n−1 for some n ≥ 1. Our goal is to show that if (ξ j ) j∈N is a minimizing sequence in H i ( ) of the functional E, then (max {u n , ξ j }) j∈N is also a minimizing sequence. To this end, we begin by observing that u n ∈ W 1,2 0 ( )∩ L 2 ( ; V + dx) is the minimizer of the functional F n defined on W 1,2 0 ( ) ∩ L 2 ( ; V + dx) by In particular, for every ξ ∈ H i ( ), We next show the following consequence of (7.2) for the functional E : Indeed, for every η ∈ H i ( ), Applying (7.2) and u n ≥ min {u n , ξ}, we get Thus, which gives (7.3).
Using a standard property of the maximum and minimum between two functions, we then get from (7.3) that As we mentioned above, we now apply this estimate to a minimizing sequence (ξ j ) j∈N in H i ( ) of the functional E. We first recall that, by Proposition 7.1, we have ξ j → θ i,h in H i ( ). Since the functions u n and u n−1 are bounded and belong to L 1 ( ; V − dx), they also belong to L 2 ( ; V − dx). By the Dominated Convergence Theorem and the induction assumption u n−1 ≤ θ i,h in D i , we then have We then deduce from (7.4) that Hence, (max {u n , ξ j }) j∈N is also a minimizing sequence of E as claimed. Therefore, max {u n , ξ j } → θ i,h in H i ( ) as j → ∞ and then, by (1.8), Thus, u n ≤ θ i,h almost everywhere in D i .
We now identify the limit of the sequence (u n ) n∈N with the minimizer θ i,h : Proposition 7.6 Let h ∈ L 2 (D i ; w i dx) be a nonnegative function and let (u n ) n∈N be the sequence defined in Proposition 6.
To this end, we recall that, for every n ∈ N, we have u n ∈ H i ( ) and for every ξ ∈ W 1,2 0 ( ) ∩ L 2 ( ; V + dx). On the other hand, since (u n ) n∈N is nondecreasing, Taking ξ = u n in (7.5), we get Since h ∈ L 2 (D i ; w i dx) and (u n ) n∈N is bounded in L 2 (D i ; w i dx), it follows from this estimate that (u n ) n∈N is also bounded in H i ( ). Therefore, there exists a subsequence (u n j ) j∈N such that Since (u n ) n∈N is nondecreasing, it converges pointwise and then, by Lemma 7.4, u = lim j→∞ u n j = lim n→∞ u n almost everywhere in D i .
Recalling that E has a unique minimizer θ i,h in H i ( ), to conclude the proof of the proposition it suffices to show that u satisfies the Euler-Lagrange equation: (7.6) where (·|·) i denotes the inner product of H i ( ) associated to its norm. Since we do not know whether u ∈ L 2 (D i ; V dx), one should be careful when letting n → ∞ in (7.5).
To overcome this difficulty, we proceed with ξ ∈ H i ( ) ∩ L ∞ ( ). We rewrite (7.5) with n = n j as As j → ∞ in (7.7), we then get For a general ξ ∈ H i ( ), we apply this identity to T k (ξ ) and then let k → ∞. Equation (7.6) then follows from the density of H i ( ) in H i ( ). Since the minimizer θ i,h satisfies the same equation, it follows by uniqueness that u = θ i,h . In particular, u = θ i,h almost everywhere in D i .

Proof of Theorem 1.3 Let z i be the duality solution of
We claim that, for every n ∈ N, where u is the distributional solution of (1.1) and (u n ) n∈N is the sequence of duality solutions defined in Proposition 6.1 with datum Q(z i ), that is for n ≥ 1, Since for every nonnegative f ∈ L ∞ ( ), ζ f is also nonnegative, we havê We prove (8.3) by induction. Firstly, we have u 0 = 0 ≤ uχ D i by assumption on u. Assume now that, for some n ≥ 1, the inequality holds for n − 1. Using the duality formulation of z i and the induction assumption in (8.4), for every nonnegative f ∈ L ∞ ( ) we havê Using (8.2) and the fact that z i and uχ D i are duality solutions of (8.1) and (1.1) with datum μ D i , respectively, we then get Since f is an arbitrary nonnegative function in L ∞ ( ), estimate (8.3) then follows.
As a consequence of (8.3), we have u n = 0 almost everywhere in \ D i . (8.5) Let us now show that, for every n ≥ 1, Indeed, we observe that, by the representation formula (3.6), for almost every x ∈ we have Since μ(D i ) > 0 and G x > 0 in D i for x ∈ D i , we then have Thus, Q(z i ) > 0 almost everywhere in D i . Applying again (3.6), for almost every x ∈ we have For x ∈ D i , the integrand is almost everywhere positive in D i and (8.6) follows.
where f α,β ∈ L 1 (B 1 ) is a nonnegative function and f α,β (x) > 0 for x = 0. Indeed, by an explicit computation, As the function , for β < α in this range we then have Observe that in Example 8.1 one has u β / ∈ L 2 (B 1 ; V − dx) if and only if β ≥ (N − 2)/2. We conclude this section showing that the expected spectral property can be recovered on each component D i provided the solution has enough integrability with respect to V − . for some diffuse measure τ such that |τ |( \ S) = 0. By [7], for every k > 0 we then havê Since |T k (ū i )| ≤ |u| and u = 0 quasi-everywhere on S by Proposition 3.3, we have Therefore, from |τ |( \ S) = 0 it follows that Using (1.8) with T k (u) and this identity, we get Since u ∈ L 2 (D i ; V − dx), by the Dominated Convergence Theorem the right-hand side converges to 0 as k → ∞. Therefore, applying Fatou's lemma we get Applying the representation formula (3.6) and the fact that G x = 0 in \ D i for x ∈ D i , we get Since u is nontrivial in D i , the integral in the right-hand side is positive for some x ∈ D i . We then must have´D i V − u dx > 0 and (9.1) follows. We now let ξ ∈ W 1,2 0 ( ) ∩ L 2 ( ; V + dx). Since ν α (D i ) > 0, from Theorem 1.3 there exists a measurable function w α : D i → (0, +∞) such that Taking the limit as α → 1, by the Monotone Convergence Theorem it follows that D i (|∇ξ | 2 + V ξ 2 ) ≥ 0.
Proof of Theorem 1. 4 Using the notation of the proof of Theorem 1.3, we take a bounded measurable function w i : D i → (0, +∞) such that u ∈ L 2 (D i ; w i dx) and Since 0 < u ≤ u almost everywhere in D i , this weight satisfies (8.8) and then (1.8) holds. Denote by (v n ) n∈N the sequence defined in Proposition 6.1 with datum w i hχ D i , where h ∈ L 2 (D i ; w i dx) and 0 ≤ h ≤ u. We claim that, for every n ∈ N, v n ≤ uχ D i almost everywhere in . (9.3) Indeed, by (9.2) and the fact that h ≤ u in , By comparison of solutions, we then get v n ≤ u n ≤ uχ D i almost everywhere in , where (u n ) n∈N is the sequence used in the proof of Theorem 1.3 and the second inequality is given by (8.3).
As a consequence of (9.3), the sequence (v n ) n∈N is bounded in L 2 (D i ; w i dx) and L 1 (D i ; V − dx). Then, by Proposition 7.6, we have lim n→∞ v n = θ i,h almost everywhere in D i . To conclude, since v n is a duality solution we havê Since ζ f is bounded, 0 ≤ w i h ≤ Cu ∈ L 1 (D i ) and 0 ≤ T n (V − )v n−1 ≤ V − u ∈ L 1 (D i ), it then follows from the Dominated Convergence Theorem that Hence, θ i,h χ D i is a duality solution of (5.1) with datum w i hχ D i . As we have θ i,h = θ i,h χ D i in and w i h + V − θ i,h ∈ L 1 (D i ), the conclusion follows from Proposition 5.4.
The next example shows that the validity of a Poincaré inequality (1.8) is not enough to ensure that minimizers θ i,h are distributional solutions of an equation of the form (1.7) for any μ ∈ M( ). where d ∂ω denotes the distance to ∂ω. Since S depends only on V + , one shows that S = ∂ω, whence \ S has two connected components, namely D 1 := ω and D 2 := \ ω. The Poincaré inequality (1.8) with positive weight holds for both of them. This is clear on D 2 , while on D 1 there exists > 1/4 diam 2 (ω) such that ω (|∇ξ | 2 + V ξ 2 ) ≥ ˆω ξ 2 for every ξ ∈ W 1,2 0 (ω) ; see [2, Theorem II]. Hence, for any h ∈ L 2 (D 1 ), the functional E with i = 1 has a minimizer θ 1,h in H 1 ( ) ⊂ L 2 (D 1 ). We claim that if h is nonnegative and´D 1 h > 0, then θ 1,h cannot be a distributional solution of (1.1) for any μ ∈ M( ), thus extending the nonexistence of continuous supersolutions from [2,Theorem III]. To this end, it suffices to show that V θ 1,h / ∈ L 1 (D 1 ). Since h is nonnegative, we have θ 1,h ≥ 0 almost everywhere in ω. Moreover, We deduce that V θ 1,h / ∈ L 1 (D 1 ) and then θ 1,h cannot be a distributional solution in .