Introduction
We are concerned with positive solutions of the system
(−Δ)mu=∣x∣αvp,(−Δ)mv=∣x∣βuqin Rn,(1)
where m≥2 is an integer and p,q>0, α,β∈R. The scaling uR(x)=Rσ1u(Rx), vR(x)=Rσ2v(Rx) with
σ1=pq−12m+α+p(2m+β),σ2=pq−12m+β+q(2m+α)(2)
leaves (1) invariant when pq=1, and it singles out the critical hyperbola
p+1n+α+q+1n+β=n−2m.(3)
Phan [14] conjectured that, for p,q≥1, pq=1, and α,β≥0, the system (1) has no positive classical solution if and only if (p,q) lies below the critical hyperbola. The nonexistence half, which is [14 Conjecture A] and is also stated by Fazly [6 Conjecture 1], asserts that there is no positive classical solution when
p+1n+α+q+1n+β>n−2m.(4)
Fazly proved it for radial solutions in every dimension and for bounded solutions in dimension n=2m+1 [6]. Phan proved it for bounded solutions when, in addition, max{2m(p+1),2m(q+1)}≥(n−2m−1)(pq−1), and without the boundedness assumption in a smaller range of exponents [14]. Liouville theorems for integral and polyharmonic systems have also been obtained by the method of scaling spheres [5, 7]. Arthur and Yan obtained further results under a slow-decay assumption [2]. More recently, Xu and Luo proved a super polyharmonic property and Liouville theorems for radial solutions subject to the growth condition u,v=O(∣x∣2) [20]. These restrictions are absent from the result below. In higher dimensions the general unrestricted case has remained open.
The second-order case m=1 has a longer history. For α=β=0 it is the Lane–Emden conjecture. Mitidieri introduced a Rellich identity for systems and obtained the first nonexistence regions [10, 11]. Serrin and Zou settled dimension three under a growth condition [16]. Poláčik, Quittner, and Souplet removed that condition [15], and Souplet proved the conjecture in dimension four [17]. Phan formulated and partly proved the weighted version [13], and Li and Zhang proved it in dimension three [8]. In each of these works the decisive step is an estimate for the boundary terms of a Rellich–Pohozaev identity, and the available Sobolev-type estimates control them only in low dimensions. A recent preprint [12] removes this restriction for m=1 and every n. Its main estimate does not bound the boundary term of the localized Pohozaev identity. Instead, it shows that this term is dominated by an interaction term of favorable sign, up to an error εE+Cε relative to the localized energy E. The domination comes from an elementary comparison between pairs of intervals on parallel lines, applied to the level sets of the sources.
That argument uses the equation only through the Newton potential representation u=Γ∗(∣x∣αvp), v=Γ∗(∣x∣βuq), together with the fact that the kernel is a power of ∣x∣. Our first observation is that it goes through for a Riesz kernel of any order μ∈(1,n). Mei and Wei carried out this extension for the unweighted fractional system, that is, for Riesz kernels of order 2s∈(0,2) and α=β=0 [9]. When 2s≤1 the kernel gradient is not locally integrable, and they combine the two equations before removing the diagonal. The present problem requires weights and orders μ=2m≥4. The natural route for m≥2 is therefore to prove that positive solutions of (1) are Riesz potentials of order 2m.
This is where the polyharmonic problem differs from the second-order one. When m=1, a positive solution is superharmonic and the representation follows from the Riesz decomposition of superharmonic functions. When m≥2, positivity of u says nothing directly about the sign of −Δu, and the representation can fail. In fact, for m even and pq<1, or for p=q=1, the system (1) with α=β=0 has positive radial solutions with Δu>0 everywhere, although (4) holds (Proposition 1.4 below). The property that has to be proved is the super polyharmonic property
(−Δ)iu>0,(−Δ)iv>0in Rn,1≤i≤m−1.
For the scalar equation it goes back to Wei and Xu [21]. For systems (−Δ)kui=fi(u1,…,ul) without weights, Chen and Li proved it and deduced the equivalence with the corresponding integral system when 2k<n [4], and Xu and Luo obtained it for radial solutions of weighted systems under a growth condition [20]. The weights ∣x∣α and ∣x∣β are not centered at the points where one averages, and we give a self-contained proof, without symmetry or growth assumptions, that also covers the endpoint cases p=1 or q=1. Spherical averages and a parity argument give the property at once except in one case: m even and (−Δ)m−1u negative somewhere. There the averages of u and v grow, but the parity count produces no contradiction. We show instead that the growth exponents feed each other and increase geometrically when pq>1, which is impossible for a finite function. The example above shows that pq>1 cannot be dropped here.
Our main result settles the nonexistence half of the conjecture.
Theorem 1.1. Let m≥2 and n≥1 be integers, let p,q≥1 with pq>1, and let α,β≥0. If (4) holds, then (1) has no solution u,v∈C2m(Rn) with u>0 and v>0 in Rn.
Together with the existence theorem of Villavert [19], this settles the conjecture of Phan for m≥2.
Corollary 1.2. Let m≥2 and n≥1 be integers, let p,q≥1 with pq=1, and let α,β≥0. Then (1) has a solution u,v∈C2m(Rn) with u>0 and v>0 in Rn if and only if
p+1n+α+q+1n+β≤n−2m. In that case the solution can be chosen radial.
The radial solutions of [19] are constructed by shooting and are stated there to be of class C2m away from the origin. We check in § 3 that they are classical solutions on all of Rn.
No boundedness, radial symmetry, decay, or energy assumption is made in Theorem 1.1. Taking u=v, p=q, and α=β gives the scalar statement: for m≥2, p>1, α≥0, and n≥1, the equation (−Δ)mu=∣x∣αup has no positive classical solution in Rn when p<n−2mn+2m+2α, with the convention that the bound is +∞ when n≤2m. For n>2m this scalar statement was proved earlier, for α>−2m and 0<p below the same bound, by Dai and Qin with the method of scaling spheres [5]. When n≤2m, condition (4) holds automatically, and Theorem 1.1 says that there is no positive solution at all. The hypotheses p,q≥1 and α,β≥0 are those of the conjecture. They enter the proof through the super polyharmonic property, which is the next statement, and α,β≥0 also makes the right sides of (1) continuous at the origin.
Theorem 1.3. Let m≥2 and n≥1 be integers, let p,q≥1 and α,β≥0, and let u,v∈C2m(Rn) be positive solutions of (1). If m is even, assume also pq>1. Then
(−Δ)iu>0and(−Δ)iv>0in Rn,1≤i≤m−1.
Proposition 1.4. Let m≥2 be even, n≥1, and α=β=0. Suppose that either p,q>0 with pq<1, or p=q=1. Then (1) has a positive radial solution u,v∈C2m(Rn) with Δu≥1 and Δv≥1 in Rn. Moreover, (p,q) satisfies (4).
We do not know whether Theorem 1.3 holds when pq>1 and min{p,q}<1. The proof below uses Jensen’s inequality for the averages of uq and vp, and this is the essential reason for the restriction p,q≥1.
Once Theorem 1.3 is available, u and v solve an integral system, and Theorem 1.1 follows from the next result with μ=2m. We state it for a general Riesz order because the proof is the same.
Theorem 1.5. Let n≥2, 1<μ<n, c0>0, p,q>0, and α,β∈R. Suppose that
p+1n+α+q+1n+β>n−μ.(5) Then there are no functions u,v∈C(Rn), positive in Rn, such that for every x∈Rn the integrals below are finite and
u(x)=c0∫Rn∣x−y∣n−μ∣y∣αv(y)pdy,v(x)=c0∫Rn∣x−y∣n−μ∣y∣βu(y)qdy.(6)
For μ=2 and n≥3, Theorem 1.5 is the integral form of the main theorem of [12], and its proof is the proof given there. We give in full the steps in which the order of the kernel enters. The restriction μ>1 is used only to differentiate the potentials in the virial identity, where ∇∣z∣μ−n must be locally integrable. For α=β=0 the system (6) is the integral system that Chen and Li related to the Lane–Emden conjecture [3], and for α=β=0 and μ∈(1,2) Theorem 1.5 is contained in [9], after normalizing c0 by scaling. We do not know whether the method of [9] for μ≤1 extends to the weighted case.
Our integral-system argument builds on the recent preprint [12]. For the Riesz kernel ∣z∣μ−n, we establish the interval comparison and the localized estimates in § 4, followed by the weighted pressure estimate in Appendix § 6. We also give the virial identity with the required kernel-gradient integrability arguments. By contrast, the super polyharmonic property, the reduction to (6), and Proposition 1.4 do not depend on [12].
Theorem 1.3 is proved in § 2, the reduction to the integral system and the proofs of Theorem 1.1 and Corollary 1.2 are in § 3, Theorem 1.5 is proved in § 4, and Proposition 1.4 in § 5.
The super polyharmonic property
Throughout this section u,v are as in Theorem 1.3, and we write
ui=(−Δ)iu,vi=(−Δ)iv(0≤i≤m−1),um=∣x∣αvp,vm=∣x∣βuq.
Thus ui,vi∈C2m−2i(Rn) and −Δui=ui+1, −Δvi=vi+1 for 0≤i≤m−1. The functions um and vm are continuous and positive away from the origin.
Fix x0∈Rn. For a continuous function w we write
wˉ(r)=∣∂Br(x0)∣1∫∂Br(x0)wdS(r>0),wˉ(0)=w(x0),
with the counting measure on ∂Br(x0)={x0±r} when n=1. If w∈C2(Rn), the divergence theorem gives rn−1wˉ′(r)=∣∂B1∣−1∫Br(x0)Δwdx. Hence, for 0≤i≤m−1,
rn−1uˉi′(r)=−∫0rtn−1uˉi+1(t)dt,uˉi′(0)=0,(7)
and the same holds for v. Two further facts are used repeatedly. First, if r≥R≥max{2,2∣x0∣} and x∈∂Br(x0), then ∣x∣≥r/2≥1. Since α,β≥0 and p,q≥1, Jensen’s inequality gives
uˉm(r)≥vˉ(r)p,vˉm(r)≥uˉ(r)q(r≥max{2,2∣x0∣}).(8)
Second, uˉm≥0 and vˉm≥0 everywhere. We first describe how a negative value propagates down a chain of averages.
Lemma 2.1. Let ℓ≥2, let W0,…,Wℓ−1∈C2([0,∞)) and Wℓ∈C([0,∞)) satisfy
rn−1Wi′(r)=−∫0rtn−1Wi+1(t)dt(0≤i≤ℓ−1), and suppose Wℓ≥0 and Wℓ−1(r1)<0 for some r1≥0. Put ϵi=(−1)ℓ−i. Then there is R∗ such that, for 0≤i≤ℓ−1 and r≥R∗,
ϵiWi(r)>0,ϵiWi′(r)≥0, and ϵiWi(r)→∞ as r→∞ for 0≤i≤ℓ−2.
Proof. Since Wℓ≥0, the function Wℓ−1 is nonincreasing, so ϵℓ−1Wℓ−1=−Wℓ−1≥c:=−Wℓ−1(r1)>0 on [r1,∞) and ϵℓ−1Wℓ−1′≥0 everywhere. Suppose that for some i≤ℓ−2 we have ϵi+1Wi+1≥c′>0 on [R,∞). Since ϵi=−ϵi+1,
ϵirn−1Wi′(r)=∫0rtn−1ϵi+1Wi+1(t)dt≥−∫0Rtn−1∣Wi+1∣dt+nc′(rn−Rn),
so ϵiWi′(r)≥c′r/(2n) for large r, and ϵiWi(r)→∞. Descending from i=ℓ−2 to i=0 proves the lemma. □
The second lemma converts a lower bound for the top of a chain into a lower bound for its bottom on a fixed window [R,2R].
Lemma 2.2. Let ℓ≥1, R>0, and let W0,…,Wℓ be as in the first display of Lemma 2.1. Let ϵi∈{−1,1} with ϵi=−ϵi+1, and suppose ϵiWi(R)≥0 and ϵiWi′(R)≥0 for 0≤i≤ℓ−1. If ϵℓWℓ(τ)≥C(τ−R)s on [R,2R] for some C>0 and s≥0, then
ϵ0W0(r)≥C2ℓ(1−n)−2ℓ(s+2ℓ)−2ℓ(r−R)s+2ℓ(R≤r≤2R).
Proof. For 0≤i≤ℓ−1 and r≥R, integrating the relation for Wi from R gives
ϵiWi(r)=ϵiWi(R)+ϵiRn−1Wi′(R)∫Rrt1−ndt+∫Rrt1−n∫Rtτn−1ϵi+1Wi+1(τ)dτdt.
The first two terms are nonnegative. On [R,2R] we have τn−1t1−n≥21−n for τ≤t, so a lower bound ϵi+1Wi+1(τ)≥C′(τ−R)s′ on [R,2R] gives
ϵiWi(r)≥(s′+1)(s′+2)21−nC′(r−R)s′+2(R≤r≤2R).
Applying this ℓ times, from i=ℓ−1 down to i=0, and using (s+2j−1)(s+2j)≤(s+2ℓ)2 for j≤ℓ proves the claim. □
Proof of Theorem 1.3. Step 1: nonnegativity. Suppose that some ui or vi with 1≤i≤m−1 takes a negative value, and let k be the largest such index. Interchanging (u,p,α) and (v,q,β) if necessary, we may assume uk(x0)<0 for some x0. By maximality, ui≥0 and vi≥0 in Rn for k<i≤m−1. In particular uˉk+1≥0, whether k<m−1 or k=m−1.
Apply Lemma 2.1 to Wi=uˉi, 0≤i≤k, with ℓ=k+1≥2 and r1=0. It gives (−1)k+1uˉ(r)→∞. If k is even, then uˉ(r)→−∞, which is impossible because u>0.
Let k be odd. Then uˉ(r)→∞, and (8) gives vˉm(r)≥1 for r large. By (7), rn−1vˉm−1′(r)≤−c(rn−Rn) for r≥R if R is large, so vˉm−1(r)→−∞. If k<m−1, this contradicts vm−1≥0.
It remains to treat k=m−1 with m even. Here we use pq>1. Lemma 2.1, applied to (uˉ0,…,uˉm) with ℓ=m and r1=0, and to (vˉ0,…,vˉm) with ℓ=m and r1 chosen with vˉm−1(r1)<0, gives R∗ such that, with ϵi=(−1)m−i,
ϵiuˉi, ϵiuˉi′, ϵivˉi, ϵivˉi′≥0on [R∗,∞),0≤i≤m−1.
Put c1=−um−1(x0)>0. Since uˉm−1 is nonincreasing, we have −uˉm−1≥c1 on [0,∞). Choose R≥max{R∗,2,2∣x0∣}, to be fixed below. Applying Lemma 2.2 to uˉ0,…,uˉm−1, with ℓ=m−1 and s=0, gives
uˉ(r)≥C0(r−R)s0(R≤r≤2R),s0=2m−2,
C0=c12(m−1)(1−n)(2m−2)−(2m−2).
Suppose that uˉ(r)≥Cj(r−R)sj on [R,2R]. By (8), vˉm≥Cjq(r−R)qsj there. Lemma 2.2 with ℓ=m gives a lower bound for vˉ, then (8) gives one for uˉm, and Lemma 2.2 once more gives uˉ(r)≥Cj+1(r−R)sj+1 on [R,2R], where
sj+1=p(qsj+2m)+2m,Cj+1=Cjpq2m(1−n)(p+1)sj+1−2m(p+1).
Here we used qsj+2m≤sj+1, which holds because p≥1. The constants C0, s0, and the recursion do not depend on R.
Set Lj=log(CjRsj), so that uˉ(2R)≥eLj. Since sj+1−pqsj=2m(p+1) and R≥1,
Lj+1≥pqLj−b−2m(p+1)logsj+1,b=m(n−1)(p+1)log2.
From sj+1≤pqsj+2m(p+1) we get sj≤S(pq)j with S=s0+2m(p+1)/(pq−1). Hence
Lj+1≥pqLj−a(j+1)−b′,a=2m(p+1)log(pq),b′=b+2m(p+1)logS.
Let θ=(1+pq)/2∈(1,pq) and M=max{a+b′,a/logθ}/(pq−θ). If Lj≥Mθj, then, since θj≥1+jlogθ,
Lj+1≥θj+1M+(pq−θ)Mθj−a(j+1)−b′≥θj+1M.
Now L0=logC0+(2m−2)logR and 2m−2>0. Fix R so large that L0≥M. Then Lj≥Mθj for every j, and uˉ(2R)=∞, which is absurd. This proves ui≥0 and vi≥0 for $1\le i\le m-1.
Step 2: strict positivity. Let 1≤i≤m−1. The function ui is nonnegative and −Δui=ui+1≥0. If ui(x1)=0 at some point, the strong minimum principle for superharmonic functions [1] gives ui≡0, hence ui+1≡0, and descending in this way we reach ∣x∣αvp≡0, which is false. The same argument applies to vi. □
Remark 2.3. The case k=m−1 with m even is the only place where pq>1 is used. For m odd, Theorem 1.3 holds for all p,q≥1. For m even it fails for p=q=1, by Proposition 1.4.
Remark 2.4. Fazly states the conjecture for nonnegative solutions [6]. Step 1 uses only u,v≥0. In Step 2, the same minimum principle applied to u and v shows that a nonnegative solution is either positive or identically zero. Hence Theorems 1.3 and 1.1 hold for nonnegative solutions, with the conclusion u≡v≡0 in Theorem 1.1.
Reduction to the integral system
We need the Riesz decomposition of a nonnegative superharmonic function on Rn in the following elementary form. Let Γ(z)=∣z∣2−n/((n−2)∣∂B1∣) for n≥3.
Lemma 3.1. Let w∈C2(Rn) satisfy w≥0 and −Δw=h≥0 with h continuous.
If n≤2, then h≡0.
If n≥3, then Γ∗h is finite everywhere and w=c+Γ∗h for a constant c≥0.
Proof. (i) Average about a point x0 with h(x0)>0. Then (rn−1wˉ′)′=−rn−1hˉ≤0, and rn−1wˉ′(r)≤−c<0 for r≥r1. For n=2 this gives wˉ(r)≤wˉ(r1)−clog(r/r1), and for n=1 it gives wˉ(r)≤wˉ(r1)−c(r−r1). Both contradict w≥0.
(ii) For every x and R>0, Green’s representation in BR(x) gives
w(x)−wˉ(R)=∫BR(x)(Γ(x−y)−Γ(R))h(y)dy.
where wˉ is the average about x. Since wˉ(R)≥0, monotone convergence as R→∞ gives Γ∗h≤w. The rest of the argument is that of [12 Lemma 2.2]: Γ∗h is locally integrable and solves −Δ(Γ∗h)=h in the sense of distributions, so w−Γ∗h is a nonnegative harmonic function on Rn, hence a nonnegative constant c by Liouville’s theorem for positive harmonic functions [1]. A priori this gives w=c+Γ∗h almost everywhere. Since Γ∗h is superharmonic, it equals at every point the limit of its averages over small balls, and so does the continuous function w−c. Hence w=c+Γ∗h everywhere. □
For integers j≥1 with 2j<n let Γ2j=Γ∗⋯∗Γ be the j-fold convolution. It is the Riesz kernel Γ2j(z)=γn,2j∣z∣2j−n with a constant γn,2j>0 [18 Chapter V, §1], and Γ∗(Γ2j−2∗h)=Γ2j∗h for every nonnegative h, by Tonelli’s theorem.
Proposition 3.2. Let m, n, p, q, α, β be as in Theorem 1.1, and let u,v∈C2m(Rn) be positive solutions of (1). Then n>2m, and for every x∈Rn
u(x)=∫RnΓ2m(x−y)∣y∣αv(y)pdy,v(x)=∫RnΓ2m(x−y)∣y∣βu(y)qdy.(9)
Proof. By Theorem 1.3, ui>0 and vi>0 for 0≤i≤m away from the origin, with the notation of Section 2. Put g=um=∣x∣αvp.
If n≤2, Lemma 3.1(i) applied to w=um−1 gives g≡0, which is false. Let n≥3. By Lemma 3.1(ii), um−1=cm−1+Γ∗g with cm−1≥0. We claim that for 1≤j≤m with 2j<n,
um−j=cm−j+Γ2j∗g,cm−j≥0.(10)
Suppose (10) holds for j−1≥1. Lemma 3.1(ii) applied to w=um−j and h=um−j+1 gives Γ∗um−j+1≤um−j<∞. Since Γ∗1=∞, this forces cm−j+1=0, and then Γ∗um−j+1=Γ2j∗g, which is (10) for j.
Suppose 2j≥n for some j≤m, and take the smallest such j. Then j≥2 because n≥3. As above, Γ∗(Γ2j−2∗g)≤um−j<∞. On the other hand, for ∣y∣≥1 we have ∣y−z∣≤2∣y∣ when ∣z∣≤1, so
Γ2j−2∗g(y)≥γn,2j−2(2∣y∣)2j−2−n∫∣z∣≤1g=c∣y∣2j−2−n,c>0.
The integral of ∣x−y∣2−n∣y∣2j−2−n over {∣y∣≥2∣x∣+1} diverges, because the integrand is comparable to ∣y∣2j−2n and 2j−2n≥−n. This contradiction shows n>2m. Then (10) holds for j=m, and u=cu+Γ2m∗g with cu≥0. These identities hold at every point, since each Γ2j∗g=Γ∗(Γ2j−2∗g) is superharmonic and the argument at the end of the proof of Lemma 3.1 applies. In the same way v=cv+Γ2m∗f with f=∣x∣βuq.
It remains to show cu=cv=0. Since Γ2m(x−y)≥γn,2m(2R)2m−n for x,y∈BR, we have
u≥γn,2m(2R)2m−n∫BRg,v≥γn,2m(2R)2m−n∫BRfin BR.
Exactly as in [12 Proposition 2.4], the case R=1 gives ∫B1f≥c(∫B1g)q and ∫B1g≥c(∫B1f)p with c depending only on n,m,p,q,α,β. As pq>1, this bounds ∫B1f by a constant independent of the solution. The pair uR(x)=Rσ1u(Rx), vR(x)=Rσ2v(Rx) is again a positive C2m solution of (1), and changing variables yields ∫BRf≤CRn−2m−σ2 for all R>0. If cu>0, then ∫BRf≥cuq∫BR∣x∣βdx=cRn+β, and n+β>n−2m−σ2, a contradiction for large R. Thus cu=0, and likewise cv=0. □
Proof of Theorem 1.1. Let u,v be as in Theorem 1.1. By Proposition 3.2, n>2m≥4 and (u,v) satisfies (6) with μ=2m∈(1,n) and c0=γn,2m. Condition (4) is (5) with μ=2m. Theorem 1.5 shows that no such pair exists. □
Proof of Corollary 1.2. Since p,q≥1 and pq=1, we have pq>1, and Theorem 1.1 shows that the condition is necessary. Conversely, assume the condition. Its left side is positive, so n>2m. We apply [19 Corollary 2], which rests on [19 Theorems 1 and 3]: a shooting argument, and the nonexistence of radial solutions of the Navier problem on balls, proved there by a Pohozaev identity. With its order taken equal to 2m, it concerns the system (−Δ)mu=∣x∣−s1vq, (−Δ)mv=∣x∣−s2up with s1,s2<2, where the weights are denoted σ1,σ2 in [19]. We take s1=−α, s2=−β, and interchange the names of p and q. The exponent condition there then becomes ours, and the hypothesis 2m<n of [19] holds. It gives a radial pair u,v, positive in Rn and of class C2m(Rn∖{0}), solving (1) away from the origin.
It remains to check the origin. In [19] the pair is obtained from the second-order system satisfied by wi=(−Δ)iu and zi=(−Δ)iv, 0≤i≤m−1, with initial data wi(0),zi(0)>0 and wi′(0)=zi′(0)=0. Hence each component W satisfies
W(r)=W(0)−∫0rt1−n∫0tτn−1F(τ)dτdt,
where F is the next component, or ∣x∣αvp, or ∣x∣βuq. All of these are continuous on [0,∞) because α,β≥0. As noted in Section 5, W(∣x∣) is then of class C1 and solves −ΔW=F in the sense of distributions on Rn. In particular u,v∈C1(Rn), so ∣x∣αvp and ∣x∣βuq are locally Hölder continuous. Applying Schauder estimates to wm−1,wm−2,…,w0 in turn, and likewise to the zi, gives u,v∈C2m(Rn). □
Remark 3.3. For m odd, the system (1) with p=q=1 has no positive classical solution, for any n≥1 and α,β≥0. Indeed, by Remark 2.3, Theorem 1.3 applies. The proof of Proposition 3.2, up to the determination of cu and cv, does not use pq>1. It gives n>2m, u≥Γ2m∗(∣x∣αvp), and v≥Γ2m∗(∣x∣βuq). The argument of Step 1 in Section 4 uses only these inequalities and forces pq>1, which is false. By Proposition 1.4, for m even and α=β=0 such solutions exist. So, at least without weights, the parity of m decides the linear case.
The integral system
In this section n≥2, 1<μ<n, and (u,v) is a positive continuous solution of (6). We write
K(z)=c0∣z∣μ−n,f=∣x∣βuq,g=∣x∣αvp,a=q+11,b=p+11,
so that u=K∗g and v=K∗f at every point. We follow [12]. We follow the strategy of [12] but give the argument for the general Riesz kernel, including the virial identity and the pressure estimate in Appendix A. The second-order kernel there has n−2 in place of n−μ, and a statement of [12] quoted below is meant with this replacement. The weights α,β and the scaling exponents σ1,σ2 are denoted A,B and α,β in [12]. To avoid a further clash, we write m0=n+3 for the exponent denoted m in [12 (3.2)]. Write k=n(m0+1), N=k+n+2, and take the radial cutoffs
d(x)=min{1,(2−∣x∣)+},η(x)=[(1−(∣x∣−1)3)+]N,w(x)=−x⋅∇η(x),X(x)=xη(x).
These coincide with the cutoffs of [12 (3.2)]; their differential bounds are collected in Appendix § 6.
Step 1: origin and exponents. Since v(0)>0 and v is continuous, v≥v(0)/2 in some ball Bρ. Finiteness of u(0) then gives ∫Bρ∣y∣−μ+αdy<∞, that is, α>−μ. Likewise β>−μ. In particular f,g∈Lloc1(Rn). For x,y∈BR we have K(x−y)≥c0(2R)μ−n. With mu(R)=minBRu and mv(R)=minBRv, this gives
mu(R)≥cRμ+αmv(R)p,mv(R)≥cRμ+βmu(R)q.
Hence mu(R)1−pq≥cRμ+α+p(μ+β), with a positive exponent. If pq≤1, the left side stays bounded as R→∞, so pq>1. We set
σ1=pq−1μ+α+p(μ+β),σ2=pq−1μ+β+q(μ+α).
Then pσ2=σ1+μ+α and qσ1=σ2+μ+β, and the pair uR(x)=Rσ1u(Rx), vR(x)=Rσ2v(Rx) is again a positive continuous solution of (6). A direct computation gives
γμ:=a(n+β)+b(n+α)−(n−μ)=(1−a−b)(σ1+σ2+μ−n),1−a−b=(p+1)(q+1)pq−1,(11)
so (5) means γμ>0 and implies σ1+σ2+μ−n>0. Finally,
∫B1(fRuR+gRvR)=Rσ1+σ2+μ−n∫BR(fu+gv).(12)
Step 2: universal bounds. We make the mass estimate explicit, since it is used repeatedly below. Put F=∫B1f and G=∫B1g. Both numbers are positive and finite. The lower bound K(x−y)≥c02μ−n for x,y∈B1 gives u≥cG and v≥cF on B1, with c>0 depending only on the kernel. Consequently,
F=∫B1∣x∣βuqdx≥c1Gq,G=∫B1∣x∣αvpdx≥c2Fp.
Since pq>1, we have Fpq−1≤(c1c2q)−1 and, symmetrically, Gpq−1≤(c2c1p)−1. Thus F+G≤C independently of the solution. Apply this bound to (uR,vR) from Step 1. The identities fR(x)=Rσ2+μf(Rx) and gR(x)=Rσ1+μg(Rx) give, for every R>0,
∫BRf≤CRn−μ−σ2,∫BRg≤CRn−μ−σ1,(13)
∫B4(f+g+u+v)≤C,∫∣x−y∣≥ℓK(x−y)(f(y)+g(y))dy≤Cℓμ−n(x∈B2, 0<ℓ≤1).(14)
In the dyadic tail estimate, the annulus 2j≤∣y∣<2j+1 contributes at most C2j(μ−n)2(j+1)(n−μ−σ2)≤C2−jσ2, and the series converges because σ2>0. The same dyadic estimate holds for x∈B4, because ∣x−y∣≥2j−1 there for j≥3. Together with K∈Lloc1 and Tonelli’s theorem, it gives the bound for ∫B4(u+v). For the second estimate in (14), the part of the integral over y∈B8 is at most c0ℓμ−n∫B8(f+g)≤Cℓμ−n by (13), and the part over y∈/B8 is at most C≤Cℓμ−n by the dyadic estimate, since ℓμ−n≥1 for 0<ℓ≤1.
Step 3: differentiability. Here μ>1 is used. The function ∣∇K(z)∣=(n−μ)c0∣z∣μ−n−1 is locally integrable, and g is bounded near every x0=0. Split g into a bounded part with compact support near x0 and a remainder, as in the last paragraph of the proof of [12 Lemma 2.2]. Since K∈Wloc1,1, the convolution of K with the first part is C1 with gradient given by ∇K, and the remainder is handled by dominated convergence. This shows that u,v∈C1(Rn∖{0}) and
∇u(x)=∫Rn∇K(x−y)g(y)dy,∇K(z)=−(n−μ)∣z∣2zK(z),
with an absolutely convergent integral at every x=0. The same holds for v.
Step 4: virial identity and localization. Define the interaction measure dπ(x,y)=f(x)g(y)K(x−y)dxdy, whose marginals are fudx and gvdy, and let
Ef=∫ηfu,Eg=∫ηgv,E=Ef+Eg,Wf=∫wfu,Wg=∫wgv.
Then
a(n+β)Ef+b(n+α)Eg−aWf−bWg=(n−μ)∫R2n∣x−y∣2(X(x)−X(y))⋅(x−y)dπ(x,y).(15)
We justify the identity directly, including all kernel-gradient integrals. Fix x=0, set r=∣x∣, and first suppose r is sufficiently small. Since v is bounded near the origin, g(y)≤Cu,vrα on Br/2(x), even when α<0. Using ∣X(x)∣≤Cr, the local part satisfies
∣X(x)∣∫∣x−y∣≤r/2∣∇K(x−y)∣g(y)dy≤Cu,vr1+α∫0r/2sμ−2ds≤Cu,vrμ+α.
This is bounded as r↓0 because μ>1 and α>−μ. On ∣x−y∣>r/2 we have ∣X(x)∣≤Cr≤2C∣x−y∣, and therefore
∣X(x)∣∫∣x−y∣>r/2∣∇K(x−y)∣g(y)dy≤C∫K(x−y)g(y)dy=Cu(x).
The right side remains bounded near zero. Away from zero, the nearby source g is locally bounded and ∇K∈Lloc1, while the far part is dominated by a multiple of K∗g. Since X has compact support and f∈Lloc1, these bounds prove
∫Rnf(x)∣X(x)∣∫Rn∣∇K(x−y)∣g(y)dydx<∞.
Interchanging (f,u,β) and (g,v,α) gives the analogous estimate. We may therefore use Fubini’s theorem in both differentiated potential identities.
For 0<ρ<1, let Ωρ=Rn∖Bρ. Integration by parts, using f=∣x∣βuq and div(∣x∣βX)=∣x∣β((n+β)η−w), yields
−∫ΩρfX⋅∇u=a∫Ωρ((n+β)η−w)fu+aρβ+1∫∂Bρuq+1dS.
The last term is Ou,v(ρn+β) and tends to zero since β>−μ>−n. The corresponding calculation with g and v has a vanishing remainder Ou,v(ρn+α). Taking ρ↓0 thus gives
−∫fX⋅∇u−∫gX⋅∇v=a(n+β)Ef+b(n+α)Eg−aWf−bWg.
On the other hand, ∇K(z)=−(n−μ)zK(z)/∣z∣2 and the absolute convergence proved above permit us to exchange x and y in the second kernel-gradient integral. The resulting sum is (n−μ)∫Φdπ, establishing (15).
We next prove the localization estimate, following the geometric decomposition of [12 Lemma 3.2]. Write
Φ(x,y)=∣x−y∣2(X(x)−X(y))⋅(x−y),D(x,y)=∣x−y∣2∣x∣2((x−y)⋅x)2,
with w(x)D(x,y)=0 where w(x)=0. Fix 0<δ<1/4 and, for d(z)>0, set ℓ(z)=δd(z)m0. Call (x,y) close if ∣x−y∣≤ℓ(z) for at least one endpoint z∈{x,y} with d(z)>0, and far otherwise. On close pairs, since d is 1-Lipschitz, d is comparable to d(z) along the segment [x,y]. The derivative bounds for η and X in (20) (proved independently in Appendix § 6) and the identity
∇X=ηI−wωx⊗ωx
show by the mean value theorem and Taylor’s formula that
∣Φ−η(x)+w(x)D(x,y)∣+∣η(x)−η(y)∣≤Cδd(z)N≤Cδ(η(x)+η(y)).
The marginal identities therefore bound the integral over close pairs by CδE.
For a far pair, every endpoint z∈B2 satisfies ∣x−y∣>ℓ(z). Using ∣X∣≤CdN, w≤CdN−1, and 0≤D≤1, we obtain
∣Φ−η(x)+w(x)D∣+∣η(x)−η(y)∣≤Cz∈{x,y}:d(z)>0∑(d(z)N−1+ℓ(z)d(z)N).
For terms associated with x, the tail bound (14) gives
C∫B2f(x)(d(x)N−1+ℓ(x)d(x)N)ℓ(x)μ−ndx≤C(δμ−n+δμ−n−1)∫B2f(x)(d(x)A1+d(x)A2)dx≤Cδ,
where A1=N−1−m0(n−μ)=μn+3μ+2n+1>0 and A2=N−m0(n−μ+1)=μn+3μ+n−1>0. The terms associated with y are bounded in the same way, using g. Finally, Ef−Eg=∫(η(x)−η(y))dπ by the marginal identities. Adding the close- and far-pair contributions and then choosing δ sufficiently small yields, for every ε>0,
∣Ef−Eg∣+∫R2n∣Φ(x,y)−η(x)+w(x)D(x,y)∣dπ≤εE+Cε.(16)
Step 5: the interval comparison. Fix a unit vector e and distinct z,z′∈e⊥. This is possible because n≥2. Let λ=∣z−z′∣ and h=s−s′ for x=z+se, y=z′+s′e, and put
K(s,s′)=c0(λ2+h2)(μ−n)/2,D(s,s′)=λ2+h2h2.
For a bounded interval I=(s−,s+) with midpoint mI, let KI(s′)=21(K(s−,s′)+K(s+,s′)), and define KJ(s) in the second variable. The comparison [12 Lemma 4.1] holds with n−2 replaced by n−μ: for nonempty open intervals I,J and coefficients cI,cJ≥0, with I bounded when cI>0 and J bounded when cJ>0,
∫I×J(cI+cJ)K≤∫I×J[cIKI+cJKJ+((n−μ)min{cI,cJ}D+∣cI−cJ∣)K].(17)
We include the short proof, which uses only two properties of the kernel. The first is ∂sK=−(n−μ)h(λ2+h2)−1K=−∂s′K, which with integration by parts against s−mI and s′−mJ gives
∫I×J(2K−KI−KJ)=(n−μ)∫I×Jλ2+h2h2−(mI−mJ)hK.
The second is that K depends on h only through ∣h∣. Let d0=mI−mJ. The image of dsds′ under (s,s′)↦h has density ρ(h)=∣I∩(J+h)∣, the overlap of two intervals whose centers are ∣h−d0∣ apart. So ρ(h)=F(∣h−d0∣) with F nonincreasing. If d0≥0 and h>0, then ∣h−d0∣≤∣h+d0∣ and ρ(h)≥ρ(−h), with the reverse inequality when d0≤0. Pairing h with −h against the odd function hK/(λ2+h2) gives d0∫I×Jh(λ2+h2)−1K≥0, so the left side of the last display is at most (n−μ)∫I×JDK. This proves (17) when cI=cJ. For general coefficients, if at most one of them is positive, then ∣cI−cJ∣=cI+cJ and (17) holds pointwise. Otherwise both intervals are bounded, all integrals are finite, and with c=min{cI,cJ}
∫IX∖I?[(cI+cJ)K−cIKI−cJKJ]=c∫IX×I?(2K−KI−KJ)+(cI−c)∫IX×I?(K−KI)+(cJ−c)∫IX×I?(K−KJ).
The first term is at most c(n−μ)∫DK, and the last two are at most ∣cI−cJ∣∫K because KI,KJ≥0. When n−μ≤1, the integral of K over an unbounded interval may diverge. This is why (17) is written with nonnegative terms only and no subtraction is made on an interval with zero coefficient.
Step 6: the pressure estimate. For every ε>0 there is Cε independent of the solution such that
aWf+bWg≤(n−μ)∫R2nw(x)D(x,y)dπ(x,y)+εE+Cε.(18)
This is [12 Proposition 5.1] with n−2 replaced by n−μ. Since it is the longest step and the one most specific to the method, we give a complete proof for the kernel K in Appendix A, given (14) and (17).
Step 7: conclusion. Write T=∫w(x)D(x,y)dπ. Inserting (16) and ∫η(x)dπ=Ef into (15) gives
c1Ef+c2Eg≤aWf+bWg−(n−μ)T+(n−μ)(εE+Cε),c1=a(n+β)−(n−μ),c2=b(n+α).
By (18), the right side is at most (n−μ+1)εE+Cε. Since c1+c2=γμ, we have c1Ef+c2Eg=2γμE+2c1−c2(Ef−Eg), and (16) yields
2γμE≤(n−μ+1+2∣c1−c2∣)εE+Cε.
The energy E is finite for each solution, because u,v are bounded on B2 and α,β>−μ>−n. Choosing ε small, we obtain E≤C with C independent of the solution. As η=1 on B1,
∫B1(fu+gv)≤C
for every positive continuous solution of (6). Applying this to (uR,vR) and using (12) with σ1+σ2+μ−n>0, we get
∫BR(fu+gv)≤CRn−μ−σ1−σ2→0as R→∞.
But ∫BR(fu+gv)≥∫B1(fu+gv)>0 for R≥1. This contradiction proves Theorem 1.5.
Positive solutions without the super polyharmonic property
Proof of Proposition 1.4. Since m is even, (−Δ)m=Δm. We look for radial functions wi=Δiu and zi=Δiv, 0≤i≤m−1, solving
wi(r)=1+∫0rt1−n∫0tτn−1wi+1(τ)dτdt,zi(r)=1+∫0rt1−n∫0tτn−1zi+1(τ)dτdt,(19)
with wm=z0p and zm=w0q. For a continuous function h on [0,∞), the function W(r)=1+∫0rt1−n∫0tτn−1h(τ)dτdt satisfies W(0)=1 and W′(0)=0, and W(∣x∣) solves ΔW=h in the sense of distributions in Rn. Hence a continuous solution of (19) on [0,ρ) with w0,z0≥1 gives u=w0, v=z0 with Δmu=vp and Δmv=uq in Bρ. Since u,v≥1, the nonlinearities are smooth functions of u and v, and Schauder estimates give u,v∈C∞(Bρ).
The right sides of (19) are Lipschitz in the unknowns on the set where w0,z0≥21. The contraction principle gives a solution on some [0,ρ), and it can be continued as long as all wi,zi stay bounded and w0,z0 stay away from zero. On the existence interval, wm,zm>0, so by (19) the functions wm−1,zm−1 are nondecreasing and at least 1. Descending, every wi and zi is nondecreasing and at least 1. In particular u,v≥1 and Δu=w1≥1, Δv=z1≥1.
To exclude a finite maximal radius ρ, we bound the solution on [0,ρ). Since z0p is nondecreasing,
∫0tτn−1z0p≤tnz0(r)p/nfor t≤r,
and, iterating in (19),
w0(r)≤(1+r2)m+r2mz0(r)p,z0(r)≤(1+r2)m+r2mw0(r)q(0≤r<ρ).
If pq<1, then with P=(1+ρ2)m we obtain w0≤P+2ρ2m(Pp+ρ2mpw0pq) on [0,ρ). This bounds w0, hence z0, hence every wi and zi on [0,ρ), which is impossible. If p=q=1, then (19) is a linear Volterra system with continuous kernels, and its solution exists on [0,∞). In both cases the solution is global.
Finally, pq≤1 gives 1−p+11−q+11=(p+1)(q+1)pq−1≤0. Hence p+1n+q+1n≥n>n−2m, which is (4). □
Appendix A. Proof of the pressure estimate
In this appendix we prove (18). The argument is that of [12 Section 5], written for the kernel K(z)=c0∣z∣μ−n. We keep the notation of Section 4. In particular (u,v) is a positive continuous solution of (6) with n≥2 and 1<μ<n, and f=∣x∣βuq, g=∣x∣αvp, with the understanding that, in layer-set constructions only, the source functions have been redefined to vanish at the single point 0. Every constant C below depends only on n,μ,c0,p,q,α,β and on the parameters indicated as subscripts. It never depends on the solution or on the direction e.
The cutoff. Recall the cutoffs d, η, w, X and the integers m0=n+3, k=n(m0+1), N=k+n+2 from Section 4. On 1<∣x∣<2 we have d=1−s with s=∣x∣−1, and 1−s3=d(1+s+s2). Differentiating gives
dN≤η≤CdN,∣∇η∣+w≤CdN−1,∣∇2η∣+∣∇w∣+∣∇2X∣≤CdN−2,dw≤Cη,(20)
while η=1 and w=0 on B1, and η=w=0 outside B2. Since w vanishes to second order at ∣x∣=1, the function w is C1 on Rn.
Cap weights. Fix κ∈(0,1) and δ∈(0,1/4), and put ℓ(x)=δd(x)m0 when d(x)>0. Let Ee and Pe denote expectation and probability for the normalized surface measure on the unit sphere Sn−1. Fix ψ∈C∞([0,∞)) with 0≤ψ≤1, ψ=1 near 0, and ψ=0 on [1,∞), and set Z(l)=Eeψ(∣e−θ∣/l) for 0<l≤1. By rotation invariance, Z does not depend on θ∈Sn−1, and polar coordinates around θ give cln−1≤Z(l)≤Cln−1 and ∣Z′(l)∣≤Cln−2. For e∈Sn−1 define
he(x)=w(x)Z(κd(x))ψ(∣e−ωx∣/(κd(x)))(1<∣x∣<2),he(x)=0otherwise,
where ωx=x/∣x∣. Then Eehe=w. Differentiating the cap profile or Z(κd)−1 costs at most a relative factor Cκd−1, so by (20)
0≤he≤CκdN−n,∣∇he∣≤CκdN−n−1on 1<∣x∣<2.(21)
The profile is constant near its center, so he is C1 there. Moreover, he tends to zero at both boundary spheres, because w vanishes at ∣x∣=1 and N−n>1. Hence he is continuous on Rn, and ∣∇he∣≤CκdN−n−1 at every point off the spheres ∣x∣=1 and ∣x∣=2, since he=0 outside the annulus. A segment meets these spheres in at most four points, so this bound may be integrated along segments.
Lemma A.1. For every direction e,
∫R2n∣he(x)−he(y)∣dπ(x,y)≤CκE+Cκ,δ.(22) Moreover, with De(x,y)=∣x−y∣2((x−y)⋅e)2,
Ee∫R2nhe(x)De(x,y)dπ(x,y)≤∫R2nw(x)D(x,y)dπ(x,y)+CκE,(23) with C independent of κ and $\delta.
Proof. Call a pair (x,y) close if ∣x−y∣≤ℓ(z) for some endpoint z∈{x,y} with d(z)>0, and far otherwise. Since d is 1-Lipschitz and ℓ(z)≤δd(z), for a close pair we have (1−δ)d(z)≤d≤(1+δ)d(z) on the segment [x,y]. Integrating the gradient bound along the segment gives
∣he(x)−he(y)∣≤Cκd(z)N−n−1ℓ(z)=Cκδd(z)N+2≤Cκδ(η(x)+η(y)),
since m0=n+3 and d(z)N≤η(z). By the marginal identities, close pairs contribute at most Cκδ(Ef+Eg). For a far pair we use ∣he(x)−he(y)∣≤he(x)+he(y). The term he(x) vanishes unless 1<∣x∣<2, and then ∣x−y∣>ℓ(x). By (21) and (14), its contribution is at most
Cκ∫B2f(x)d(x)N−n∫∣x−y∣>ℓ(x)K(x−y)g(y)dydx≤Cκδμ−n∫B2f(x)d(x)N−n−m0(n−μ)dx≤Cκ,δ,
because N−n−m0(n−μ)=μn+3μ+n+2>0 and ∫B2f≤C. The term he(y) is treated in the same way with g in place of f. This proves (22).
For unit vectors z^, e, ω we have ∣(z^⋅e)2−(z^⋅ω)2∣≤∣e−ω∣∣e+ω∣≤2∣e−ω∣. On the support of he(x) we have ∣e−ωx∣≤κd(x), so heDe≤heD+2κdhe. Averaging in e, using Eehe=w and dw≤Cη, gives (23). □
Layer intervals. Fix e. For x∈Rn and 0<t<f(x), let I=I(e,x,t) be the connected component containing x of the open set {x+se:f(x+se)>t} on the line x+Re. For the purpose of defining the layers, we assign the value f(0)=0 (and similarly g(0)=0), even if a weight exponent is zero. This modification at one point changes none of the integrals or potentials. Since f is continuous away from the origin, the resulting layer set is open. Define
χf(e,x,t)={infIhe,0,if I is bounded and ∣I∣≤2δinfId,otherwise,(24)
and define χg(e,y,t′) in the same way from the layers {g>t′}. Thus χf(e,x,t) depends on x only through I, and 0≤χf(e,x,t)≤he(x).
These functions are jointly measurable in (e,x,t). Indeed, the forward exit distance of x from its component exceeds r≥0 exactly when the compact segment {x+se:0≤s≤r} lies in {f>t}, and this is an open condition in (e,x,t). The same holds for the backward exit distance. The length of I is their sum, possibly +∞. Since d and he are continuous, the infimum in (24) may be taken over rational points between the forward and backward exit points. When I is bounded, its endpoints are x plus or minus the exit distances times e, so they are measurable as well, and so is the endpoint average uI defined below, because u is continuous. Define the losses
Lf(e)=∫Rnu(x)∫0f(x)(he(x)−χf(e,x,t))dtdx,Lg(e)=∫Rnv(y)∫0g(y)(he(y)−χg(e,y,t′))dt′dy.
They are finite because he is bounded and supported in B2∖B1, where u and v are bounded and f,g are integrable.
Lemma A.2. For every P≥1,
Ee(Lf(e)+Lg(e))≤(Cκδ+Cκ,δP−1)E+CP, where the last constant does not depend on κ, δ, or P.
Proof. We treat Lf. Only points with $1<|x|<2$ contribute. Fix such an x and write d=d(x) and ℓ=δdm0. If ∣I∣≤ℓ, then infId≥d−ℓ≥(1−δ)d and ∣I∣≤δd≤2δinfId, so the component is retained. Integrating (21) along I as in Lemma A.1 gives he(x)−χf(e,x,t)≤CκδdN. If ∣I∣>ℓ, then one of the two segments of length ℓ/2 starting at x in the directions ±e lies in {f>t}. Let F be the set of e for which the forward segment does. Polar coordinates centered at $x give
∫Bℓ/2(x)f≥n∣Sn−1∣t(2ℓ)nPe(F).
Since Bℓ/2(x)⊂B4, (14) gives Pe{∣I∣>ℓ}≤2Pe(F)≤C/(tℓn). For t≥Pd−k we bound the loss on this event by supehe(x)≤CκdN−n, and obtain
Ee[he(x)−χf(e,x,t)]≤CκδdN+CκdN−ntδndm0nC≤(Cκδ+Cκ,δP−1)dN,
because k−m0n=n. For t<Pd−k we use only Ee[he(x)−χf]≤Eehe(x)=w(x), so
∫0min{f(x),Pd−k}Ee[he(x)−χf(e,x,t)]dt≤Pd−kw(x)≤CPdN−1−k=CPdn+1≤CP.
By Tonelli’s theorem we may integrate these bounds in t and then against u(x)dx. The first gives at most (Cκδ+Cκ,δP−1)∫B4fu≤(Cκδ+Cκ,δP−1)Ef, and the second at most CP∫B4u≤CP. The same argument applies to Lg. □
Endpoint values. Suppose χf(e,x,t)>0, and let x± be the endpoints of I. Then I is bounded and he≥χf>0 on I. Hence, by continuity, he(x±)≥χf>0, so x± lie in the open annulus 1<∣x∣<2. There f is continuous, and maximality of I gives f(x±)=t, that is, u(x±)=(t∣x±∣−β)1/q. Put uI=21(u(x−)+u(x+)) and τ=(t∣x∣−β)1/q. Since ∣x±−x∣≤∣I∣≤2δd(x) and the function r↦r−β/q has a bounded logarithmic derivative on [1,2],
∣uI−τ∣≤Cδd(x)τ≤Cδd(x)u(x),(25)
where we used t<f(x), that is, τ<u(x). The same holds for g, with vJ and (t′∣y∣−α)1/p. When a component weight vanishes, every product involving its endpoint average is defined to be zero.
Assembly over layers. Fix e and t,t′>0. Write x=z+se and y=z′+s′e with z,z′∈e⊥, and take z=z′. For each component I of the f-layer on the line through z and each component J of the g-layer on the line through z′, apply (17) with cI=χI and cJ=χJ, the common values of χf on I and of χg on J. Since min{χI,χJ}≤χI≤he(x) for x∈I,
∫I×J(χI+χJ)K≤∫I×J[χIKI+χJKJ+(n−μ)he(x)De(x,y)+∣χI−χJ∣K]dsds′.(26)
Every term is nonnegative. Each open layer on a line has at most countably many components. We sum (26) over all pairs of components and integrate over z,z′∈e⊥ and t,t′>0. The set {z=z′} is null because n≥2. By Tonelli’s theorem and ∫0∞1{g(y)>t′}dt′=g(y), the identity u=K∗g turns the two terms on the left into
Uf(e)=∫Rn∫0f(x)χf(e,x,t)u(x)dtdx,Ug(e)=∫Rn∫0g(y)χg(e,y,t′)v(y)dt′dy.
For the endpoint terms, integrating KI over all g-layers recovers 21((K∗g)(x−)+(K∗g)(x+))=uI. For each endpoint, the omitted points y form the line z+Re, which is a null set. Hence these terms become
Vf(e)=∫Rn∫0f(x)χf(e,x,t)uIdtdx,Vg(e)=∫Rn∫0g(y)χg(e,y,t′)vJdt′dy.
The pressure term becomes (n−μ)∫he(x)De(x,y)dπ. We check finiteness before subtracting anything. Let Hf(e)=∫hefu and Hg(e)=∫hegv. Both are finite for each solution. We have Uf≤Hf, Ug≤Hg, and by (25) also Vf≤CHf and Vg≤CHg. The pressure term is at most (n−μ)Hf. Finally, ∣χf(e,x,t)−χg(e,y,t′)∣≤he(x)+he(y), so the integrated mismatch term is at most Hf+Hg. All integrated terms are therefore finite, even when ∫KK=∞ for some unbounded component. Subtracting Vf+Vg and using
∣χf(e,x,t)−χg(e,y,t′)∣≤∣he(x)−he(y)∣+(he(x)−χf(e,x,t))+(he(y)−χg(e,y,t′)),
we find that Se=Uf(e)−Vf(e)+Ug(e)−Vg(e) satisfies
Se≤(n−μ)∫he(x)De(x,y)dπ+∫∣he(x)−he(y)∣dπ+Lf(e)+Lg(e).(27)
Recovery of the pressure. For x=0,
∫0f(x)[u(x)−(t∣x∣−β)1/q]dt=f(x)u(x)−q+1q∣x∣−β/qf(x)1+1/q=af(x)u(x),
since ∣x∣−β/qf(x)1/q=u(x). With τ=(t∣x∣−β)1/q≤u(x) and (25),
he(x)(u−τ)−χf(u−uI)=(he(x)−χf)(u−τ)+χf(uI−τ)≤(he(x)−χf)u+Cδd(x)he(x)u,
where u=u(x) and χf=χf(e,x,t). Integrating over 0<t<f(x) and then in x, and doing the same for g, we obtain
aHf(e)+bHg(e)≤Se+Lf(e)+Lg(e)+Cδ∫dhe(fu+gv)dx.
We insert (27), which contains a second copy of Lf+Lg, and average over e. Since EeHf=Wf, EeHg=Wg, and Ee(dhe)=dw≤Cη, Lemmas A.1 and A.2 give
aWf+bWg≤(n−μ)∫w(x)D(x,y)dπ+(Cκ+Cκδ+Cκ,δP−1+Cδ)E+Cκ,δ+CP.
The constant multiplying κ does not depend on κ, δ, or P. The one multiplying δ depends only on κ, and the one multiplying P−1 does not depend on P. We first choose κ, then δ, then P, so that the coefficient of E is at most $\varepsilon. This proves (18).
Declarations
Conflict of interest. The author declares that he has no conflict of interest.
Data availability. Data sharing is not applicable to this article, as no datasets were generated or analyzed during the current study.
Use of AI tools. AI-assisted tools were used during the preparation of the manuscript. The author reviewed and edited all AI-assisted output and takes full responsibility for the content of the work.
References
Phuong Le1,2 (ORCID: 0000-0003-4724-7118)
1Faculty of Economic Mathematics, University of Economics and Law, Ho Chi Minh City, Vietnam; 2Vietnam National University, Ho Chi Minh City, Vietnam
Email address: phuongl@uel.edu.vn