STABLE SOLUTIONS FOR THE BILAPLACIAN WITH EXPONENTIAL NONLINEARITY. 8 0 JUANDA´VILA,LOUISDUPAIGNE,IGNACIOGUERRA,ANDMARCELOMONTENEGRO 0 2 Abstract. Letλ∗>0denotethelargestpossiblevalueofλsuchthat n ∆2u=λeu inB a 8 J < u= ∂u =0 on∂B ∂n 6 : hasasolution,whereBistheunitballinRN andnistheexteriorunitnormal 1 vector. Weshowthatforλ=λ∗thisproblempossessesauniqueweaksolution u∗. Weprovethatu∗issmoothifN ≤12andsingularwhenN ≥13,inwhich ] P case u∗(r) = −4logr+log(8(N −2)(N −4)/λ∗)+o(1) as r → 0. We also A considertheproblemwithgeneralconstant Dirichletboundaryconditions. . h t a m 1. Introduction [ We study the fourth order problem 1 ∆2u=λeu in B v u=a on ∂B 1 (1)  4  ∂u =b on ∂B 4 ∂n 2 . where a, b R, B is the unit ball in RN, N 1, n is the exterior unit normal 1 ∈ ≥ vector and λ 0 is a parameter. 0 ≥ Recently higherorderequationshaveattractedthe interestofmanyresearchers. 8 0 In particular fourth order equations with an exponential non-linearity have been : studiedin4dimensions,inasettinganalogoustoLiouville’sequation,in[3,12,24] v and in higher dimensions by [1, 2, 4, 5, 13]. i X We shall pay special attention to (1) in the case a = b = 0, as it is the natural r fourth order analogue of the classical Gelfand problem a ∆u=λeu in Ω (2) − u=0 on ∂Ω (cid:26) (Ω is a smooth bounded domain in RN) for which a vast literature exists [7, 8, 9, 10, 18, 19, 20, 21]. From the technical point of view, one of the basic tools in the analysis of (2) is the maximum principle. As pointed out in [2], in general domains the maximum principlefor∆2 withDirichletboundaryconditionisnotvalidanymore. Oneofthe reasons to study (1) in a ball is that a maximum principle holds in this situation, see [6]. In this simpler setting, though there are some similarities between the two problems, severaltools that are well suited for (2) no longer seem to work for (1). 1991 Mathematics Subject Classification. Primary35J65, Secondary35J40. Key words and phrases. Biharmonic,singularsolutions,stability. 1 2 JUANDA´VILA,LOUISDUPAIGNE,IGNACIOGUERRA,ANDMARCELOMONTENEGRO Asastart,letusintroducetheclassofweaksolutionsweshallbeworkingwith: we say that u H2(B) is a weak solution to (1) if eu L1(B), u = a on ∂B, ∈ ∈ ∂u =b on ∂B and ∂n ∆u∆ϕ=λ euϕ, for all ϕ C∞(B). ∈ 0 ZB ZB The following basic result is a straightforwardadaptation of Theorem 3 in [2]. Theorem1.1. ([2])Thereexistsλ∗ suchthatif0 λ<λ∗ then (1)has aminimal ≤ smooth solution u and if λ>λ∗ then (1) has no weak solution. λ The limit u∗ = limλրλ∗uλ exists pointwise, belongs to H2(B) and is a weak solution to (1). It is called the extremal solution. The functions u , 0 λ < λ∗ and u∗ are radially symmetric and radially de- λ ≤ creasing. The branchofminimal solutionsof (1) hasanimportantproperty,namely u is λ stable in the sense that (3) (∆ϕ)2 λ euλϕ2, ϕ C∞(B), ≥ ∀ ∈ 0 ZB ZB see [2, Proposition 37]. Theauthorsin[2]poseseveralquestions,someofwhichweaddressinthiswork. Firstweshowthattheextremalsolutionu∗ istheuniquesolutionto(1)intheclass of weak solutions. Actually the statement is stronger, asserting that for λ = λ∗ there are no strict super-solutions. Theorem 1.2. If (4) v H2(B), ev L1(B), v =a, ∂v b ∈ ∈ |∂B ∂n|∂B ≤ and (5) ∆v∆ϕ λ∗ evϕ ϕ C∞(B), ϕ 0, ≥ ∀ ∈ 0 ≥ ZB ZB then v =u∗. In particular for λ=λ∗ problem (1) has a unique weak solution. This result is analogous to work of Martel [19] for more general versions of (2) where the exponential function is replaced by a positive, increasing, convex and superlinear function. Next, we discuss the regularity of the extremal solution u∗. In dimensions N = 5,...,16 the authors of [2] find, with a computer assisted proof, a radial singular solutionU to(1) witha=b=0 associatedto a parameterλ >8(N 2)(N 4). σ σ − − They show that λ < λ∗ if N 10 and claim to have numerical evidence that σ ≤ this holds for N 12. They leave open the question of whether u∗ is singular in ≤ dimension N 12. We prove ≤ Theorem 1.3. If N 12 then the extremal solution u∗ of (1) is smooth. ≤ The method introduced in [10, 20] to prove the boundedness of u∗ in low di- mensions for (2) seems not useful for (1), thus requiring a new strategy. A first indicationthattheborderlinedimensionfortheboundednessofu∗ is12isRellich’s inequality [23], which states that if N 5 then ≥ N2(N 4)2 ϕ2 (6) (∆ϕ)2 − ϕ C∞(RN), ZRN ≥ 16 ZRN |x|4 ∀ ∈ 0 STABLE SOLUTIONS FOR ∆2u=λeu 3 where the constant N2(N 4)2/16 is known to be optimal. The proof of Theo- rem1.3isbasedontheobse−rvationthatifu∗ issingularthenλ∗eu∗ 8(N 2)(N ∼ − − 4)x−4 near the origin. But 8(N 2)(N 4) > N2(N 4)2/16 if N 12 which | | − − − ≤ would contradict the stability condition (3). In view of Theorem 1.3, it is natural to ask whether u∗ is singular in dimension N 13. If a=b=0, we prove ≥ Theorem 1.4. Let N 13 and a = b = 0. Then the extremal solution u∗ to (1) ≥ is unbounded. Forgeneralboundaryvalues,itseemsmoredifficulttodeterminethedimensions forwhichtheextremalsolutionissingular. Weobservefirstthatgivenanya,b R, ∈ u∗ is the extremal solution of (1) if and only if u∗ a is the extremal solution of − the same equation with boundary condition u=0 on ∂B. In particular, if λ∗(a,b) denotestheextremalparameterforproblem(1),onehasthatλ∗(a,b)=e−aλ∗(0,b). So the value of a is irrelevant. But one may ask if Theorem 1.4 still holds for any N 13andanyb R. Thesituationturnsouttobesomewhatmorecomplicated: ≥ ∈ Proposition 1.5. a) Fix N 13 and take any a R. Assume b 4. There exists a critical ≥ ∈ ≥ − parameter bmax >0, depending only on N, such that the extremal solution u∗ is singular if and only if b bmax. b) Fix b 4 and take any a R≤. There exists a critical dimension Nmin ≥− ∈ ≥ 13,dependingonlyonb,suchthattheextremalsolutionu∗ to (1)issingular if N Nmin. ≥ Remark 1.6. We have not investigated the case b< 4. • − If follows from item a) that for b [ 4,0], the extremal solution is singular • ∈ − if and only if N 13. ≥ Italsofollows fromitema)thatthereexistvaluesofbforwhichNmin >13. • We do not know whether u∗ remains bounded for 13 N <Nmin. ≤ Our proof of Theorem 1.4 is related to an idea that Brezis and Va´zquez ap- pliedfor the Gelfand problemandis basedon a characterizationof singularenergy solutions through linearized stability (see Theorem 3.1 in [8]). In our context we show Proposition 1.7. Assume that u H2(B) is an unbounded weak solution of (1) ∈ satisfying the stability condition (7) λ euϕ2 (∆ϕ)2, ϕ C∞(B). ≤ ∀ ∈ 0 ZB ZB Then λ=λ∗ and u=u∗. We do not use Proposition 1.7 directly but some variants of it – see Lemma 2.6 and Remark 2.7 in Section 2 – because we do not have at our disposal an explicit solutiontotheequation(1). Instead,weshowthatitisenoughtofindasufficiently good approximation to u∗. When N 32 we are able to construct such an ap- ≥ proximation by hand. However, for 13 N 31 we resort to a computer assisted ≤ ≤ generation and verification. 4 JUANDA´VILA,LOUISDUPAIGNE,IGNACIOGUERRA,ANDMARCELOMONTENEGRO Only in very few situations one may take advantage of Proposition 1.7 directly. For instance for problem (1) with a=0 and b= 4 we have an explicit solution − u¯(x)= 4log x − | | associatedto λ¯ =8(N 2)(N 4). Thanks to Rellich’s inequality (6) the solution − − u¯ satisfies condition (7) when N 13. Therefore, by Theorem 1.3 and a direct ≥ application of Proposition 1.7 we obtain Theorem 1.4 in the case b= 4. − In [2] the authors say that a radial weak solution u to (1) is weakly singular if limru′(r) exists. r→0 For example, the singular solutions U of [2] verify this condition. σ As a corollary of Theorem 1.2 we show Proposition 1.8. The extremal solution u∗ to (1) with b 4 is always weakly ≥ − singular. A weakly singular solution either is smooth or exhibits a log-type singularity at theorigin. Moreprecisely,ifuisanon-smoothweaklysingularsolutionof (1)with parameter λ then (see [2]) 8(N 2)(N 4) limu(r)+4logr =log − − , r→0 λ limru′(r)= 4. r→0 − In Section 2 we describe the comparison principles we use later on. Section 3 is devotedtotheproofoftheuniquenessofu∗ andPropositions1.7and1.8. Weprove Theorem1.3,theboundednessofu∗ inlowdimensions,inSection4. Theargument for Theorem 1.4 is contained in Section 5 for the case N 32 and Section 6 for ≥ 13 N 31. In Section 7 we give the proof of Proposition 1.5. ≤ ≤ Notation. B : ball of radius R in RN centered at the origin. B =B . R 1 • n: exterior unit normal vector to B R • All inequalities or equalities for functions in Lp spaces are understood to • be a.e. 2. Comparison principles Lemma 2.1. (Boggio’s principle, [6]) If u C4(B ) satisfies R ∈ ∆2u 0 in B R ≥ ∂u  u= =0 on ∂B  ∂n R then u 0 in B . ≥ R  Lemma 2.2. Let u L1(B ) and suppose that R ∈ u∆2ϕ 0 ≥ ZBR for all ϕ C4(B ) such that ϕ 0 in B , ϕ = 0 = ∂ϕ . Then u 0 in ∈ R ≥ R |∂BR ∂n|∂BR ≥ B . Moreover u 0 or u>0 a.e. in B . R R ≡ For a proof see Lemma 17 in [2]. STABLE SOLUTIONS FOR ∆2u=λeu 5 Lemma 2.3. If u H2(B ) is radial, ∆2u 0 in B in the weak sense, that is R R ∈ ≥ ∆u∆ϕ 0 ϕ C∞(B ), ϕ 0 ≥ ∀ ∈ 0 R ≥ ZBR and u 0, ∂u 0 then u 0 in B . |∂BR ≥ ∂n|∂BR ≤ ≥ R Proof. We only deal with the case R=1 for simplicity. Solve ∆2u =∆2u in B 1 1 ∂u  u = 1 =0 on ∂B  1 ∂n 1 in the sense u1 ∈H02(B1) and B1∆u1∆ϕ= B1∆u∆ϕ for all ϕ∈C0∞(B1). Then u 0 in B by Lemma 2.2. 1 ≥ 1 R R Let u = u u so that ∆2u =0 in B . Define f =∆u . Then ∆f =0 in B 2 1 2 1 2 1 − andsince f is radialwe find that f is constant. It follows that u =ar2+b. Using 2 the boundary conditions we deduce a+b 0 and a 0, which imply u 0. (cid:3) 2 ≥ ≤ ≥ Similarly we have Lemma 2.4. If u H2(B ) and ∆2u 0 in B in the weak sense, that is R R ∈ ≥ ∆u∆ϕ 0 ϕ C∞(B ), ϕ 0 ≥ ∀ ∈ 0 R ≥ ZBR and u =0, ∂u 0 then u 0 in B . |∂BR ∂n|∂BR ≤ ≥ R ThenextlemmaisaconsequenceofadecompositionlemmaofMoreau[22]. For a proof see [14, 15]. Lemma 2.5. Let u H2(B ). Then there exist unique w,v H2(B ) such that ∈ 0 R ∈ 0 R u=w+v, w 0, ∆2v 0 in B and ∆w∆v =0. ≥ ≤ R BR We need the following comparison prRinciple. Lemma 2.6. Let u , u H2(B ) with eu1, eu2 L1(B ). Assume that 1 2 R R ∈ ∈ ∆2u λeu1 in B 1 R ≤ in the sense (8) ∆u ∆ϕ λ eu1ϕ ϕ C∞(B ), ϕ 0, 1 ≤ ∀ ∈ 0 R ≥ ZBR ZBR and ∆2u λeu2 in B in the similar weak sense. Suppose also 2 R ≥ ∂u ∂u 1 2 u =u and = . 1|∂BR 2|∂BR ∂n |∂BR ∂n |∂BR Assume furthermore that u is stable in the sense that 1 (9) λ eu1ϕ2 (∆ϕ)2, ϕ C∞(B ). ≤ ∀ ∈ 0 R ZBR ZBR Then u u in B . 1 2 R ≤ 6 JUANDA´VILA,LOUISDUPAIGNE,IGNACIOGUERRA,ANDMARCELOMONTENEGRO Proof.Let u = u u . By Lemma 2.5 there exist w,v H2(B ) such that 1 − 2 ∈ 0 R u=w+v, w 0 and ∆2v 0. Observe that v 0 so w u u . 1 2 ≥ ≤ ≤ ≥ − By hypothesis we have for all ϕ C∞(B ), ϕ 0, ∈ 0 R ≥ ∆(u u )∆ϕ λ (eu1 eu2)ϕ λ (eu1 eu2)ϕ 1 2 − ≤ − ≤ − ZBR ZBR ZBR∩[u1≥u2] and by density this holds also for w: (10) (∆w)2 = ∆(u u )∆w 1 2 − ZBR ZBR λ (eu1 eu2)w =λ (eu1 eu2)w, ≤ − − ZBR∩[u1≥u2] ZBR where the first equality holds because ∆w∆v =0. By density we deduce from BR (9): R (11) λ eu1w2 (∆w)2. ≤ ZBR ZBR Combining (10) and (11) we obtain eu1w2 (eu1 eu2)w. ≤ − ZBR ZBR Since u u w the previous inequality implies 1 2 − ≤ (12) 0 (eu1 eu2 eu1(u u ))w. 1 2 ≤ − − − ZBR But by convexity of the exponential function eu1 eu2 eu1(u u ) 0 and we 1 2 − − − ≤ deduce from (12)that (eu1 eu2 eu1(u u ))w =0. Recalling that u u w 1 2 1 2 we deduce that u u . − − − − ≤(cid:3) 1 2 ≤ Remark 2.7. The following variant of Lemma 2.6 also holds: Let u , u H2(B ) be radial with eu1, eu2 L1(B ). Assume ∆2u λeu1 1 2 R R 1 ∈ ∈ ≤ in B in the sense of (8) and ∆2u λeu2 in B . Suppose u u and R 2 ≥ R 1|∂BR ≤ 2|∂BR ∂u1 ∂u2 and that thestability condition (9) holds. Then u u in B . ∂n|∂BR ≥ ∂n|∂BR 1 ≤ 2 R Proof. We solve for u˜ H2(B ) such that ∈ 0 R ∆u˜∆ϕ= ∆(u u )∆ϕ ϕ C∞(B ). 1− 2 ∀ ∈ 0 R ZBR ZBR By Lemma 2.3 it follows that u˜ u u . Next we apply the decomposition of 1 2 ≥ − Lemma2.5to u˜,thatis u˜=w+v withw,v H2(B ), w 0,∆2v 0inB and ∆w∆v =0. Then the argument follows∈that0of LRemm≥a 2.6. ≤ R (cid:3) BR R Finally, in several places we will need the method of sub and supersolutions in the context of weak solutions. Lemma 2.8. Let λ > 0 and assume that there exists u¯ H2(B ) such that R ∈ eu¯ L1(B ), R ∈ ∆u¯∆ϕ λ eu¯ϕ for all ϕ C∞(B ),ϕ 0 ≥ ∈ 0 R ≥ ZBR ZBR STABLE SOLUTIONS FOR ∆2u=λeu 7 and ∂u¯ u¯=a, b on ∂B . 1 ∂n ≤ Then there exists a weak solution to (1) such that u u¯. ≤ The proof is similar to that of Lemma 19 in [2]. 3. Uniqueness of the extremal solution: proof of Theorem 1.2 Proof of Theorem 1.2. Suppose that v H2(B) satisfies (4), (5) and v u∗. ∈ 6≡ Notice that we do not need v to be radial. The idea of the proof is as follows : Step 1. The function 1 u = (u∗+v) 0 2 is a super-solution to the following problem ∆2u=λ∗eu+µηeu in B u=a on ∂B (13)   ∂u =b on ∂B ∂n forsomeµ=µ0 >0,whereη ∈C0∞(B),0≤η ≤1isafixedradialcut-offfunction such that η(x)=1 for x 1, η(x)=0 for x 3. | |≤ 2 | |≥ 4 Step 2. Using a solution to (13) we construct, for some λ > λ∗, a super-solution to (1). This provides a solution u for some λ>λ∗, which is a contradiction. λ Proof of Step 1. Observe that given 0 < R < 1 we must have for some c = 0 c (R)>0 0 (14) v(x) u∗(x)+c x R. 0 ≥ | |≤ To prove this we recallthe Green’s function for ∆2 with Dirichlet boundary condi- tions ∆2G(x,y)=δ x B x y ∈ G(x,y)=0 x ∂B  ∈  ∂G(x,y)=0 x ∂B, ∂n ∈ where δy is the Dirac massat y B. Boggio gave an explicit formula for G(x,y) ∈ which was used in [16] to prove that in dimension N 5 (the case 1 N 4 can ≥ ≤ ≤ be treated similarly) d(x)2d(y)2 (15) G(x,y) x y 4−Nmin 1, ∼| − | x y 4 (cid:18) | − | (cid:19) where d(x)=dist(x,∂B)=1 x. −| | 8 JUANDA´VILA,LOUISDUPAIGNE,IGNACIOGUERRA,ANDMARCELOMONTENEGRO anda b meansthat forsomeconstantC >0we haveC−1a b Ca (uniformly ∼ ≤ ≤ for x,y B). Formula (15) yields ∈ (16) G(x,y) cd(x)2d(y)2 ≥ forsomec>0andthis inturnimpliesthatforsmoothfunctionsv˜andu˜suchthat v˜ u˜ H2(B) and ∆2(v˜ u˜) 0, − ∈ 0 − ≥ ∂∆ G ∂(v˜ u˜) x v˜(y) u˜(y)= (x,y)(v˜ u˜) ∆ G(x,y) − dx x − ∂n − − ∂n Z∂B(cid:16) x (cid:17) + G(x,y)∆2(v˜ u˜)dx − ZB cd(y)2 (∆2v˜ ∆2u˜)d(x)2dx. ≥ − ZB Using a standard approximation procedure, we conclude that v(y) u∗(y) cd(y)2λ∗ (ev eu∗)d(x)2dx. − ≥ − ZB Since v u∗, v u∗ we deduce (14). ≥ 6≡ Let u =(u∗+v)/2. Then by Taylor’s theorem 0 1 1 1 (17) ev =eu0 +(v u )eu0 + (v u )2eu0 + (v u )3eu0 + (v u )4eξ2 0 0 0 0 − 2 − 6 − 24 − for some u ξ v and 0 2 ≤ ≤ (18) eu∗ =eu0 +(u∗ u )eu0 + 1(u∗ u )2eu0 + 1(u∗ u )3eu0 + 1 (u∗ u )4eξ1 0 0 0 0 − 2 − 6 − 24 − for some u∗ ξ u . Adding (17) and (18) yields 1 0 ≤ ≤ (19) 1(ev+eu∗) eu0 + 1(v u∗)2eu0. 2 ≥ 8 − From (14) with R = 3/4 and (19) we see that u = (u∗+v)/2 is a super-solution 0 of (13) with µ :=c /8. 0 0 Proof of Step 2. Let us show now how to obtaina weaksuper-solutionof (1) for some λ>λ∗. Given µ>0, let u denote the minimal solutionto (13). Define ϕ as 1 the solution to ∆2ϕ =µηeu in B 1 ϕ =0 on ∂B  1  ∂ϕ1 =0 on ∂B, ∂n and ϕ2 be the solution of  ∆2ϕ =0 in B 2 ϕ =a on ∂B  2  ∂ϕ2 =b on ∂B. ∂n If N 5 (the case 1 N 4can be treated similarly), relation (16) yields ≥ ≤ ≤ (20) ϕ (x) c d(x)2 for all x B, 1 1 ≥ ∈ STABLE SOLUTIONS FOR ∆2u=λeu 9 for some c > 0. But u is a radial solution of (13) and therefore it is smooth in 1 B B . Thus 1/4 \ (21) u(x) Mϕ +ϕ for all x B , 1 2 1/2 ≤ ∈ for some M >0. Therefore, from(20) and (21), for λ>λ∗ with λ λ∗ sufficiently − small we have ( λ 1)u ϕ +( λ 1)ϕ in B. λ∗ − ≤ 1 λ∗ − 2 Let w = λ u ϕ ( λ 1)ϕ . The inequality just stated guaranteesthat w u. λ∗ − 1− λ∗ − 2 ≤ Moreover λµ ∆2w =λeu+ ηeu µηeu λeu λew in B λ∗ − ≥ ≥ and ∂w w =a =b on ∂B. ∂n Therefore w is a super-solution to (1) for λ. By the method of sub and super- solutions a solution to (1) exists for some λ>λ∗, which is a contradiction. (cid:3) Proof of Proposition 1.7. Let u H2(B), λ>0 be a weak unbounded solution ∈ of (1). If λ < λ∗ from Lemma 2.6 we find that u u where u is the minimal λ λ ≤ solution. Thisisimpossiblebecauseu issmoothanduunbounded. Ifλ=λ∗ then λ necessarily u=u∗ by Theorem 1.2. (cid:3) ProofofProposition1.8. Letudenotetheextremalsolutionof (1)withb 4. ≥− If u is smooth, then the result is trivial. So we restrict to the case where u is singular. By Theorem1.3 we have inparticular thatN 13. We may alsoassume ≥ that a=0. If b= 4 by Theorem 1.2 we know that if N 13 then u= 4log x − ≥ − | | so that the desiredconclusion holds. Henceforth we assume b> 4 in this section. − For ρ>0 define u (r)=u(ρr)+4logρ, ρ so that ∆2u =λ∗euρ in B . ρ 1/ρ Then du ρ =u′(1)+4>0. dρ ρ=1,r=1 (cid:12) Hence, there is δ >0 such that(cid:12) (cid:12) u (r)<u(r) for all 1 δ <r 1, 1 δ <ρ 1. ρ − ≤ − ≤ This implies (22) u (r)<u(r) for all 0<r 1, 1 δ <ρ 1. ρ ≤ − ≤ Otherwise set r =sup 0<r <1 u (r) u(r) . 0 ρ { | ≥ } This definition yields (23) u (r )=u(r ) and u′(r ) u′(r ). ρ 0 0 ρ 0 ≤ 0 10 JUANDA´VILA,LOUISDUPAIGNE,IGNACIOGUERRA,ANDMARCELOMONTENEGRO Write α=u(r ), β =u′(r ). Then u satisfies 0 0 ∆2u=λeu on B r0 (24) u(r )=α  0 u′(r )=β. 0 Observe that u is an unbounded H2(Br0) solution to (24), which is also stable. ThusProposition1.7showsthatuistheextremalsolutiontothisproblem. Onthe other hand u is a supersolution to (24), since u′(r ) β by (23). We may now ρ ρ 0 ≤ use Theorem 1.2 and we deduce that u(r)=u (r) for all 0<r r , ρ 0 ≤ which in turn implies by standard ODE theory that u(r)=u (r) for all 0<r 1, ρ ≤ a contradiction with (22). This proves estimate (22). From (22) we see that du ρ (25) (r) 0 for all 0<r 1. dρ ρ=1 ≥ ≤ (cid:12) But (cid:12) (cid:12) du ρ (r)=u′(r)r+4 for all 0<r 1 dρ ρ=1 ≤ (cid:12) and this together with ((cid:12)25) implies (cid:12) du 1 1 (26) ρ(r)= (u′(ρr)ρr+4) 0 for all 0<r , 0<ρ 1. dρ ρ ≥ ≤ ρ ≤ whichmeans that u (r) is non-decreasingin ρ. We wish to show that lim u (r) ρ ρ→0 ρ exists for all 0<r 1. For this we shall show ≤ 8(N 2)(N 4) 1 (27) u (r) 4log(r)+log − − for all 0<r , 0<ρ 1. ρ ≥− λ∗ ≤ ρ ≤ (cid:18) (cid:19) Set 8(N 2)(N 4) u (r)= 4log(r)+log − − . 0 − λ∗ (cid:18) (cid:19) and suppose that (27) is not true for some 0<ρ<1. Let r =sup 0<r <1/ρ u (r)<u (r) . 1 ρ 0 { | } Observe that (28) λ∗ >8(N 2)(N 4). − − Otherwise w = 4lnr would be a strict supersolution of the equation satisfied by − u, which is not possible by Theorem 1.2. In particular, r <1/ρ and 1 u (r )=u (r ) and u′(r ) u′(r ). ρ 1 0 1 ρ 1 ≥ 0 1 It follows that u is a supersolution of 0 ∆2u=λ∗eu in B r1 u=A on ∂B (29)  r1  ∂u =B on ∂B , ∂n r1 