STABILITY THEORY FOR NEUMANN EIGENFUNCTIONS 1. A Sobolev inequality Lemma 1.1. Let ABCD be a parallelogram, and let u be a C 2 function on ABCD. Let α be the angle subtended by A. Then Z 1 |u(A) − u(B) + u(C) − u(D)| 6 |∇2 u|. sin(α) ABCD Proof. We may normalise A = 0, so that C = B + D. From two applications of the fundamental theorem of calculus one has Z 1Z 1 ∂st u(sB + tD) = u(A) − u(B) + u(C) − u(D). 0 0 The left-hand side can be rewritten as Z 1 (B · ∇)(D · ∇)u. |ABCD| ABCD Since |ABCD| = |B||D| sin(α) the claim follows. Lemma 1.2. Let ABC be a triangle with angles α, β, γ, and let u be a C 2 function on ABC that obeys the Neumann boundary condition n · ∇u = 0 on the boundary of ABC. Then for any P ∈ ABC, one has Z 4 |∇2 u|. |u(P ) − u(Q)| 6 3 min(sin(α), sin(β), sin(γ)) ABC Proof. Let P ∈ ABC. Let D, E ∈ AB, F, G ∈ BC, H, I ∈ AC be the points such that ADXI, BF XE, CHXG are parallelograms (thus D, G are the intersections with AB, BC respectively of the line through X parallel to AC, and so forth. Then from the preceding lemma one has Z 1 |u(X) + u(A) − u(D) − u(I)| 6 |∇2 u| sin(α) ADXI Z 1 |u(X) + u(B) − u(F ) − u(E)| 6 |∇2 u| sin(β) BF XE Z 1 |u(X) + u(C) − u(H) − u(G)| 6 |∇2 u|. sin(γ) CHXG 1 2 STABILITY THEORY FOR NEUMANN EIGENFUNCTIONS Also, by reflecting the triangles DEX, F GX, XHI across the Neumann boundary and using the previous lemma, we see that Z 2 |u(X) + u(X) − u(D) − u(E)| 6 |∇2 u| sin(γ) DEX Z 2 |u(X) + u(X) − u(F ) − u(G)| 6 |∇2 u| sin(α) F GX Z 2 |∇2 u|. |u(X) + u(X) − u(H) − u(I)| 6 sin(β) XHI Summing the latter three combinations of u and subtracting the former three using the triangle inequality, we conclude that Z 2 |3u(X) − u(A) − u(B) − u(C)| 6 |∇2 u|. min(sin(α), sin(β), sin(γ)) ABC A similar inequality holds with X replaced by Y . Subtracting the two inequailties, we obtain the claim. Corollary 1.3. Let ABC be a triangle with angles α, β, γ, and let u be a C 2 function on ABC which is smooth up to the boundary except possibly at the vertices A, B, C, and which obeys the Neumann boundary condition n · ∇u = 0 on the boundary of ABC, and has mean zero on ABC. Then 4 kukL∞ (ABC) 6 |ABC|1/2 k∆ukL2 (ABC) . 3 min(sin(α), sin(β), sin(γ)) Proof. If u has mean zero, then kukL∞ (ABC) is bounded by |u(P ) − u(Q)| for some P, Q ∈ ABC. From the previous lemma we thus have kukL∞ (ABC) 6 4 |ABC|1/2 k∇2 ukL2 (ABC) . 3 min(sin(α), sin(β), sin(γ)) It will thus suffice to show the Bochner-Weitzenbock identity Z Z 2 2 |∇ u| = |∆u|2 . ABC ABC But this can be accomplished by two integration by parts, using the smoothness and Neumann boundary hypotheses on u (and a regularisation argument if necessary to cut away from the vertices) more details needed here. 2. Schwarz-Christoffel Let 0 < α, β, γ < π be angles adding up to π, then we can define a Schwarz-Christoffel map Φα,β : H → ABC from the half-plane H := {z : =(z) > 0} to a triangle ABC with angles α, β, γ by the formula Z z dζ Φα,β (z) := , 1−α/π (1 − ζ)1−β/π 0 ζ STABILITY THEORY FOR NEUMANN EIGENFUNCTIONS 3 where the integral is over any contour from 0 to z in H, and one chooses the branch cut to make both factors in the denominator positive real on the interval [0, 1]. Thus the vertices of the triangle are given by A := Φα,β (0) = 0 Z B := Φα,β (1) = 1 dt Γ(α/π)Γ(β/π) = 1−α/π (1 − t)1−β/π Γ((α + β)/π) 0 t Z −∞ dt C := Φα,β (∞) = −eiα 1−α/π |t| (1 − t)1−β/π Z ∞0 ds = eiα 1−α/π (s − 1) s1−β/π 1 Z 1 dv = eiα 1−α/π v 1−γ/π 0 (v − 1) Γ(α/π)Γ(γ/π) = eiα Γ((α + γ)/π) where we have used the beta function identity Z 1 Γ(x)Γ(y) dt = 1−x 1−y (1 − t) Γ(x + y) 0 t and the changes of variable s = 1 − t, v = 1/s. In particular, the area of the triangle ABC can be expressed as 1 Γ(α/π)2 Γ(β/π)Γ(γ/π) |ABC| = |B||C| sin(α) = sin(α) 2 2Γ((α + β)/π)Γ((α + γ)/π) which can be simplified using the formula Γ(z)Γ(1 − z) = expression |ABC| = π sin(πz) as the more symmetric 1 Γ(α/π)2 Γ(β/π)2 Γ(γ/π)2 sin(α) sin(β) sin(γ). 2π 2 (2.1) We write |Φ0α,β (z)| = eω(z) where ω = ωα,β is the harmonic function ω(z) := ( If u : ABC → H defined by α β − 1) log |z| + ( − 1) log |1 − z|. π π (2.2) R is a smooth function, and ũ : H → R is its pullback to the half-plane ũ := u ◦ Φα,β then we have Z Z u= ABC H e2ω ũ. 4 STABILITY THEORY FOR NEUMANN EIGENFUNCTIONS In a similar vein we have the conformal invariance of the two-dimensional Dirichlet energy Z Z 2 |∇u| = |∇ũ|2 ABC H and the conformal transformation of the Laplacian: f ∆ũ(z) = e2ω ∆u. In particular, the Rayleigh quotient Z Z 2 |u|2 |∇u| / ABC with mean zero condition quotient R ABC ABC u = 0 becomes, when pulled back to Z 2 Z |∇ũ| / H with mean zero condition H e ũ = 0. R H H, the Rayleigh e2ω |ũ|2 2ω Let u2 , u3 , . . . be an L2 -normalised eigenbasis for the Neumann Laplacian −∆ on ABC with eigenvalues λ2 6 λ3 6 . . ., thus −∆uk = λk uk on ABC with Neumann boundary data n · ∇uk = 0 and orthonormality Z uj uk = δjk ABC and mean zero condition Z uj = 0. ABC One can show that when ABC is acute-angled, these eigenfunctions are smooth except possibly at the vertices A, B, C, and are uniformly C 2 . add details here Pulling all this back to H, we obtain transformed eigenfunctions ũ2 , ũ3 , . . . on conformal eigenfunction equation −∆ũk = λk e2ω ũk on H to the (2.3) H with Neumann boundary data n · ∇ũk = 0 (2.4) e2ω ũj ũk = δjk (2.5) e2ω ũj = 0. (2.6) and orthonormality Z and mean zero condition H Z H STABILITY THEORY FOR NEUMANN EIGENFUNCTIONS 5 Now suppose that we vary the angle parameters α, β, γ smoothly with respect to some time parameter t, thus also varying the triangles ABC, eigenfunctions uk and transformed eigenfunctions ũk , eigenvalues λk , and conformal factor ω. We will use dots to indicate time differentiation, thus for instance α̇ = dtd α. Let us formally suppose that all of the above data vary smoothly (or at least C 1 ) in time we will eventually need to justify this, of course. Since α + β + γ = π, we have α̇ + β̇ + γ̇ = 0. The variation ω̇ of the conformal factor is explicitly computable from (2.2) as being a logarithmic weight: β̇ α̇ ω̇ = log |z| + log |1 − z|. π π Next, by (formally) differentiating (2.3) we obtain an equation for the variation ˜˙uk of the k th eigenfunction: −∆ũ˙ k = λ̇k e2ω ũk + 2λk ω̇e2ω ũk + λk e2ω ũ˙ k . (2.7) To solve this equation for ũ˙ k , we observe from differentiating (2.4), (2.5), (2.6) that Z e2ω ũ˙ k = 0 (2.8) H and Z e2ω ũk ũ˙ k = 0 H and n · ∇ũ˙ k = 0. By eigenfunction expansion, we thus have XZ ˙ũk = ( e2ω ũl ũ˙ k )ũl H l6=k (2.9) in a suitable sense (L2 with weight eω ). Now we evaluate the expression in parentheses. Integrating (2.7) against ũl and using (2.5) reveals that Z Z Z 2ω ˙ − ∆ũk ũl = 2λk ω̇e ũk ũl + λk e2ω ũ˙ k ũl . (2.10) H H H By Green’s theorem and the Neumann conditions on ũ˙ k and ũl , the left-hand side is Z − ũ˙ k ∆ũl H which by (2.3) is equal to Z λl e2ω ũ˙ k ũl . H Inserting this into (2.10) we see that Z Z 2λk 2ω ˙ e ũk ũl = ω̇e2ω ũk ũl λ − λ l k H H 6 STABILITY THEORY FOR NEUMANN EIGENFUNCTIONS and thus by (2.9) Z 2λ k ũ˙ k = ( ω̇e2ω ũk ũl )ũl . λ − λ l k H l6=k X (2.11) We can take Laplacians and conclude that X 2λk λl Z 2ω ˙ ω̇e2ω ũk ũl )ũl . −∆ũk = e ( λ − λ l k H l6=k 2 λ3 k λl is bounded in magnitude by λ2λ3 −λ . From the orthonormality (2.5) Set k = 2, then λ2λl −λ 2 k and the Bessel inequality, we conclude that Z Z 2λ2 λ3 −2ω 2 1/2 ˙ ( e |∆ũ2 | ) 6 ( |ω̇|2 e2ω ũ22 )1/2 . λ3 − λ2 H H If we change coordinates by writing ũ˙ = u̇ ◦ Φ we conclude that Z Z 2λ2 λ3 ( |∆u̇2 | ) 6 ( |ω̇|2 e2ω ũ22 )1/2 . λ3 − λ2 H ABC Also, u̇2 has mean zero on ABC by (2.8). We conclude from Corollary 1.3 that Z 4 1/2 2λ2 λ3 ku̇2 kL∞ 6 |ABC| ( |ω̇|2 e2ω ũ22 )1/2 . 3 min(sin(α), sin(β), sin(γ)) λ3 − λ2 H Pulling back to H, and estimating ũ2 in L∞ norm, we conclude that kũ˙ 2 kL∞ (H) 6 Xkũ2 kL∞ (H) 2 1/2 where X is the explicit (but somewhat messy) quantity Z 4 1/2 2λ2 λ3 |ABC| ( |ω̇|2 e2ω )1/2 . X := 3 min(sin(α), sin(β), sin(γ)) λ3 − λ2 H This gives stability of the second eigenfunction in L∞ norm, as long as there is an eigenvalue gap λ3 − λ2 > 0. One can compute one factor in the quantity X as follows. Observe that Z 1 e2ω = |ABC| = 2 Γ(α/π)2 Γ(β/π)2 Γ(γ/π)2 sin(α) sin(β) sin(γ). 2π H If we view α, β, γ (and hence ω) as varying linearly in time, and differentiate the above equation under the integral sign twice in time, we conclude that Z Z d2 2 2ω e2ω 4 |ω̇| e = 2 dt H H 2 ∂ ∂2 ∂2 = (α̇2 2 + 2α̇β̇ + 2α̇γ̇ ∂α ∂α∂β ∂α∂γ 2 2 ∂ ∂ + β̇ 2 2 + 2β̇ γ̇ + γ̇ 2 ) ∂β ∂βγ 1 ( 2 Γ(α/π)2 Γ(β/π)2 Γ(γ/π)2 sin(α) sin(β) sin(γ)). 2π STABILITY THEORY FOR NEUMANN EIGENFUNCTIONS 7 R This, in principle, expresses the factor ( H |ω̇|2 e2ω )1/2 in X as an explicit combination of trigonometric functions, gamma functions, and the first two derivatives of the gamma function. However, this formula is somewhat messy, to say the least.