A D V E R T |
I S E M E N T |
| Math Sites: lean ages 13-∞ readme referees parents | >>> MAITH GAMES <<< | all 372 compute stand |
|
LEVEL 1 OF 1 · The critical dimension for the one-phase Bernoulli problem
The critical dimension for one-phase Bernoulli minimizers
expertly designed by an internal OpenAI model · released 2026-09-24
· original PDF
IntroductionFor a nonnegative function \(v\in H^1(B)\), where \(B\subset\mathbb R^d\) is a ball, the one-phase Bernoulli energy is \[ J(v;B)=\int_B\bigl(|\nabla v|^2+\mathbf 1_{\{v>0\}}\bigr)\,dx. \tag{1}\] A nonnegative \(u\in H^1_{\mathrm{loc}}(\mathbb R^d)\) is a global minimizer if \(J(u;B)\leq J(v;B)\) for every ball \(B\) and every nonnegative \(v\in H^1(B)\) with \(v-u\in H^1_0(B)\). The competitors in these comparisons are arbitrary Sobolev functions with the prescribed trace. A function is one-homogeneous if \(u(rx)=r u(x)\) for every \(r>0\), almost everywhere in \(x\). The nonzero flat solutions are \[u(x)=(x\cdot e)_+,\qquad |e|=1.\] Let \(d_*\) be the least dimension containing a nonzero, nonflat, one-homogeneous global minimizer. Theorem 1. Every nonzero one-homogeneous global minimizer of (1) in \(\mathbb R^d\), \(1\leq d\leq6\), is flat. There is a nonflat one-homogeneous global minimizer in \(\mathbb R^7\). Consequently, \[d_*=7.\] A nonnegative \(u\in H^1_{\mathrm{loc}}(D)\) is a local minimizer in an open set \(D\subset\mathbb R^n\) if it satisfies the same comparison on every ball compactly contained in \(D\). At an interior free-boundary point \(x_0\), a blowup is a locally uniform limit of \(u(x_0+rx)/r\) along radii tending to zero. A point is regular if one such limit is a flat solution, and singular otherwise. Write \(\operatorname{Sing}(u)\) for the set of singular points in \(D\). Corollary 2 (Sharp singular-set bound). Let \(n\ge2\), let \(D\subset\mathbb R^n\) be open, and let \(u\in H^1_{\mathrm{loc}}(D)\) be a nonnegative local minimizer of the one-phase Bernoulli energy in \(D\). Its interior free boundary is smooth outside \(\operatorname{Sing}(u)\). This set is empty when \(n\le6\), is locally finite when \(n=7\), and satisfies \[\dim_{\mathrm H}\operatorname{Sing}(u)\le n-7\qquad(n\ge7).\] For every \(n\ge7\), a global minimizer attains this dimension bound. The upper bound is the standard dimension-reduction consequence of the critical dimension, and sharpness follows by taking Cartesian products of the De Silva–Jerison cone with Euclidean space. We verify these deductions in Section 8. The variational theory of the one-phase Bernoulli problem begins with Alt and Caffarelli’s existence and regularity theory [1]. Weiss’s monotonicity formula and dimension reduction [18] connect the singularities of general minimizers to homogeneous global minimizers; see also [17]. Thus the first dimension admitting a nonflat minimizing cone determines both the threshold for singular free boundaries and the codimension bound furnished by this theory. Our main result answers positively the critical-dimension question whether \(d_*=7\). A second strand of the theory passes from weak free boundaries to classical ones. Caffarelli’s Harnack-inequality approach established regularity for Lipschitz free boundaries [2]. De Silva developed a different improvement-of-flatness argument allowing a nonzero right-hand side [4]; this is the approach presented in the regularity chapters of [17]. Higher regularity has classical antecedents in Kinderlehrer and Nirenberg [14]. The version used here, for Lipschitz viscosity solutions, is due to De Silva, Ferrari and Salsa [6]. These results make a classification of homogeneous minimizing blowups applicable to general local minimizers. As recalled by Jerison and Savin [13], the dimension-three rigidity result for stable homogeneous solutions with smooth cross-section is due to Caffarelli, Jerison and Kenig [3]. Jerison and Savin proved that stable homogeneous solutions with smooth cross-section are flat in dimensions at most four [13]; their regularity consequence for minimizers [13] gives \(d_*\geq5\). Their argument constructs functions of the Hessian satisfying inequalities for the linearized free-boundary problem. In the other direction, De Silva and Jerison constructed a nonflat homogeneous global minimizer in dimension seven [7], proving \(d_*\leq7\). The minimizing property in that construction is with respect to all Sobolev competitors with the prescribed boundary trace. Firester, Tsiamis and Wang also construct further symmetric minimizing one-phase cones, including a second example in dimension seven [11]. Recent work further describes cones with symmetry. Firester, Tsiamis and Wang [10] prove that their cohomogeneity-one cones with \(O(n-k)\times O(k)\) symmetry, \(1\leq k\leq n-2\), are unstable for \(n\leq6\) and strictly stable for \(n\geq7\). The exclusion proved here applies to globally minimizing cones without a symmetry assumption. Cone classification also differs from rigidity of arbitrary entire stable solutions. Fernández-Real and Serra prove that every global classical stable solution in dimension four is one-dimensional [8]. Their result removes homogeneity in that dimension. For an admissible increasing family of nonnegative boundary data, Fernández-Real and Yu prove regularity through dimension six for almost every parameter [9]. This is a generic-data result: the increase is required at a fixed linear rate on the positive set of the smaller boundary datum. The obstacle and the proof strategy.The stability inequality compares the Dirichlet energy of a test function with a curvature term on the free boundary. The use of the harmonic Hessian has a geometric precedent in Simons’s second-fundamental-form argument for minimal cones [16], a connection developed explicitly by Jerison and Savin. A natural Bernoulli test is a power of the Hessian norm. Jerison and Savin explain why this unweighted choice does not give a uniform instability criterion in dimensions four through six: the interior and boundary estimates impose different restrictions on the power, and their sharp configurations need not agree [13]. Their four-dimensional argument uses a more refined function of the Hessian eigenvalues [13]. Our proof uses the normalized shape of the Hessian to construct both a scalar test and a separate vector field whose divergence controls the test’s energy. The two constructions are chosen together, so the interior and boundary estimates retain the same tensor data. The proof has three stages. First, suppose the first nonflat minimizing cone occurs in dimension five or six. Blowups away from its vertex reduce the dimension and are therefore flat. Regularity theory makes each component of its spherical positive set smooth, and at least one component has nonzero Hessian. A five-dimensional cone can be made independent of one extra coordinate without losing minimality. Its spherical link then has precisely two exceptional tips. Thus both possible dimensions lead to a domain \(\Omega\subset S^5\) with a smooth regular boundary and, possibly, these two tips (Section [sec:reductions]). Second, let \(g\) be the restriction of the resulting cone to the sphere, and set \(T=\nabla^2g+g\mathop{\mathrm{Id}}\) and \(w=|T|\). Where \(w>0\), we construct a positive scalar \(h\) from \(w\), the normalized tensor \(T/w\), \(g\), and \(\nabla g\), and a vector field \(Z\) from the same data and \(\nabla T/w\). The target inequalities are \[\mathop{\mathrm{div}}Z\ge |\nabla h|^2+4h^2+\epsilon_*w^{2a}, \qquad N\cdot Z+Hh^2\ge0,\] where \(a\in(1/2,1)\) is fixed, \(\epsilon_*>0\), \(N\) points into \(\Omega\), and \(H=-T(N,N)\). Spherical stability supplies the reverse integrated comparison between the test’s energy and its boundary curvature term. The extra positive mass \(\epsilon_*w^{2a}\) will give the contradiction. The geometry and stability inequality are developed in Section [sec:geometry]. We then state the universal tensor inequality and derive the two geometric comparisons; the full exact verification follows in Section 6. Two cancellations make this construction possible. A Laplacian contraction and an antisymmetric curvature contraction eliminate fourth derivatives from \(\mathop{\mathrm{div}}Z\). At the boundary, reflection in the normal direction removes the derivative components not determined by the boundary conditions. The remaining inequalities involve only finitely many normalized tensors. We prove them by weighted sums of tensor squares and exact rational remainder bounds, with all coefficients and contraction conventions supplied in Appendix 9. Finally, neither the normalized Hessian nor its derivatives are defined at \(w=0\), and suspension introduces singular tips. These sets cannot be discarded without an estimate. A regularized Bochner identity gives an integrable weight controlling the derivatives near every Hessian zero. The explicit suspension formulas give an integrable power of the distance parameter at both tips. Smooth product cutoffs therefore remove both obstructions and make all error terms vanish. Section [sec:cutoffs] completes the stability contradiction. Organization and conventions.Section [sec:reductions] reduces a hypothetical first nonflat minimizer to a smooth spherical component in dimension six, allowing two suspension tips. Section [sec:geometry] derives the tensor identities and the stability inequality. Section 4 states the universal algebraic estimate, and Section [sec:flux] derives the pointwise geometric comparison from it. Section 6 gives the complete exact verification; Section [sec:cutoffs] then removes the cutoffs to obtain the contradiction. Section 8 gives the regularity and sharpness consequences. Appendix 9 gives all coefficient data. All tensor norms and contractions use orthonormal frames. Repeated indices are summed, \(\Delta=\mathop{\mathrm{tr}}\nabla^2\), and a boundary normal denoted by \(N\) points into the positive phase. The curvature \(H=-T(N,N)\) used below is not divided by the dimension of the boundary. Reduction to a smooth link and its suspension
We first record the consequences of energy minimality needed to put the argument on a smooth spherical domain. Throughout Section [sec:reductions], a minimizer has the full ballwise Sobolev comparison property defined in Section 1. The regularity inputsThe ballwise convention also gives comparison on every bounded open set \(O\Subset\mathbb R^d\): extend an \(H^1_0(O)\) competitor difference by zero into a ball containing \(\overline O\), apply minimality there, and cancel the energy outside \(O\). Conversely, comparison on bounded open sets includes comparison on balls. It is even enough to have comparison against arbitrary admissible competitors in every origin-centered ball, by the same extension argument. There is no restriction to homogeneous competitors. The convention that competitors are nonnegative makes no difference to the standard one-phase results. If \(v-u\in H^1_0(O)\) and \(u\geq0\), then \(v_+-u\in H^1_0(O)\) and \[\int_O\bigl(|\nabla v_+|^2+\mathbf 1_{\{v_+>0\}}\bigr) \leq \int_O\bigl(|\nabla v|^2+\mathbf 1_{\{v>0\}}\bigr).\] For completeness, the assertion about \(H^1_0(O)\) follows by approximating \(v-u\) in \(H^1\) by compactly supported smooth functions and applying the continuity of positive-part truncation in \(H^1\); the truncated differences in the approximation have compact support in \(O\). Proposition 3 (Regularity and blowups of minimizers). Let \(d\geq2\) and let \(u\geq0\) be a local minimizer of the one-phase energy with volume coefficient \(1\) in an open set \(D\subset\mathbb R^d\).
Sources and applicability. We use the constant-coefficient theory with \(\Lambda=1\) in [17]. For each ball \(B\Subset D\), the restriction of \(u\) to \(B\) minimizes against every \(H^1_0(B)\) competitor difference. Zero extension of differences therefore gives comparison on every \(O\Subset B\), as required by Definitions 2.11–2.12 there. We apply the local results inside such balls. For global minimizers, the containing-ball argument above gives the comparison convention on all bounded open sets. Local Lipschitz continuity is Theorem 3.1. Harmonicity also follows directly: for a smooth test supported compactly in the positive phase, both signs of a sufficiently small additive variation preserve that phase, so the Dirichlet first variation vanishes. The Lipschitz bound and \(u(x_0)=0\) give locally uniform compactness of the rescalings. Proposition 4.1 gives \(\sup_{B_r(x_0)}u\geq\kappa r\) for small \(r\), with \(\kappa>0\), so a limit cannot vanish identically. Proposition 6.2 gives strong Sobolev and positivity-indicator convergence and the full minimizing property of the limit; Proposition 9.12 gives its one-homogeneity. All these statements can be applied in a bounded ball compactly contained in \(D\) before taking the rescalings. In particular, the assertion about minimality is the minimizing compactness result, including the volume term. Definition 6.10 and Corollary 8.2, with the one-sided graph conclusion in Theorem 8.1, give a \(C^{1,\alpha}\) regular neighborhood from a single half-space blowup. Proposition 7.1 supplies the viscosity Bernoulli condition. To justify all subsequent boundary derivatives, use [6] with \(f_+=f_-=0\) and \(G(b)=1+b\). We check the two-phase viscosity condition of [5] when \(u\geq0\). Let \(v\) be a comparison touching \(u\) at a free-boundary point \(x_0\), with a \(C^2\) interface and \(C^2\) branches having nonnegative normal slopes \(a\) and \(b\) on its positive and negative sides. If \(v\) touches from below, then \(v_+\leq u\) in a neighborhood. For \(a>0\), extend its positive branch to a signed \(C^2\) function \(\phi\), negative just across the interface; then \(\phi\leq u\) and touches at \(x_0\). The one-phase condition gives \(a\leq1\leq G(b)\); the case \(a=0\) is immediate. If \(v\) touches from above, the neighborhood inequality \(v\geq u\geq0\) forces \(v=u=0\) throughout its nonpositive side, so \(b=0\). A \(C^2\) extension \(\phi\) of its positive branch satisfies \(\phi_+\geq u\) on both sides and touches at \(x_0\). Thus the one-phase condition gives \(a\geq1=G(0)\). These uses of the one-phase condition are justified by [17]: to meet its smooth-test convention, replace each \(C^2\) extension by its quadratic Taylor polynomial minus, respectively plus, \(\varepsilon|x-x_0|^2\). For any \(\varepsilon>0\), these are smooth lower and upper supports in a sufficiently small neighborhood, with the same contact value and gradient. The interior equations hold because \(u\) is harmonic where positive and identically zero on the interior of its zero set. The function \(G\) is smooth, strictly increasing and unbounded, with \(G(0)=1\); the cited results allow the negative phase to vanish. The preceding \(C^{1,\alpha}\) one-sided graph satisfies the slab flatness hypothesis after rotation and sufficiently small dilation, while the local Lipschitz bound is preserved. Theorem 1.1 gives \(C^{2,\gamma}\) regularity, and Corollary 1.2 gives smoothness. Local harmonic boundary regularity with zero Dirichlet data [12] then gives smoothness of \(u\) from the positive side. These estimates apply after flattening a smooth boundary patch and localizing away from its artificial boundary; iteration and Sobolev embedding give every required derivative. The Bernoulli condition and nonnegativity give (3). ◻ As in Section 1, a free-boundary point is regular if it has a half-space blowup and singular otherwise. Proposition 3 explains the geometric meaning of regularity. Conversely, a smooth one-sided free boundary with (3) has a half-space blowup by its first-order Taylor expansion. We henceforth use the continuous representative of a minimizer. If it is one-homogeneous almost everywhere, it is one-homogeneous everywhere: for each fixed \(r>0\), both sides of \(u(rx)=ru(x)\) are continuous and agree almost everywhere. Dimension reduction at a nonzero pointThe next two lemmas give the dimension-reduction steps explicitly; compare [17]. Lemma 4 (Translation invariance of an off-origin blowup). Let \(u\) be a one-homogeneous global minimizer, let \(0\ne x_0\in\partial\{u>0\}\), and let \(U\) be any limit in (2). Then \[U(y+t x_0)=U(y)\qquad(y\in\mathbb R^d,\ t\in\mathbb R).\] Proof. Write \(u_r=u_{x_0,r}\). For fixed \(t\in\mathbb R\) and small \(r>0\), set \(r'=r/(1+rt)>0\). Homogeneity gives the exact identity \[ u_r(y+t x_0) =\frac{u\bigl((1+rt)(x_0+r'y)\bigr)}r =u_{r'}(y). \tag{4}\] Let \(L\) be a Lipschitz constant for \(u\) on a fixed neighborhood of \(x_0\). Since \(u(x_0)=0\), for \(|y|\leq M\) and small \(r\), \[\begin{align*} |u_{r'}(y)-u_r(y)| &\leq\frac{L|r-r'|M}{r'} +LrM\left|\frac1{r'}-\frac1r\right|\\ &=2LM\frac{|r-r'|}{r'}=2LM|t|r. \end{align*}\] Thus the \(r'\) rescalings along the same chosen sequence have the same locally uniform limit as the \(r\) rescalings. Passing to that limit in (4) proves the assertion for every fixed \(y,t\). No uniqueness of blowups is required. ◻ Lemma 5 (Minimality of a translation-invariant factor). Suppose \(U(y,s)=v(y)\) is a global minimizer in \(\mathbb R^k\times\mathbb R\), where \(v\geq0\) belongs to \(H^1_{\mathrm{loc}}(\mathbb R^k)\). Then \(v\) is a global minimizer in \(\mathbb R^k\). Proof. Suppose that a ball \(B\subset\mathbb R^k\) admits a nonnegative \(q\in H^1(B)\) with \(q-v\in H^1_0(B)\) and \[J(v;B)-J(q;B)=\delta>0.\] Let \(\psi\) be the zero extension of \(q-v\). For \(L>1\), choose \(\eta_L:\mathbb R\to[0,1]\) equal to \(1\) on \([-L,L]\), equal to \(0\) outside \([-L-1,L+1]\), and linear on each intervening interval. Define \[V_L(y,s)=U(y,s)+\eta_L(s)\psi(y).\] This function is nonnegative: on \(B\times\mathbb R\) it is the convex combination \((1-\eta_L)v+\eta_Lq\). Its difference from \(U\) belongs to \(H^1\) and has compact support in a sufficiently large ball \(\mathcal B_L\subset\mathbb R^{k+1}\), so it belongs to \(H^1_0(\mathcal B_L)\). The energy difference on \(B\times[-L,L]\) is \(-2L\delta\). On either transition cylinder, convexity gives \[|\nabla_y V_L|^2\leq(1-\eta_L)|\nabla v|^2+\eta_L|\nabla q|^2, \qquad |\partial_sV_L|^2\leq|q-v|^2.\] The difference of the positivity indicators is at most \(1\). Hence the sum of the two transition contributions is bounded above by \[C=2\int_B\bigl(|\nabla q|^2+|q-v|^2+1\bigr),\] independently of \(L\). Outside these three cylinders the energy is unchanged. Thus \[J(V_L;\mathcal B_L)-J(U;\mathcal B_L)\leq-2L\delta+C<0\] for large \(L\), contradicting minimality of \(U\). ◻ In dimension one, every nonnegative one-homogeneous Sobolev function has the form \(u(x)=b_+x_++b_-(-x)_+\) with \(b_+,b_-\ge0\). If both slopes are positive, a nonnegative smooth bump \(\phi\) supported in \((-R,R)\) with \(\phi(0)>0\) preserves positivity volume and gives \[J(u+\tau\phi;(-R,R))-J(u;(-R,R)) =-2\tau(b_++b_-)\phi(0)+\tau^2\int_{-R}^R|\phi'|^2<0\] for small \(\tau>0\). Thus a nonzero minimizing cone has just one positive slope. After reflection write \(u(x)=b x_+\), \(b>0\). The competitors \(v_t(x)=bR(x-t)_+/(R-t)\) on \((-R,R)\), \(|t|<R\), have the same trace and energy \(b^2R^2/(R-t)+(R-t)\). Its derivative at the minimizing parameter \(t=0\) is \(b^2-1=0\), so \(b=1\). This proves flatness in dimension one directly. The lower-dimensional classification of Jerison–Savin [13] and the dimension-seven minimizing example of De Silva–Jerison [7] give \(5\leq d_*\leq7\). De Silva–Jerison’s Theorem 1.1 uses the same functional and arbitrary ballwise Dirichlet competitors. For the exclusion argument, suppose from now on that \[ d=d_*\in\{5,6\}, \tag{5}\] and fix a nonzero nonflat one-homogeneous global minimizer \(u\) in \(\mathbb R^d\). Proposition 6 (Smooth original link). Every point of \(\partial\{u>0\}\setminus\{0\}\) is regular. Put \(m_o=d-1\) and \(g_o=u|_{S^{m_o}}\). The set \(\{g_o>0\}\) is nonempty and proper, has finitely many connected components, and their closures are pairwise disjoint compact smooth manifolds with boundary. On each such component \(\Omega_o\), the restriction of \(g_o\) is smooth up to the boundary from within \(\Omega_o\) and satisfies \[ \Delta_{S^{m_o}}g_o=-m_o g_o\quad\hbox{in }\Omega_o, \qquad g_o=0,\quad \nabla_{S^{m_o}}g_o=N_o \quad\hbox{on }\partial\Omega_o, \tag{6}\] where \(N_o\) is the inward spherical unit normal. In particular, \(\partial\Omega_o\ne\varnothing\). Proof. At a nonzero free-boundary point \(x_0\), choose a blowup \(U\) from Proposition 3. By Lemma 4, in orthogonal coordinates \(\mathbb R^d=x_0^\perp\times\mathbb R\) it has the form \(U(y,s)=v(y)\). The factor \(v\) is locally Lipschitz, nonnegative, nonzero and one-homogeneous. Lemma 5 shows that it is a global minimizer. Since its dimension is \(d-1<d_*\), it is flat. Therefore \(U\) itself is a half-space solution, and Proposition 3 makes \(x_0\) regular. The spherical positive set is nonempty by nontriviality. The polar-coordinate formula for the Euclidean Laplacian of \(u(ry)=r g_o(y)\) gives \[ \Delta_{\mathbb R^d}u(ry)=r^{-1} \bigl(\Delta_{S^{m_o}}g_o(y)+m_o g_o(y)\bigr) \tag{7}\] in the positive phase. If \(g_o>0\) on the entire sphere, integrating the harmonicity identity over the sphere would give \(m_o\int g_o=0\), a contradiction. Thus its positive set is proper. The conical free boundary has the radial direction tangent to it at every nonzero point: the curve \(r\mapsto rx\) lies in that free boundary. Consequently its smooth hypersurface is transverse to the unit sphere, and their intersection is a smooth hypersurface of \(S^{m_o}\). Its regular charts have the spherical positive phase exactly on one side. Each point of the sphere therefore has a neighborhood whose intersection with \(\{g_o>0\}\) is either connected or empty (using a small ball for points away from the free boundary). A finite subcover of the sphere shows that there are only finitely many positive components. If two component closures met, their common point would be a free-boundary point; the connected positive side of a regular chart there would meet both components, forcing them to coincide. The same charts show that each closure is a smooth compact domain. Each component has nonempty boundary because the sphere is connected and the positive set is proper. Smoothness of \(u\) from its positive side gives smoothness of \(g_o\) on these closures. Equation (7) gives the interior equation in (6). On the boundary, the radial derivative of \(u\) is \(g_o=0\), so the Euclidean gradient is tangent to the sphere. Equation (3) then gives the remaining boundary conditions. ◻ Lemma 7 (Selection of a nonflat component). There is a component \(\Omega_o\) in Proposition 6 such that \[ T_o:=\nabla^2_{S^{m_o}}g_o+g_o\mathop{\mathrm{Id}}\not\equiv0 \quad\hbox{on }\Omega_o. \tag{8}\] Here \(\mathop{\mathrm{Id}}\) denotes the spherical metric, viewed as a symmetric two-tensor. Proof. For \(y\) in the spherical positive set and \(r>0\), \[\nabla_{\mathbb R^d}u(ry)=g_o(y)y+\nabla_{S^{m_o}}g_o(y).\] Differentiating this formula, or using Euler’s identity for the radial entries, gives \[ D^2u(ry)[y,\cdot]=0, \qquad D^2u(ry)[v_1,v_2]=r^{-1}T_o(y)[v_1,v_2] \quad(v_1,v_2\perp y). \tag{9}\] Suppose that \(T_o\) vanished identically on every positive component. Then \(D^2u=0\) on each corresponding connected open cone \(D_i\), so \(u(x)=b_i\cdot x+c_i\) there. Homogeneity forces \(c_i=0\), and positivity forces \(b_i\ne0\). Thus \(D_i\subset H_i:=\{b_i\cdot x>0\}\). The set \(D_i\) is also closed relative to \(H_i\): if \(x\in\overline{D_i}\cap H_i\), continuity gives \(u(x)=b_i\cdot x>0\); a small positive neighborhood of \(x\) meets \(D_i\), hence belongs to the same component. Since \(H_i\) is connected, \(D_i=H_i\). The Bernoulli condition at any nonzero point of its boundary gives \(|b_i|=1\). Two disjoint open half-spaces through the origin must have opposite unit normals: if their normals are \(e,f\) with \(f\ne-e\), the vector \(e+f\) lies in both. But opposite half-spaces have spherical closures meeting along the equator \(e^\perp\cap S^{m_o}\), contrary to the pairwise disjoint closures in Proposition 6. Thus there is exactly one positive component, and \(u=(x\cdot e)_+\), contrary to nonflatness. This proves (8). ◻ Suspension and the working domainProposition 8 (Suspension preserves global minimality). If \(u\geq0\) is a global minimizer in \(\mathbb R^k\), then \(\widetilde u(x,s)=u(x)\) is a global minimizer in \(\mathbb R^{k+1}\). If \(u\) is the cone in (5) with \(d=5\), the singular free-boundary set of \(\widetilde u\) is exactly \(\{0\}\times\mathbb R\). For the component chosen in Lemma 7, the corresponding positive component on \(S^5\) is \[ \widetilde\Omega =\bigl\{\Phi(t,y):=(\sin t\,y,\cos t): 0<t<\pi,\ y\in\Omega_o\bigr\}, \qquad \widetilde g(\Phi(t,y))=\sin t\,g_o(y). \tag{10}\] Its boundary consists of the smooth side \(\Phi((0,\pi)\times\partial\Omega_o)\) and the two singular tips \(p_+=(0,1)\) and \(p_-=(0,-1)\). Proof. Fix a ball \(\mathcal B=B_R((x_c,s_c))\subset\mathbb R^{k+1}\) and a nonnegative \(V\in H^1(\mathcal B)\) with \(V-\widetilde u\in H^1_0(\mathcal B)\). Extend that difference by zero to \(F\in H^1(\mathbb R^{k+1})\). For \(|s-s_c|<R\), let \[B_s=B_{\sqrt{R^2-(s-s_c)^2}}(x_c)\] be the horizontal slice. For almost every such \(s\), \(F(\cdot,s)|_{B_s}\in H^1_0(B_s)\). To see the zero-boundary assertion without a trace regularity assumption on \(V\), approximate \(F\) in \(H^1(\mathbb R^{k+1})\) by the zero extensions of functions in \(C_c^\infty(\mathcal B)\). After passing to a subsequence, Fubini’s Theorem gives convergence of their slices to \(F(\cdot,s)\) in \(H^1(\mathbb R^k)\) for almost every \(s\). Each approximating slice lies in \(C_c^\infty(B_s)\), so the assertion follows by closure. For almost every \(s\), \(V_s=V(\cdot,s)\) is therefore an admissible nonnegative competitor for \(u\) in \(B_s\). Integrating its minimizing inequality and using the nonnegativity of the additional derivative energy, we obtain \[\begin{align*} J(\widetilde u;\mathcal B) &=\int_{s_c-R}^{s_c+R}J(u;B_s)\,ds\\ &\leq\int_{s_c-R}^{s_c+R}J(V_s;B_s)\,ds \leq J(V;\mathcal B). \end{align*}\] This proves the full ballwise minimizing property, for arbitrary centers and arbitrary admissible Sobolev competitors. One-homogeneity and nontriviality are plainly preserved. Flatness of \(\widetilde u\) would force its half-space normal to have zero last coordinate, and would therefore force \(u\) to be flat. In the conical case, \(u(0)=0\) and a positive ray approaches the origin, so \(0\in\partial\{u>0\}\). The product identity \[\partial\{\widetilde u>0\}=\partial\{u>0\}\times\mathbb R\] shows that all its off-axis points are regular: their positive phase is the product of a smooth one-sided regular chart for \(u\) and an interval. At any axis point \(z_0=(0,s_0)\), every rescaling is exactly \[\frac{\widetilde u(z_0+r(x,s))}{r} =\frac{u(rx)}r=\widetilde u(x,s).\] This is nonflat, so no axis point is regular. The axis meets \(S^5\) only in \(p_+\) and \(p_-\). Away from these two points, \(\Phi\) is a smooth product coordinate map from \((0,\pi)\times S^4\) to \(S^5\setminus\{p_+,p_-\}\). Positivity in these coordinates is precisely \(g_o(y)>0\), so connected components are as in (10). Since \(\overline{\Omega_o}\) is compact, its suspended closure is obtained by adding the two endpoints of the interval, with each endpoint collapsed to its tip. Its remaining boundary is exactly the stated smooth side. ◻ Notation for the remaining argument.Fix the original component \(\Omega_o\subset S^{m_o}\) selected in Lemma 7, where \(m_o\in\{4,5\}\). If \(m_o=5\), set \(\Omega=\Omega_o\), \(g=g_o\), and \(\Sigma=\varnothing\). If \(m_o=4\), set \(\Omega=\widetilde\Omega\), \(g=\widetilde g\), and \(\Sigma=\{p_+,p_-\}\). In either case \(\Omega\subset S^5\) is connected, and \(g\) comes from a nonnegative one-homogeneous global minimizer in \(\mathbb R^6\). The closure is smooth away from \(\Sigma\), with positive phase on one side of its regular boundary \[\partial_{\mathrm{reg}}\Omega:=\partial\Omega\setminus\Sigma.\] All spherical boundary values below are taken from the positive side. Tests on \(\overline\Omega\) will be smooth and supported away from \(\Sigma\). Every compact portion of the selected Euclidean cone’s closure that avoids the origin and, in the suspended case, the axis is separated from all other positive components. This follows from the disjoint compact original closures, and in the suspension from the same fact on any closed subinterval of \((0,\pi)\). Consequently the local variations on this component can be made as variations of the global minimizer. No minimizing property of the selected component extended by zero is needed. Spherical geometry and the stability inequality
Let \(\Omega_o\subset S^{m_o}\), \(m_o\in\{4,5\}\), be the original component supplied by Proposition 6 and Lemma 7. Its closure is a compact smooth manifold with nonempty boundary. The function \(g_o\) is smooth up to that boundary, is positive in \(\Omega_o\), and satisfies \[\Delta g_o=-m_o g_o,\qquad g_o=0,\quad \nabla g_o=N_o\quad\hbox{on }\partial\Omega_o,\] where \(N_o\) is the inward unit normal. Moreover, \(\nabla^2g_o+g_o\mathop{\mathrm{Id}}\) does not vanish identically. We give the identities in a general tangent dimension \(m\). The compact original component first provides strict gradient and boundary-sign bounds; we then transport these bounds to the working link and derive its stability inequality. All spherical derivatives below are covariant, and \(\Delta\) on tensors is the rough Laplacian \(\mathop{\mathrm{tr}}\nabla^2\). Repeated indices in an orthonormal frame are summed. The spherical Hessian and its boundary valuesFor a smooth solution of \(\Delta g=-mg\) on a domain in \(S^m\), set \[ p=\nabla g,\qquad T=\nabla^2g+g\mathop{\mathrm{Id}},\qquad B_{ijk}=\nabla_iT_{jk},\qquad w=|T|,\qquad s^2=g^2+|p|^2. \tag{11}\] The notation \(s^2\) denotes the displayed smooth function, including at points where it vanishes. If \(u(ry)=rg(y)\) is the homogeneous extension, then in the radial and tangential orthonormal frame \[ D^2u(ry)=r^{-1} \begin{pmatrix}0&0\\0&T(y)\end{pmatrix}. \tag{12}\] Indeed, Euler’s identity gives \(D^2u\,y=0\), and differentiation of \(\nabla_{\mathbb R^{m+1}}u=g y+p\) along a tangent vector gives the tangential block \(\nabla^2g+g\mathop{\mathrm{Id}}\) at radius one. Homogeneity then gives the factor \(r^{-1}\). The same gradient formula identifies \(s^2\) with \(|\nabla_{\mathbb R^{m+1}}u|^2\) on the unit sphere. Proposition 9 (Interior identities). The tensor \(T\) is symmetric and trace-free, and \(B\) is fully symmetric and trace-free in every pair of slots. In particular \(\mathop{\mathrm{div}}T=0\). The following identities hold: \[\begin{align*} \nabla_lB_{ijk}-\nabla_iB_{ljk} &=\delta_{lj}T_{ik}+\delta_{lk}T_{ij} -\delta_{ij}T_{lk}-\delta_{ik}T_{lj}, \tag{13}\\ \Delta T&=mT, &\Delta w^2&=2m w^2+2|B|^2, \tag{14}\\ \nabla_l p_i&=T_{li}-g\delta_{li}, &\nabla_l g&=p_l, \tag{15}\\ \nabla s^2&=2Tp, &\Delta s^2&=2w^2. \tag{16}\end{align*}\] On the open set \(\{w>0\}\) one also has \[ \nabla_l w=w^{-1}T_{jk}B_{ljk},\qquad |\nabla w|\le |B|. \tag{17}\] The inequality \(|\nabla(w^2)|^2\le4w^2|B|^2\) holds everywhere, including at \(w=0\). Proof. The trace equation is \(\mathop{\mathrm{tr}}T=\Delta g+mg=0\). On the unit sphere the commutation rule for the covector \(p\) gives \[\nabla_i(\nabla^2g)_{jk}-\nabla_j(\nabla^2g)_{ik} =\delta_{ik}p_j-\delta_{jk}p_i.\] Adding the derivatives of \(g\mathop{\mathrm{Id}}\) cancels the right side. Thus \(T\) is Codazzi: \(\nabla_iT_{jk}=\nabla_jT_{ik}\). Together with symmetry in \(j,k\), this proves full symmetry of \(B\). Differentiating \(\mathop{\mathrm{tr}}T=0\) shows \(B_{ijj}=0\); full symmetry gives all other traces and hence \(\mathop{\mathrm{div}}T=0\). For clarity, the same curvature convention on a covariant two-tensor is \[([\nabla_l,\nabla_i]T)_{jk} =\delta_{lj}T_{ik}+\delta_{lk}T_{ij} -\delta_{ij}T_{lk}-\delta_{ik}T_{lj}.\] This proves (13). Substitute \((l,i,j,k)=(i,j,i,k)\) there and sum over \(i\). By symmetry and the zero traces of \(B\), the left side is \(\Delta T_{jk}\), whereas the right side is \[mT_{jk}+T_{jk}-T_{jk}-\delta_{jk}\mathop{\mathrm{tr}}T=mT_{jk}.\] The product rule for \(|T|^2\) now proves (14). The definitions give (15) and \[\nabla_l s^2=2gp_l+2p_i(T_{li}-g\delta_{li})=2T_{li}p_i.\] Taking the divergence and using \(\mathop{\mathrm{div}}T=0\) and \(\mathop{\mathrm{tr}}T=0\) gives \(\Delta s^2=2T_{li}(T_{li}-g\delta_{li})=2w^2\). Finally, differentiating \(w^2\) and applying Cauchy–Schwarz proves both gradient estimates. ◻ Lemma 10 (Strict gradient bound on the original component). On \(\Omega_o\) one has \(s_o^2<1\). At every boundary point, put \[ H_o=-T_o(N_o,N_o). \tag{18}\] Then \[ H_o>0,\qquad T_op_o=-H_op_o,\qquad w_o>0 \quad\hbox{on }\partial\Omega_o. \tag{19}\] Consequently \(w_o\) has a positive minimum on \(\partial\Omega_o\), and its zero set is a compact subset of \(\Omega_o\). Proof. The boundary conditions give \(s_o^2=1\) on \(\partial\Omega_o\). By (16), \(s_o^2\) is subharmonic. The maximum principle on the compact closure gives \(s_o^2\le1\). If equality held at an interior point, the strong maximum principle would make \(s_o^2\) constant on the connected component, and then \(\Delta s_o^2=2w_o^2\) would imply \(T_o\equiv0\). Thus \(s_o^2<1\) inside. The smooth boundary has the interior sphere property at each point. The Hopf boundary lemma therefore gives \(\partial_{N_o}s_o^2<0\). Since \(p_o=N_o\) there, (16) identifies this derivative as \(2T_o(N_o,N_o)\), proving \(H_o>0\). For a tangent vector \(\tau\) to \(\partial\Omega_o\), tangential differentiation of \(s_o^2=1\) similarly gives \(T_o(\tau,N_o)=0\). This proves the eigenvector assertion and \(w_o\ge H_o>0\). Continuity and compactness of the boundary give a positive minimum of \(w_o\) and a neighborhood of the boundary disjoint from its zero set. ◻ The remaining boundary formulas are local. Thus they apply on the original boundary and, once the strict sign has been transported, on the regular side of the suspension. Proposition 11 (Boundary jets). Suppose \(g=0\), \(p=N\), and \(T(N,N)=-H<0\) on a smooth boundary portion, where \(N\) is inward. Let Greek indices be tangent to this boundary and put \[\mathfrak h_{\alpha\beta} =\langle\nabla_{\alpha}N,e_\beta\rangle.\] Then \(T_{N\alpha}=0\), \(T_{\alpha\beta}=\mathfrak h_{\alpha\beta}\), and \(H=\mathop{\mathrm{tr}}\mathfrak h\); in particular \(H\) is a sum of principal curvatures with the specified convention, without division by \(m-1\). If \(\nabla^{\partial}\) is the boundary connection, all components of \(B\) are described, up to permutation of slots, by \[\begin{align*} B_{\alpha\beta\gamma} &=(\nabla^{\partial}_{\alpha}\mathfrak h)_{\beta\gamma}, &B_{NN\alpha}&=-\partial_\alpha H, \tag{20}\\ B_{N\alpha\beta} &=T_{NN}T_{\alpha\beta}-(T^2)_{\alpha\beta}, &B_{NNN}&=w^2. \tag{21}\end{align*}\] The components with an odd number of normal slots are thus determined by the pointwise tensor \(T\); the other components involve tangential derivatives of curvature. The tensor \(V\) below encodes this odd-normal part after normalization. Here \(w>0\). Define the normalized tensors and eigenvalue \[A=T/w,\qquad \lambda=T_{NN}/w=-H/w, \qquad S_0=\lambda A-A^2,\] and the symmetric tensor \[ V_{ijk}=p_i(S_0)_{jk}+p_j(S_0)_{ki}+p_k(S_0)_{ij} +p_ip_jp_k. \tag{22}\] Then \(|p|=1\), \(Ap=\lambda p\), \(\lambda<0\), \(S_0p=0\), \(\mathop{\mathrm{tr}}S_0=-1\), and \(V\) is trace-free. The tensors \(B/w^2\) and \(V\) agree in every component with an odd number of normal slots. Moreover, \[ V_{Njk}=(S_0+p\otimes p)_{jk},\qquad \partial_N w=w^2\bigl(2\lambda-\mathop{\mathrm{tr}}A^3\bigr). \tag{23}\] Proof. On the boundary, tangential differentiation of \(|p|^2=1\) gives \(T_{N\alpha}=0\), and \(\nabla_\alpha p=T e_\alpha\) because \(g=0\). Since \(p=N\) along the boundary, this proves \(T_{\alpha\beta}=\mathfrak h_{\alpha\beta}\). The trace equation for \(T\) gives \(\mathop{\mathrm{tr}}\mathfrak h=-T_{NN}=H\). In differentiating three tangential entries of \(T\), the normal connection terms pair with \(T(N,e_\alpha)=0\). This yields the first formula in (20). Likewise, \[\partial_\alpha T(N,N) =B_{\alpha NN}+2T(\nabla_\alpha N,N)=B_{\alpha NN},\] which proves the second formula by symmetry of \(B\). For a tangent field \(e_\alpha\), differentiate \(T(N,e_\alpha)=0\) along \(e_\beta\). Since \(\langle N,\nabla_\beta e_\alpha\rangle =-\mathfrak h_{\beta\alpha}\), this gives \[0=B_{\beta N\alpha}+(T^2)_{\beta\alpha} -T_{NN}T_{\beta\alpha}.\] This proves the first formula in (21). Taking its tangential trace gives \[\sum_\alpha B_{N\alpha\alpha} =T_{NN}\sum_\alpha T_{\alpha\alpha} -\sum_{\alpha,\beta}T_{\alpha\beta}^2 =-T_{NN}^2-\sum_{\alpha,\beta}T_{\alpha\beta}^2=-w^2.\] The zero trace of \(B\) then gives \(B_{NNN}=w^2\). Normalization gives \(S_0p=0\) and \(\mathop{\mathrm{tr}}S_0=-|A|^2=-1\). Consequently \(V_{ijj}=p_i\mathop{\mathrm{tr}}S_0+2(S_0p)_i+p_i=0\). In an orthonormal frame with first vector \(p=N\), one has \(V_{N\alpha\beta}=(S_0)_{\alpha\beta}\) and \(V_{NNN}=1\). These are exactly the normalized identities (21), and \(V_{Njk}=S_{0,jk}+p_jp_k\). Finally, \(T\) has only its normal-normal and tangential-tangential blocks, so the odd-normal formulas account for the entire contraction \(T_{jk}B_{Njk}\). Thus \[\partial_Nw=w^{-1}T_{jk}B_{Njk} =w^2 A_{jk}(S_0+p\otimes p)_{jk} =w^2(2\lambda-\mathop{\mathrm{tr}}A^3),\] as asserted. ◻ Transport to a suspensionProposition 12 (Suspended data). If \(m_o=4\), write \(\rho=\sin t\), \(c=\cos t\) and parameterize the regular part of the suspended component by \[(t,y)\longmapsto (\rho y,c),\qquad 0<t<\pi,\quad y\in\Omega_o.\] Its metric is \(dt^2+\rho^2 ds_{S^4}^2\), and \(\widetilde g=\rho g_o\). With \(E_0=\partial_t\) and \(E_\alpha=\rho^{-1}e_\alpha\) for a lifted orthonormal frame on \(S^4\), the data satisfy \[\begin{align*} \widetilde p_0&=c g_o, &\widetilde p_\alpha&=(p_o)_\alpha, &\widetilde s^2&=s_o^2, \tag{24}\\ \widetilde T_{00}&=0, &\widetilde T_{0\alpha}&=0, &\widetilde T_{\alpha\beta}&=\rho^{-1}(T_o)_{\alpha\beta}. \tag{25}\end{align*}\] The only possibly nonzero components of \(\widetilde B\) are \[ \widetilde B_{\alpha\beta\gamma} =\rho^{-2}(B_o)_{\alpha\beta\gamma},\qquad \widetilde B_{0\alpha\beta} =\widetilde B_{\alpha0\beta} =\widetilde B_{\alpha\beta0} =-c\rho^{-2}(T_o)_{\alpha\beta}. \tag{26}\] In particular, \[ \widetilde w=\rho^{-1}w_o,\qquad |\widetilde B|^2 =\rho^{-4}\bigl(|B_o|^2+3c^2w_o^2\bigr),\qquad \widetilde H=\rho^{-1}H_o \quad\hbox{on the regular side.} \tag{27}\] The volume and side-boundary measures are \[ dV=\rho^4\,dt\,dV_o,\qquad d\sigma=\rho^3\,dt\,d\sigma_o. \tag{28}\] Every local identity of Proposition 9 holds with \(m=5\), and Proposition 11 holds on the regular side. Also \(\widetilde s^2<1\) inside and \(\widetilde H,\widetilde w>0\) on that side. Proof. At a point where the base frame is normal, the warped metric has connection \[\nabla_{E_0}E_0=\nabla_{E_0}E_\alpha=0,\qquad \nabla_{E_\alpha}E_0=(c/\rho)E_\alpha,\qquad \nabla_{E_\alpha}E_\beta=-(c/\rho)\delta_{\alpha\beta}E_0.\] The horizontal part of the last expression at other points is the base connection with factor \(\rho^{-1}\). Differentiating \(\widetilde g=\rho g_o\) gives (24). Its Hessian has entries \[(\nabla^2\widetilde g)_{00}=-\rho g_o, \quad (\nabla^2\widetilde g)_{0\alpha}=0, \quad (\nabla^2\widetilde g)_{\alpha\beta} =\rho^{-1}(\nabla_o^2 g_o)_{\alpha\beta} +(c^2/\rho)g_o\delta_{\alpha\beta}.\] Adding \(\widetilde g\mathop{\mathrm{Id}}\) proves (25). Differentiating that tensor with the displayed connection gives (26); entries with two or three zero slots vanish. Each term with exactly one zero slot occurs in three positions, which proves the factor \(3\) in (27). The inward normal to the side is the horizontal lift of \(N_o\). Thus the normal-normal entry of (25) gives \(\widetilde H=\rho^{-1}H_o\). The metric on the side is \(dt^2+\rho^2 ds_{\partial\Omega_o}^2\), proving (28). The trace of (25) is zero, so \(\Delta\widetilde g=-5\widetilde g\). The interior identities follow locally from Proposition 9. The bounds for \(s^2\), \(H\) and \(w\) follow from the original compact component and (24)–(27); no maximum principle at a tip is used. The local boundary hypotheses and hence Proposition 11 hold on the regular side. ◻ We now use the unadorned notation \(\Omega,g,p,T,B,w,H\) for the working component in \(S^5\) defined at the end of Section [sec:reductions]. It is either \(\Omega_o\) when \(m_o=5\), or the above suspension when \(m_o=4\). All boundary integrals mean integrals over the regular boundary, with the two tips omitted in the suspended case. The original data retain the subscript \(o\) when needed. Second variation under compact perturbationsWe derive stability directly from the minimizing property, including its sign and test class. The argument works in either original ambient dimension and in the six-dimensional suspension. Proposition 13 (Euclidean stability on a component). Let \(D\) be the selected positivity component of a global minimizing cone in \(\mathbb R^{m+1}\). Let \(\phi\) be any real smooth function up to the regular boundary of \(D\), with compact support in that regular closure away from the vertex and from any other singular points. Then \[ \int_D |\nabla\phi|^2\,dx \ge \int_{\partial D}H_E\phi^2\,d\sigma_E, \qquad H_E=-D^2u(N_E,N_E), \tag{29}\] where \(N_E\) is the inward Euclidean unit normal. No restriction is imposed on the boundary trace or normal derivative of \(\phi\). Proof. Every regular free-boundary point has a neighborhood where the positive phase is the single side belonging to \(D\). Compactness of the boundary portion meeting \(\mathop{\mathrm{supp}}\phi\) gives finitely many such neighborhoods and a collar separated from every other positivity component and every singular point. Choose the collar over a slightly larger boundary portion, so that the extension of \(\phi\) can vanish near its lateral edge. Extend \(\phi\) smoothly across this boundary with compact support in these neighborhoods. Also extend the positive side of \(u\) smoothly to a function \(F\) on the collar, with its interior jets unchanged. Since \(u=0\) and \(\nabla u=N_E\) on the boundary, shrinking the collar makes \(F\) negative on its exterior side and positive on its interior side. Such an extension need only be smooth to finite order sufficient for the expansions below; smooth boundary regularity supplies this, and a partition of unity joins the local extensions while preserving the interior function and its jets. For a real parameter \(\varepsilon\) of either sufficiently small sign, use \((F+\varepsilon\phi)_+\) in the collar, \(u+\varepsilon\phi\) on the rest of the affected portion of \(D\), and \(u\) elsewhere. These definitions agree: \(F=u\) on the interior of the collar, and on the compact interior support outside a smaller collar \(u\) has a positive minimum. Near the exterior and lateral edges of the modification the function agrees with \(u\). The resulting competitor \(u_\varepsilon\) is nonnegative, belongs to \(H^1_{\mathrm{loc}}\), and has compactly supported difference from \(u\) in a containing ball. It is therefore admissible for the original global minimizer. This construction uses no separate minimizing assertion for \(D\). We compute the change in energy in such a ball. First integrate the perturbed density over the fixed original domain \(D\), and then account for the signed boundary layer where its positivity set changes. The fixed-domain contribution is \[\begin{align*} &2\varepsilon\int_D\langle\nabla u,\nabla\phi\rangle\,dx +\varepsilon^2\int_D|\nabla\phi|^2\,dx\\ &\hspace{15mm}=-2\varepsilon\int_{\partial D}\phi\,d\sigma_E +\varepsilon^2\int_D|\nabla\phi|^2\,dx. \tag{30}\end{align*}\] Here harmonicity eliminates the interior term and the outward normal is \(-N_E\), so \(\partial_{-N_E}u=-1\). All integrations by parts have compact support on the regular portion. Use coordinates \((y,\xi)\) obtained by flowing from the original boundary under \[V_F=\frac{\nabla F}{|\nabla F|^2},\qquad \xi=F.\] The denominator is nonzero on a sufficiently small compact collar. At \(\xi=0\), \(V_F=N_E\). If \(j(y,\xi)\) is the volume Jacobian relative to \(d\sigma_E(y)\,d\xi\), then \[ j(y,0)=1,\qquad \partial_\xi j(y,0)=\mathop{\mathrm{div}}V_F(y,0) =\Delta F-2F_{N_EN_E}=2H_E. \tag{31}\] The equality \(\Delta F=0\) on the original boundary follows from matching the interior second derivatives; harmonicity outside \(D\) is unnecessary. Moreover, \[\left.\partial_\xi|\nabla F|^2\right|_{\xi=0} =2F_{N_EN_E}=-2H_E,\] and hence \[ \left.\partial_\xi\bigl[(1+|\nabla F|^2)j\bigr] \right|_{\xi=0}=2H_E. \tag{32}\] Write \(\phi_0=\phi(y,0)\) and \(\phi_N=\partial_{N_E}\phi(y,0)\). The new zero surface has coordinate \(b_\varepsilon(y)\) determined by \(b_\varepsilon+\varepsilon\phi(y,b_\varepsilon)=0\). The implicit function theorem and \(\partial_\xi\phi(y,0)=\phi_N\) give, uniformly on the compact patch, \[ b_\varepsilon=-\varepsilon\phi_0 +\varepsilon^2\phi_0\phi_N+O(|\varepsilon|^3). \tag{33}\] Throughout the layer \(|\xi|\le C|\varepsilon|\), its perturbed energy density is \[ (1+|\nabla(F+\varepsilon\phi)|^2)j =2+2\varepsilon\phi_N+2H_E\xi +O(\varepsilon^2+|\varepsilon\xi|+\xi^2). \tag{34}\] Integrating this density from \(b_\varepsilon\) to \(0\) is the signed correction, whether the phase expands or contracts. The terms of order at most two are \[\begin{align*} \int_{b_\varepsilon}^{0} (1+|\nabla(F+\varepsilon\phi)|^2)j\,d\xi &=-2b_\varepsilon-2\varepsilon\phi_N b_\varepsilon -H_E b_\varepsilon^2+O(|\varepsilon|^3)\\ &=2\varepsilon\phi_0 +\varepsilon^2\bigl(-2\phi_0\phi_N +2\phi_0\phi_N-H_E\phi_0^2\bigr) +O(|\varepsilon|^3)\\ &=2\varepsilon\phi_0-\varepsilon^2H_E\phi_0^2 +O(|\varepsilon|^3). \tag{35}\end{align*}\] This displays the cancellation of the normal-derivative terms. The uniform remainder is integrable over the compact affected boundary. Adding (30) also cancels the linear terms, leaving \[J(u_\varepsilon)-J(u) =\varepsilon^2\left( \int_D|\nabla\phi|^2\,dx -\int_{\partial D}H_E\phi^2\,d\sigma_E\right) +o(\varepsilon^2).\] The energies here are in a ball containing the modification. Minimality and division by \(\varepsilon^2\) prove (29). ◻ The angular constant comes from the sharp radial Hardy threshold, which also enters the cone instability criteria in [13]. We include the radial calculation to specify the compact supports and the order of limits. Corollary 14 (Spherical stability). For every real function \(v\) smooth up to the regular boundary of the working component \(\Omega\subset S^5\), with compact support in its closure away from the tips if present, \[ \int_{\partial\Omega}H v^2\,d\sigma \le \int_\Omega\bigl(|\nabla v|^2+4v^2\bigr)\,dV. \tag{36}\] The trace of \(v\) on the regular boundary is arbitrary. Proof. More generally, consider a link component \(\Omega\subset S^m\) with angular tests supported in its regular closure, and write \(D=\{ry:r>0,\ y\in\Omega\}\). On its regular side the normal is tangent to \(S^m\), and (12) gives \(H_E(ry)=r^{-1}H(y)\). Polar volume has factor \(r^m\,dr\) and side area has factor \(r^{m-1}\,dr\). For \(\zeta\in C_c^\infty((0,\infty))\) and an allowed angular test \(v\), put \(\phi(ry)=\zeta(r)v(y)\) in Proposition 13. It gives \[\begin{align*} \left(\int_0^\infty r^{m-2}\zeta^2\,dr\right) \int_{\partial\Omega}Hv^2\,d\sigma &\le \left(\int_0^\infty r^{m-2}\zeta^2\,dr\right) \int_\Omega|\nabla v|^2\,dV\\ &\quad+ \left(\int_0^\infty r^m|\zeta'|^2\,dr\right) \int_\Omega v^2\,dV. \end{align*}\] For each fixed \(\zeta\) and \(v\), the Euclidean support lies in a finite annulus and is separated from the singular set and other components, so the preceding second variation applies. Set \(\beta=(m-1)/2\). For \(f_\zeta(r)=r^\beta\zeta(r)\), integration of the cross term gives the radial Hardy identity \[\int_0^\infty r^m|\zeta'|^2\,dr -\beta^2\int_0^\infty r^{m-2}\zeta^2\,dr =\int_0^\infty r|f_\zeta'(r)|^2\,dr\ge0.\] To approach its constant, choose nonzero \(\psi\in C_c^\infty(\mathbb R)\) and set \(\zeta_k(r)=r^{-\beta}\psi((\log r)/k)\). The substitution \(\tau=(\log r)/k\) yields \[ \frac{\int_0^\infty r^m|\zeta_k'|^2\,dr} {\int_0^\infty r^{m-2}\zeta_k^2\,dr} =\beta^2+\frac1{k^2} \frac{\int_\mathbb R|\psi'|^2\,d\tau} {\int_\mathbb R\psi^2\,d\tau}. \tag{37}\] Indeed the numerator before division is \(k\int_\mathbb R(-\beta\psi+k^{-1}\psi')^2\,d\tau\), and \(\int\psi\psi'=0\) by compact support. Divide the separated stability inequality by its positive radial factor and let \(k\to\infty\). In tangent dimension \(m=5\), one has \(\beta^2=4\), proving (36). The same computation on an original link with \(m_o=4\) would give \(9/4\); the working link always has \(m=5\). The order of limits is fixed: for each angular test and each finite \(k\), first take \(\varepsilon\to0\) in the second variation, and then take \(k\to\infty\). Angular cutoffs used later are removed only after (36) has been obtained for each fixed smooth test. The collar and the allowable size of \(\varepsilon\) may depend on these fixed supports. ◻ A universal algebraic certificate
The comparison needed for the stability inequality is a statement about finite-dimensional tensors. We state it with free algebraic data before deriving its geometric consequences. All tensor spaces here are over \(\mathbb R^5\), with the Euclidean inner product and full tensor norm; repeated indices range from \(1\) to \(5\) and are summed. A symmetric tensor is called trace-free if contraction of any two slots is zero. We write \(\mathop{\mathrm{STF}}^r(\mathbb R^5)\) for the space of symmetric trace-free tensors of rank \(r\). Fix \[ a=\frac{29}{50},\qquad Q=250000,\qquad \epsilon_* =\frac{438907}{12000000},\qquad \delta_* =\frac{6984137359}{31250000000}. \tag{38}\] Let \(A\) be a symmetric trace-free matrix with \(|A|^2=\mathop{\mathrm{tr}}A^2=1\), and let \(X\) be a symmetric trace-free three-tensor. Set \[q_l=X_{lij}A_{ij},\qquad \chi=1-|p|^2-g^2,\] where \(p,\ell\in\mathbb R^5\) and \(g,w,z\in\mathbb R\) are otherwise free. In particular, \(w\) and \(z\) need not be positive. On tensor polynomials in \(A,p,g,\delta\), define \(D_l\) by the product rule and \[ D_l A_{ij}=X_{lij}-q_lA_{ij},\qquad D_l p_i=wA_{li}-gz\delta_{li},\qquad D_l g=zp_l,\qquad D_l\delta_{ij}=0. \tag{39}\] This is a first derivative with values in the enlarged polynomial algebra containing \(X,w,z\); no derivative of these new variables is used. Proposition 15 (Universal inequalities). There are rational tensor polynomials \(f,U_{jk},K_{lijk},M_l,P_l\) in \(A,p,g,\delta\), with the following properties. They are equivariant under orthogonal changes of coordinates, with \(f\) invariant. Their coefficients are specified by Table I in Appendix 9, using the decoding in Section 6.1. Put \[C_{lijk}=K_{lijk}-K_{iljk},\qquad G_{lijk}=\delta_{li}U_{jk}+C_{lijk}.\] For every choice of the data above with \(\chi\ge0\), \[\begin{align*} \mathcal I:={}&z^2\left[5U_{jk}A_{jk} +C_{lijk}(\delta_{lj}A_{ik}+\delta_{lk}A_{ij})-4f\right]\\ &+\left[D_lG_{lijk}+(2a-1)q_lG_{lijk}\right]X_{ijk}\\ &+w\left[D_lM_l+(2a+1)q_lM_l\right] +z\left[D_lP_l+2aq_lP_l\right]\\ &+f|\ell|^2-(D_lf+2aq_lf)\ell_l \ \ge\ \epsilon_*(z^2+|\ell|^2). \tag{40}\end{align*}\] In particular, \(f\ge\epsilon_*\) whenever \(g^2+|p|^2\le1\). For every symmetric trace-free unit matrix \(A\) and vector \(p\) satisfying \[g=0,\qquad |p|=1,\qquad Ap=\lambda p,\qquad\lambda<0,\] define \[S_0=\lambda A-A^2,\qquad V_{ijk}=p_i(S_0)_{jk}+p_j(S_0)_{ki}+p_k(S_0)_{ij}+p_ip_jp_k.\] Then \[ \mathcal S:=-\lambda f+U_{jk}(S_0+p\otimes p)_{jk} +C_{lijk}p_lV_{ijk}+p_lM_l \ \ge\ \delta_*. \tag{41}\] Moreover, \(U\) and \(K\) are even in \(p\), and \(P=0\) when \(g=0\). We prove Proposition 15 in Section 6. First, Section [sec:flux] derives its geometric consequences using only the stated properties. The calculation explains the Laplacian and skew contractions in \(G\), the three coefficients of \(q\), and the free vector \(\ell\). From the algebraic certificate to a geometric flux
We work on the regular part of the spherical component \(\Omega\subset S^5\) constructed in Section [sec:reductions], and put \[\Omega_+=\{x\in\Omega:w(x)>0\},\qquad a=\frac{29}{50}.\] All assertions up to the side boundary use derivatives from the positive side. In the suspended case the two tips are excluded throughout Section [sec:flux]. We apply the universal inequalities of Proposition 15, whose proof follows in Section 6. For its coefficient polynomials \(f,U,K,M,P\), with their stated slot orders, set \[ C_{lijk}=K_{lijk}-K_{iljk},\qquad G_{lijk}=\delta_{li}U_{jk}+C_{lijk}. \tag{42}\] Thus \(C_{iljk}=-C_{lijk}\). No other slot symmetry of \(U\) or \(K\) will be needed. The orthogonal equivariance asserted in Proposition 15 ensures that all the fields defined below are independent of the orthonormal frame. Normalized differentiation and the fourth derivativesOn \(\Omega_+\) define \[ A=\frac{T}{w},\qquad X=\frac{B}{w},\qquad q_l=X_{ljk}A_{jk}. \tag{43}\] The identities of Proposition 9 show that \(A\in\mathop{\mathrm{STF}}^2(\mathbb R^5)\), \(|A|=1\), and \(X\in\mathop{\mathrm{STF}}^3(\mathbb R^5)\). Moreover, \(g^2+|p|^2\leq1\), so these are admissible data for the interior certificate. The homogenizing variable in that certificate is set to \(z=1\) in every geometric application. Lemma 16. At the data (43), the derivation \(D_l\) of (39), evaluated at \(z=1\), equals covariant differentiation on every coefficient polynomial in \(A,p,g,\delta\). In particular, \[ \nabla_lw=wq_l,\qquad \nabla_lA_{ij}=X_{lij}-q_lA_{ij},\qquad \nabla_lp_i=wA_{li}-g\delta_{li},\qquad \nabla_lg=p_l. \tag{44}\] Furthermore, \[ \frac1wG_{lijk}\nabla_lB_{ijk} =5U_{jk}A_{jk} +C_{lijk}(\delta_{lj}A_{ik}+\delta_{lk}A_{ij}). \tag{45}\] Proof. Differentiating \(w^2=T_{jk}T_{jk}\) gives \(\nabla_lw=T_{jk}B_{ljk}/w=wq_l\). The quotient rule then gives the formula for \(\nabla A\), and \(T=\nabla^2g+g\mathop{\mathrm{Id}}\) gives the formula for \(\nabla p\). In a covariantly constant orthonormal frame at the point under consideration these are exactly the generator rules for \(D_l\). The product rule and compatibility of the connection with contractions prove the assertion for every coefficient polynomial. The rules also preserve the normalization constraints: \[D_l(\mathop{\mathrm{tr}}A)=0,\qquad D_l(|A|^2)=2q_l-2q_l=0\] on the admissible data. Here \(D_l\) is applied only to coefficient polynomials in \(A,p,g,\delta\). Differentiation of the factors \(w\) and \(X\), which are not coefficient variables, is performed separately. In particular, no rule for \(D_lX\) or \(D_lq\) is being assumed. For the \(U\) part of (45), \[\frac1w\delta_{li}U_{jk}\nabla_lB_{ijk} =\frac1wU_{jk}\Delta T_{jk}=5U_{jk}A_{jk}.\] For the skew part, the curvature identity in Proposition 9 yields \[\begin{align*} \frac1w C_{lijk}\nabla_lB_{ijk} &=\frac1{2w}C_{lijk} (\nabla_lB_{ijk}-\nabla_iB_{ljk})\\ &=\frac12 C_{lijk} (\delta_{lj}A_{ik}+\delta_{lk}A_{ij} -\delta_{ij}A_{lk}-\delta_{ik}A_{lj})\\ &=C_{lijk}(\delta_{lj}A_{ik}+\delta_{lk}A_{ij}). \end{align*}\] The last equality follows by interchanging the summed indices \(l,i\) in the last two terms and using \(C_{iljk}=-C_{lijk}\). Equivalently, \(C_{lijk}\nabla_lB_{ijk}\) equals \(K_{lijk}(\nabla_lB_{ijk}-\nabla_iB_{ljk})\); thus the factor \(1/2\) belongs to the formula with \(C\), not to the formula with \(K\). This calculation eliminates every undetermined fourth derivative of \(g\). Indeed, if two possible tensors \(\nabla_lB_{ijk}\) have the same Laplacian contraction and the same skew difference in \(l,i\), their difference has zero contraction with \(\delta_{li}U_{jk}\) and with \(C_{lijk}\). No constraint on the remaining fourth derivative entries is required. ◻ The square root and the interior comparisonProposition 15 gives \[ f(A,p,g)\geq\epsilon_* =\frac{438907}{12000000}>0 \tag{46}\] on the entire normalized coefficient domain. The verification in Section 6.5 obtains this bound using the free algebraic data \(X=w=z=0\). Since \(f\) is independent of \(w\), this permitted algebraic substitution does not impose \(w=0\) on the geometric data: the fields below are defined on \(\Omega_+\), where \(w>0\). Define on \(\Omega_+\) \[ h=w^a\sqrt f,\qquad Z_l=w^{2a}\bigl(G_{lijk}X_{ijk}+wM_l+P_l\bigr). \tag{47}\] These fields are smooth on \(\Omega_+\) and extend smoothly to each regular side boundary point, since \(w>0\) there by Proposition 11. Proposition 17. The scalar \(h\) and vector field \(Z\) satisfy \[ \mathop{\mathrm{div}}Z\geq 4h^2+|\nabla h|^2+\epsilon_*w^{2a} \qquad\text{on }\Omega_+. \tag{48}\] Proof. First, \[\nabla_lX_{ijk}=\frac1w\nabla_lB_{ijk}-q_lX_{ijk}.\] Applying the product rule to each of the three summands of \(Z\), and then using (45), gives the exact identity \[\begin{align*} w^{-2a}\mathop{\mathrm{div}}Z ={}&5U_{jk}A_{jk} +C_{lijk}(\delta_{lj}A_{ik}+\delta_{lk}A_{ij})\\ &+\bigl[D_lG_{lijk}+(2a-1)q_lG_{lijk}\bigr]X_{ijk}\\ &+w\bigl[D_lM_l+(2a+1)q_lM_l\bigr] +D_lP_l+2aq_lP_l. \tag{49}\end{align*}\] Every \(D_l\) in Equation (49) is evaluated at \(z=1\). The three coefficients of \(q\) come respectively from differentiating \(w^{2a}X\), \(w^{2a+1}\), and \(w^{2a}\); in the first expression the quotient \(X=B/w\) contributes \(-q_lX\). Thus their values are exactly \[2a-1=\frac4{25},\qquad 2a+1=\frac{54}{25},\qquad 2a=\frac{29}{25}.\] Put \[\mathcal F_l=D_lf+2aq_lf,\qquad \ell_* =\frac{\mathcal F}{2f}.\] Differentiating \(h\) and completing the square give, respectively, \[ \nabla_lh=\frac{w^a}{2\sqrt f}\mathcal F_l, \qquad |\nabla h|^2=\frac{w^{2a}}{4f}|\mathcal F|^2, \tag{50}\] and \[f|\ell|^2-\mathcal F\cdot\ell =f|\ell-\ell_*|^2-\frac{|\mathcal F|^2}{4f}.\] In particular, substituting (49) into the interior polynomial \(\mathcal I\) of (40) gives the identity, for every \(\ell\), \[ \mathcal I\big|_{z=1} =w^{-2a}\bigl(\mathop{\mathrm{div}}Z-4h^2-|\nabla h|^2\bigr) +f|\ell-\ell_*|^2. \tag{51}\] The division by \(f\) is justified uniformly by (46). Evaluating (40) at \(\ell=\ell_*\) therefore gives the slightly stronger estimate \[\mathop{\mathrm{div}}Z-4h^2-|\nabla h|^2 \geq\epsilon_*w^{2a}(1+|\ell_*|^2),\] which implies (48). ◻ Boundary reflection and the signed fluxAt a regular side boundary point, Proposition 11 gives \[g=0,\quad p=N,\quad |p|=1,\quad Ap=\lambda p, \qquad \lambda=-\frac H w<0.\] As in the boundary certificate, define \[ S_0=\lambda A-A^2,\qquad V_{ijk}=p_i(S_0)_{jk}+p_j(S_0)_{ki}+p_k(S_0)_{ij} +p_ip_jp_k. \tag{52}\] Lemma 18. At such a point, \[ p_lG_{lijk}\frac{B_{ijk}}{w^2} =p_lG_{lijk}V_{ijk} =U_{jk}(S_0+p\otimes p)_{jk}+C_{lijk}p_lV_{ijk}. \tag{53}\] Proof. Let \(R=\mathop{\mathrm{Id}}-2p\otimes p\) be reflection in \(p^\perp\). Since \(A\) is symmetric and \(Ap=\lambda p\), it preserves \(p^\perp\), and hence \(RAR^T=A\). Also \(Rp=-p\). Orthogonal equivariance and the even parity in \(p\) of \(U,K\), and therefore of \(G\), give \[R^{\otimes4}G(A,p,0)=G(RAR^T,Rp,0) =G(A,-p,0)=G(A,p,0).\] For the three-tensor \(F_{ijk}=p_lG_{lijk}\), contraction commutes with orthogonal transformations, so \[R^{\otimes3}F=(Rp)_l(R^{\otimes4}G)_{lijk}=-F.\] Choose any orthonormal basis with its last vector \(p\). A component of a three-tensor with \(k\) normal slots is multiplied by \((-1)^k\) under \(R^{\otimes3}\). It follows that \(F\) has zero components whenever the number of normal slots is even. This assertion concerns the full unsymmetrized tensor \(F\); no permutation of its slots, and no extra symmetry of \(U\) or \(K\), has been used. For tangent indices \(\alpha,\beta\), the boundary jet identities give \[\frac{B_{N\alpha\beta}}{w^2} =\lambda A_{\alpha\beta}-(A^2)_{\alpha\beta} =(S_0)_{\alpha\beta},\qquad \frac{B_{NNN}}{w^2}=1.\] Because \(S_0p=0\), these are precisely the components of \(V\) with one or three normal slots, including all their permutations. Thus the part of \(B/w^2\) with an odd number of normal slots equals \(V\); \(V\) itself has no even-normal components. The even-normal components of \(B\), which are not determined by \(A,p,g,w\), have zero contraction with \(F\). This proves the first equality in (53). Finally, \(S_0p=0\) and \(|p|=1\) imply \[p_iV_{ijk}=(S_0)_{jk}+p_jp_k.\] Splitting \(G\) according to (42) proves the second equality. The argument used only the orthogonal splitting \(\mathbb R^5=p^\perp\oplus\mathbb Rp\). It therefore applies without any assumption on multiplicities, signs, or nonvanishing of the tangential eigenvalues of \(A\). ◻ Proposition 19. On the regular side boundary, \[ N\cdot Z+Hh^2=w^{2a+1}\mathcal S \geq\delta_*w^{2a+1}>0, \qquad \delta_* =\frac{6984137359}{31250000000}, \tag{54}\] where \(\mathcal S\) is the boundary polynomial in (41). In particular \(N\cdot Z+Hh^2\geq0\). Proof. Proposition 15 gives \(P=0\) at the boundary, where \(g=0\). Since \(X=B/w\), Lemma 18 gives \[\begin{align*} N\cdot Z &=w^{2a+1}\left( U_{jk}(S_0+p\otimes p)_{jk} +C_{lijk}p_lV_{ijk}+p_lM_l\right),\\ Hh^2&=-\lambda w^{2a+1}f. \end{align*}\] Their sum is exactly \(w^{2a+1}\mathcal S\). The boundary data satisfy all hypotheses of (41), proving the stated lower bound. The sign is expressed using the inward normal \(N\); the outward normal used by the divergence theorem is \(-N\). Only the boundary value of \(P\) vanishes. Its derivative term \(D_lP_l\) remains in (49). For example, a monomial \(P_l=gQ_l\) has \(D_iP_l=p_iQ_l+gD_iQ_l\) at \(z=1\), so its derivative need not vanish when \(g=0\). ◻ Uniform estimates for the cutoff argumentLemma 20. There is a finite constant \(C\), depending only on the coefficient polynomials, the dimension, \(a\), and \(\epsilon_*\), such that on \(\Omega_+\) \[ |h|\leq Cw^a,\qquad |\nabla h|\leq Cw^a(1+w+|X|),\qquad |Z|\leq Cw^{2a}(1+w+|X|). \tag{55}\] The constant is independent of \(w\), \(|X|\), and all subsequent cutoff parameters. Proof. The coefficient domain \[\{(A,p,g):A\in\mathop{\mathrm{STF}}^2(\mathbb R^5),\ |A|=1,\ |p|^2+g^2\leq1\}\] is compact. The coefficient polynomials and their first derivatives with respect to their coefficient variables are therefore uniformly bounded there. Also \(|q|\leq|X|\) by Cauchy–Schwarz. The rules (44) consequently imply \[|D f|+|q|\,|f|\leq C(1+w+|X|).\] Use (46) and (50) for the estimates on \(h\) and \(\nabla h\), and use (47) and the boundedness of \(G,M,P\) for the estimate on \(Z\). ◻ Exact verification of the algebraic certificateWe return to the free algebraic data of Section 4 to prove Proposition 15. All tensors are again over \(\mathbb R^5\). The scalars \(w,z\) and vector \(\ell\) are free, and \(X\) is an arbitrary symmetric trace-free three-tensor. Table I specifies the polynomials \(f,U,K,M,P\), and Table II specifies nonnegative comparison forms. For the interior inequality, we compare \(\mathcal I\) with a positive diagonal quadratic form in \((w,z,\ell,X)\) plus projected tensor squares, and bound the remaining error term by term. For the boundary inequality, the analogous comparison uses scalar squares in the four tangential eigenvalues and a uniform bound on eleven remainder coefficients. We give the contraction rules and rational bounds in full. Tensor monomials and the fixed polynomialsHere and in Appendix 9, a rank-\(r\) tensor monomial is encoded by a word of decimal digits, partitioned into triples. Its first triple \(hbj\) specifies \[g^h(\mathop{\mathrm{tr}}A^3)^b w^{\mathbf1_{\{j=1\}}} z^{\mathbf1_{\{j=2\}}}.\] Each later triple \(uvk\) specifies an edge carrying the matrix \(A^k\), with \(A^0=\delta\). An endpoint labeled \(i\in\{0,\ldots,r-1\}\) is the \(i\)th free slot, starting with slot zero. An endpoint labeled \(5\) carries one copy of \(p\), and one labeled \(6\) carries one copy of \(\ell\). Three incidences of label \(7\) are contracted against the three slots of one \(X\); three incidences of label \(8\), if present, use a second copy of \(X\). Each occurrence of a vector endpoint is a separate vector factor. Matrix entries, vector entries, and tensor entries are multiplied, and all nonfree slots are summed. Edges are unoriented because \(A\) is symmetric. Permuting the three incidences at an \(X\) does not change its value because \(X\) is symmetric. There are no implicit multiplicities or factorials in this convention. Write \(\mathcal M_r(s)\) for the tensor encoded by the word \(s\). For example, \[\mathcal M_0(\texttt{200550})=g^2|p|^2,\qquad \mathcal M_4(\texttt{000500520131})_{lijk}=p_lp_jA_{ik},\qquad \mathcal M_1(\texttt{000771700})_l=q_l.\] The digit labels encode slots and must be distinguished from the coordinate indices that are summed from \(1\) to \(5\). Table IFor \(F\in\{f,U,K,M,P\}\), let \(r_F\) be respectively \(0,2,4,1,1\), with slot orders respectively empty, \((j,k)\), \((l,i,j,k)\), \((l)\), \((l)\). If a row of Table I consists of \(F,s,c\), its contribution is \(c\mathcal M_{r_F}(s)/Q\); thus each polynomial is given by \[ F=\frac1Q\sum_{(F,s,c)\ \text{in Table I}}c\,\mathcal M_{r_F}(s). \tag{56}\] Table I contains \(34,57,34,74,32\) rows for \(f,U,K,M,P\), respectively. Every free slot occurs exactly once in every row, and these rows have no \(X,\ell,w,z\) factors. Every \(P\) row has a positive \(g\) exponent. Because an edge has two endpoints and \(U,K\) have even rank, their numbers of \(p\) endpoints are even. This proves the parity assertions of Proposition 15 directly from the definitions. The same definitions show that all five polynomials are equivariant under orthogonal changes of coordinates, with \(f\) invariant. The nonnegative comparison polynomialsThe positive tensor squares below have the form of Gram-matrix certificates for polynomial positivity; compare [15]. Here all rational polynomials, projection identities, and remainder bounds are explicit, and we verify the certificate directly. A header \(\texttt{@ }r\ \mathrm{kind}\ \mathrm{mode}\) in Table II begins a new block; consecutive equal headers still designate different blocks. For a block of nonnegative rank, write its monomial rows as \(s_1,\ldots,s_N\) and its rectangular integer coefficient matrix as \((c_{tc})\). Define \[J_c=\frac1Q\sum_{t=1}^N c_{tc}\mathcal M_r(s_t).\] Its contribution to \(\Gamma\) is \[ \rho\sum_c\langle J_c,\Pi J_c\rangle,\qquad \rho=\begin{cases}1,&\mathrm{mode}=\texttt{I},\\ \chi,&\mathrm{mode}=\texttt{B}. \end{cases} \tag{57}\] Here \(\Pi\) is specified below by \(r\) and \(\mathrm{kind}\). There are nineteen such blocks with a total of \(529\) monomial rows. Every row has degree one in \((w,z,\ell,X)\), assigning degree one to each component of these variables and degree zero to \(A,p,g\). Hence \(\Gamma\) has degree two in these variables. The two remaining blocks have rank \(-1\) and fourteen rows each. In these blocks the four digits of a row word are an exponent \(\beta=(\beta_1,\ldots,\beta_4)\), encoding \(x^\beta\) for \(x\in\mathbb R^4\). Both coefficient matrices have five columns. Set \[J_c(x)=\frac1Q\sum_t c_{tc}x^{\beta_t},\qquad \operatorname{Av}F(x)=\frac1{24}\sum_{\pi\in S_4} F(x_{\pi(1)},\ldots,x_{\pi(4)}).\] Their sum defines the scalar polynomial \[ \Lambda(x)=\sum_{\text{rank }-1\text{ blocks}} \operatorname{Av}\left(\rho(x)\sum_c J_c(x)^2\right), \qquad \rho(x)=\begin{cases}1,&\mathrm{mode}=\texttt{I},\\ x_1+x_2+x_3+x_4,&\mathrm{mode}=\texttt{H}. \end{cases} \tag{58}\] Consequently \(\Lambda\ge0\) when \(\sum x_i\ge0\). Scaled projection operatorsFor an arbitrary rank-\(r\) tensor \(F\), let \(F^\pi\) denote slot permutation by \(\pi\). We use the averaged symmetrization \(\mathop{\mathrm{Sym}}F=(r!)^{-1}\sum_{\pi\in S_r}F^\pi\). The operator \(\operatorname{Tr}_{ab}\) contracts slots \(a,b\) with the Euclidean metric, leaving the remaining slots in their original order. The label \(\texttt{none}\) means \(\Pi=\mathop{\mathrm{Id}}\). The label \(\texttt{anti}\) means \[\Pi F=\sum_{\pi\in S_r}\operatorname{sgn}(\pi)F^\pi.\] For \(\texttt{sym}\), put \(F'=\sum_{\pi\in S_r}F^\pi\) and \(v=\operatorname{Tr}_{12}F'\). The operators used in Table II are \[\begin{align*} (\Pi F)_{ij}&=5F'_{ij}-\delta_{ij}v &&(r=2),\tag{59}\\ (\Pi F)_{ijk}&=7F'_{ijk} -(\delta_{ij}v_k+\delta_{ik}v_j+\delta_{jk}v_i) &&(r=3). \tag{60}\end{align*}\] For \(r=3\) and \(\texttt{hook}\), set \[\begin{align*} F'&=3(F+F^{(12)})-\sum_{\pi\in S_3}F^\pi, &v&=\operatorname{Tr}_{12}F',\\ (\Pi F)_{ijk}&=8F'_{ijk} -(2\delta_{ij}v_k-\delta_{ik}v_j-\delta_{jk}v_i). \tag{61}\end{align*}\] Lemma 21. All these operators are self-adjoint and positive semidefinite on their entire tensor spaces. In particular, \(\Gamma\ge0\) whenever \(\chi\ge0\). Proof. Each slot permutation is orthogonal, with adjoint its inverse. Averaging over all permutations, with or without their signs, gives the orthogonal projection onto symmetric or alternating tensors. Thus \(\texttt{anti}\) is \(r!\) times an orthogonal projection. For a symmetric matrix \(S\), \(S-\delta\mathop{\mathrm{tr}}S/5\) is its orthogonal trace-free projection. Since \(F'=2\mathop{\mathrm{Sym}}F\) in (59), that operator is ten times the symmetric trace-free projection. For rank three define \[(\mathcal Jv)_{ijk}=\delta_{ij}v_k+\delta_{ik}v_j+\delta_{jk}v_i.\] Then \(\operatorname{Tr}_{12}\mathcal Jv=7v\), and for symmetric \(S\), \(\langle S,\mathcal Jv\rangle=3\langle\operatorname{Tr}_{12}S,v\rangle\). It follows that \(S-\mathcal J(\operatorname{Tr}_{12}S)/7\) is the orthogonal projection onto symmetric trace-free tensors. Since now \(F'=6\mathop{\mathrm{Sym}}F\), (60) is \(42\) times that projection. For completeness, the trace correction in (61) can be checked without any representation-theoretic assumptions. Let \[S_{12}=\frac{\mathop{\mathrm{Id}}+(12)}2,\qquad S=\frac16\sum_{\pi\in S_3}\pi,\qquad P_0=S_{12}-S.\] The identities \(S_{12}S=SS_{12}=S\) show that \(P_0\) is an orthogonal projection. Its image \(\mathcal W\) consists of first-pair symmetric tensors orthogonal to fully symmetric tensors. For \(H\in\mathcal W\), \[H_{ijk}=H_{jik},\qquad H_{ijk}+H_{ikj}+H_{jki}=0.\] If \(v=\operatorname{Tr}_{12}H\), contraction of the second identity gives \(\operatorname{Tr}_{13}H=\operatorname{Tr}_{23}H=-v/2\). Define \[(\mathcal Bv)_{ijk}=2\delta_{ij}v_k-\delta_{ik}v_j-\delta_{jk}v_i.\] Direct substitution gives \(\mathcal Bv\in\mathcal W\), \(\operatorname{Tr}_{12}\mathcal Bv=8v\), and \[\langle H,\mathcal Bz\rangle =2\langle v,z\rangle -\langle\operatorname{Tr}_{13}H,z\rangle -\langle\operatorname{Tr}_{23}H,z\rangle =3\langle v,z\rangle.\] Therefore \(Q_0H=H-\mathcal B(\operatorname{Tr}_{12}H)/8\) has all traces zero, and \(H-Q_0H\) is orthogonal to every trace-free tensor in \(\mathcal W\). Thus \(Q_0\) is the orthogonal projection in \(\mathcal W\) onto its trace-free part. Since \(F'=6P_0F\), the hook operator is \(48Q_0P_0\); \(Q_0P_0\) is the orthogonal projection in the full tensor space onto that subspace. In particular \[\langle F,\Pi F\rangle=48|Q_0P_0F|^2\ge0.\] Finally the weights in (57) are nonnegative under the stated constraint. ◻ Exact contraction and differentiation rulesWe give explicit collection rules both to specify the finite calculation and to justify that it is an identity on all the constrained tensors. A graph key records the rank, powers of \(g,w,z\), a multiset of trace factors \(\mathop{\mathrm{tr}}A^k\), and a multiset of edges with their powers. For purposes of ordering keys, replace endpoint labels \(5,6,7,8\) by \(-1,-2,-3,-4\). Sort the endpoints of each edge, sort the edge list and trace list, and take the lexicographically larger edge list obtained by retaining or exchanging \(-3,-4\). A lone cubic tensor is thereby labeled \(-3\). These choices are only ordering conventions: exchange of two identical copies of \(X\) preserves the value of every graph. Tensor multiplication takes the disjoint union of graphs, shifts the free slot labels of the second factor, and distinguishes any two cubic tensor copies. A contraction of two free slots joins their incident edges. If these are different edges, their powers add; if they are the same edge, it becomes a trace \(\mathop{\mathrm{tr}}A^k\). The removed free slots are deleted and the others renumbered in order. The identities underlying these operations are precisely \[\sum_i(A^h)_{ai}(A^k)_{ib}=(A^{h+k})_{ab},\qquad \sum_i(A^k)_{ii}=\mathop{\mathrm{tr}}A^k.\] Reduce the trace factors of powers \(0,1,2\) to \(5,0,1\), respectively. An edge of power zero connecting two slots of the same \(X\) makes the monomial zero. Thus every reduction is an equality on the allowed data. Other identities need not be imposed: retaining two different keys for functions that happen to agree cannot invalidate either the expansion or the later triangle inequality. The derivative of a path and of a trace is \[\begin{align*} D_l(A^k)&=\sum_{h=0}^{k-1}A^hX_lA^{k-1-h}-kq_lA^k, & (X_l)_{ij}&=X_{lij},\tag{62}\\ D_l\mathop{\mathrm{tr}}A^k&=k\,\mathop{\mathrm{tr}}(A^{k-1}X_l)-kq_l\mathop{\mathrm{tr}}A^k. \tag{63}\end{align*}\] For \(k=0\) the sums and derivatives are zero. These formulas follow by differentiating each matrix factor using (39); cyclic invariance of trace makes all \(k\) insertion positions on a trace equal. In graph form, inserting \(X_l\) on an edge of power \(k\) splits it into two paths of lengths \(h,k-1-h\) ending at a new \(X\), whose last slot is the derivative index \(l\). The negative terms from all \(A\) factors combine into minus the total \(A\) degree times \(q_l\) times the original graph. Each \(p\) endpoint is in turn replaced by \(wA_{l\cdot}-gz\delta_{l\cdot}\), and a factor \(g^h\) gives \(hzg^{h-1}p_l\). If an edge has two \(p\) endpoints, these are two product-rule terms, including their multiplicity when they coincide. These operations are also sound if traces were simplified first, since \[D_l\mathop{\mathrm{tr}}A=0,\qquad D_l\mathop{\mathrm{tr}}A^2=2X_{lij}A_{ij}-2q_l|A|^2=0.\] Only the original Table I monomials are differentiated, exactly once. No already differentiated graph, or graph involving \(X,w,z\), is differentiated again. Thus no unspecified higher derivative enters the calculation. Lemma 22 (Closure of the graph calculation). The preceding rules evaluate all terms of \(\mathcal I\) and \(\Gamma\) by a finite sum of scalar graphs. In every intermediate graph a free slot has exactly one incidence, each cubic tensor has exactly three incidences, and there are at most two cubic tensors. Every resulting scalar graph has degree two in \((w,z,\ell,X)\). Proof. The incidence assertions hold for each table row by its decoding. A disjoint union preserves them. A permutation only renames free slots. Contracting two free slots removes their unique incidences and joins the remaining ends, or closes a trace; it does not change the number of slots of any cubic tensor. The derivative insertion described above replaces one edge by two incidences of a new \(X\) and supplies its third incidence as a new free slot. Trace differentiation does the same with a self-edge. The \(q_l\) term supplies two incidences on its self-edge and the third at \(l\). Derivatives of \(p\) and \(g\) supply a free derivative slot without introducing a cubic tensor. These statements prove the incidence assertions by induction over the operations. Table I has degree zero and its derivative has degree one in the indicated variables. The formula for \(\mathcal I\) multiplies such derivatives by one further factor, or a degree-zero coefficient by two factors. Each Table II row has degree one; projection, contraction, and multiplication by \(\chi\) preserve that degree. Squaring produces degree two. In particular no operation used here needs three copies of \(X\). Termination is immediate from the finite tables, finite permutation sums, and the finite product rule; every contraction reduces the number of free slots by two. ◻ The linear contribution of each Table I row is obtained by substitution into (40), forming \(C=K-K^{(12)}\) for a \(K\) row. One check on the slot order and coefficients is the single row \(U=A\) with coefficient one. Its interior contribution is \[5z^2+|X|^2+(2a-2)|q|^2,\] and its boundary contribution is \(2\lambda-\mathop{\mathrm{tr}}A^3\). To expand a Gram block, sum over unordered pairs of rows \(s,t\): the coefficient is \(\sum_c c_{sc}c_{tc}/Q^2\), doubled when \(s\ne t\), and the graph polynomial is the contraction of row \(t\) with \(\Pi\) applied to row \(s\), multiplied by \(\rho\). This uses self-adjointness of \(\Pi\) and accounts for every term of (57) exactly once. Bounds for scalar graphsPut \(\theta=4/5\). We first establish the tensor estimates that will bound each scalar graph, retaining the dependence between \(p\) and \(g\). Lemma 23. For the constrained matrix \(A\) and symmetric trace-free tensor \(X\), \[\begin{align*} \|A\|_{\mathrm{op}}^2&\le\theta,& |A^k|^2&\le\theta^{k-1}\quad(k\ge1),\tag{64}\\ |\mathop{\mathrm{tr}}A^k|^2&\le\theta^{k-2}\quad(k\ge2),& |\mathop{\mathrm{tr}}A^3|^2&\le\frac9{20},\tag{65}\\ |X:A|^2&\le\frac57|X|^2,& (X:A)_i&=X_{ijk}A_{jk}. \tag{66}\end{align*}\] If \(h,b>0\) and \(g^2+|p|^2\le1\), then \[ |g|^{2h}|p|^{2b} \le \left(\frac{h}{h+b}\right)^h \left(\frac{b}{h+b}\right)^b. \tag{67}\] Proof. Let the eigenvalues of \(A\) be \(\alpha_1,\ldots,\alpha_5\). For each \(i\), \[\alpha_i^2=\left(\sum_{j\ne i}\alpha_j\right)^2 \le4\sum_{j\ne i}\alpha_j^2=4(1-\alpha_i^2).\] This proves the operator bound, and \(\sum_i\alpha_i^{2k}\le\theta^{k-1}\sum_i\alpha_i^2\) proves the Frobenius bound. Moreover \(|\sum_i\alpha_i^k|\le\sum_i|\alpha_i|^k \le\theta^{(k-2)/2}\sum_i\alpha_i^2\) for \(k\ge2\). For the sharper cubic bound maximize and minimize \(\sum_i\alpha_i^3\) on the compact set \(\sum_i\alpha_i=0\), \(\sum_i\alpha_i^2=1\). The gradients of the two constraints are independent there. Lagrange multipliers give \(3\alpha_i^2=\mu+2\nu\alpha_i\), so an extremum has two distinct eigenvalues (a single value would violate the constraints). If their multiplicities are \(j,5-j\), the values are, up to simultaneous sign, \[\sqrt{\frac{5-j}{5j}},\qquad -\sqrt{\frac{j}{5(5-j)}}.\] Their cubic trace has absolute value \(|5-2j|/\sqrt{5j(5-j)}\). Its square is \(9/20\) for \(j=1,4\) and \(1/30\) for \(j=2,3\), proving (65). To prove (66), fix a vector \(\zeta\) and symmetrize \(\zeta\otimes A\): \[B_{ijk}=\frac{\zeta_iA_{jk}+\zeta_jA_{ik}+\zeta_kA_{ij}}3, \qquad v=\operatorname{Tr}_{12}B=\frac23 A\zeta.\] Expanding the nine terms in \(|B|^2\) gives \(|B|^2=|\zeta|^2/3+2|A\zeta|^2/3\). The symmetric trace-free projection is \(F=B-\mathcal Jv/7\), in the notation of Lemma 21. Since \(|\mathcal Jv|^2=21|v|^2\) and \(\langle B,\mathcal Jv\rangle=3|v|^2\), \[|F|^2=|B|^2-\frac37|v|^2 =\frac13|\zeta|^2+\frac{10}{21}|A\zeta|^2 \le\frac57|\zeta|^2.\] Symmetry and trace-freeness of \(X\) give \(\langle X:A,\zeta\rangle=\langle X,F\rangle\); Cauchy–Schwarz and duality prove the claim. Finally put \(s=g^2,t=|p|^2\). For \(h,b>0\), the maximum of \(s^ht^b\) on \(s,t\ge0\), \(s+t\le1\) lies on \(s+t=1\). Differentiating \(h\log s+b\log(1-s)\) gives the unique interior maximum \(s=h/(h+b)\), proving (67). If either exponent is zero, the corresponding bound is simply \(1\). ◻ For a nonzero scalar graph key \(t\) of degree two, let \[y=(|w|,|z|,|\ell|,|X|)^T\] and let \(i,j\in\{1,2,3,4\}\) be its two degree-one types, counted with multiplicity and ordered so \(i\le j\). Let \(h\) be its \(g\) exponent and \(b\) its number of \(p\) endpoints. Define \(L_t^2\) as a product of the following rational factors:
Here separate occurrences of a vector endpoint count as distinct vertices; thus a \(p\)–\(p\) edge is governed by Item [cert:graph-rule-edge], not Item [cert:graph-rule-self]. Power-zero self-edges at \(X\) have already vanished in the collection rules. Lemma 24 (Complete graph bound). Every scalar graph retained by Lemma 22 satisfies \[ |\mathcal M(t)|\le L_t y_i y_j \tag{68}\] on \(g^2+|p|^2\le1\). Proof. First separate all scalar trace factors and use Lemma 23. A cubic tensor has at most one self-edge, since such an edge consumes two of its three incidences. An edge of power \(k\ge2\) reduces that tensor to the vector \(v_i=X_{ijk}(A^k)_{jk}\), for which \(|v|\le|X|\,|A^k|\le|X|\theta^{(k-1)/2}\). For \(k=1\) use (66). These reductions leave at most two cubic vertices and no self-edge on either. We list all possible connected components of the remaining graph. With zero cubic vertices, a component is one edge joining two vectors; its contraction is bounded by the operator norm of its matrix times the two vector norms. With one cubic vertex, all its three slots end at vector leaves. Move each edge matrix onto its vector, and use tensor Cauchy–Schwarz, obtaining the norm of \(X\) times the three vector norms and the product of the three operator norms. If a component contains two cubic vertices, they are joined by \(r=1,2,\) or \(3\) edges; every other slot ends at a vector leaf. Contract these leaves first, obtaining rank-\(r\) tensors \(Y,Z\) with the same preceding norm bounds. Up to a permutation of their slots, the remaining contraction is \[\left\langle Y, (A^{k_1}\otimes\cdots\otimes A^{k_r})Z\right\rangle.\] Its absolute value is at most \(|Y||Z|\prod_{s=1}^r\|A^{k_s}\|_{\mathrm{op}}\). This includes all three parallel edges when \(r=3\), without a dimension factor. If the two cubic vertices have no edge between them, they lie in two components of the one-vertex type already treated. These cases are exhaustive: all other vertices have valence one, so cannot occur internally on a path, and there are no further cubic vertices. The reduced vectors from self-edges are ordinary vector leaves in this classification. Multiply the estimates over components. Each original \(\ell\) or \(X\) norm appears once per occurrence, all ordinary edge matrices contribute \(\theta^{k/2}\), and the \(p\) endpoints contribute \(|p|^b\). The scalar factors \(g^h\) and \(|p|^b\) are bounded jointly by (67). Finally the total degree is two, so all remaining \(w,z,\ell,X\) factors give exactly \(y_i y_j\). ◻ We can now bound every term left after subtracting the nonnegative comparison forms. The graph estimates reduce this remainder to a quadratic form in four magnitudes. The next calculation shows that its error is smaller than the reserved positive diagonal terms. The interior remainder and its coercive marginSet \[ d=\left(\frac2{25},\frac2{25},\frac14,1\right),\qquad E=\Gamma+\sum_{i=1}^4d_i y_i^2-\mathcal I. \tag{69}\] The squared norms in Equation (69) are polynomial contractions. Expand and collect \(E=\sum_t e_t\mathcal M(t)\) by Section 6.3, discarding zero coefficients only after their signed contributions have been summed. All coefficients are rational. The denominator \(Q^2\) suffices: Gram terms have that denominator, and the denominators of \(a,d_i\), and all Table I coefficients divide it after multiplication by \(Q^2\). For each retained key let \(n_t\) be the least nonnegative integer satisfying \[ n_t^2\ge10^{10}L_t^2, \qquad \widehat L_t=\frac{n_t}{10^5}. \tag{70}\] There is no floating-point square root in this prescription. For a rational \(s=10^{10}L_t^2\), compute \(n=\lfloor\sqrt{\lfloor s\rfloor}\rfloor\) by integer arithmetic and increase it by one exactly when \(n^2<s\). Form a symmetric matrix \(D_0\) by adding \[ \frac{|e_t|\widehat L_t}{1+\mathbf1_{\{i\ne j\}}} \tag{71}\] to its \(ij\) entry and also its \(ji\) entry if \(i\ne j\), starting with zero. The factor \(1/2\) in an off-diagonal entry is required because the associated term in \(y^TD_0y\) is \(2(D_0)_{ij}y_iy_j\). Lemma 24 and the triangle inequality therefore give \[ |E|\le y^TD_0y. \tag{72}\] For clarity, the exact finite calculation needed here is the following. Decode all rows of Tables I and II; expand the nineteen Gram blocks by unordered row pairs and the finite projections (59)–(61); subtract each Table I contribution to (40); and add the four contractions with coefficients \(d_i\). Use the contraction, derivative, and key ordering rules above after each operation. Combine all signed coefficients at equal keys. For each nonzero key compute its rational \(L_t^2\), perform (70), and add (71). This gives all ten upper-triangular entries as follows: \[
\begin{array}{c|r|r}
(i,j)&\text{nonzero keys}&2\cdot10^5Q^2(D_0)_{ij}\\\hline
(1,1)&89&321268243322644\\
(1,2)&53&96616799870252\\
(1,3)&123&294707669526536\\
(1,4)&170&385792585960684\\
(2,2)&64&269583307398596\\
(2,3)&85&115717471031956\\
(2,4)&106&125155603107368\\
(3,3)&238&605509501361296\\
(3,4)&414&618424270356936\\
(4,4)&372&1004504025266456
\end{array}
\tag{73}\] In particular there are \(1714\) nonzero keys. These finite sums use only integer and rational operations and comparisons; the complete literal input is in Appendix 9. The accompanying program Entrywise, (73) implies \(D_0\le\overline D\), where \[ \overline D=\frac1{10^5} \begin{pmatrix} 2571&773&2358&3087\\ 773&2157&926&1002\\ 2358&926&4845&4948\\ 3087&1002&4948&8037 \end{pmatrix}. \tag{74}\] For example, this entrywise check consists of comparing each last-column integer in (73) with the corresponding integer in (74) multiplied by \(2Q^2\). Let \(\nu=(2484,1440,996,304)^T\). The exact diagonal slacks are \[ \left(d_i-\frac{(\overline D\nu)_i}{\nu_i}\right)_{i=1}^4 =\left(\frac{673}{18400},\frac{438907}{12000000}, \frac{2844869}{24900000},\frac{3479389}{7600000}\right). \tag{75}\] For \(y_i\ge0\), weighted Young inequalities give \[2y_iy_j\le\frac{\nu_j}{\nu_i}y_i^2+ \frac{\nu_i}{\nu_j}y_j^2, \qquad y^TD_0y\le y^T\overline D y \le\sum_i\frac{(\overline D\nu)_i}{\nu_i}y_i^2.\] Use \(\Gamma\ge0\) in (69) and then (72). It follows that \(\mathcal I\) is at least the sum of \(y_i^2\) times the four positive slacks in (75). In particular the coefficients of \(z^2\) and \(|\ell|^2\) are at least \(\epsilon_*\), proving (40). To obtain the stated positivity of \(f\), keep \(A,p,g\) fixed, set \(X=w=z=0\), and choose \(|\ell|=1\). Then \(q=0\), every \(D\) term is zero, and \(\mathcal I=f\). Boundary reduction and the eleven coefficientsIt remains to verify (41). Orthogonal equivariance allows us to take \(p=e_5\) and to diagonalize \(A\) on \(p^\perp\). Write \[ A=\operatorname{diag}(x_1,x_2,x_3,x_4,\lambda),\qquad \lambda=-\sum_{i=1}^4x_i, \qquad R(x):=\sum_{i=1}^4x_i^2+\left(\sum_{i=1}^4x_i\right)^2=1. \tag{76}\] No sign or distinctness condition is imposed on the \(x_i\); the normal eigenvalue condition is exactly \(\sum_i x_i>0\). In these coordinates \((S_0)_{55}=0\) and \((S_0)_{\alpha\alpha}=b_\alpha:=\lambda x_\alpha-x_\alpha^2\). The only nonzero entries of \(V\) are \(V_{555}=1\) and the three permutations of \(V_{5\alpha\alpha}=b_\alpha\). Since \(C_{55jk}=0\), a direct component formula for the boundary polynomial is therefore \[ \mathcal S=-\lambda f+U_{55}+M_5+ \sum_{\alpha=1}^4 b_\alpha (U_{\alpha\alpha}+C_{5\alpha5\alpha} +C_{5\alpha\alpha5}), \tag{77}\] where the fixed polynomials are evaluated at (76), \(p=e_5,g=0\). This formula does not assume additional slot symmetries of \(U\) or \(K\). Equivalently, one may contract the boundary expression as a scalar graph and then substitute \[ p\cdot A^kp=\left(-\sum_i x_i\right)^k,\qquad \mathop{\mathrm{tr}}A^k=\sum_i x_i^k+\left(-\sum_i x_i\right)^k. \tag{78}\] For \(k=0\), the latter is \(5\) and the former is \(1\). The trace reductions \(\mathop{\mathrm{tr}}A=0\), \(\mathop{\mathrm{tr}}A^2=1\) remain valid on \(R=1\). Thus either (77) or (78) evaluates the same scalar on all the allowed data. Orthogonal invariance under permutations of \(p^\perp\) shows that \(\mathcal S\) is symmetric in \(x_1,\ldots,x_4\). Define the boundary remainder on this constraint set by \[E_b=\Lambda+\frac14-\mathcal S.\] Expanding the two boundary Gram blocks and the Table I expression gives a polynomial representative with monomials of degree at most five. A finite way to check the degree assertion is to use (77): discard positive \(g\) powers, substitute the diagonal matrix and normal vector in each of the \(231\) rows, and add the displayed terms. All remaining monomials have degree at most five. Each boundary \(J_c\) has degrees one and two, so its squared terms have degree at most four, or five after the \(\texttt{H}\) weight. The following convention gives a small exact list of its coefficients. For an ordinary monomial \(x^\alpha\) of degree \(j\le5\), replace it by \[ x^\alpha R^{\lfloor(5-j)/2\rfloor}. \tag{79}\] This preserves its value on \(R=1\) and gives degree four if \(j\) is even and degree five if \(j\) is odd. Perform the replacement before collecting exponents. Because \(R\) is symmetric, it commutes with permutation averaging. For a nondecreasing exponent vector \(\beta\) put \[m_\beta(x)=\operatorname{Av}(x^\beta).\] Expand (79) and sum the coefficients of all exponent vectors whose nondecreasing rearrangement is \(\beta\). Their sum is the coefficient of \(m_\beta\) after averaging: indeed \(\operatorname{Av}(x^\alpha)=m_\beta\) for every member \(\alpha\) of that orbit, including when exponents repeat. Thus sorting exponents and summing coefficients introduces no extra factor of \(24\). Since \(\mathcal S\) and the constant are symmetric and \(\Lambda\) was defined by averaging, the resulting averaged polynomial still equals \(E_b\) on \(R=1\). Precisely, using the row-pair expansion of Section 6.3 for the two rank-\(-1\) blocks, subtracting the boundary contributions from Table I with denominator \(Q\), and adding the homogenized constant \(1/4\), one obtains \[ E_b(x)=\sum_\beta e^{(b)}_\beta m_\beta(x)\quad\text{on }R=1, \tag{80}\] with exactly the following eleven nonzero coefficients: \[ \begin{array}{c|r@{\qquad}c|r} \beta&Q^2 e^{(b)}_\beta&\beta&Q^2 e^{(b)}_\beta\\\hline (0,0,0,4)&-13479110 &(0,0,0,5)&17712712\\ (0,0,1,3)&-93201124 &(0,0,1,4)&151038620\\ (0,0,2,2)&-78543848 &(0,0,2,3)&244636976\\ (0,1,1,2)&-139986162&(0,1,1,3)&310091962\\ (0,1,2,2)&445177656 &(1,1,1,1)&-4861942\\ (1,1,1,2)&157995170&& \end{array} \tag{81}\] This coefficient computation consists solely of the explicitly stated substitutions, multiplication of ordinary four-variable polynomials, and coefficient addition. It is also included in the accompanying exact reproducer. On \(R=1\) every \(|x_i|\le1\), hence \(|m_\beta(x)|\le1\). Adding the eleven absolute integer coefficients gives \[Q^2\sum_\beta|e^{(b)}_\beta|=1656725282.\] Since \(\sum_i x_i=-\lambda>0\), (58) is nonnegative. It follows from (80) that \[\mathcal S=\Lambda+\frac14-E_b \ge\frac14-\frac{1656725282}{62500000000} =\frac{6984137359}{31250000000}=\delta_*.\] This completes the proof of Proposition 15. The estimate also persists at \(\lambda=0\) by continuity; repeated tangent eigenvalues cause no singularity in any of the formulas. The pointwise comparisons and field bounds of Section [sec:flux] are therefore established. Section [sec:cutoffs] now justifies their integration across Hessian zeros and near the suspension tips. Hessian zeros, suspension tips, and the contradiction
We now justify the integration of the pointwise comparison. Retain the original smooth component \(\Omega_o\subset S^{m_o}\), with \(m_o\in\{4,5\}\), from Proposition 6 and Lemma 7. Its tensor \(T_o\) is not identically zero. Write \[B_o=\nabla T_o,\qquad w_o=|T_o|,\qquad \mathcal Z_o=\{y\in\overline\Omega_o:w_o(y)=0\},\qquad \Omega_{o,+}=\Omega_o\setminus\mathcal Z_o.\] The tensor \(T_o\) is smooth on the compact closure; its norm \(w_o\) is continuous there and smooth where it is positive. The working domain \(\Omega\subset S^5\) is \(\Omega_o\) when \(m_o=5\), and its suspension when \(m_o=4\). On the regular part of this working domain we use the unadorned notation of Sections [sec:geometry] and [sec:flux]. In particular, \(Z\) denotes the flux vector field, whereas \(\mathcal Z_o\) denotes the original Hessian zero set. Set \(M_o=\|w_o\|_{L^\infty(\Omega_o)}\). We fix \[a=\frac{29}{50},\qquad \epsilon_* =\frac{438907}{12000000}>0.\] Interior and boundary integrals use their respective spherical volume and area measures; in the suspension, boundary integrals refer to its regular side boundary. Weighted integrability on the original componentLemma 25 (Weighted Bochner estimate). The set \(\mathcal Z_o\) is a compact subset of \(\Omega_o\), and \[ \int_{\Omega_{o,+}}w_o^{2a-2}|B_o|^2\,dV_o<\infty. \tag{82}\] Consequently the function \[ D_o(y)= \begin{cases} w_o(y)^{2a}\bigl(1+|B_o(y)|/w_o(y)\bigr)^2, &y\in\Omega_{o,+},\\ 0,&y\in\mathcal Z_o, \end{cases} \tag{83}\] belongs to \(L^1(\Omega_o)\). Proof. The boundary conclusion of Lemma 10 gives \(w_o>0\) at every point of \(\partial\Omega_o\). Compactness and continuity therefore give a boundary neighborhood on which \(w_o\) has a positive lower bound. The closed set \(\mathcal Z_o\) is thus compactly contained in the interior. If it is empty, \(w_o\) has a positive minimum on the whole compact closure, and (82) follows immediately from the boundedness of \(B_o\). Otherwise choose \(\eta\in C_c^\infty(\Omega_o)\) with \(0\leq\eta\leq1\), equal to one on an open neighborhood of \(\mathcal Z_o\). For \(0<\delta\leq1\) put \(W_\delta=w_o^2+\delta\). Although \(w_o\) need not be smooth at its zeros, \(W_\delta\) is smooth and strictly positive. The identities in Proposition 9 give \[\Delta W_\delta=2m_o w_o^2+2|B_o|^2\geq2|B_o|^2, \qquad |\nabla W_\delta|^2 =4|\langle T_o,B_o\rangle|^2 \leq4W_\delta|B_o|^2.\] Set \(I_\delta=\int_{\Omega_o}\eta W_\delta^{a-1}|B_o|^2\,dV_o\). Applying the chain rule to \(W_\delta^a\) and integrating twice by parts against the compactly supported \(\eta\) yields \[\begin{align*} 2I_\delta &\leq \int_{\Omega_o}\eta W_\delta^{a-1}\Delta W_\delta\,dV_o\\ &=\frac1a\int_{\Omega_o}(\Delta\eta)W_\delta^a\,dV_o +(1-a)\int_{\Omega_o} \eta W_\delta^{a-2}|\nabla W_\delta|^2\,dV_o\\ &\leq\frac1a\int_{\Omega_o}(\Delta\eta)W_\delta^a\,dV_o +4(1-a)I_\delta. \end{align*}\] Since \(2-4(1-a)=4a-2=8/25\), this proves the explicit uniform estimate \[ \frac8{25}I_\delta \leq \frac1a(M_o^2+1)^a\|\Delta\eta\|_{L^1(\Omega_o)}. \tag{84}\] As \(\delta\downarrow0\), the nonnegative integrands restricted to \(\Omega_{o,+}\) increase to \(\eta w_o^{2a-2}|B_o|^2\) because \(a<1\). Monotone convergence and (84) control that integral. The compact set \(\mathop{\mathrm{supp}}(1-\eta)\subset\overline\Omega_o\) is disjoint from \(\mathcal Z_o\); writing \(b_\eta>0\) for the minimum of \(w_o\) there, the remaining integral is bounded by \[\int_{\Omega_{o,+}}(1-\eta)w_o^{2a-2}|B_o|^2\,dV_o \leq b_\eta^{2a-2}\|B_o\|_{L^2(\Omega_o)}^2.\] Thus all constants in this argument depend only on the fixed original component and its smooth data, on \(a\), and on the chosen interior cutoff \(\eta\), and are independent of \(\delta\). Finally, on \(\Omega_{o,+}\), \[D_o\leq2w_o^{2a}+2w_o^{2a-2}|B_o|^2.\] The first term is integrable by compactness, and the second by (82). No measure, dimension, or regularity property of \(\mathcal Z_o\) has been used. ◻ Smooth angular cutoffs and the integrated comparisonLet \(\vartheta\in C^\infty([0,\infty))\) satisfy \(0\leq\vartheta\leq1\), \(\vartheta=0\) on \([0,1]\), and \(\vartheta=1\) on \([2,\infty)\). Fix this profile once and for all, and set \[L_\vartheta=\|\vartheta'\|_\infty, \qquad \kappa=2L_\vartheta.\] For integers \(j\geq3\) define \[ \alpha_j(y)=\vartheta(jw_o(y)),\qquad \xi_j= \begin{cases} \alpha_j,&m_o=5,\\ \alpha_j(y)\beta_j(t),\quad \beta_j(t)=\vartheta(j\sin t),&m_o=4. \end{cases} \tag{85}\] Here \(\alpha_j\) is defined to be zero at \(\mathcal Z_o\), consistently with the displayed formula. It is smooth on \(\overline\Omega_o\): near any zero of \(w_o\) it is identically zero, and all its transitions occur where \(w_o\) is smooth and positive. In the suspension, \(\beta_j\) vanishes on whole neighborhoods of both tips. Figure 1 displays the two exclusions made by the product cutoff in the suspended case. Its transition bands shrink toward the two tips and the original Hessian zero set. The weighted estimate above will control the errors produced in all these bands at once. On \(\Omega_+=\{x\in\Omega:w(x)>0\}\), use the fields \[h=w^a\sqrt{f(A,p,g)},\qquad Z_l=w^{2a}(G_{lijk}X_{ijk}+wM_l+P_l)\] constructed in Section [sec:flux]. By Lemma 20, there is a fixed \(C_F\geq1\) such that \[ |h|\leq C_Fw^a,\qquad |\nabla h|\leq C_Fw^a(1+w+|X|),\qquad |Z|\leq C_Fw^{2a}(1+w+|X|). \tag{86}\] The constant depends only on \(a\), the dimension \(5\), the finitely many coefficients defining \(f,G,M,P\), and the positive lower bound \(f\geq\epsilon_*\). In particular \(C_F\) is independent of the cone, \(w\), \(X\), and \(j\). For each fixed \(j\), define \(v_j=\xi_jh\) and \(F_j=\xi_j^2Z\) on \(\Omega_+\), and extend them by zero where \(w=0\). These are smooth on the regular closure of \(\Omega\) and are supported away from the tips when present. The same neighborhood argument as for \(\alpha_j\) proves smoothness of these extensions; it does not require \(h\) or \(Z\) themselves to extend smoothly across \(w=0\). All products below containing one of these cutoff factors or its derivative are likewise understood to be zero on \(w=0\). The mass \(w^{2a}\), whose exponent is positive, has its ordinary value zero there and involves no normalized tensor. Lemma 26 (Integrated comparison). For each fixed \(j\geq3\), \[ \epsilon_*\int_\Omega\xi_j^2w^{2a}\,dV \leq\int_\Omega\left( h^2|\nabla\xi_j|^2 +2\xi_jh\langle\nabla h,\nabla\xi_j\rangle -2\xi_j\langle Z,\nabla\xi_j\rangle\right)\,dV. \tag{87}\] Proof. The interior comparison of Proposition 17 gives \[\int_\Omega\xi_j^2 (4h^2+|\nabla h|^2+\epsilon_*w^{2a})\,dV \leq\int_\Omega\xi_j^2\mathop{\mathrm{div}}Z\,dV.\] Apply the divergence theorem to the smooth field \(F_j\). In the suspension one may first truncate the domain inside the two neighborhoods where \(F_j\) vanishes; the artificial end boundaries have zero flux. On the regular side boundary the outward normal is \(-N\). Thus the boundary comparison of Proposition 19, \(N\cdot Z+Hh^2\geq0\), gives \[\begin{align*} \int_\Omega\xi_j^2\mathop{\mathrm{div}}Z\,dV &=-\int_{\partial\Omega}\xi_j^2\langle N,Z\rangle\,d\sigma -2\int_\Omega\xi_j\langle Z,\nabla\xi_j\rangle\,dV\\ &\leq\int_{\partial\Omega}H(\xi_jh)^2\,d\sigma -2\int_\Omega\xi_j\langle Z,\nabla\xi_j\rangle\,dV. \end{align*}\] Corollary 14 applies to \(v_j\) and bounds the boundary integral by \[\int_\Omega\bigl(|\nabla(\xi_jh)|^2+4\xi_j^2h^2\bigr)\,dV.\] Expanding the square and canceling the common terms gives (87). ◻ Every cancellation above is made before removing the angular cutoff. For fixed \(j\) and radial index \(k\), the tests \(\zeta_k(r)v_j\) from Corollary 14 have compact support on the regular Euclidean cone. First take the variation parameter \(\varepsilon\) to zero on this fixed support; then let \(k\to\infty\) in Equation (37) to obtain the constant \(4\). The smooth extension collar and the allowed variation size may depend on \(j\) and \(k\). We now take \(j\to\infty\) only in Equation (87), which contains the positive mass and three cutoff errors. The energy and boundary terms have already canceled for each fixed \(j\). Vanishing of all three errorsOn \(\Omega_{o,+}\), differentiation of \(\alpha_j\) and \(|\nabla w_o|\leq|B_o|\) give \[ |\nabla_o\alpha_j| \leq\kappa\frac{|B_o|}{w_o}. \tag{88}\] To see the constant, wherever \(\vartheta'(jw_o)\ne0\) one has \(1<jw_o<2\), so \(j\leq2/w_o\); outside this transition region the derivative is zero. The same argument gives \[ |\beta_j'(t)|\leq\frac{\kappa}{\sin t}. \tag{89}\] At every fixed point of \(\Omega_{o,+}\), \(\alpha_j=1\) and \(\nabla_o\alpha_j=0\) for all sufficiently large \(j\). At every fixed \(t\in(0,\pi)\) the corresponding statements hold for \(\beta_j\). Write \(b_o=|B_o|/w_o\) on \(\Omega_{o,+}\). Let \(E_{1,j}\), \(E_{2,j}\), and \(E_{3,j}\) denote, in order, the three signed integrands on the right of (87). First suppose \(m_o=5\), so \(\Omega=\Omega_o\), and set \(K_n=1+M_o\). Since \(X=B_o/w_o\), \[1+w+|X|\leq K_n(1+b_o),\qquad |\nabla\xi_j|\leq\kappa(1+b_o).\] Equations (86) and the preceding bounds give \[ |E_{1,j}|+|E_{2,j}|+|E_{3,j}| \leq C_F^2(\kappa^2+4K_n\kappa)D_o. \tag{90}\] Indeed the three coefficient bounds are respectively \(C_F^2\kappa^2\), \(2C_F^2K_n\kappa\), and \(2C_FK_n\kappa\), and \(C_F\geq1\). The dominator is integrable by Lemma 25. Each error is eventually zero at every point of \(\Omega_{o,+}\), and is defined to be zero on \(\mathcal Z_o\). Dominated convergence therefore makes the right side of (87) tend to zero. Now suppose \(m_o=4\). Write \(s_t=\sin t\) and use the suspension coordinates \((t,y)\in(0,\pi)\times\Omega_o\). The exact formulas in Proposition 12 give \[ w(t,y)=\frac{w_o(y)}{s_t},\qquad |B(t,y)|^2 =\frac{|B_o(y)|^2+3\cos^2(t)w_o(y)^2}{s_t^4},\qquad dV=s_t^4\,dt\,dV_o. \tag{91}\] In particular \(\Omega_+=(0,\pi)\times\Omega_{o,+}\), and there \[|X|=\frac{\sqrt{b_o^2+3\cos^2t}}{s_t} \leq\frac{\sqrt3}{s_t}(1+b_o).\] The warped metric gives the exact cutoff identity \[|\nabla\xi_j|^2 =\alpha_j^2|\beta_j'|^2 +s_t^{-2}\beta_j^2|\nabla_o\alpha_j|^2,\] so (88)–(89) imply \[ |\nabla\xi_j|\leq\frac{\kappa}{s_t}(1+b_o),\qquad 1+w+|X|\leq\frac{K_s}{s_t}(1+b_o),\qquad K_s=1+M_o+\sqrt3. \tag{92}\] Multiplying the bounds for the three errors by the volume factor in (91) now gives \[ s_t^4\bigl(|E_{1,j}|+|E_{2,j}|+|E_{3,j}|\bigr) \leq C_F^2(\kappa^2+4K_s\kappa) s_t^{2-2a}D_o(y). \tag{93}\] This inequality is initially read on \(\Omega_{o,+}\), and both sides are assigned zero on \(\mathcal Z_o\). Its constant depends only on the fixed certificate, \(a\), \(\vartheta\), and \(M_o\), and is independent of \(j\), \(t\), and \(y\). It also shows explicitly why the original weighted \(|B_o|^2\) estimate suffices: by (83), \[s_t^{2-2a}D_o(y) \leq 2s_t^{21/25} \bigl(w_o^{29/25}+w_o^{-21/25}|B_o|^2\bigr) \quad\hbox{on }\Omega_{o,+}.\] The right side is integrable with respect to \(dt\,dV_o\), because \(21/25>-1\) and (82) holds. Every error is eventually zero at each fixed \((t,y)\in(0,\pi)\times\Omega_{o,+}\). Dominated convergence applied to (93) therefore makes the right side of (87) tend to zero also in the suspension. Theorem 27. The first dimension admitting a nonflat one-homogeneous global minimizer cannot be \(5\) or \(6\). Proof. If the first dimension were \(5\) or \(6\), the reductions would provide exactly one of the two working domains just considered. In either case, the right side of (87) tends to zero. On the original component, however, \[0<I_o:=\int_{\Omega_o}w_o^{2a}\,dV_o<\infty.\] Finiteness follows from boundedness and compactness. Positivity follows from \(T_o\not\equiv0\) and continuity, which give an open subset of \(\Omega_o\) on which \(w_o>0\). In the nonsuspended case dominated convergence gives \[\lim_{j\to\infty}\int_\Omega\xi_j^2w^{2a}\,dV=I_o.\] In the suspended case (91) gives instead \[\lim_{j\to\infty}\int_\Omega\xi_j^2w^{2a}\,dV = I_o\int_0^\pi(\sin t)^{4-2a}\,dt = I_o\int_0^\pi(\sin t)^{71/25}\,dt>0.\] Here the dominating mass density is the integrable, everywhere defined function \((\sin t)^{71/25}w_o^{29/25}\); its value is zero wherever \(w_o=0\). These strictly positive limits contradict (87), since \(\epsilon_*>0\). ◻ The known lower-dimensional classification gives \(d_*\geq5\), and the dimension-seven globally minimizing example gives \(d_*\leq7\), as stated in Section 1. Theorem 27 excludes the only two remaining possibilities below \(7\). Hence \(d_*=7\), completing the proof of Theorem 1. Regularity and sharpness of the dimension boundWe now pass from the classification of homogeneous global minimizers to local free boundaries. This step uses the dimension-reduction theorem of Weiss [18], in the formulation [17]. Proof of Corollary 2. Apply the cited dimension-reduction theorem with volume coefficient one on balls compactly contained in \(D\). Its inputs are uniform flatness regularity, linear nondegeneracy, compactness of blowups, and the fact that all blowup limits are homogeneous global minimizers. These are the constant-coefficient results used in Proposition 3; flatness regularity is [17]. The theorem gives an empty singular set below \(d_*\), local finiteness in dimension \(d_*\), and Hausdorff dimension at most \(n-d_*\) above it. Theorem 1 sets \(d_*=7\). Smoothness at each regular point follows from Proposition 3(iii). A countable cover by interior balls gives the assertions on \(D\). For sharpness, let \(U\) be the nonflat global minimizer in \(\mathbb R^7\) of De Silva and Jerison [7], and set \(U_n(x,y)=U(x)\) on \(\mathbb R^7\times\mathbb R^{n-7}\). Repeated application of Proposition 8 shows that \(U_n\) is globally minimizing. At each point \((0,y_0)\), every rescaling of \(U_n\) is \(U_n\) itself, by homogeneity and translation invariance in \(y\). It cannot be flat: a half-space normal would have zero \(y\)-component, forcing \(U\) to be flat. Thus \(\{0\}\times\mathbb R^{n-7}\subset\operatorname{Sing}(U_n)\). Its dimension is \(n-7\), and the proved upper bound gives equality. ◻ Complete coefficient tablesThe following two tables specify the polynomials used in Section 6.1.1 and the comparison forms in Section 6.2. Every coefficient printed below is an integer numerator with denominator \[Q=250000.\] Leading zeroes in monomial codes are significant. All rows and all block headers are included. Decoding and block conventionsFor nonnegative rank, read a monomial code in consecutive triples. Its first triple \(h\,b\,j\) contributes \(g^h(\mathop{\mathrm{tr}}A^3)^b w^{[j=1]}z^{[j=2]}\). Each subsequent triple \(u\,v\,k\) is an unoriented edge carrying \(A^k\), with \(A^0=\mathop{\mathrm{Id}}\). Endpoints \(0,\ldots,r-1\) denote the free slots in order, \(5\) denotes a copy of \(p\), and \(6\) a copy of \(\ell\); three occurrences of \(7\) are contracted with one \(X\), and three occurrences of \(8\) with a second copy. Multiply all factors and contract all nonfree slots. There are no additional multiplicity factors; see Section 6.1. In Table I the first field specifies the polynomial, the second its monomial, and the third its coefficient numerator. The free-slot orders are \(U_{jk}\), \(K_{lijk}\), \(M_l\), and \(P_l\), while \(f\) has rank zero. The numbers of rows for \((f,U,M,P,K)\) are \((34,57,74,32,34)\), for a total of \(231\). Each line of Table II beginning with For rank \(-1\), the four code digits instead give the four individual exponents of \(x_1,x_2,x_3,x_4\). The block contributes to \(\Lambda\) the average over all \(24\) permutations of \(\rho\sum_c J_c^2\). Here Table I: the multiplier coefficientsThe columns below read from top to bottom, first on the left and then on the right. 2
Table II: the comparison coefficientsThe block headers are part of the data.
Reproduction of the finite arithmeticIn the complete source directory accompanying this article, run
|
| ||||||||
|