Introduction
A Ricci soliton on a smooth manifold is a pair ( g , X ) (g, X) ( g , X ) consisting of a Riemannian metric g g g and a vector field X X X satisfying
Ric g + 1 2 L X g = λ g , (1) \operatorname{Ric}_g + \frac{1}{2}\mathcal{L}_X g = \lambda g, \tag*{(1)} Ric g + 2 1 L X g = λ g , (1)
for some constant λ ∈ R \lambda\in\mathbb{R} λ ∈ R . Under Ricci flow, Ricci solitons evolve only by scaling and pullback by diffeomorphisms and thus provide models for singularities of the Ricci flow. A Ricci soliton is called shrinking , steady , or expanding if the constant λ \lambda λ in (1) is positive, zero, or negative. If X X X is a gradient field of some smooth scalar function u u u , then the Ricci soliton equation (1) becomes
Ric g + Hess g u = λ g . (2) \operatorname{Ric}_g + \operatorname{Hess}_g u = \lambda g. \tag*{(2)} Ric g + Hess g u = λ g . (2)
In this case, we call the metric g g g a gradient Ricci soliton with soliton potential u u u . On a compact Ricci soliton, X X X is gradient up to addition of a Killing field. If the Ricci soliton is non-Einstein, it is shrinking [21 §§2–3] .
Many examples of compact non-Einstein Ricci solitons have been constructed using Kähler metrics [17 , 7 , 10 ] . Tian and Cao asked whether there exists a compact Ricci soliton that is nowhere locally Kähler, non-Einstein, and locally irreducible [11 Question 9.1] . Here nowhere locally Kähler means that no nonempty open subset carries a compatible parallel complex structure, and *locally irreducible means that no nonempty open subset is isometric to a product of two positive-dimensional Riemannian manifolds. Let T 2 ⊂ S U ( 3 ) T^2 \subset\mathrm{SU}(3) T 2 ⊂ SU ( 3 ) act on C P 1 \mathbb{CP}^1 CP 1 by diag ( z 1 , z 2 , z 3 ) ⋅ [ a : b ] = [ a : ( z 3 / z 1 ) b ] \operatorname{diag}(z_1,z_2,z_3) \cdot[a:b] = [a:(z_3/z_1)b] diag ( z 1 , z 2 , z 3 ) ⋅ [ a : b ] = [ a : ( z 3 / z 1 ) b ] , and define
M = S U ( 3 ) × T 2 C P 1 . (3) M = \mathrm{SU}(3) \times_{T^2} \mathbb{CP}^1. \tag*{(3)} M = SU ( 3 ) × T 2 CP 1 . (3)
Main Theorem. The compact simply-connected manifold M M M admits a smooth gradient shrinking Ricci soliton whose metric is nowhere locally Kähler, non-Einstein, and locally irreducible.
The manifold M M M already carries a Kähler shrinking Ricci soliton from the construction of Dancer and Wang [10 ] . Their work extends the projective-line bundle constructions of Koiso [17 ] and Cao [7 ] to bundles over products of Fano Kähler–Einstein manifolds and generalised flag manifolds. We construct a non-Kähler shrinking soliton on M M M by solving (2) for metrics and potentials invariant under the cohomogeneity one action of S U ( 3 ) × S 1 \mathrm{SU}(3) \times S^1 SU ( 3 ) × S 1 on M M M . This reduces the construction to a shooting problem for a system of ordinary differential equations, with smoothness conditions at the two singular orbits.
Böhm [2 ] used this approach to construct infinitely many inhomogeneous Einstein metrics on S n S^n S n for n = 5 , … , 9 n = 5,\ldots,9 n = 5 , … , 9 and on certain products of spheres with homogeneous Einstein manifolds. Nienhaus and Wink [20 ] later used shooting to construct three non-round, pairwise non-isometric Einstein metrics on S 10 S^{10} S 10 . More recently, Buttsworth and Hodgkinson [5 ] obtained a non-round Einstein metric on S 12 S^{12} S 12 using a computer-assisted proof.
Dancer, Hall and Wang [9 §5] numerically studied compact shrinking solitons among cohomogeneity one metrics determined by two functions. Within the families and parameter ranges tested, their shooting searches recovered the Koiso–Cao soliton and known Einstein metrics.
On noncompact manifolds, Wink [26 , 25 ] constructed complete steady and expanding solitons. Donovan [12 ] constructed complete non-Einstein asymptotically conical expanding Ricci solitons in dimension four with S U ( 2 ) \mathrm{SU}(2) SU ( 2 ) symmetry. These constructions require completeness at an infinite end, whereas our compact problem requires a second smooth collapse at finite distance.
The manifold M M M is an S 2 S^2 S 2 bundle over the six-dimensional full flag manifold B = S U ( 3 ) / T 2 B = \mathrm{SU}(3)/T^2 B = SU ( 3 ) / T 2 . The S U ( 3 ) × S 1 \mathrm{SU}(3) \times S^1 SU ( 3 ) × S 1 action has orbit space [ 0 , T ] [0,T] [ 0 , T ] , and its principal orbits are circle bundles over B B B . The circle fibres collapse at the two singular orbits, each of which is diffeomorphic to B B B . The isotropy representation of a principal orbit decomposes into four pairwise inequivalent irreducible summands: the one-dimensional tangent space to the circle fibre and three two-dimensional horizontal summands. An invariant metric is therefore determined by four functions on [ 0 , T ] [0,T] [ 0 , T ] .
For the known Kähler shrinker on M M M [10 ] , the horizontal metric is a multiple of a fixed Kähler–Einstein metric on B B B . We allow the relative scales of the horizontal summands to vary. The proof uses a subfamily preserved by the soliton equations in which three functions determine the metric (see § 2 ).
Smooth local solutions in this subfamily are determined by three parameters at each singular orbit (see § 3 ). In a two-ended shooting method, the two local solutions are integrated towards a common principal orbit. Fixing the integration distance from one end leaves the six endpoint parameters and the integration distance from the other to vary. The matching conditions require agreement of the orbit metrics and their first normal derivatives, together with the normal derivatives of the soliton potentials. These seven conditions define a matching map whose zeros give smooth solitons on M M M (see § 4 ).
We first find an approximate solution by numerical shooting. Convergent series at the singular orbits and interval integration give rigorous bounds on the matching map and its derivatives near the numerical candidate. We use these bounds in a Newton–Kantorovich argument to prove that an exact zero exists nearby. The verification procedure and code are described in Appendix A.
Computer-assisted fixed-point arguments have been used to construct Einstein metrics [5 , 23 ] , including the Kähler–Einstein metric on the third del Pezzo surface [4 ] . For a one-parameter shooting problem, Wang [Wan26b] used validated trajectories and an intermediate-value argument. Cao-Labora and Rodríguez-Vázquez [CLRV26] also use a finite-dimensional matching argument in their Einstein construction.
Estimates on the potential and curvature show that the resulting metric is non-Einstein and obstruct compatible parallel complex structures and local product decompositions.
AI disclosure. The construction, proof arguments and computational verification code were initially generated by ChatGPT Sol 6.1 on 30 October 2026 in response to a prompt supplied by the second-named author. AI assistance was also used to draft and revise the manuscript and to check the proof arguments. The authors have checked the mathematical arguments and reviewed the verification code to confirm that it implements the estimates used in the proof. A complete replay of the verification program passed all computational checks. The source code, recorded outputs and reproduction instructions are supplied with the manuscript. The authors take full responsibility for the manuscript and the computer-assisted proof.
Acknowledgements. The second-named author gratefully acknowledges access to OpenAI’s ChatGPT for Academic Researchers program through a research workspace awarded to Kyle Broder. The second-named author was supported by an Australian Government Research Training Program Scholarship.
Geometric preliminaries
We construct the manifold from its isotropy groups and normal discs, following the cohomogeneity one description of Grove–Ziller [15 §1] and Hoelscher [16 §1] . We then describe the invariant metrics and compute their Ricci curvature from the homogeneous-space formula in [1 Corollary 7.38] . A similar cohomogeneity one setup was used by the first-named author to construct non-compact gradient steady Ricci solitons in [But25].
The group diagram and the two disc bundles
Let G G G be a compact Lie group acting on a connected smooth manifold M M M with cohomogeneity one. If the orbit space is a closed interval, the action is encoded by a group diagram
G ↑ ↑ K − K + ↑ ↑ H (4) \begin{array}{c}
G\\
\uparrow\ \ \uparrow\\
K_{-}\ \ \ K_{+}\\
\uparrow\ \ \uparrow\\
H \tag*{(4)}
\end{array} G ↑ ↑ K − K + ↑ ↑ H (4)
or, more briefly, by
H ⊂ K − , K + ⊂ G . H \subset K_{-},K_{+} \subset G. H ⊂ K − , K + ⊂ G .
Here H H H is the principal isotropy subgroup and K ± K_{\pm} K ± are the isotropy subgroups of the two singular orbits. The principal and singular orbits are, respectively,
P = G / H , B ± = G / K ± . P = G/H,\qquad B_{\pm} = G/K_{\pm}. P = G / H , B ± = G / K ± .
For a point on B ± B_{\pm} B ± , the slice representation of K ± K_{\pm} K ± is transitive on the unit sphere in the normal space, with stabiliser H H H . Thus K ± / H ≅ S ℓ ± K_{\pm}/H \cong S^{\ell_{\pm}} K ± / H ≅ S ℓ ± , where ℓ ± + 1 \ell_{\pm}+1 ℓ ± + 1 is the codimension of B ± B_{\pm} B ± . The slice theorem then gives the equivariant decomposition
M = ( G × K − D − ) ∪ G / H ( G × K + D + ) , (5) M = (G \times_{K_{-}} D_{-}) \mathop{\cup}_{G/H} (G \times_{K_{+}} D_{+}), \tag*{(5)} M = ( G × K − D − ) ∪ G / H ( G × K + D + ) , (5)
where D ± D_{\pm} D ± is the normal disc with its slice action. Conversely, any such group diagram determines a cohomogeneity one manifold by this construction.
For the manifold in Main Theorem, we take
G = S U ( 3 ) × S 1 , K − = K + = K = T 2 × S 1 , G = \mathrm{SU}(3) \times S^{1},\qquad K_{-}=K_{+}=K=T^{2}\times S^{1}, G = SU ( 3 ) × S 1 , K − = K + = K = T 2 × S 1 ,
where T 2 T^{2} T 2 is the diagonal maximal torus in S U ( 3 ) \mathrm{SU}(3) SU ( 3 ) , and
H = { ( diag ( z 1 , z 2 , z 3 ) , z 1 z 3 ) : z i ∈ S 1 , z 1 z 2 z 3 = 1 } . H = \left\{\left(\operatorname{diag}(z_{1},z_{2},z_{3}),\frac{z_{1}}{z_{3}}\right): z_{i}\in S^{1},\ z_{1}z_{2}z_{3}=1\right\}. H = { ( diag ( z 1 , z 2 , z 3 ) , z 3 z 1 ) : z i ∈ S 1 , z 1 z 2 z 3 = 1 } .
Thus H ≅ T 2 H\cong T^{2} H ≅ T 2 , while K / H ≅ S 1 K/H\cong S^{1} K / H ≅ S 1 . Both singular orbits are the full flag manifold
G / K ≅ S U ( 3 ) / T 2 . G/K \cong\mathrm{SU}(3)/T^{2}. G / K ≅ SU ( 3 ) / T 2 .
Invariant metrics and the reductive decomposition
Write
g = s u ( 3 ) ⊕ R Z . \mathfrak{g}=\mathfrak{su}(3)\oplus\mathbb{R}Z. g = su ( 3 ) ⊕ R Z .
where Z Z Z generates the S 1 S^{1} S 1 -factor with period 2 π 2\pi 2 π . Let B s u ( 3 ) B_{\mathfrak{su}(3)} B su ( 3 ) be the Killing form of s u ( 3 ) \mathfrak{su}(3) su ( 3 ) . We use the bi-invariant inner product Q Q Q on g \mathfrak{g} g determined by
Q ∣ s u ( 3 ) = − B s u ( 3 ) , Q ( Z , Z ) = 1 , Q ( Z , s u ( 3 ) ) = 0. (6) Q|_{\mathfrak{su}(3)}=-B_{\mathfrak{su}(3)},\qquad Q(Z,Z)=1,\qquad Q(Z,\mathfrak{su}(3))=0. \tag*{(6)} Q ∣ su ( 3 ) = − B su ( 3 ) , Q ( Z , Z ) = 1 , Q ( Z , su ( 3 )) = 0. (6)
If E i j E_{ij} E ij denotes the elementary matrix, set
X i j = 1 12 ( E i j − E j i ) , Y i j = i 12 ( E i j + E j i ) , i < j . X_{ij}=\frac{1}{\sqrt{12}}(E_{ij}-E_{ji}),\qquad Y_{ij}=\frac{\mathrm{i}}{\sqrt{12}}(E_{ij}+E_{ji}),\qquad i<j. X ij = 12 1 ( E ij − E j i ) , Y ij = 12 i ( E ij + E j i ) , i < j .
These six vectors are Q Q Q -orthonormal. Together with Z Z Z and any Q Q Q -orthonormal basis of the diagonal Cartan algebra t \mathfrak{t} t , they form a Q Q Q -orthonormal basis of g \mathfrak{g} g .
Define χ : T 2 → S 1 \chi:T^{2}\to S^{1} χ : T 2 → S 1 by
χ ( diag ( z 1 , z 2 , z 3 ) ) = z 3 z 1 , (7) \chi(\operatorname{diag}(z_{1},z_{2},z_{3}))=\frac{z_{3}}{z_{1}}, \tag*{(7)} χ ( diag ( z 1 , z 2 , z 3 )) = z 1 z 3 , (7)
and let γ : t → R \gamma:\mathfrak{t}\to\mathbb{R} γ : t → R be determined by χ ( exp A ) = exp ( i γ ( A ) ) \chi(\exp A)=\exp(\mathrm{i}\gamma(A)) χ ( exp A ) = exp ( i γ ( A )) . Then
h = { A − γ ( A ) Z : A ∈ t } . \mathfrak{h}=\{A-\gamma(A)Z:A\in\mathfrak{t}\}. h = { A − γ ( A ) Z : A ∈ t } .
We choose the following Ad ( H ) \operatorname{Ad}(H) Ad ( H ) -invariant reductive complement:
g = h ⊕ m , m = m 0 ⊕ m 1 ⊕ m 2 ⊕ m 3 , \mathfrak{g}=\mathfrak{h}\oplus\mathfrak{m},\qquad\mathfrak{m}=\mathfrak{m}_{0}\oplus\mathfrak{m}_{1}\oplus\mathfrak{m}_{2}\oplus\mathfrak{m}_{3}, g = h ⊕ m , m = m 0 ⊕ m 1 ⊕ m 2 ⊕ m 3 ,
m 0 = R Z , \mathfrak{m}_{0}=\mathbb{R}Z, m 0 = R Z ,
m 1 = span { X 13 , Y 13 } , \mathfrak{m}_{1}=\operatorname{span}\{X_{13},Y_{13}\}, m 1 = span { X 13 , Y 13 } ,
m 2 = span { X 12 , Y 12 } , \mathfrak{m}_{2}=\operatorname{span}\{X_{12},Y_{12}\}, m 2 = span { X 12 , Y 12 } ,
m 3 = span { X 23 , Y 23 } . (8) \mathfrak{m}_{3}=\operatorname{span}\{X_{23},Y_{23}\}. \tag*{(8)} m 3 = span { X 23 , Y 23 } . (8)
The four summands in (8) are irreducible and pairwise inequivalent real Ad ( H ) \operatorname{Ad}(H) Ad ( H ) -modules. The tangent space of a singular orbit is m 1 ⊕ m 2 ⊕ m 3 \mathfrak{m}_{1}\oplus\mathfrak{m}_{2}\oplus\mathfrak{m}_{3} m 1 ⊕ m 2 ⊕ m 3 , and m 0 \mathfrak{m}_{0} m 0 is the collapsing direction. Since
[ X i j , Y i j ] = i 6 ( E i i − E j j ) , [X_{ij},Y_{ij}]=\frac{\mathrm{i}}{6}(E_{ii}-E_{jj}), [ X ij , Y ij ] = 6 i ( E ii − E j j ) ,
projection along g = h ⊕ m \mathfrak{g}=\mathfrak{h}\oplus\mathfrak{m} g = h ⊕ m gives
[ X 13 , Y 13 ] m = − 1 3 Z , [ X 12 , Y 12 ] m = − 1 6 Z , [ X 23 , Y 23 ] m = − 1 6 Z . (9) [X_{13},Y_{13}]_{\mathfrak{m}}=-\frac{1}{3}Z,\qquad[X_{12},Y_{12}]_{\mathfrak{m}}=-\frac{1}{6}Z,\qquad[X_{23},Y_{23}]_{\mathfrak{m}}=-\frac{1}{6}Z. \tag*{(9)} [ X 13 , Y 13 ] m = − 3 1 Z , [ X 12 , Y 12 ] m = − 6 1 Z , [ X 23 , Y 23 ] m = − 6 1 Z . (9)
For later use, write
( b 1 , b 2 , b 3 ) = ( 1 3 , 1 6 , 1 6 ) , [ X i , Y i ] m = − b i Z , (10) (b_{1},b_{2},b_{3})=\left(\frac{1}{3},\frac{1}{6},\frac{1}{6}\right),\qquad[X_{i},Y_{i}]_{\mathfrak{m}}=-b_{i}Z, \tag*{(10)} ( b 1 , b 2 , b 3 ) = ( 3 1 , 6 1 , 6 1 ) , [ X i , Y i ] m = − b i Z , (10)
where ( X 1 , Y 1 ) = ( X 13 , Y 13 ) (X_{1},Y_{1})=(X_{13},Y_{13}) ( X 1 , Y 1 ) = ( X 13 , Y 13 ) , ( X 2 , Y 2 ) = ( X 12 , Y 12 ) (X_{2},Y_{2})=(X_{12},Y_{12}) ( X 2 , Y 2 ) = ( X 12 , Y 12 ) , and ( X 3 , Y 3 ) = ( X 23 , Y 23 ) (X_{3},Y_{3})=(X_{23},Y_{23}) ( X 3 , Y 3 ) = ( X 23 , Y 23 ) .
The following Ad ( H ) \operatorname{Ad}(H) Ad ( H ) -invariant covector on m \mathfrak{m} m determines a G G G -invariant one-form θ \theta θ on P = G / H P=G/H P = G / H :
θ ( Z ) = 1 , θ ∣ m 1 ⊕ m 2 ⊕ m 3 = 0. (11) \theta(Z)=1,\qquad\theta|_{\mathfrak{m}_{1}\oplus\mathfrak{m}_{2}\oplus\mathfrak{m}_{3}}=0. \tag*{(11)} θ ( Z ) = 1 , θ ∣ m 1 ⊕ m 2 ⊕ m 3 = 0. (11)
Schur’s lemma now shows that every G G G -invariant metric on P P P has the form
h = f 0 2 θ 2 + f 1 2 Q ∣ m 1 + f 2 2 Q ∣ m 2 + f 3 2 Q ∣ m 3 , f i > 0. h=f_{0}^{2}\theta^{2}+f_{1}^{2}Q|_{\mathfrak{m}_{1}}+f_{2}^{2}Q|_{\mathfrak{m}_{2}}+f_{3}^{2}Q|_{\mathfrak{m}_{3}},\qquad f_{i}>0. h = f 0 2 θ 2 + f 1 2 Q ∣ m 1 + f 2 2 Q ∣ m 2 + f 3 2 Q ∣ m 3 , f i > 0.
Now let g g g be a G G G -invariant metric on M M M , and choose a geodesic meeting every principal orbit orthogonally, parametrised by g g g -arc length t t t . Its G G G -translates identify the regular part with M r e g ≅ ( 0 , T ) × P M_{\mathrm{reg}}\cong(0,T)\times P M reg ≅ ( 0 , T ) × P ; in this geodesic gauge, the metric has no mixed d t dt d t -orbit terms, and therefore takes the form
g = d t 2 + f 0 ( t ) 2 θ 2 + f 1 ( t ) 2 Q ∣ m 1 + f 2 ( t ) 2 Q ∣ m 2 + f 3 ( t ) 2 Q ∣ m 3 , f i ( t ) > 0. (12) g=dt^{2}+f_{0}(t)^{2}\theta^{2}+f_{1}(t)^{2}Q|_{\mathfrak{m}_{1}}+f_{2}(t)^{2}Q|_{\mathfrak{m}_{2}}+f_{3}(t)^{2}Q|_{\mathfrak{m}_{3}},\qquad f_{i}(t)>0. \tag*{(12)} g = d t 2 + f 0 ( t ) 2 θ 2 + f 1 ( t ) 2 Q ∣ m 1 + f 2 ( t ) 2 Q ∣ m 2 + f 3 ( t ) 2 Q ∣ m 3 , f i ( t ) > 0. (12)
Smoothness at the singular orbits
We use the smoothness criterion of Verdiani–Ziller [22 Theorem 1 and §§3.1–3.3] . In this action the normal slice is two-dimensional, m 0 \mathfrak{m}_0 m 0 is the only collapsing orbit direction, and the central circle acts with weight one on the normal slice and trivially on the tangent space of the singular orbit. Consequently the general criterion takes a particularly simple form.
Proposition 2.1. Let s s s denote inward distance from one of the singular orbits, and write the metric coefficients as functions of s s s near that orbit. The metric in (12) and an invariant function u u u extend smoothly across the orbit if and only if, for s s s sufficiently small,
f 0 ( s ) = s ϕ 0 ( s 2 ) , ϕ 0 ( 0 ) = 1 , f i ( s ) = ϕ i ( s 2 ) , ϕ i ( 0 ) > 0 ( 1 ≤ i ≤ 3 ) , u ( s ) = ϕ u ( s 2 ) , (13) \begin{aligned}
f_0(s)&=s\phi_0(s^2), \qquad\phi_0(0)=1,\\
f_i(s)&=\phi_i(s^2), \qquad\phi_i(0)>0 \quad(1\leq i\leq3),\\
u(s)&=\phi_u(s^2), \tag*{(13)}
\end{aligned} f 0 ( s ) f i ( s ) u ( s ) = s ϕ 0 ( s 2 ) , ϕ 0 ( 0 ) = 1 , = ϕ i ( s 2 ) , ϕ i ( 0 ) > 0 ( 1 ≤ i ≤ 3 ) , = ϕ u ( s 2 ) , (13) for smooth functions ϕ 0 , ϕ 1 , ϕ 2 , ϕ 3 , ϕ u \phi_0,\phi_1,\phi_2,\phi_3,\phi_u ϕ 0 , ϕ 1 , ϕ 2 , ϕ 3 , ϕ u .
Proof. The subgroup generated by Z Z Z has period 2 π 2\pi 2 π , intersects H H H trivially, and generates K / H K/H K / H . Thus the restriction of the metric to the normal disc is smooth at its centre precisely when f 0 f_0 f 0 extends as an odd function with f 0 ′ ( 0 ) = 1 f_0'(0)=1 f 0 ′ ( 0 ) = 1 . The central circle acts trivially on each tangential summand m i \mathfrak{m}_i m i , so the remaining metric coefficients extend as positive even functions. There are no mixed conditions because the four principal-isotropy summands are pairwise inequivalent. Finally, a rotation-invariant smooth function on the normal disc is a smooth function of s 2 s^2 s 2 . These statements are exactly the conditions in (13) . □
At the left-hand singular orbit these conditions apply with s = t s=t s = t ; at the right-hand singular orbit they apply to f i ( T − s ) f_i(T-s) f i ( T − s ) and u ( T − s ) u(T-s) u ( T − s ) , where s = T − t s=T-t s = T − t . In particular,
f 0 ′ ( 0 ) = 1 , f 0 ′ ( T ) = − 1 , f i ′ ( 0 ) = f i ′ ( T ) = u ′ ( 0 ) = u ′ ( T ) = 0 ( 1 ≤ i ≤ 3 ) . f_0'(0)=1,\qquad f_0'(T)=-1,\qquad f_i'(0)=f_i'(T)=u'(0)=u'(T)=0\quad(1\leq i\leq3). f 0 ′ ( 0 ) = 1 , f 0 ′ ( T ) = − 1 , f i ′ ( 0 ) = f i ′ ( T ) = u ′ ( 0 ) = u ′ ( T ) = 0 ( 1 ≤ i ≤ 3 ) .
We use the full parity conditions in (13) below.
Ricci curvature of G-invariant metrics
Write g = d t 2 + h ( t ) g=dt^2+h(t) g = d t 2 + h ( t ) , where
h ( t ) = f 0 ( t ) 2 θ 2 + f 1 2 Q ∣ m 1 + f 2 2 Q ∣ m 2 + f 3 2 Q ∣ m 3 . (14) h(t)=f_0(t)^2\theta^2+f_1^2Q|_{\mathfrak{m}_1}+f_2^2Q|_{\mathfrak{m}_2}+f_3^2Q|_{\mathfrak{m}_3}. \tag*{(14)} h ( t ) = f 0 ( t ) 2 θ 2 + f 1 2 Q ∣ m 1 + f 2 2 Q ∣ m 2 + f 3 2 Q ∣ m 3 . (14)
We first compute the Ricci curvature of the homogeneous metric h ( t ) h(t) h ( t ) directly from the Lie brackets. For a fixed t t t , write h = h ( t ) h=h(t) h = h ( t ) , let B g B_{\mathfrak{g}} B g denote the Killing form of g \mathfrak{g} g , and choose an h h h -orthonormal basis { E a } \{E_a\} { E a } of m \mathfrak{m} m . For X ∈ m X\in\mathfrak{m} X ∈ m , the Ad ( H ) \operatorname{Ad}(H) Ad ( H ) -invariance of m \mathfrak{m} m gives [ X , h ] ⊂ m [X,\mathfrak{h}]\subset\mathfrak{m} [ X , h ] ⊂ m , so the h \mathfrak{h} h -diagonal block of ad X \operatorname{ad}X ad X is zero. Hence
tr m ( pr m ad X ) = tr g ( ad X ) = 0 , \operatorname{tr}_{\mathfrak{m}}(\operatorname{pr}_{\mathfrak{m}}\operatorname{ad}X)=\operatorname{tr}_{\mathfrak{g}}(\operatorname{ad}X)=0, tr m ( pr m ad X ) = tr g ( ad X ) = 0 ,
where pr m \operatorname{pr}_{\mathfrak{m}} pr m is projection along g = h ⊕ m \mathfrak{g}=\mathfrak{h}\oplus\mathfrak{m} g = h ⊕ m , and the last equality uses the unimodularity of the compact group G G G . Thus the mean-curvature vector in [1 Corollary 7.38] vanishes, and the remaining formula gives, for X ∈ m X\in\mathfrak{m} X ∈ m ,
Ric h ( X , X ) = ⟨ M m X , X ⟩ h − 1 2 B g ( X , X ) . (15) \operatorname{Ric}_h(X,X)=\langle M_{\mathfrak{m}}X,X\rangle_h-\frac{1}{2}B_{\mathfrak{g}}(X,X). \tag*{(15)} Ric h ( X , X ) = ⟨ M m X , X ⟩ h − 2 1 B g ( X , X ) . (15)
where M m M_{\mathfrak{m}} M m is the moment map given by
⟨ M m X , X ⟩ h = − 1 2 ∑ a ∥ [ X , E a ] m ∥ h 2 + 1 4 ∑ a , b ⟨ [ E a , E b ] m , X ⟩ h 2 . (16) \langle M_{\mathfrak{m}}X,X\rangle_h=-\frac{1}{2}\sum_a\lVert[X,E_a]_{\mathfrak{m}}\rVert_h^2+\frac{1}{4}\sum_{a,b}\langle[E_a,E_b]_{\mathfrak{m}},X\rangle_h^2. \tag*{(16)} ⟨ M m X , X ⟩ h = − 2 1 a ∑ ∥[ X , E a ] m ∥ h 2 + 4 1 a , b ∑ ⟨[ E a , E b ] m , X ⟩ h 2 . (16)
The centrality of Z Z Z gives [ Z , m ] = 0 [Z,\mathfrak{m}]=0 [ Z , m ] = 0 , thus B g ( Z , g ) = 0 B_{\mathfrak{g}}(Z,\mathfrak{g})=0 B g ( Z , g ) = 0 . Besides the brackets in (9) , the only remaining nonzero projected brackets are between distinct summands m 1 , m 2 , m 3 \mathfrak{m}_1,\mathfrak{m}_2,\mathfrak{m}_3 m 1 , m 2 , m 3 . For example,
[ X 13 , X 12 ] = 1 12 X 23 , [ X 13 , Y 12 ] = − 1 12 Y 23 , [X_{13},X_{12}]=\frac{1}{\sqrt{12}}X_{23},\qquad[X_{13},Y_{12}]=-\frac{1}{\sqrt{12}}Y_{23}, [ X 13 , X 12 ] = 12 1 X 23 , [ X 13 , Y 12 ] = − 12 1 Y 23 ,
[ Y 13 , X 12 ] = 1 12 Y 23 , [ Y 13 , Y 12 ] = 1 12 X 23 . [Y_{13},X_{12}]=\frac{1}{\sqrt{12}}Y_{23},\qquad[Y_{13},Y_{12}]=\frac{1}{\sqrt{12}}X_{23}. [ Y 13 , X 12 ] = 12 1 Y 23 , [ Y 13 , Y 12 ] = 12 1 X 23 .
Each of these four brackets has squared Q Q Q -norm 1 / 12 1/12 1/12 . Cyclically permuting the matrix indices gives the same squared norms for the other pairs of distinct summands. Together with (9) , these identities provide all the bracket data needed in Besse’s formula.
Since Ric h ♯ \operatorname{Ric}^{\sharp}_{h} Ric h ♯ is G G G -invariant and the four isotropy summands are pairwise inequivalent, it is diagonal with respect to (8). Let R i R_i R i be the eigenvalue of its Ricci endomorphism on m i \mathfrak{m}_i m i :
Ric h ♯ ∣ m i = R i Id . \left.\operatorname{Ric}^{\sharp}_{h}\right|_{\mathfrak{m}_i}=R_i\operatorname{Id}. Ric h ♯ m i = R i Id .
Proposition 2.2. The Ricci curvature of the homogeneous metric in (14) is given by
R 0 = f 0 2 2 ∑ i = 1 3 b i 2 f i 4 , R i = 1 2 f i 2 + f i 4 − f j 4 − f k 4 12 f 1 2 f 2 2 f 3 2 − b i 2 f 0 2 2 f i 4 , { i , j , k } = { 1 , 2 , 3 } , (17) \begin{aligned}
R_0&=\frac{f_0^2}{2}\sum_{i=1}^{3}\frac{b_i^2}{f_i^4},\\
R_i&=\frac{1}{2f_i^2}+\frac{f_i^4-f_j^4-f_k^4}{12f_1^2f_2^2f_3^2}-\frac{b_i^2f_0^2}{2f_i^4},\qquad\{i,j,k\}=\{1,2,3\}, \tag*{(17)}
\end{aligned} R 0 R i = 2 f 0 2 i = 1 ∑ 3 f i 4 b i 2 , = 2 f i 2 1 + 12 f 1 2 f 2 2 f 3 2 f i 4 − f j 4 − f k 4 − 2 f i 4 b i 2 f 0 2 , { i , j , k } = { 1 , 2 , 3 } , (17) where ( b 1 , b 2 , b 3 ) = ( 1 3 , 1 6 , 1 6 ) (b_1,b_2,b_3)=\left(\frac{1}{3},\frac{1}{6},\frac{1}{6}\right) ( b 1 , b 2 , b 3 ) = ( 3 1 , 6 1 , 6 1 ) .
Proof. Using the notation in (10), take the h h h -orthonormal basis of m \mathfrak{m} m
E 0 = Z f 0 , E i , 1 = X i f i , E i , 2 = Y i f i , 1 ≤ i ≤ 3. E_0=\frac{Z}{f_0},\qquad E_{i,1}=\frac{X_i}{f_i},\qquad E_{i,2}=\frac{Y_i}{f_i},\qquad1\leq i\leq3. E 0 = f 0 Z , E i , 1 = f i X i , E i , 2 = f i Y i , 1 ≤ i ≤ 3.
By (9),
[ E i , 1 , E i , 2 ] m = − b i f 0 f i 2 E 0 . [E_{i,1},E_{i,2}]_{\mathfrak{m}}=-\frac{b_if_0}{f_i^2}E_0. [ E i , 1 , E i , 2 ] m = − f i 2 b i f 0 E 0 .
Since Z Z Z is central, B g ( E 0 , E 0 ) = 0 B_g(E_0,E_0)=0 B g ( E 0 , E 0 ) = 0 and [ E 0 , m ] = 0 [E_0,\mathfrak{m}]=0 [ E 0 , m ] = 0 . Equations (15) and (16) therefore give
R 0 = 1 4 ∑ a , b ⟨ [ E a , E b ] m , E 0 ⟩ h 2 = 1 4 ∑ i = 1 3 2 ( b i f 0 f i 2 ) 2 = f 0 2 2 ∑ i = 1 3 b i 2 f i 4 . R_0=\frac{1}{4}\sum_{a,b}\langle[E_a,E_b]_{\mathfrak{m}},E_0\rangle_h^2=\frac{1}{4}\sum_{i=1}^{3}2\left(\frac{b_if_0}{f_i^2}\right)^2=\frac{f_0^2}{2}\sum_{i=1}^{3}\frac{b_i^2}{f_i^4}. R 0 = 4 1 a , b ∑ ⟨[ E a , E b ] m , E 0 ⟩ h 2 = 4 1 i = 1 ∑ 3 2 ( f i 2 b i f 0 ) 2 = 2 f 0 2 i = 1 ∑ 3 f i 4 b i 2 .
For { i , j , k } = { 1 , 2 , 3 } \{i,j,k\}=\{1,2,3\} { i , j , k } = { 1 , 2 , 3 } , the same bracket identities and the displayed brackets between distinct summands yield
⟨ M m E i , 1 , E i , 1 ⟩ h = − 1 2 ∑ a ∥ [ E i , 1 , E a ] m ∥ h 2 + 1 4 ∑ a , b ⟨ [ E a , E b ] m , E i , 1 ⟩ h 2 = − 1 2 ( b i 2 f 0 2 f i 4 + f k 2 6 f i 2 f j 2 + f j 2 6 f i 2 f k 2 ) + f i 2 12 f j 2 f k 2 = − b i 2 f 0 2 2 f i 4 + f i 4 − f j 4 − f k 4 12 f i 2 f j 2 f k 2 . \begin{aligned}
\langle\mathcal{M}_{\mathfrak{m}}E_{i,1},E_{i,1}\rangle_h
&=-\frac{1}{2}\sum_a\|[E_{i,1},E_a]_{\mathfrak{m}}\|_h^2+\frac{1}{4}\sum_{a,b}\langle[E_a,E_b]_{\mathfrak{m}},E_{i,1}\rangle_h^2\\
&=-\frac{1}{2}\left(\frac{b_i^2f_0^2}{f_i^4}+\frac{f_k^2}{6f_i^2f_j^2}+\frac{f_j^2}{6f_i^2f_k^2}\right)+\frac{f_i^2}{12f_j^2f_k^2}\\
&=-\frac{b_i^2f_0^2}{2f_i^4}+\frac{f_i^4-f_j^4-f_k^4}{12f_i^2f_j^2f_k^2}.
\end{aligned} ⟨ M m E i , 1 , E i , 1 ⟩ h = − 2 1 a ∑ ∥ [ E i , 1 , E a ] m ∥ h 2 + 4 1 a , b ∑ ⟨[ E a , E b ] m , E i , 1 ⟩ h 2 = − 2 1 ( f i 4 b i 2 f 0 2 + 6 f i 2 f j 2 f k 2 + 6 f i 2 f k 2 f j 2 ) + 12 f j 2 f k 2 f i 2 = − 2 f i 4 b i 2 f 0 2 + 12 f i 2 f j 2 f k 2 f i 4 − f j 4 − f k 4 .
Since B g ∣ m i = − Q ∣ m i B_g|_{\mathfrak{m}_i}=-Q|_{\mathfrak{m}_i} B g ∣ m i = − Q ∣ m i ,
− 1 2 B g ( E i , 1 , E i , 1 ) = 1 2 f i 2 . -\frac{1}{2}B_g(E_{i,1},E_{i,1})=\frac{1}{2f_i^2}. − 2 1 B g ( E i , 1 , E i , 1 ) = 2 f i 2 1 .
Substitution into (15) gives
R i = 1 2 f i 2 − b i 2 f 0 2 2 f i 4 − f k 2 12 f i 2 f j 2 − f j 2 12 f i 2 f k 2 + f i 2 12 f j 2 f k 2 = 1 2 f i 2 + f i 4 − f j 4 − f k 4 12 f 1 2 f 2 2 f 3 2 − b i 2 f 0 2 2 f i 4 , \begin{aligned}
R_i&=\frac{1}{2f_i^2}-\frac{b_i^2f_0^2}{2f_i^4}-\frac{f_k^2}{12f_i^2f_j^2}-\frac{f_j^2}{12f_i^2f_k^2}+\frac{f_i^2}{12f_j^2f_k^2}\\
&=\frac{1}{2f_i^2}+\frac{f_i^4-f_j^4-f_k^4}{12f_1^2f_2^2f_3^2}-\frac{b_i^2f_0^2}{2f_i^4},
\end{aligned} R i = 2 f i 2 1 − 2 f i 4 b i 2 f 0 2 − 12 f i 2 f j 2 f k 2 − 12 f i 2 f k 2 f j 2 + 12 f j 2 f k 2 f i 2 = 2 f i 2 1 + 12 f 1 2 f 2 2 f 3 2 f i 4 − f j 4 − f k 4 − 2 f i 4 b i 2 f 0 2 ,
as required. □
We now pass from the homogeneous orbit metric h ( t ) h(t) h ( t ) to the cohomogeneity one metric
g = d t 2 + h ( t ) . g=dt^2+h(t). g = d t 2 + h ( t ) .
Consider the normal vector field to the principal orbits N = ∂ t N=\partial_t N = ∂ t , and define the shape operator of the principal orbits by
L X = ∇ X g N , L = 1 2 h − 1 h ′ . LX=\nabla^g_XN,\qquad L=\frac{1}{2}h^{-1}h'. L X = ∇ X g N , L = 2 1 h − 1 h ′ .
where ∇ g \nabla^g ∇ g is the Levi-Civita connection of g g g . The shape operator is diagonal with L ∣ m i = L i Id L|_{\mathfrak{m}_i}=L_i\operatorname{Id} L ∣ m i = L i Id and
L i = f i ′ f i . L_i=\frac{f_i'}{f_i}. L i = f i f i ′ .
Proposition 2.3. The Ricci curvature of the cohomogeneity one metric in (12) is
Ric g ( X , Y ) = Ric h ( X , Y ) − h ( L ′ X , Y ) − ( tr L ) h ( L X , Y ) , \operatorname{Ric}_g(X,Y)=\operatorname{Ric}_h(X,Y)-h(L'X,Y)-(\operatorname{tr}L)h(LX,Y), Ric g ( X , Y ) = Ric h ( X , Y ) − h ( L ′ X , Y ) − ( tr L ) h ( L X , Y ) , Ric g ( N , X ) = 0 , \operatorname{Ric}_g(N,X)=0, Ric g ( N , X ) = 0 , Ric g ( N , N ) = − ( tr L ) ′ − tr ( L 2 ) , \operatorname{Ric}_g(N,N)=-(\operatorname{tr}L)'-\operatorname{tr}(L^2), Ric g ( N , N ) = − ( tr L ) ′ − tr ( L 2 ) , for X , Y X,Y X , Y tangent to the principal orbits. In terms of the metric coefficients,
Ric g ( N , N ) = − f 0 ′ ′ f 0 − 2 ∑ j = 1 3 f j ′ f j , \operatorname{Ric}_g(N,N)=-\frac{f_0''}{f_0}-2\sum_{j=1}^{3}\frac{f_j'}{f_j}, Ric g ( N , N ) = − f 0 f 0 ′′ − 2 j = 1 ∑ 3 f j f j ′ , and, for unit X ∈ m i X\in\mathfrak{m}_i X ∈ m i (0 ≤ i ≤ 3 0\leq i\leq3 0 ≤ i ≤ 3 ),
Ric g ( X , X ) = R i − f i ′ ′ f i − f i ′ f i ( f 0 ′ f 0 + 2 ∑ j = 1 3 f j ′ f j − f i ′ f i ) , \operatorname{Ric}_g(X,X)=R_i-\frac{f_i''}{f_i}-\frac{f_i'}{f_i}\left(\frac{f_0'}{f_0}+2\sum_{j=1}^{3}\frac{f_j'}{f_j}-\frac{f_i'}{f_i}\right), Ric g ( X , X ) = R i − f i f i ′′ − f i f i ′ ( f 0 f 0 ′ + 2 j = 1 ∑ 3 f j f j ′ − f i f i ′ ) , where R i R_i R i are given by (17). The components between distinct isotropy summands vanish.
Proof. The normal and tangential formulas follow from [10 (1.2), (1.4)] . The mixed components vanish by [10 Remark 1.20] , since the isotropy representation is multiplicity-free. Substituting L i = f i ′ / f i L_i=f_i'/f_i L i = f i ′ / f i gives the formulas in metric coefficients. □
Since b 2 = b 3 b_2=b_3 b 2 = b 3 , the curvature formulas are symmetric under interchange of ( f 2 , L 2 ) (f_2,L_2) ( f 2 , L 2 ) and ( f 3 , L 3 ) (f_3,L_3) ( f 3 , L 3 ) . Hence the Ricci equations preserve the subfamily f 2 = f 3 f_2=f_3 f 2 = f 3 , L 2 = L 3 L_2=L_3 L 2 = L 3 . On this subfamily,
R 0 = f 0 2 18 f 1 4 + f 0 2 36 f 2 4 , R 1 = 1 3 f 1 2 + f 1 2 12 f 2 4 − f 0 2 18 f 1 4 , R 2 = R 3 = 1 2 f 2 2 − f 1 2 12 f 2 4 − f 0 2 72 f 2 4 , tr L = L 0 + 2 L 1 + 4 L 2 . (18) \begin{aligned}
R_0&=\frac{f_0^2}{18f_1^4}+\frac{f_0^2}{36f_2^4},\\
R_1&=\frac{1}{3f_1^2}+\frac{f_1^2}{12f_2^4}-\frac{f_0^2}{18f_1^4},\\
R_2&=R_3=\frac{1}{2f_2^2}-\frac{f_1^2}{12f_2^4}-\frac{f_0^2}{72f_2^4},\\
\operatorname{tr}L&=L_0+2L_1+4L_2. \tag*{(18)}
\end{aligned} R 0 R 1 R 2 tr L = 18 f 1 4 f 0 2 + 36 f 2 4 f 0 2 , = 3 f 1 2 1 + 12 f 2 4 f 1 2 − 18 f 1 4 f 0 2 , = R 3 = 2 f 2 2 1 − 12 f 2 4 f 1 2 − 72 f 2 4 f 0 2 , = L 0 + 2 L 1 + 4 L 2 . (18)
The singular initial value problem
We first derive the soliton equations on the regular set and state the boundary value problem on the disc-bundle manifold. We then solve the equations locally at either singular orbit. The local existence theory for singular cohomogeneity one initial value problems was developed by Eschenburg and Wang [14 ] in the Einstein case and extended by Buzano [6 ] to gradient Ricci solitons. Here we use the explicit isotropy data and smoothness conditions to obtain convergent series and quantitative bounds for the initial solution data in the shooting argument.
The soliton equation and boundary conditions
We now restrict the cohomogeneity one metrics in (12) to the family f 2 = f 3 f_2=f_3 f 2 = f 3 . Let u = u ( t ) u=u(t) u = u ( t ) be the invariant gradient soliton potential, and define the weighted mean curvature by
ξ = tr L − u ′ . (19) \xi=\operatorname{tr}L-u'. \tag*{(19)} ξ = tr L − u ′ . (19)
From the Ricci curvature formulas of § 2 , the gradient soliton equation in (2) with λ = 1 \lambda=1 λ = 1 yields the following first-order system
f i ′ = f i L i , i = 0 , 1 , 2 , L 0 ′ = − ξ L 0 + f 0 2 36 f 2 4 + f 0 2 18 f 1 4 − 1 , L 1 ′ = − ξ L 1 + 1 3 f 1 2 + f 1 2 12 f 2 4 − f 0 2 18 f 1 4 − 1 , L 2 ′ = − ξ L 2 + 1 2 f 2 2 − f 1 2 12 f 2 4 − f 1 2 72 f 2 4 − 1 , ξ ′ = − 1 − L 0 2 − 2 L 1 2 − 4 L 2 2 . (20) \begin{aligned}
f_i' &= f_iL_i,\qquad i=0,1,2,\\
L_0' &= -\xi L_0+\frac{f_0^2}{36f_2^4}+\frac{f_0^2}{18f_1^4}-1,\\
L_1' &= -\xi L_1+\frac{1}{3f_1^2}+\frac{f_1^2}{12f_2^4}-\frac{f_0^2}{18f_1^4}-1,\\
L_2' &= -\xi L_2+\frac{1}{2f_2^2}-\frac{f_1^2}{12f_2^4}-\frac{f_1^2}{72f_2^4}-1,\\
\xi' &= -1-L_0^2-2L_1^2-4L_2^2. \tag*{(20)}
\end{aligned} f i ′ L 0 ′ L 1 ′ L 2 ′ ξ ′ = f i L i , i = 0 , 1 , 2 , = − ξ L 0 + 36 f 2 4 f 0 2 + 18 f 1 4 f 0 2 − 1 , = − ξ L 1 + 3 f 1 2 1 + 12 f 2 4 f 1 2 − 18 f 1 4 f 0 2 − 1 , = − ξ L 2 + 2 f 2 2 1 − 12 f 2 4 f 1 2 − 72 f 2 4 f 1 2 − 1 , = − 1 − L 0 2 − 2 L 1 2 − 4 L 2 2 . (20)
for Y = ( f 0 , f 1 , f 2 , L 0 , L 1 , L 2 , ξ ) Y=(f_0,f_1,f_2,L_0,L_1,L_2,\xi) Y = ( f 0 , f 1 , f 2 , L 0 , L 1 , L 2 , ξ ) . The vector field in (20) is real analytic on { f 1 f 2 ≠ 0 } \{f_1f_2\ne0\} { f 1 f 2 = 0 } ; we use the metric region f 0 , f 1 , f 2 > 0 f_0,f_1,f_2>0 f 0 , f 1 , f 2 > 0 .
Proposition 3.1. Let g g g be a metric of the form given in (12) on I × P ⊆ M r e g I\times P\subseteq M_{\mathrm{reg}} I × P ⊆ M reg , with f 2 = f 3 f_2=f_3 f 2 = f 3 , and let u = u ( t ) u=u(t) u = u ( t ) be an invariant function. Then ( g , u ) (g,u) ( g , u ) satisfies (2) with λ = 1 \lambda=1 λ = 1 if and only if
Y = ( f 0 , f 1 , f 2 , L 0 , L 1 , L 2 , ξ ) Y=(f_0,f_1,f_2,L_0,L_1,L_2,\xi) Y = ( f 0 , f 1 , f 2 , L 0 , L 1 , L 2 , ξ ) solves (20) . Conversely, any solution Y Y Y of (20) with f 0 , f 1 , f 2 > 0 f_0,f_1,f_2>0 f 0 , f 1 , f 2 > 0 determines such a metric and a potential u u u , up to an additive constant, by
u ′ = L 0 + 2 L 1 + 4 L 2 − ξ . u'=L_0+2L_1+4L_2-\xi. u ′ = L 0 + 2 L 1 + 4 L 2 − ξ .
Proof. The identities L i = f i ′ / f i L_i=f_i'/f_i L i = f i ′ / f i , for i = 0 , 1 , 2 i=0,1,2 i = 0 , 1 , 2 , give the first three equations in (20) . By Theorem 2.3, the tangential soliton equation is equivalent to
L ′ = Ric h ( t ) ♯ − Id − ξ L . L'=\operatorname{Ric}^{\sharp}_{h(t)}-\operatorname{Id}-\xi L. L ′ = Ric h ( t ) ♯ − Id − ξ L .
Using (18) , this gives exactly the three equations for L 0 , L 1 , L 2 L_0,L_1,L_2 L 0 , L 1 , L 2 in (20) . Similarly, Theorem 2.3 gives the normal soliton equation as
− ( tr L ) ′ − tr ( L 2 ) + u ′ ′ = 1. -(\operatorname{tr}L)'-\operatorname{tr}(L^2)+u''=1. − ( tr L ) ′ − tr ( L 2 ) + u ′′ = 1.
Since ξ ′ = ( tr L ) ′ − u ′ ′ \xi'=(\operatorname{tr}L)'-u'' ξ ′ = ( tr L ) ′ − u ′′ and L 3 = L 2 L_3=L_2 L 3 = L 2 , this is equivalent to
ξ ′ = − 1 − L 0 2 − 2 L 1 2 − 4 L 2 2 . \xi'=-1-L_0^2-2L_1^2-4L_2^2. ξ ′ = − 1 − L 0 2 − 2 L 1 2 − 4 L 2 2 .
The mixed soliton equation holds because the mixed Ricci components vanish, as established in § 2 , and ( Hess g u ) ( N , X ) = 0 (\operatorname{Hess}_g u)(N,X)=0 ( Hess g u ) ( N , X ) = 0 . Conversely, let Y Y Y solve (20) with f 0 , f 1 , f 2 > 0 f_0,f_1,f_2>0 f 0 , f 1 , f 2 > 0 . Define the metric g g g by (12) , setting f 3 = f 2 f_3=f_2 f 3 = f 2 , and choose u u u satisfying
u ′ = L 0 + 2 L 1 + 4 L 2 − ξ . u'=L_0+2L_1+4L_2-\xi. u ′ = L 0 + 2 L 1 + 4 L 2 − ξ .
The first three equations ensure that the shape operator of g g g has eigenvalues L i L_i L i , so that ξ = tr L − u ′ \xi=\operatorname{tr}L-u' ξ = tr L − u ′ . The preceding calculations therefore show that Ric g + Hess g u = g \operatorname{Ric}_g+\operatorname{Hess}_g u=g Ric g + Hess g u = g . □
We seek T > 0 T>0 T > 0 and a solution of (20) on ( 0 , T ) (0,T) ( 0 , T ) with f 0 , f 1 , f 2 > 0 f_0,f_1,f_2>0 f 0 , f 1 , f 2 > 0 such that the metric and potential from Theorem 3.1 satisfy the smoothness conditions of Theorem 2.1 at both singular orbits.
Parameters at the singular orbits and leading Taylor series
At the left singular orbit, the smoothness conditions in Theorem 2.1 give
f 0 ( 0 ) = 0 , f 0 ′ ( 0 ) = 1 , f 1 ′ ( 0 ) = f 2 ′ ( 0 ) = u ′ ( 0 ) = 0. f_0(0)=0,\qquad f_0'(0)=1,\qquad f_1'(0)=f_2'(0)=u'(0)=0. f 0 ( 0 ) = 0 , f 0 ′ ( 0 ) = 1 , f 1 ′ ( 0 ) = f 2 ′ ( 0 ) = u ′ ( 0 ) = 0.
We prescribe the data at this orbit
α 0 = u ′ ′ ( 0 ) , α 1 = f 1 ( 0 ) > 0 , α 2 = f 2 ( 0 ) = f 3 ( 0 ) > 0. (21) \alpha_0=u''(0),\qquad\alpha_1=f_1(0)>0,\qquad\alpha_2=f_2(0)=f_3(0)>0. \tag*{(21)} α 0 = u ′′ ( 0 ) , α 1 = f 1 ( 0 ) > 0 , α 2 = f 2 ( 0 ) = f 3 ( 0 ) > 0. (21)
Fix u ( 0 ) = 0 u(0)=0 u ( 0 ) = 0 . Although f 1 ′ ( 0 ) = f 2 ′ ( 0 ) = 0 f_1'(0)=f_2'(0)=0 f 1 ′ ( 0 ) = f 2 ′ ( 0 ) = 0 , one cannot obtain the next coefficients by simply setting t = 0 t=0 t = 0 in (20) : both L 0 L_0 L 0 and ξ \xi ξ have leading term t − 1 t^{-1} t − 1 . The finite limits needed below are
lim t → 0 L i ( t ) t = f i ′ ′ ( 0 ) f i ( 0 ) , i = 1 , 2. \lim_{t\to0}\frac{L_i(t)}{t}=\frac{f_i''(0)}{f_i(0)},\qquad i=1,2. t → 0 lim t L i ( t ) = f i ( 0 ) f i ′′ ( 0 ) , i = 1 , 2.
They are determined by inserting the smooth even expansions of f 1 f_1 f 1 and f 2 f_2 f 2 into the equations for L 1 L_1 L 1 and L 2 L_2 L 2 and balancing their constant terms. For r 1 , r 2 > 0 r_1,r_2>0 r 1 , r 2 > 0 , define
a 1 ( r 1 , r 2 ) = 1 6 r 1 2 + r 1 2 24 r 2 4 − 1 2 , a 2 ( r 1 , r 2 ) = 1 4 r 2 2 − r 1 2 24 r 4 2 − 1 2 . (22) \begin{aligned}
a_1(r_1,r_2)&=\frac{1}{6r_1^2}+\frac{r_1^2}{24r_2^4}-\frac{1}{2},\\
a_2(r_1,r_2)&=\frac{1}{4r_2^2}-\frac{r_1^2}{24r_4^2}-\frac{1}{2}. \tag*{(22)}
\end{aligned} a 1 ( r 1 , r 2 ) a 2 ( r 1 , r 2 ) = 6 r 1 2 1 + 24 r 2 4 r 1 2 − 2 1 , = 4 r 2 2 1 − 24 r 4 2 r 1 2 − 2 1 . (22)
The resulting second derivatives are
f i ′ ′ ( 0 ) = α i a i ( α 1 , α 2 ) , i = 1 , 2. f_i^{\prime\prime}(0)=\alpha_i a_i(\alpha_1,\alpha_2),\qquad i=1,2. f i ′′ ( 0 ) = α i a i ( α 1 , α 2 ) , i = 1 , 2.
Substituting the Taylor expansions allowed by the smoothness conditions in Theorem 2.1 into (20) and comparing coefficients gives
f 0 ( t ) = t + α 0 − 2 a 1 − 4 a 2 − 1 6 t 3 + O ( t 5 ) , f i ( t ) = α i + α i a i 2 t 2 + O ( t 4 ) , u ( t ) = α 0 2 t 2 + O ( t 4 ) , (23) \begin{aligned}
f_0(t)&=t+\frac{\alpha_0-2a_1-4a_2-1}{6}t^3+O(t^5),\\
f_i(t)&=\alpha_i+\frac{\alpha_i a_i}{2}t^2+O(t^4),\\
u(t)&=\frac{\alpha_0}{2}t^2+O(t^4), \tag*{(23)}
\end{aligned} f 0 ( t ) f i ( t ) u ( t ) = t + 6 α 0 − 2 a 1 − 4 a 2 − 1 t 3 + O ( t 5 ) , = α i + 2 α i a i t 2 + O ( t 4 ) , = 2 α 0 t 2 + O ( t 4 ) , (23)
where i = 1 , 2 i=1,2 i = 1 , 2 and a i = a i ( α 1 , α 2 ) a_i=a_i(\alpha_1,\alpha_2) a i = a i ( α 1 , α 2 ) . In Theorem 3.2, we prove that ( α 0 , α 1 , α 2 ) (\alpha_0,\alpha_1,\alpha_2) ( α 0 , α 1 , α 2 ) uniquely determine a local real analytic solution. Its Taylor coefficients are determined recursively by (33) .
At the right singular orbit, we use s = T − t s=T-t s = T − t and prescribe the data
ω 0 = u ′ ′ ( T ) , ω 1 = f 1 ( T ) > 0 , ω 2 = f 2 ( T ) = f 3 ( T ) > 0. (24) \omega_0=u^{\prime\prime}(T),\qquad\omega_1=f_1(T)>0,\qquad\omega_2=f_2(T)=f_3(T)>0. \tag*{(24)} ω 0 = u ′′ ( T ) , ω 1 = f 1 ( T ) > 0 , ω 2 = f 2 ( T ) = f 3 ( T ) > 0. (24)
The functions f i ( T − s ) f_i(T-s) f i ( T − s ) and u ( T − s ) − u ( T ) u(T-s)-u(T) u ( T − s ) − u ( T ) have the expansions in (23) , with t t t replaced by s s s and ( α 0 , α 1 , α 2 ) (\alpha_0,\alpha_1,\alpha_2) ( α 0 , α 1 , α 2 ) replaced by ( ω 0 , ω 1 , ω 2 ) (\omega_0,\omega_1,\omega_2) ( ω 0 , ω 1 , ω 2 ) . In particular,
u ( T − s ) = u ( T ) + ω 0 2 s 2 + O ( s 4 ) . u(T-s)=u(T)+\frac{\omega_0}{2}s^2+O(s^4). u ( T − s ) = u ( T ) + 2 ω 0 s 2 + O ( s 4 ) .
Reduction to a regular-singular system
The expansions in (23) show that L 0 L_0 L 0 and ξ \xi ξ have leading term t − 1 t^{-1} t − 1 at the left singular orbit, while f 0 f_0 f 0 , L 1 L_1 L 1 , L 2 L_2 L 2 vanish to at least first order. We therefore set x = t 2 x=t^2 x = t 2 and write
f 0 ( t ) = t f ~ 0 ( x ) , L 0 ( t ) = t − 1 + t ℓ 0 ( x ) , L i ( t ) = t ℓ i ( x ) ( i = 1 , 2 ) , ξ ( t ) = t − 1 + t ξ ^ ( x ) . \begin{aligned}
f_0(t)&=t\widetilde{f}_0(x),\qquad L_0(t)=t^{-1}+t\ell_0(x),\\
L_i(t)&=t\ell_i(x)\quad(i=1,2),\qquad\xi(t)=t^{-1}+t\widehat{\xi}(x).
\end{aligned} f 0 ( t ) L i ( t ) = t f 0 ( x ) , L 0 ( t ) = t − 1 + t ℓ 0 ( x ) , = t ℓ i ( x ) ( i = 1 , 2 ) , ξ ( t ) = t − 1 + t ξ ( x ) .
Under the identities L i = f i ′ / f i L_i=f_i'/f_i L i = f i ′ / f i , ξ = tr L − u ′ \xi=\operatorname{tr}L-u' ξ = tr L − u ′ and f 3 = f 2 f_3=f_2 f 3 = f 2 , the smoothness conditions for the metric and potential in Theorem 2.1 are equivalent to smooth extension in x x x of
f ~ 0 , f 1 2 , f 2 2 , ℓ 0 , ℓ 1 , ℓ 2 , ξ ^ , \widetilde{f}_0,\ f_1^2,\ f_2^2,\ \ell_0,\ \ell_1,\ \ell_2,\ \widehat{\xi}, f 0 , f 1 2 , f 2 2 , ℓ 0 , ℓ 1 , ℓ 2 , ξ ,
with f ~ 0 ( 0 ) = 1 \widetilde{f}_0(0)=1 f 0 ( 0 ) = 1 and f i 2 ( 0 ) > 0 f_i^2(0)>0 f i 2 ( 0 ) > 0 for i = 1 , 2 i=1,2 i = 1 , 2 . Indeed, writing f 0 ( t ) = t ϕ 0 ( x ) f_0(t)=t\phi_0(x) f 0 ( t ) = t ϕ 0 ( x ) , f i ( t ) = ϕ i ( x ) f_i(t)=\phi_i(x) f i ( t ) = ϕ i ( x ) and u ( t ) = ϕ u ( x ) u(t)=\phi_u(x) u ( t ) = ϕ u ( x ) as in (13) gives
ℓ 0 = 2 ϕ 0 ′ ϕ 0 , ℓ i = 2 ϕ i ′ ϕ i ( i = 1 , 2 ) , ξ ^ = ℓ 0 + 2 ℓ 1 + 4 ℓ 2 − 2 ϕ u ′ , \ell_0=2\frac{\phi_0'}{\phi_0},\qquad\ell_i=2\frac{\phi_i'}{\phi_i}\quad(i=1,2),\qquad\widehat{\xi}=\ell_0+2\ell_1+4\ell_2-2\phi_u', ℓ 0 = 2 ϕ 0 ϕ 0 ′ , ℓ i = 2 ϕ i ϕ i ′ ( i = 1 , 2 ) , ξ = ℓ 0 + 2 ℓ 1 + 4 ℓ 2 − 2 ϕ u ′ ,
where the derivatives in this display are with respect to x x x . These expressions are smooth near x = 0 x=0 x = 0 . Conversely, positive square roots of f 1 2 f_1^2 f 1 2 and f 2 2 f_2^2 f 2 2 , together with f 0 = t f ~ 0 ( t 2 ) f_0=t\widetilde{f}_0(t^2) f 0 = t f 0 ( t 2 ) , give the required metric parity and endpoint values. The identity defining ξ \xi ξ gives
u ′ ( t ) = t ( ℓ 0 + 2 ℓ 1 + 4 ℓ 2 − ξ ^ ) ( t 2 ) . u'(t)=t\left(\ell_0+2\ell_1+4\ell_2-\widehat{\xi}\right)(t^2). u ′ ( t ) = t ( ℓ 0 + 2 ℓ 1 + 4 ℓ 2 − ξ ) ( t 2 ) .
Consequently, for some constant c c c ,
u ( t ) = c + 1 2 ∫ 0 t 2 ( ℓ 0 + 2 ℓ 1 + 4 ℓ 2 − ξ ^ ) ( z ) d z , u(t)=c+\frac{1}{2}\int_0^{t^2}\left(\ell_0+2\ell_1+4\ell_2-\widehat{\xi}\right)(z)\,dz, u ( t ) = c + 2 1 ∫ 0 t 2 ( ℓ 0 + 2 ℓ 1 + 4 ℓ 2 − ξ ) ( z ) d z ,
which is a smooth function of t 2 t^2 t 2 and supplies the required extension of the potential. By (20), the system of ODEs in these coordinates is
d f ~ 0 d x = 1 2 f ~ 0 ℓ 0 , d ( f i 2 ) d x = f i 2 ℓ i , i = 1 , 2 , x d ℓ 0 d x = − ℓ 0 − ξ ^ + 1 2 + x 2 ( − ξ ^ ℓ 0 + f ~ 0 2 36 f 2 4 + f ~ 0 2 18 f 1 4 ) , x d ℓ 1 d x = a 1 ( f 1 , f 2 ) − ℓ 1 − x 2 ( ξ ^ ℓ 1 + f ~ 0 2 18 f 1 4 ) , x d ℓ 2 d x = a 2 ( f 1 , f 2 ) − ℓ 2 − x 2 ( ξ ^ ℓ 2 + f ~ 0 2 72 f 2 4 ) , x d ξ ^ d x = − ℓ 0 − ξ ^ + 1 2 − x 2 ( ℓ 0 2 + 2 ℓ 1 2 + 4 ℓ 2 2 ) . (25) \begin{aligned}
\frac{d\widetilde{f}_0}{dx} &= \frac{1}{2}\widetilde{f}_0\ell_0,\\
\frac{d(f_i^2)}{dx} &= f_i^2\ell_i,\qquad i=1,2,\\
x\frac{d\ell_0}{dx} &= -\ell_0-\frac{\widehat{\xi}+1}{2}+\frac{x}{2}\left(-\widehat{\xi}\ell_0+\frac{\widetilde{f}_0^2}{36f_2^4}+\frac{\widetilde{f}_0^2}{18f_1^4}\right),\\
x\frac{d\ell_1}{dx} &= a_1(f_1,f_2)-\ell_1-\frac{x}{2}\left(\widehat{\xi}\ell_1+\frac{\widetilde{f}_0^2}{18f_1^4}\right),\\
x\frac{d\ell_2}{dx} &= a_2(f_1,f_2)-\ell_2-\frac{x}{2}\left(\widehat{\xi}\ell_2+\frac{\widetilde{f}_0^2}{72f_2^4}\right),\\
x\frac{d\widehat{\xi}}{dx} &= -\ell_0-\frac{\widehat{\xi}+1}{2}-\frac{x}{2}\left(\ell_0^2+2\ell_1^2+4\ell_2^2\right). \tag*{(25)}
\end{aligned} d x d f 0 d x d ( f i 2 ) x d x d ℓ 0 x d x d ℓ 1 x d x d ℓ 2 x d x d ξ = 2 1 f 0 ℓ 0 , = f i 2 ℓ i , i = 1 , 2 , = − ℓ 0 − 2 ξ + 1 + 2 x ( − ξ ℓ 0 + 36 f 2 4 f 0 2 + 18 f 1 4 f 0 2 ) , = a 1 ( f 1 , f 2 ) − ℓ 1 − 2 x ( ξ ℓ 1 + 18 f 1 4 f 0 2 ) , = a 2 ( f 1 , f 2 ) − ℓ 2 − 2 x ( ξ ℓ 2 + 72 f 2 4 f 0 2 ) , = − ℓ 0 − 2 ξ + 1 − 2 x ( ℓ 0 2 + 2 ℓ 1 2 + 4 ℓ 2 2 ) . (25)
In (25), a i ( f 1 , f 2 ) a_i(f_1,f_2) a i ( f 1 , f 2 ) denotes the rational function defined in (22), applied pointwise to the positive functions f 1 ( x ) f_1(x) f 1 ( x ) and f 2 ( x ) f_2(x) f 2 ( x ) . Although the new variables extend smoothly to x = 0 x=0 x = 0 , the equations remain singular there because x x x multiplies the derivatives of ℓ 0 , ℓ 1 , ℓ 2 , ξ ^ \ell_0,\ell_1,\ell_2,\widehat{\xi} ℓ 0 , ℓ 1 , ℓ 2 , ξ .
The expansions in (23) give
f ~ 0 ( 0 ) = 1 , f i 2 ( 0 ) = α i 2 , ℓ i ( 0 ) = a i ( α 1 , α 2 ) , i = 1 , 2. \widetilde{f}_0(0)=1,\qquad f_i^2(0)=\alpha_i^2,\qquad\ell_i(0)=a_i(\alpha_1,\alpha_2),\qquad i=1,2. f 0 ( 0 ) = 1 , f i 2 ( 0 ) = α i 2 , ℓ i ( 0 ) = a i ( α 1 , α 2 ) , i = 1 , 2.
and
ℓ 0 ( 0 ) = 1 3 ( α 0 − 2 a 1 ( α 1 , α 2 ) − 4 a 2 ( α 1 , α 2 ) − 1 ) , ξ ^ ( 0 ) = − 1 − 2 ℓ 0 ( 0 ) . \ell_0(0)=\frac{1}{3}\left(\alpha_0-2a_1(\alpha_1,\alpha_2)-4a_2(\alpha_1,\alpha_2)-1\right),\qquad\widehat{\xi}(0)=-1-2\ell_0(0). ℓ 0 ( 0 ) = 3 1 ( α 0 − 2 a 1 ( α 1 , α 2 ) − 4 a 2 ( α 1 , α 2 ) − 1 ) , ξ ( 0 ) = − 1 − 2 ℓ 0 ( 0 ) .
Thus ( α 0 , α 1 , α 2 ) (\alpha_0,\alpha_1,\alpha_2) ( α 0 , α 1 , α 2 ) determines the initial values of the seven functions
( f ~ 0 , f 1 2 , f 2 2 , ℓ 0 , ℓ 1 , ℓ 2 , ξ ^ ) . (\widetilde{f}_0,f_1^2,f_2^2,\ell_0,\ell_1,\ell_2,\widehat{\xi}). ( f 0 , f 1 2 , f 2 2 , ℓ 0 , ℓ 1 , ℓ 2 , ξ ) .
Diagonalising the singular part of (25) leads to the change of variables
W = ( f ~ 0 f 1 2 f 2 2 ℓ 0 − ξ ^ ℓ 1 − a 1 ( f 1 , f 2 ) ℓ 2 − a 2 ( f 1 , f 2 ) 2 ℓ 0 + ξ ^ + 1 ) , W 0 = ( 1 α 1 2 α 2 2 α 0 − 2 a 1 ( α 1 , α 2 ) − 4 a 2 ( α 1 , α 2 ) 0 0 0 ) . (26) W=
\begin{pmatrix}
\widetilde{f}_0\\
f_1^2\\
f_2^2\\
\ell_0-\widehat{\xi}\\
\ell_1-a_1(f_1,f_2)\\
\ell_2-a_2(f_1,f_2)\\
2\ell_0+\widehat{\xi}+1
\end{pmatrix},
\qquad
W_0=
\begin{pmatrix}
1\\
\alpha_1^2\\
\alpha_2^2\\
\alpha_0-2a_1(\alpha_1,\alpha_2)-4a_2(\alpha_1,\alpha_2)\\
0\\
0\\
0
\end{pmatrix}. \tag*{(26)} W = f 0 f 1 2 f 2 2 ℓ 0 − ξ ℓ 1 − a 1 ( f 1 , f 2 ) ℓ 2 − a 2 ( f 1 , f 2 ) 2 ℓ 0 + ξ + 1 , W 0 = 1 α 1 2 α 2 2 α 0 − 2 a 1 ( α 1 , α 2 ) − 4 a 2 ( α 1 , α 2 ) 0 0 0 . (26)
By the smoothness conditions at the singular orbit, every component of W W W extends smoothly to t = 0 t=0 t = 0 , with W ( 0 ) = W 0 W(0)=W_0 W ( 0 ) = W 0 . The original variables are recovered from W = ( W 1 , … , W 7 ) W=(W_1,\ldots,W_7) W = ( W 1 , … , W 7 ) by
f ~ 0 = W 1 , f 1 2 = W 2 , f 2 2 = W 3 , ℓ 0 = 1 3 ( W 4 + W 7 − 1 ) , ξ ^ = 1 3 ( W 7 − 2 W 4 − 1 ) \widetilde{f}_0=W_1,\qquad f_1^2=W_2,\qquad f_2^2=W_3,\qquad\ell_0=\frac{1}{3}(W_4+W_7-1),\qquad\widehat{\xi}=\frac{1}{3}(W_7-2W_4-1) f 0 = W 1 , f 1 2 = W 2 , f 2 2 = W 3 , ℓ 0 = 3 1 ( W 4 + W 7 − 1 ) , ξ = 3 1 ( W 7 − 2 W 4 − 1 )
and
ℓ 1 = a 1 ( f 1 , f 2 ) + W 5 , ℓ 2 = a 2 ( f 1 , f 2 ) + W 6 . \ell_1=a_1(f_1,f_2)+W_5,\qquad\ell_2=a_2(f_1,f_2)+W_6. ℓ 1 = a 1 ( f 1 , f 2 ) + W 5 , ℓ 2 = a 2 ( f 1 , f 2 ) + W 6 .
Define the rational map F l o c \mathcal{F}_{\mathrm{loc}} F loc by
F l o c ( W ) = ( f ~ 0 ℓ 0 2 f 1 2 ℓ 1 f 2 2 ℓ 2 − ξ ^ ℓ 0 2 + f ~ 0 2 ( 1 72 f 2 4 + 1 36 f 1 4 ) + ℓ 0 2 + 2 ℓ 1 2 + 4 ℓ 2 2 2 − ξ ^ ℓ 1 2 − f ~ 0 2 36 f 1 4 + ℓ 1 6 f 1 2 + f 1 2 ( 2 ℓ 2 − ℓ 1 ) 24 f 2 4 − ξ ^ ℓ 2 2 − f ~ 0 2 144 f 2 4 + ℓ 2 4 f 2 2 + f 1 2 ( ℓ 1 − 2 ℓ 2 ) 24 f 2 2 − ξ ^ ℓ 0 + f ~ 0 2 ( 1 36 f 2 4 + 1 18 f 1 4 ) − ℓ 0 2 + 2 ℓ 1 2 + 4 ℓ 2 2 2 ) (27) \mathcal{F}_{\mathrm{loc}}(W)=\begin{pmatrix}
\frac{\tilde{f}_0\ell_0}{2}\\
f_1^2\ell_1\\
f_2^2\ell_2\\
-\frac{\hat{\xi}\ell_0}{2}+\tilde{f}_0^2\left(\frac{1}{72f_2^4}+\frac{1}{36f_1^4}\right)+\frac{\ell_0^2+2\ell_1^2+4\ell_2^2}{2}\\
-\frac{\hat{\xi}\ell_1}{2}-\frac{\tilde{f}_0^2}{36f_1^4}+\frac{\ell_1}{6f_1^2}+\frac{f_1^2(2\ell_2-\ell_1)}{24f_2^4}\\
-\frac{\hat{\xi}\ell_2}{2}-\frac{\tilde{f}_0^2}{144f_2^4}+\frac{\ell_2}{4f_2^2}+\frac{f_1^2(\ell_1-2\ell_2)}{24f_2^2}\\
-\hat{\xi}\ell_0+\tilde{f}_0^2\left(\frac{1}{36f_2^4}+\frac{1}{18f_1^4}\right)-\frac{\ell_0^2+2\ell_1^2+4\ell_2^2}{2}
\end{pmatrix} \tag*{(27)} F loc ( W ) = 2 f ~ 0 ℓ 0 f 1 2 ℓ 1 f 2 2 ℓ 2 − 2 ξ ^ ℓ 0 + f ~ 0 2 ( 72 f 2 4 1 + 36 f 1 4 1 ) + 2 ℓ 0 2 + 2 ℓ 1 2 + 4 ℓ 2 2 − 2 ξ ^ ℓ 1 − 36 f 1 4 f ~ 0 2 + 6 f 1 2 ℓ 1 + 24 f 2 4 f 1 2 ( 2 ℓ 2 − ℓ 1 ) − 2 ξ ^ ℓ 2 − 144 f 2 4 f ~ 0 2 + 4 f 2 2 ℓ 2 + 24 f 2 2 f 1 2 ( ℓ 1 − 2 ℓ 2 ) − ξ ^ ℓ 0 + f ~ 0 2 ( 36 f 2 4 1 + 18 f 1 4 1 ) − 2 ℓ 0 2 + 2 ℓ 1 2 + 4 ℓ 2 2 (27)
The ODE to be solved is then
x d W d x = S ( W − W 0 ) + x F l o c ( W ) , S = diag ( 0 , 0 , 0 , 0 , − 1 , − 1 , − 3 / 2 ) , (28) x\frac{dW}{dx}=S(W-W_0)+x\mathcal{F}_{\mathrm{loc}}(W),\qquad S=\operatorname{diag}(0,0,0,0,-1,-1,-3/2), \tag*{(28)} x d x d W = S ( W − W 0 ) + x F loc ( W ) , S = diag ( 0 , 0 , 0 , 0 , − 1 , − 1 , − 3/2 ) , (28)
subject to W ( 0 ) = W 0 W(0)=W_0 W ( 0 ) = W 0 , with W 0 W_0 W 0 as in (26) .
Local existence, uniqueness, and analyticity.
Lemma 3.2. For each ( α 0 , α 1 , α 2 ) ∈ R × ( 0 , ∞ ) 2 (\alpha_0,\alpha_1,\alpha_2)\in\mathbb{R}\times(0,\infty)^2 ( α 0 , α 1 , α 2 ) ∈ R × ( 0 , ∞ ) 2 , there exists ε > 0 \varepsilon>0 ε > 0 such that (28) has a unique smooth solution W = W ( x ) W=W(x) W = W ( x ) on [ 0 , ε ] [0,\varepsilon] [ 0 , ε ] with W ( 0 ) = W 0 W(0)=W_0 W ( 0 ) = W 0 . The solution is real analytic in x x x . Near any fixed triple, ε \varepsilon ε may be chosen uniformly and W W W depends real analytically, and hence smoothly, on ( α 0 , α 1 , α 2 ) (\alpha_0,\alpha_1,\alpha_2) ( α 0 , α 1 , α 2 ) .
Proof. We give the fixed-point formulation because the same quantities will be estimated explicitly below. Put V = W − W 0 V=W-W_0 V = W − W 0 . Since S W 0 = 0 SW_0=0 S W 0 = 0 , a regular solution of (28) corresponds to a fixed point V V V of the integral operator
( T V ) j ( x ) = x ∫ 0 1 s − S j j ( F l o c ( W 0 + V ( s x ) ) ) j d s , 1 ≤ j ≤ 7. (29) (\mathcal{T}V)_j(x)=x\int_0^1s^{-S_{jj}}\bigl(\mathcal{F}_{\mathrm{loc}}(W_0+V(sx))\bigr)_j\,ds,\qquad1\leq j\leq7. \tag*{(29)} ( T V ) j ( x ) = x ∫ 0 1 s − S j j ( F loc ( W 0 + V ( s x )) ) j d s , 1 ≤ j ≤ 7. (29)
Choose positive radii d = ( d 1 , … , d 7 ) \mathbf{d}=(d_1,\ldots,d_7) d = ( d 1 , … , d 7 ) such that the box centred at W 0 W_0 W 0
P ( W 0 , d ) = { W 0 + v ∈ C 7 : ∣ v j ∣ ≤ d j } \mathcal{P}(W_0,\mathbf{d})=\{W_0+v\in\mathbb{C}^7:|v_j|\leq d_j\} P ( W 0 , d ) = { W 0 + v ∈ C 7 : ∣ v j ∣ ≤ d j }
does not meet { W 2 W 3 = 0 } \{W_2W_3=0\} { W 2 W 3 = 0 } . The rational map F l o c \mathcal{F}_{\mathrm{loc}} F loc is holomorphic so there are constants B j ( 0 ) B_j^{(0)} B j ( 0 ) and B j k ( 1 ) B_{jk}^{(1)} B j k ( 1 ) satisfying
∣ ( F l o c ) j ( W ) ∣ ≤ B j ( 0 ) , ∣ ∂ k ( F l o c ) j ( W ) ∣ ≤ B j k ( 1 ) (30) |(\mathcal{F}_{\mathrm{loc}})_j(W)|\leq B_j^{(0)},\qquad|\partial_k(\mathcal{F}_{\mathrm{loc}})_j(W)|\leq B_{jk}^{(1)} \tag*{(30)} ∣ ( F loc ) j ( W ) ∣ ≤ B j ( 0 ) , ∣ ∂ k ( F loc ) j ( W ) ∣ ≤ B j k ( 1 ) (30)
on a neighbourhood of P ( W 0 , d ) \mathcal{P}(W_0,\mathbf{d}) P ( W 0 , d ) . For ρ > 0 \rho>0 ρ > 0 , let D ρ = { x ∈ C : ∣ x ∣ < ρ } D_\rho=\{x\in\mathbb{C}:|x|<\rho\} D ρ = { x ∈ C : ∣ x ∣ < ρ } , and let A ρ \mathcal{A}_\rho A ρ be the vector space of continuous maps V : D ρ ‾ → C 7 V:\overline{D_\rho}\to\mathbb{C}^7 V : D ρ → C 7 that are holomorphic on D ρ D_\rho D ρ and satisfy V ( 0 ) = 0 V(0)=0 V ( 0 ) = 0 , equipped with the weighted norm
∥ V ∥ ρ , d = max 1 ≤ j ≤ 7 sup ∣ x ∣ ≤ ρ ∣ V j ( x ) ∣ d j . \|V\|_{\rho,\mathbf{d}}=\max_{1\leq j\leq7}\sup_{|x|\leq\rho}\frac{|V_j(x)|}{d_j}. ∥ V ∥ ρ , d = 1 ≤ j ≤ 7 max ∣ x ∣ ≤ ρ sup d j ∣ V j ( x ) ∣ .
Uniform limits preserve holomorphy on D ρ D_\rho D ρ and the value at zero, so A ρ \mathcal{A}_\rho A ρ is a closed subspace of C ( D ρ ‾ , C 7 ) C(\overline{D_\rho},\mathbb{C}^7) C ( D ρ , C 7 ) . The weighted norm is equivalent to the supremum norm, hence A ρ \mathcal{A}_\rho A ρ is a Banach space. Since S j j ≤ 0 S_{jj}\leq0 S j j ≤ 0 , the kernels in (29) are integrable and ∫ 0 1 s − S j j d s = ( 1 − S j j ) − 1 \int_0^1s^{-S_{jj}}\,ds=(1-S_{jj})^{-1} ∫ 0 1 s − S j j d s = ( 1 − S j j ) − 1 . Consequently, on the closed unit ball in A ρ \mathcal{A}_\rho A ρ ,
∥ T V ∥ ρ , d ≤ ρ K , ∥ T V − T V ~ ∥ ρ , d ≤ ρ Λ ∥ V − V ~ ∥ ρ , d , \|\mathcal{T}V\|_{\rho,\mathbf{d}}\leq\rho K,\qquad\|\mathcal{T}V-\mathcal{T}\widetilde{V}\|_{\rho,\mathbf{d}}\leq\rho\Lambda\|V-\widetilde{V}\|_{\rho,\mathbf{d}}, ∥ T V ∥ ρ , d ≤ ρ K , ∥ T V − T V ∥ ρ , d ≤ ρ Λ∥ V − V ∥ ρ , d ,
where
K j = B j ( 0 ) ( 1 − S j j ) d j , Λ j = ∑ k B j k ( 1 ) d k ( 1 − S j j ) d j , K = max j K j , Λ = max j Λ j . (31) K_j=\frac{B_j^{(0)}}{(1-S_{jj})d_j},\qquad\Lambda_j=\frac{\sum_kB_{jk}^{(1)}d_k}{(1-S_{jj})d_j},\qquad K=\max_jK_j,\qquad\Lambda=\max_j\Lambda_j. \tag*{(31)} K j = ( 1 − S j j ) d j B j ( 0 ) , Λ j = ( 1 − S j j ) d j ∑ k B j k ( 1 ) d k , K = j max K j , Λ = j max Λ j . (31)
For sufficiently small ρ \rho ρ , one has ρ K < 1 \rho K < 1 ρ K < 1 and ρ Λ < 1 \rho\Lambda< 1 ρ Λ < 1 , so T \mathcal{T} T is a strict contraction of the closed unit ball into itself. Its fixed point is holomorphic in x x x . The same choices can be made uniformly for parameters in a small complex neighbourhood of any fixed triple, and the contraction iterates then show that the fixed point is holomorphic as a function of ( x , α 0 , α 1 , α 2 ) (x,\alpha_0,\alpha_1,\alpha_2) ( x , α 0 , α 1 , α 2 ) . For real parameters, uniqueness and complex conjugation show that the solution is real.
Choose 0 < ε < ρ 0 < \varepsilon< \rho 0 < ε < ρ , so the analytic fixed point gives a smooth solution on [ 0 , ε ] [0,\varepsilon] [ 0 , ε ] . To prove uniqueness among smooth solutions, fix real parameters and, for 0 < δ ≤ ε 0 < \delta\leq\varepsilon 0 < δ ≤ ε , consider the real Banach space
X δ = { V ∈ C ( [ 0 , δ ] , R 7 ) : V ( 0 ) = 0 } , ∥ V ∥ δ , d = max j sup 0 ≤ x ≤ δ ∣ V j ( x ) ∣ d j . \mathcal{X}_{\delta} = \{V \in C([0,\delta],\mathbb{R}^{7}) : V(0) = 0\}, \qquad\lVert V\rVert_{\delta,\mathbf{d}} = \max_j \sup_{0\leq x\leq\delta} \frac{\lvert V_j(x)\rvert}{d_j}. X δ = { V ∈ C ([ 0 , δ ] , R 7 ) : V ( 0 ) = 0 } , ∥ V ∥ δ , d = j max 0 ≤ x ≤ δ sup d j ∣ V j ( x )∣ .
The same estimates for (29) show that T \mathcal{T} T maps its closed unit ball into itself and is a contraction with factor at most δ Λ < 1 \delta\Lambda< 1 δ Λ < 1 .
Let W = W 0 + V W = W_0 + V W = W 0 + V be any smooth solution with the prescribed initial value. For x > 0 x > 0 x > 0 , (28) gives
d d x ( x − S j j V j ( x ) ) = x − S j j ( F l o c ( W 0 + V ( x ) ) ) j . \frac{d}{dx}\left(x^{-S_{jj}}V_j(x)\right)=x^{-S_{jj}}\left(\mathcal{F}_{\mathrm{loc}}(W_0+V(x))\right)_j. d x d ( x − S j j V j ( x ) ) = x − S j j ( F loc ( W 0 + V ( x )) ) j .
Since S j j ≤ 0 S_{jj} \leq0 S j j ≤ 0 and V ( 0 ) = 0 V(0) = 0 V ( 0 ) = 0 , the expression inside the derivative tends to zero as x ↓ 0 x \downarrow0 x ↓ 0 . Integrating from zero therefore gives (29) . After reducing δ \delta δ if necessary, both V V V and the restriction of the analytic fixed point lie in the closed unit ball of X δ \mathcal{X}_{\delta} X δ . Uniqueness of the fixed point in this real function space makes them equal there. Uniqueness for the regular ODE on x > 0 x>0 x > 0 then extends this equality throughout their common interval of definition. □
For the quantitative argument we estimate the errors in truncated Taylor series of this local solution. Regard W W W as a function of the distance coordinate t t t , with x = t 2 x=t^2 x = t 2 . Since S W 0 = 0 S W_0=0 S W 0 = 0 , (28) gives
d W d t = G ( t , W ) : = 2 t F l o c ( W ) + 2 S W t . (32) \frac{dW}{dt}=G(t,W):=2t\mathcal{F}_{\mathrm{loc}}(W)+\frac{2SW}{t}. \tag*{(32)} d t d W = G ( t , W ) := 2 t F loc ( W ) + t 2 S W . (32)
Write
W = W 0 + ∑ n ≥ 1 W [ n ] x n . W=W_0+\sum_{n\geq1}W^{[n]}x^n. W = W 0 + n ≥ 1 ∑ W [ n ] x n .
Comparing coefficients in (28) gives the recurrence
( n Id − S ) W [ n ] = [ F l o c ( W 0 + ∑ j = 1 n − 1 W [ j ] x j ) ] n − 1 . (33) (n\operatorname{Id}-S)W^{[n]}=\left[\mathcal{F}_{\mathrm{loc}}\left(W_0+\sum_{j=1}^{n-1}W^{[j]}x^j\right)\right]_{n-1}. \tag*{(33)} ( n Id − S ) W [ n ] = [ F loc ( W 0 + j = 1 ∑ n − 1 W [ j ] x j ) ] n − 1 . (33)
Here W [ n ] W^{[n]} W [ n ] is a vector Taylor coefficient, while W j W_j W j denotes the j j j th scalar component of W W W . The notation [ ⋅ ] n − 1 [\cdot]_{n-1} [ ⋅ ] n − 1 denotes the coefficient of x n − 1 x^{n-1} x n − 1 . Every diagonal entry of n Id − S n\operatorname{Id}-S n Id − S is positive for n ≥ 1 n\geq1 n ≥ 1 , so this recurrence determines W [ n ] W^{[n]} W [ n ] uniquely from the preceding coefficients. To bound the truncation error, set
P N ( x ) = W 0 + ∑ n = 1 N W [ n ] x n . P_N(x)=W_0+\sum_{n=1}^{N}W^{[n]}x^n. P N ( x ) = W 0 + n = 1 ∑ N W [ n ] x n .
Whenever the contraction conditions above hold for ρ \rho ρ and d \mathbf{d} d , the corresponding solution satisfies ∣ W j − W 0 j ∣ ≤ d j \lvert W_j-W_{0j}\rvert\leq d_j ∣ W j − W 0 j ∣ ≤ d j on ∣ x ∣ ≤ ρ \lvert x\rvert\leq\rho ∣ x ∣ ≤ ρ . For 0 < r < ρ 0<r<\rho 0 < r < ρ , Cauchy’s coefficient estimate gives
∣ ( W [ n ] ) j ∣ ≤ r − n max ∣ x ∣ = r ∣ W j ( x ) − W 0 j ∣ ≤ d j r − n , n ≥ 1. \left\lvert\left(W^{[n]}\right)_j\right\rvert\leq r^{-n}\max_{\lvert x\rvert=r}\lvert W_j(x)-W_{0j}\rvert\leq d_jr^{-n}, \qquad n\geq1. ( W [ n ] ) j ≤ r − n ∣ x ∣ = r max ∣ W j ( x ) − W 0 j ∣ ≤ d j r − n , n ≥ 1.
Letting r ↑ ρ r\uparrow\rho r ↑ ρ gives ∣ ( W [ n ] ) j ∣ ≤ d j ρ − n \left\lvert\left(W^{[n]}\right)_j\right\rvert\leq d_j\rho^{-n} ( W [ n ] ) j ≤ d j ρ − n . Summing the resulting geometric tail yields
∣ ( W − P N ) j ( x ) ∣ ≤ d j ( ∣ x ∣ / ρ ) N + 1 1 − ∣ x ∣ / ρ , ∣ x ∣ < ρ . (34) \lvert(W-P_N)_j(x)\rvert\leq d_j\frac{(\lvert x\rvert/\rho)^{N+1}}{1-\lvert x\rvert/\rho}, \qquad\lvert x\rvert<\rho. \tag*{(34)} ∣( W − P N ) j ( x )∣ ≤ d j 1 − ∣ x ∣ / ρ (∣ x ∣ / ρ ) N + 1 , ∣ x ∣ < ρ . (34)
No holomorphic extension beyond ∣ x ∣ ≤ ρ \lvert x\rvert\leq\rho ∣ x ∣ ≤ ρ is needed.
Numerical candidate and choice of verification data
The initial exploration used truncated Taylor series to start the numerical integration near each singular orbit. An adaptive Dormand–Prince 5(4) solver [13 ] continued each solution inward to its first zero of ξ \xi ξ . The strict decrease of ξ \xi ξ makes this a natural section for comparing the two trajectories. On this section, the metric coefficients must agree and the shape variables must have opposite signs. A damped least-squares iteration of Levenberg–Marquardt type [18 ] , with a central-difference Jacobian, sought to satisfy these matching conditions.
The initial numerical search varied the values of f 1 , f 2 , f 3 f_1,f_2,f_3 f 1 , f 2 , f 3 at the two singular orbits independently, without imposing f 2 = f 3 f_2=f_3 f 2 = f 3 . We checked the numerical implementation by comparing its results for the known Kähler shrinking soliton on the same bundle with an independent quadrature of that solution [10 §4] . Of sixteen starting points, four approached the candidate described below, nine approached the known Kähler soliton, and three reached imposed search bounds with a nonzero residual. The new candidate exhibited f 2 = f 3 f_2=f_3 f 2 = f 3 numerically. This suggested using the exactly invariant subfamily established in § 2 . Details of the recorded search and of the reproducible numerical refinement are given in Subsection A.3.
At the matching times, we compare the solution data in the coordinates
Y ^ = C ( Y ) = ( f 0 , f 1 2 , f 2 2 , L 0 , L 1 , L 2 , ξ ) , \widehat{Y}=C(Y)=(f_0,f_1^2,f_2^2,L_0,L_1,L_2,\xi), Y = C ( Y ) = ( f 0 , f 1 2 , f 2 2 , L 0 , L 1 , L 2 , ξ ) ,
which avoid taking square roots when converting from W W W by (39) . We parametrise the data at the two singular orbits by
α ^ = ( α 0 , α 1 2 , α 2 2 ) , ω ^ = ( ω 0 , ω 1 2 , ω 2 2 ) , \widehat{\alpha}=(\alpha_0,\alpha_1^2,\alpha_2^2), \qquad\widehat{\omega}=(\omega_0,\omega_1^2,\omega_2^2), α = ( α 0 , α 1 2 , α 2 2 ) , ω = ( ω 0 , ω 1 2 , ω 2 2 ) ,
and vary the seven parameters p = ( α ^ , ω ^ , T R ) p=(\widehat{\alpha},\widehat{\omega},T_R) p = ( α , ω , T R ) . Let p 0 p_0 p 0 denote the exact vector used for verification. Table 1 displays its rounded coordinates together with the fixed left comparison time T L T_L T L .
Table 1. Numerical candidate: decimal displays of the verification inputs, rounded to twelve decimal places.
In particular, the choice T L ≈ 1.74836775414244 T_L\approx1.74836775414244 T L ≈ 1.74836775414244 comes from the left trajectory’s numerical ξ = 0 \xi=0 ξ = 0 hitting time. For the rigorous calculation, we fix its stored floating-point value and interpret it as the exact rational number
T L = 1968487091515621 2 50 . (35) T_L=\frac{1968487091515621}{2^{50}}. \tag*{(35)} T L = 2 50 1968487091515621 . (35)
We allow T R T_R T R to vary. Fixing T L T_L T L determines the location at which the trajectories are compared and removes the freedom to move that location along a joined solution. The fixed-time matching problem does not impose ξ ( T L ) = 0 \xi(T_L)=0 ξ ( T L ) = 0 , although prescribing this condition and allowing both T L T_L T L and T R T_R T R to vary would give an alternative construction.
The matching residual compares the left trajectory at inward time T L T_L T L with the right trajectory at inward time T R T_R T R . Its first three components are the left values of f 0 , f 1 2 , f 2 2 f_0,f_1^2,f_2^2 f 0 , f 1 2 , f 2 2 minus their right values; its remaining four components are the sums of the left and right values of L 0 , L 1 , L 2 , ξ L_0,L_1,L_2,\xi L 0 , L 1 , L 2 , ξ . Let F n u m ( p ) \mathcal{F}_{\mathrm{num}}(p) F num ( p ) denote the seven-component matching residual computed by numerically continuing the solutions from the two singular orbits with parameters p p p . We evaluate this residual at p 0 p_0 p 0 using the Taylor recurrence in (33) and ordinary-time Taylor integration with 60-digit decimal arithmetic. The largest absolute component of the computed residual is
∥ F n u m ( p 0 ) ∥ ∞ = max 1 ≤ i ≤ 7 ∣ ( F n u m ( p 0 ) ) i ∣ ≈ 6.57 ⋅ 10 − 14 . \left\|\mathcal{F}_{\mathrm{num}}(p_0)\right\|_{\infty}=\max_{1\leq i\leq7}\left|\left(\mathcal{F}_{\mathrm{num}}(p_0)\right)_i\right|\approx6.57\cdot10^{-14}. ∥ F num ( p 0 ) ∥ ∞ = 1 ≤ i ≤ 7 max ∣ ( F num ( p 0 ) ) i ∣ ≈ 6.57 ⋅ 1 0 − 14 .
Table 2 shows its stability as the integration order increases. Every row uses the same parameters p 0 p_0 p 0 ; only the integration order changes.
Table 2. Numerical residuals at the fixed candidate p 0 p_0 p 0 . The second column gives the largest absolute component of the computed matching residual. The third gives the largest absolute difference between corresponding components of this residual and the reference residual. All runs start at t 0 = 1 / 32 t_0=1/32 t 0 = 1/32 using a degree-32 Taylor polynomial in x = t 2 x=t^2 x = t 2 , and use integration steps h ≤ min ( 1 / 32 , t / 4 ) h\leq\min(1/32,t/4) h ≤ min ( 1/32 , t /4 ) . The reference computation uses integration order 30 and h ≤ min ( 1 / 64 , t / 4 ) h\leq\min(1/64,t/4) h ≤ min ( 1/64 , t /4 ) . These comparisons are numerical observations, not rigorous error bounds.
Let C C C be a fixed numerical approximation to the inverse Jacobian of the matching residual at p 0 p_0 p 0 ; its exact rational entries are listed in Appendix § 7 . Starting instead with every coordinate of p 0 p_0 p 0 rounded to six decimal places, the iteration p ( k + 1 ) = p ( k ) − C F n u m ( p ( k ) ) p^{(k+1)}=p^{(k)}-CF_{\mathrm{num}}(p^{(k)}) p ( k + 1 ) = p ( k ) − C F num ( p ( k ) ) gives successive residual norms of approximately 1.73 ⋅ 10 − 6 1.73\cdot10^{-6} 1.73 ⋅ 1 0 − 6 , 2.10 ⋅ 10 − 12 2.10\cdot10^{-12} 2.10 ⋅ 1 0 − 12 , 3.94 ⋅ 10 − 22 3.94\cdot10^{-22} 3.94 ⋅ 1 0 − 22 , and 4.30 ⋅ 10 − 31 4.30\cdot10^{-31} 4.30 ⋅ 1 0 − 31 . The largest absolute difference between corresponding coordinates of the refined point and p 0 p_0 p 0 is approximately 1.55 ⋅ 10 − 13 1.55\cdot10^{-13} 1.55 ⋅ 1 0 − 13 . The rigorous verification uses the original rational inputs p 0 p_0 p 0 , specified together with the matching map in § 4 .
Figure 1. Numerical metric functions f 0 f_0 f 0 , f 1 f_1 f 1 , f 2 = f 3 f_2=f_3 f 2 = f 3 and potential derivative u ′ ( t ) u'(t) u ′ ( t ) at the fixed candidate p 0 p_0 p 0 , plotted on a common scale. The soliton vector field is ∇ g u = u ′ ( t ) ∂ t \nabla_g u=u'(t)\partial_t ∇ g u = u ′ ( t ) ∂ t , where u ′ = L 0 + 2 L 1 + 4 L 2 − ξ u'=L_0+2L_1+4L_2-\xi u ′ = L 0 + 2 L 1 + 4 L 2 − ξ . The right trajectory, computed in inward distance s s s , is plotted at t = T L + T R − s t=T_L+T_R-s t = T L + T R − s , with the sign of its potential derivative reversed. The vertical grey line marks the matching location. The two numerical pieces agree to plotting accuracy; exact matching is proved in § 4 .
Uniform estimates and Taylor remainders
Decimals used as input bounds denote exact rational numbers unless explicitly identified as 64-bit floating-point values. All root data and preconditioner entries are specified as exact rational numbers. Using the coordinates ( α 0 , α 1 2 , α 2 2 ) (\alpha_0,\alpha_1^2,\alpha_2^2) ( α 0 , α 1 2 , α 2 2 ) and ( ω 0 , ω 1 2 , ω 2 2 ) (\omega_0,\omega_1^2,\omega_2^2) ( ω 0 , ω 1 2 , ω 2 2 ) introduced in Subsection 3.5 makes the initial data W 0 W_0 W 0 rational in these parameters. We obtain the two chart centres by rounding the first and second triples of the numerical candidate p 0 p_0 p 0 to twelve decimal places, as displayed in Table 1 :
α ^ c = ( 2.296157352821 , 0.085184676795 , 0.162557980826 ) , ω ^ c = ( − 2.338419627964 , 0.415303147163 , 0.941298586351 ) . (36) \begin{aligned}
\widehat{\alpha}_c &= (2.296157352821, 0.085184676795, 0.162557980826),\\
\widehat{\omega}_c &= (-2.338419627964, 0.415303147163, 0.941298586351). \tag*{(36)}
\end{aligned} α c ω c = ( 2.296157352821 , 0.085184676795 , 0.162557980826 ) , = ( − 2.338419627964 , 0.415303147163 , 0.941298586351 ) . (36)
For each centre in (36) , allow the three coordinates to vary independently in complex discs of radius ϵ = 10 − 4 \epsilon=10^{-4} ϵ = 1 0 − 4 . We also use the products of three discs of radius δ = 10 − 6 \delta=10^{-6} δ = 1 0 − 6 about the same centres. The gap between the radii allows Cauchy derivative estimates. Recall the radius vector d = ( d 1 , … , d 7 ) \mathbf d=(d_1,\ldots,d_7) d = ( d 1 , … , d 7 ) from the proof of Theorem 3.2: its entries specify the domain ∣ W j − W 0 j ∣ ≤ d j |W_j-W_{0j}|\leq d_j ∣ W j − W 0 j ∣ ≤ d j on which the local solution is constructed. The constants K K K and Λ \Lambda Λ in (31) control the integral operator and its contraction estimate.
For both families of local solutions, take
d = ( 1 / 50 , 1 / 500 , 1 / 500 , 1 / 5 , 1 / 20 , 1 / 20 , 1 / 10 ) . (37) \mathbf d=(1/50,1/500,1/500,1/5,1/20,1/20,1/10). \tag*{(37)} d = ( 1/50 , 1/500 , 1/500 , 1/5 , 1/20 , 1/20 , 1/10 ) . (37)
These choices keep f 1 2 , f 2 2 f_1^2,f_2^2 f 1 2 , f 2 2 away from zero and permit a common radius ρ = 1 / 256 \rho=1/256 ρ = 1/256 in x x x , with ρ K < 1 \rho K<1 ρ K < 1 and ρ Λ < 1 \rho\Lambda<1 ρ Λ < 1 . Recall the bounds in (30) : B j ( 0 ) B_j^{(0)} B j ( 0 ) bounds the absolute value of the j j j th component of F l o c \mathcal F_{\mathrm{loc}} F loc , and B j k ( 0 ) B_{jk}^{(0)} B j k ( 0 ) bounds the absolute value of its derivative with respect to W k W_k W k . For each family, these bounds hold uniformly on a product of complex discs containing every W 0 + v W_0+v W 0 + v with ∣ v j ∣ ≤ d j |v_j|\leq d_j ∣ v j ∣ ≤ d j , as the three parameters vary independently in their complex discs of radius ϵ = 10 − 4 \epsilon=10^{-4} ϵ = 1 0 − 4 . We refer to this product as the domain for the solution data.
Table 3 gives rational upper bounds for K j K_j K j and Λ j \Lambda_j Λ j from (31) , uniformly over the complex discs of radius 10 − 4 10^{-4} 1 0 − 4 . Each bound is rounded upward to a multiple of 10 − 3 10^{-3} 1 0 − 3 .
Table 3. Bounds for the local estimates in each coordinate. The four numerical columns give numerators of upper bounds with common denominator 10 3 10^3 1 0 3 .
Proposition 3.3. On each product of three complex discs of radius 10 − 4 10^{-4} 1 0 − 4 , the operator in (29) has a unique fixed point with ∣ W j − W 0 j ∣ ≤ d j |W_j-W_{0j}|\leq d_j ∣ W j − W 0 j ∣ ≤ d j on ∣ x ∣ ≤ ρ = 1 / 256 |x|\leq\rho=1/256 ∣ x ∣ ≤ ρ = 1/256 . The constants in (31) satisfy
K L < 81 , K R < 144 , Λ L < 8 , Λ R < 25. K_L<81,\qquad K_R<144,\qquad\Lambda_L<8,\qquad\Lambda_R<25. K L < 81 , K R < 144 , Λ L < 8 , Λ R < 25. The real solutions have positive f 1 , f 2 f_1,f_2 f 1 , f 2 and f ~ 0 ≥ 49 / 50 \tilde f_0\geq49/50 f ~ 0 ≥ 49/50 .
Proof. We verify the bounds needed for the contraction argument in Theorem 3.2, uniformly for parameters in their complex discs of radius 10 − 4 10^{-4} 1 0 − 4 . We use complex discs so that the resulting bounds also justify the Cauchy estimates for the Taylor remainders and their parameter derivatives.
Write
D ( c , r ) = { z ∈ C : ∣ z − c ∣ ≤ r } ⊂ C , c ∈ R , r ≥ 0. D(c,r)=\{z\in\mathbb C:|z-c|\leq r\}\subset\mathbb C,\qquad c\in\mathbb R,\quad r\geq0. D ( c , r ) = { z ∈ C : ∣ z − c ∣ ≤ r } ⊂ C , c ∈ R , r ≥ 0.
For products of two such discs, we have
D ( c , r ) D ( c ′ , r ′ ) ⊂ D ( c c ′ , ∣ c ∣ r ′ + ∣ c ′ ∣ r + r r ′ ) . D(c,r)D(c',r')\subset D\left(cc',|c|r'+|c'|r+rr'\right). D ( c , r ) D ( c ′ , r ′ ) ⊂ D ( c c ′ , ∣ c ∣ r ′ + ∣ c ′ ∣ r + r r ′ ) .
Indeed, writing z = c + δ z=c+\delta z = c + δ and z ′ = c ′ + δ ′ z'=c'+\delta' z ′ = c ′ + δ ′ gives z z ′ − c c ′ = c δ ′ + c ′ δ + δ δ ′ zz'-cc'=c\delta'+c'\delta+\delta\delta' z z ′ − c c ′ = c δ ′ + c ′ δ + δ δ ′ , and the triangle inequality gives the displayed radius. Addition adds the centres and radii. If ∣ c ∣ > r |c|>r ∣ c ∣ > r , the disc avoids zero and its reciprocal is exactly
1 D ( c , r ) = D ( c c 2 − r 2 , r c 2 − r 2 ) . \frac{1}{D(c,r)}=D\left(\frac{c}{c^2-r^2},\frac{r}{c^2-r^2}\right). D ( c , r ) 1 = D ( c 2 − r 2 c , c 2 − r 2 r ) .
All centres and radii used in the computation are rational, so these operations can be performed in exact rational arithmetic.
For each family, we bound W 0 W_0 W 0 on the complex discs of radius $10^{-4}$ by a product of discs, then increase its j j j th coordinate radius by d j d_j d j from (37) . The resulting domain contains every W 0 + V ( x ) W_0+V(x) W 0 + V ( x ) with ∥ V ∥ ρ , d ≤ 1 \lVert V\rVert_{\rho,d}\leq1 ∥ V ∥ ρ , d ≤ 1 , for all allowed data. Since W 2 = f 1 2 W_2=f_1^2 W 2 = f 1 2 and W 3 = f 2 2 W_3=f_2^2 W 3 = f 2 2 , their disc radii are $\epsilon+d_2=\epsilon+d_3=0.0021. Subtracting these radii from the positive centres in (36) gives the following lower bounds for their absolute values throughout these domains:
The only possible poles of F l o c \mathcal{F}_{\mathrm{loc}} F loc and its derivatives occur at W 2 W 3 = 0 W_2W_3=0 W 2 W 3 = 0 , so these positive bounds justify the reciprocal operations.
The routine for local estimates in verify.py applies this disc arithmetic to the explicit rational formula in (27) , using automatic differentiation to bound its first partial derivatives as well. It returns rational bounds
∣ ( F l o c ) j ( W ) ∣ ≤ B j ( 0 ) , ∣ ∂ k ( F l o c ) j ( W ) ∣ ≤ B j k ( 1 ) \lvert(\mathcal{F}_{\mathrm{loc}})_j(W)\rvert\le B_j^{(0)}, \qquad\lvert\partial_k(\mathcal{F}_{\mathrm{loc}})_j(W)\rvert\le B_{jk}^{(1)} ∣( F loc ) j ( W )∣ ≤ B j ( 0 ) , ∣ ∂ k ( F loc ) j ( W )∣ ≤ B j k ( 1 )
valid throughout each domain. These give the upper bounds for K j K_j K j and Λ j \Lambda_j Λ j recorded in Table 3 . Taking the largest entry in each column and dividing by 10 3 10^3 1 0 3 gives
K L ≤ 80726 10 3 < 81 , K R ≤ 143735 10 3 < 144 , Λ L ≤ 7289 10 3 < 8 , Λ R ≤ 24164 10 3 < 25. K_L \le\frac{80726}{10^3}<81,\qquad K_R\le\frac{143735}{10^3}<144,\qquad\Lambda_L\le\frac{7289}{10^3}<8,\qquad\Lambda_R\le\frac{24164}{10^3}<25. K L ≤ 1 0 3 80726 < 81 , K R ≤ 1 0 3 143735 < 144 , Λ L ≤ 1 0 3 7289 < 8 , Λ R ≤ 1 0 3 24164 < 25.
With ρ = 1 / 256 \rho=1/256 ρ = 1/256 , the bounds on K L , K R , Λ L , Λ R K_L,K_R,\Lambda_L,\Lambda_R K L , K R , Λ L , Λ R give, for either singular orbit,
ρ K ≤ 144 256 = 9 16 < 1 , ρ Λ ≤ 25 256 < 1 10 . \rho K\le\frac{144}{256}=\frac{9}{16}<1,\qquad\rho\Lambda\le\frac{25}{256}<\frac{1}{10}. ρ K ≤ 256 144 = 16 9 < 1 , ρ Λ ≤ 256 25 < 10 1 .
By (31) , the first inequality ensures that T \mathcal{T} T maps the closed unit ball into itself, and the second makes it a strict contraction. Thus each choice of data in the discs of radius 10 − 4 10^{-4} 1 0 − 4 gives a unique fixed point in this ball. Setting W = W 0 + V W=W_0+V W = W 0 + V gives ∣ W j − W 0 j ∣ ≤ d j \lvert W_j-W_{0j}\rvert\le d_j ∣ W j − W 0 j ∣ ≤ d j on ∣ x ∣ ≤ 1 / 256 \lvert x\rvert\le1/256 ∣ x ∣ ≤ 1/256 .
For real parameters and real x x x , the solution is real by Theorem 3.2. The discs for W 2 W_2 W 2 and W 3 W_3 W 3 have positive centres larger than their radii, so their real values are positive; choosing the positive square roots gives f 1 , f 2 > 0 f_1,f_2>0 f 1 , f 2 > 0 . Finally, W 01 = 1 W_{01}=1 W 01 = 1 and d 1 = 1 / 50 d_1=1/50 d 1 = 1/50 , so
f ~ 0 = W 1 ≥ 1 − 1 50 = 49 50 . \widetilde f_0=W_1\ge1-\frac{1}{50}=\frac{49}{50}. f 0 = W 1 ≥ 1 − 50 1 = 50 49 .
For the parameter derivative estimates, restrict each parameter to the complex disc of radius δ = 10 − 6 \delta=10^{-6} δ = 1 0 − 6 about its centre. For either family, write ζ = ( ζ 1 , ζ 2 , ζ 3 ) \zeta=(\zeta_1,\zeta_2,\zeta_3) ζ = ( ζ 1 , ζ 2 , ζ 3 ) for α ^ = ( α 0 , α 1 2 , α 2 2 ) \hat{\alpha}=(\alpha_0,\alpha_1^2,\alpha_2^2) α ^ = ( α 0 , α 1 2 , α 2 2 ) or ω ^ = ( ω 0 , ω 1 2 , ω 2 2 ) \hat{\omega}=(\omega_0,\omega_1^2,\omega_2^2) ω ^ = ( ω 0 , ω 1 2 , ω 2 2 ) . Parameter derivatives below are taken with respect to these three entries, in the order written. The tail E N = W − P N E_N=W-P_N E N = W − P N is holomorphic as a function of ( x , ζ 1 , ζ 2 , ζ 3 ) (x,\zeta_1,\zeta_2,\zeta_3) ( x , ζ 1 , ζ 2 , ζ 3 ) , and (34) holds uniformly when the parameters lie in the discs of radius ϵ = 10 − 4 \epsilon=10^{-4} ϵ = 1 0 − 4 . If each parameter is within δ \delta δ of its centre, a circle of radius η < ϵ − δ \eta<\epsilon-\delta η < ϵ − δ about that parameter stays inside its disc of radius ϵ \epsilon ϵ . Applying the one-variable Cauchy estimate to E N E_N E N , holding the other parameters fixed, and then taking the limit as η \eta η approaches ϵ − δ = 99 / 10 6 \epsilon-\delta=99/10^6 ϵ − δ = 99/1 0 6 from below gives
∣ ∂ ζ i ( E N ) j ( x ) ∣ ≤ d j ϵ − δ ( ∣ x ∣ / ρ ) N + 1 1 − ∣ x ∣ / ρ , i = 1 , 2 , 3. \lvert\partial_{\zeta_i}(E_N)_j(x)\rvert\le\frac{d_j}{\epsilon-\delta}\frac{(\lvert x\rvert/\rho)^{N+1}}{1-\lvert x\rvert/\rho},\qquad i=1,2,3. ∣ ∂ ζ i ( E N ) j ( x )∣ ≤ ϵ − δ d j 1 − ∣ x ∣ / ρ (∣ x ∣ / ρ ) N + 1 , i = 1 , 2 , 3.
For the launch, choose x 0 = ρ / 4 x_0=\rho/4 x 0 = ρ /4 , so the geometric factor in the tail estimate is 4 − N / 3 4^{-N}/3 4 − N /3 . Taking N = 32 N=32 N = 32 gives the bounds for the solution data and their parameter derivatives displayed below. Thus we use
t 0 = 1 / 32 , x 0 = 1 / 1024 , N = 32. (38) t_0=1/32,\qquad x_0=1/1024,\qquad N=32. \tag*{(38)} t 0 = 1/32 , x 0 = 1/1024 , N = 32. (38)
Since x 0 / ρ = 1 / 4 x_0/\rho=1/4 x 0 / ρ = 1/4 and max j d j = 1 / 5 \max_j d_j=1/5 max j d j = 1/5 , the common coordinatewise bounds are
max j ∣ ( E 32 ) j ( x 0 ) ∣ ≤ 1 15 ⋅ 2 64 < 2 − 67 , max i , j ∣ ∂ ζ i ( E 32 ) j ( x 0 ) ∣ ≤ 10 6 1485 ⋅ 2 64 < 2 − 54 . \max_j\lvert(E_{32})_j(x_0)\rvert\le\frac{1}{15\cdot2^{64}}<2^{-67},\qquad\max_{i,j}\lvert\partial_{\zeta_i}(E_{32})_j(x_0)\rvert\le\frac{10^6}{1485\cdot2^{64}}<2^{-54}. j max ∣( E 32 ) j ( x 0 )∣ ≤ 15 ⋅ 2 64 1 < 2 − 67 , i , j max ∣ ∂ ζ i ( E 32 ) j ( x 0 )∣ ≤ 1485 ⋅ 2 64 1 0 6 < 2 − 54 .
These bounds hold whenever each parameter lies in its complex disc of radius δ = 10 − 6 \delta=10^{-6} δ = 1 0 − 6 .
To formulate and rigorously verify the matching conditions in § 4 , we use the coordinate chart C \mathcal{C} C for the solution data introduced in Subsection 3.5. Using f 1 2 f_1^2 f 1 2 and f 2 2 f_2^2 f 2 2 as coordinates allows conversion from W W W without taking square roots. This conversion is the rational map
W ⟼ C ( Y ) = ( t f ~ 0 , f 1 2 , f 2 2 , t − 1 + t ℓ 0 , t ℓ 1 , t ℓ 2 , t − 1 + t ξ ) , (39) W\longmapsto\mathcal{C}(Y)=\left(t\widetilde f_0,f_1^2,f_2^2,t^{-1}+t\ell_0,t\ell_1,t\ell_2,t^{-1}+t\xi\right), \tag*{(39)} W ⟼ C ( Y ) = ( t f 0 , f 1 2 , f 2 2 , t − 1 + t ℓ 0 , t ℓ 1 , t ℓ 2 , t − 1 + t ξ ) , (39)
with the substitutions from (26) . The validated integration is carried out in the W W W variables. The initial interval bounds include the tails for the solution data and their parameter derivatives given above. At a positive terminal time, this map and its chain rule convert the propagated interval bounds into bounds for C ( Y ) \mathcal C(Y) C ( Y ) and its parameter derivatives, using outward interval arithmetic.
The shooting problem
We seek a zero of a seven-dimensional matching map F \mathcal F F in a small box around the numerical candidate p 0 p_0 p 0 . The proof requires sharp interval bounds for F ( p 0 ) \mathcal F(p_0) F ( p 0 ) and uniform interval bounds for D F D\mathcal F D F on the box. We obtain these by validated integration of the solutions from the two singular orbits and their parameter derivatives, including the dependence on the right shooting time. The resulting bounds give an exact match by a Newton–Kantorovich argument.
The seven-dimensional matching map
Let
Y L ( t ; α 0 , α 1 , α 2 ) , Y R ( t ; ω 0 , ω 1 , ω 2 ) Y_L(t;\alpha_0,\alpha_1,\alpha_2), \qquad Y_R(t;\omega_0,\omega_1,\omega_2) Y L ( t ; α 0 , α 1 , α 2 ) , Y R ( t ; ω 0 , ω 1 , ω 2 )
be the two inward local solution families supplied by Theorem 3.2. We use the fixed left comparison time T L T_L T L from (35) and the rational coordinates
p = ( α 0 , α 1 2 , α 2 2 , ω 0 , ω 1 2 , ω 2 2 , T R ) . (40) p=(\alpha_0,\alpha_1^2,\alpha_2^2,\omega_0,\omega_1^2,\omega_2^2,T_R). \tag*{(40)} p = ( α 0 , α 1 2 , α 2 2 , ω 0 , ω 1 2 , ω 2 2 , T R ) . (40)
Its first three coordinates form α ^ \hat\alpha α ^ , and the next three form ω ^ \hat\omega ω ^ . We compare the solution data through the chart C \mathcal C C , which is injective when f 1 , f 2 > 0 f_1,f_2>0 f 1 , f 2 > 0 . Let D \mathcal D D consist of the points p p p with α 1 , α 2 , ω 1 , ω 2 , T R > 0 \alpha_1,\alpha_2,\omega_1,\omega_2,T_R>0 α 1 , α 2 , ω 1 , ω 2 , T R > 0 for which both solutions exist with f 0 , f 1 , f 2 > 0 f_0,f_1,f_2>0 f 0 , f 1 , f 2 > 0 at positive inward times through T L T_L T L and T R T_R T R , respectively. This set is open by analytic parameter dependence and openness of the regular ODE’s existence domain. On D \mathcal D D define
F ( p ) = C ( Y L ( T L ; α 0 , α 1 , α 2 ) ) − R C ( Y R ( T R ; ω 0 , ω 1 , ω 2 ) ) , R = diag ( 1 , 1 , 1 , − 1 , − 1 , − 1 , − 1 ) . (41) \mathcal F(p)=\mathcal C\bigl(Y_L(T_L;\alpha_0,\alpha_1,\alpha_2)\bigr)-R\mathcal C\bigl(Y_R(T_R;\omega_0,\omega_1,\omega_2)\bigr), \qquad R=\operatorname{diag}(1,1,1,-1,-1,-1,-1). \tag*{(41)} F ( p ) = C ( Y L ( T L ; α 0 , α 1 , α 2 ) ) − R C ( Y R ( T R ; ω 0 , ω 1 , ω 2 ) ) , R = diag ( 1 , 1 , 1 , − 1 , − 1 , − 1 , − 1 ) . (41)
The six parameters prescribed at the singular orbits and T R T_R T R give seven unknowns. We fix T L T_L T L and allow T R T_R T R to vary. The matching orbit need not satisfy ξ = 0 \xi=0 ξ = 0 . Matching the shape variables and ξ \xi ξ also matches the potential derivatives, since u ′ = L 0 + 2 L 1 + 4 L 2 − ξ u'=L_0+2L_1+4L_2-\xi u ′ = L 0 + 2 L 1 + 4 L 2 − ξ . Hence, we do not need a separate matching equation for u ′ u' u ′ .
Proposition 4.1. Every zero of F \mathcal F F in D \mathcal D D determines a smooth G G G -invariant gradient shrinking Ricci soliton ( g , u ) (g,u) ( g , u ) on the compact manifold M M M defined in (5) .
Proof. Let p ∗ p^* p ∗ be a zero of F \mathcal F F and let T = T L + T R ∗ T=T_L+T_R^* T = T L + T R ∗ . Define the solution data on ( 0 , T ) (0,T) ( 0 , T ) by
Y ( t ) = { Y L ( t ; α 0 ∗ , α 1 ∗ , α 2 ∗ ) , 0 < t ≤ T L , R Y R ( T − t ; ω 0 ∗ , ω 1 ∗ , ω 2 ∗ ) , T L ≤ t < T . Y(t)=
\begin{cases}
Y_L(t;\alpha_0^*,\alpha_1^*,\alpha_2^*), & 0<t\le T_L,\\
R Y_R(T-t;\omega_0^*,\omega_1^*,\omega_2^*), & T_L\le t<T.
\end{cases} Y ( t ) = { Y L ( t ; α 0 ∗ , α 1 ∗ , α 2 ∗ ) , R Y R ( T − t ; ω 0 ∗ , ω 1 ∗ , ω 2 ∗ ) , 0 < t ≤ T L , T L ≤ t < T .
Since F ( p ∗ ) = 0 \mathcal F(p^*)=0 F ( p ∗ ) = 0 , C ( R Y ) = R C ( Y ) \mathcal C(RY)=R\mathcal C(Y) C ( R Y ) = R C ( Y ) , and C \mathcal C C is injective when f 1 , f 2 > 0 f_1,f_2>0 f 1 , f 2 > 0 , the two expressions for Y Y Y agree at t = T L t=T_L t = T L . Write (20) as Y ′ = V ( Y ) Y'=\mathcal V(Y) Y ′ = V ( Y ) . Since V ( R Y ) = − R V ( Y ) \mathcal V(RY)=-R\mathcal V(Y) V ( R Y ) = − R V ( Y ) , the functions Y L ( t ; α 0 ∗ , α 1 ∗ , α 2 ∗ ) Y_L(t;\alpha_0^*,\alpha_1^*,\alpha_2^*) Y L ( t ; α 0 ∗ , α 1 ∗ , α 2 ∗ ) and R Y R ( T − t ; ω 0 ∗ , ω 1 ∗ , ω 2 ∗ ) RY_R(T-t;\omega_0^*,\omega_1^*,\omega_2^*) R Y R ( T − t ; ω 0 ∗ , ω 1 ∗ , ω 2 ∗ ) both satisfy Y ′ = V ( Y ) Y'=\mathcal V(Y) Y ′ = V ( Y ) as functions of t t t . Uniqueness for this regular ODE shows that they extend to the same smooth solution near T L T_L T L .
Define u u u by integrating u ′ = L 0 + 2 L 1 + 4 L 2 − ξ u'=L_0+2L_1+4L_2-\xi u ′ = L 0 + 2 L 1 + 4 L 2 − ξ with u ( T L ) = 0 u(T_L)=0 u ( T L ) = 0 . The potentials of the solutions constructed near the two singular orbits agree with u u u up to a constant. Hence u ( t ) u(t) u ( t ) is smooth and even near t = 0 t=0 t = 0 , while u ( T − s ) u(T-s) u ( T − s ) is smooth and even near s = 0 s=0 s = 0 . By Theorem 2.1, the metric and potential extend smoothly across both singular orbits, so gluing the two disc bundles by id P \operatorname{id}_P id P gives a smooth G G G -invariant pair ( g , u ) (g,u) ( g , u ) on M M M . The soliton equation holds on the regular set by Theorem 3.1 and at the singular orbits by continuity.
The box
Using the coordinate order in (40) , we take the box p 0 + [ − r , r ] 7 p_0 + [-r,r]^7 p 0 + [ − r , r ] 7 , with centre and radius
p 0 = ( 5170486699273585 / 2 51 6138202858770513 / 2 56 2928384247500027 / 2 54 − 5265652882566845 / 2 51 116897443675459 / 2 48 4239231962734261 / 2 52 2007847688405285 / 2 50 ) , r = 944473296573929 2 73 . (42) p_0 =
\begin{pmatrix}
5170486699273585/2^{51}\\
6138202858770513/2^{56}\\
2928384247500027/2^{54}\\
-5265652882566845/2^{51}\\
116897443675459/2^{48}\\
4239231962734261/2^{52}\\
2007847688405285/2^{50}
\end{pmatrix},
\qquad
r = \frac{944473296573929}{2^{73}}. \tag*{(42)} p 0 = 5170486699273585/ 2 51 6138202858770513/ 2 56 2928384247500027/ 2 54 − 5265652882566845/ 2 51 116897443675459/ 2 48 4239231962734261/ 2 52 2007847688405285/ 2 50 , r = 2 73 944473296573929 . (42)
Each of the seven parameters may vary by at most r ≈ 10 − 7 r \approx10^{-7} r ≈ 1 0 − 7 from its value at p 0 p_0 p 0 . The fractions above specify the exact values used in the verification. Table 1 gives rounded decimal values of the entries of p 0 p_0 p 0 .
The first three coordinates of p 0 + [ − r , r ] 7 p_0 + [-r,r]^7 p 0 + [ − r , r ] 7 specify the allowed data at B − B_- B − , and the next three specify those at B + B_+ B + . The computed bounds for these parameters differ from the corresponding centre in (36) by less than 2 ⋅ 10 − 7 2 \cdot10^{-7} 2 ⋅ 1 0 − 7 in every coordinate. Thus, the estimates near the singular orbits in § 3 apply to all these data.
To correct the numerical candidate, we use the map p ↦ p − C F ( p ) p \mapsto p-CF(p) p ↦ p − C F ( p ) , where C C C is a fixed rational matrix chosen to approximate D F ( p 0 ) − 1 DF(p_0)^{-1} D F ( p 0 ) − 1 . Its entries are listed in Appendix § 7 . We will show that, for every p ∈ p 0 + [ − r , r ] 7 p \in p_0 + [-r,r]^7 p ∈ p 0 + [ − r , r ] 7 , both solutions exist up to their matching times and satisfy f 0 , f 1 , f 2 > 0 f_0,f_1,f_2>0 f 0 , f 1 , f 2 > 0 at positive times. This will establish p 0 + [ − r , r ] 7 ⊂ D p_0 + [-r,r]^7 \subset\mathcal{D} p 0 + [ − r , r ] 7 ⊂ D , so that F F F is defined throughout the box. We will also prove
∥ C F ( p 0 ) ∥ ∞ < r / 10 , sup p ∈ p 0 + [ − r , r ] 7 ∥ I − C D F ( p ) ∥ ∞ < 1 / 10. \lVert CF(p_0)\rVert_\infty< r/10,\qquad
\sup_{p\in p_0+[-r,r]^7}\lVert I-CDF(p)\rVert_\infty< 1/10. ∥ C F ( p 0 ) ∥ ∞ < r /10 , p ∈ p 0 + [ − r , r ] 7 sup ∥ I − C D F ( p ) ∥ ∞ < 1/10.
These bounds ensure that the map sends p 0 + [ − r , r ] 7 p_0 + [-r,r]^7 p 0 + [ − r , r ] 7 into itself. The criterion in Subsection 4.5 then gives a unique zero of F F F in p 0 + [ − r , r ] 7 p_0 + [-r,r]^7 p 0 + [ − r , r ] 7 .
Continuation of the solutions and their derivatives
Starting from t 0 = 1 / 32 t_0=1/32 t 0 = 1/32 , we continue (32) together with its first variational equations:
∂ t W = G ( t , W ) , ∂ t ( D α ^ W ) = D W G ( t , W ) D α ^ W , α ^ = ( α 0 , α 1 2 , α 2 2 ) . \partial_t W = \mathcal{G}(t,W),\qquad
\partial_t(D_{\hat{\alpha}}W)=D_W\mathcal{G}(t,W)D_{\hat{\alpha}}W,\qquad
\hat{\alpha}=(\alpha_0,\alpha_1^2,\alpha_2^2). ∂ t W = G ( t , W ) , ∂ t ( D α ^ W ) = D W G ( t , W ) D α ^ W , α ^ = ( α 0 , α 1 2 , α 2 2 ) .
There is no additional term because G \mathcal{G} G has no explicit parameter dependence. § 3 supplies interval bounds for W W W and the 7 × 3 7\times3 7 × 3 matrix D α ^ W D_{\hat{\alpha}}W D α ^ W at t 0 t_0 t 0 , including the Taylor remainders and their parameter derivatives. The same procedure applies to the parameters ω ^ \hat{\omega} ω ^ at the right singular orbit.
We use validated Taylor integration [19 ] , together with the standard first-variation estimates described in [27 § 2.3–2.6] . The computation verifies existence and positivity of f ~ 0 , f 1 2 , f 2 2 \tilde f_0,f_1^2,f_2^2 f ~ 0 , f 1 2 , f 2 2 throughout each integration and gives interval bounds for the solution data and their parameter derivatives, uniformly for all allowed data at the respective singular orbit. The solution at p 0 p_0 p 0 is bounded separately to obtain a sharper bound on F ( p 0 ) F(p_0) F ( p 0 ) . Our implementation in verify.py and its step validation are described in Subsection A.2.
For a step from t t t to t + h t+h t + h , let Φ t , h ( W ) \Phi_{t,h}(W) Φ t , h ( W ) denote the solution data at t + h t+h t + h with initial data W W W at t t t . The 7 × 7 7\times7 7 × 7 derivative D W Φ t , h D_W\Phi_{t,h} D W Φ t , h propagates the parameter derivatives by the chain rule:
D α ^ W ( t + h , α ^ ) = D W Φ t , h ( W ( t , α ^ ) ) D α ^ W ( t , α ^ ) . D_{\hat{\alpha}}W(t+h,\hat{\alpha}) =
D_W\Phi_{t,h}(W(t,\hat{\alpha}))D_{\hat{\alpha}}W(t,\hat{\alpha}). D α ^ W ( t + h , α ^ ) = D W Φ t , h ( W ( t , α ^ )) D α ^ W ( t , α ^ ) .
For each of the left and right families, the integration times are fixed as the parameters vary over their prescribed ranges. At T L T_L T L and T R , 0 T_{R,0} T R , 0 , (39) and the chain rule give the bounds on C ( Y ) C(Y) C ( Y ) and its parameter derivatives needed for the matching map and its Jacobian.
The variable right time
The seventh parameter is the right shooting time. For the right solution family, define
Y ^ R ( t ; ω ^ ) = C ( Y R ( t ; ω 0 , ω 1 , ω 2 ) ) . \widehat{Y}_R(t;\hat{\omega})=C(Y_R(t;\omega_0,\omega_1,\omega_2)). Y R ( t ; ω ^ ) = C ( Y R ( t ; ω 0 , ω 1 , ω 2 )) .
In these coordinates, (20) becomes
∂ t Y ^ R ( t ; ω ^ ) = V ^ ( Y ^ R ( t ; ω ^ ) ) . \partial_t\widehat{Y}_R(t;\hat{\omega})=\widehat{\mathcal{V}}(\widehat{Y}_R(t;\hat{\omega})). ∂ t Y R ( t ; ω ^ ) = V ( Y R ( t ; ω ^ )) .
Differentiating the matching map in (41) with respect to T R T_R T R therefore gives
∂ T R F ( p ) = − R V ^ ( Y ^ R ( T R ; ω ^ ) ) . \partial_{T_R}\mathcal{F}(p)=-R\,\widehat{\mathcal{V}}(\widehat{Y}_R(T_R;\hat{\omega})). ∂ T R F ( p ) = − R V ( Y R ( T R ; ω ^ )) .
To obtain interval bounds for this column and the derivatives with respect to ω ^ \hat{\omega} ω ^ for all T R = T R , 0 + σ T_R = T_{R,0} + \sigma T R = T R , 0 + σ , ∣ σ ∣ ≤ r |\sigma| \le r ∣ σ ∣ ≤ r , we first continue the solutions constructed near the right singular orbit B + B_+ B + to inward distance T R , 0 T_{R,0} T R , 0 , the seventh coordinate of p 0 p_0 p 0 . Let X R ( 0 ) X_R^{(0)} X R ( 0 ) be a box containing all resulting values Y ^ R ( T R , 0 ; ω ^ ) \hat{Y}_R(T_{R,0};\hat{\omega}) Y ^ R ( T R , 0 ; ω ^ ) as ω ^ \hat{\omega} ω ^ varies over its prescribed ranges, and choose a larger compact box X R ⊂ { f 0 > 0 , f 1 2 > 0 , f 2 2 > 0 } X_R \subset\{f_0 > 0, f_1^2 > 0, f_2^2 > 0\} X R ⊂ { f 0 > 0 , f 1 2 > 0 , f 2 2 > 0 } . The strict inclusion
X R ( 0 ) + [ − r , r ] V ^ ( X R ) ⊂ int X R X_R^{(0)} + [-r,r]\hat{V}(X_R) \subset\operatorname{int} X_R X R ( 0 ) + [ − r , r ] V ^ ( X R ) ⊂ int X R
keeps every solution starting in X R ( 0 ) X_R^{(0)} X R ( 0 ) inside X R X_R X R for both earlier and later times with ∣ σ ∣ ≤ r |\sigma| \le r ∣ σ ∣ ≤ r . Here and below, V ^ ( X R ) \hat{V}(X_R) V ^ ( X R ) denotes a vector of intervals computed by interval arithmetic, containing all values of V ^ \hat{V} V ^ on X R X_R X R .
For initial solution data Y ^ ∈ X R ( 0 ) \hat{Y} \in X_R^{(0)} Y ^ ∈ X R ( 0 ) , let Ψ ^ σ ( Y ^ ) \hat{\Psi}_{\sigma}(\hat{Y}) Ψ ^ σ ( Y ^ ) be the solution data obtained by evolving this ODE for time σ \sigma σ . This is the analogue of Φ t , h \Phi_{t,h} Φ t , h in the coordinates Y ^ \hat{Y} Y ^ instead of W W W . No starting-time argument is needed because the equation does not depend explicitly on time. The computation verifies r Λ r t < 1 r\Lambda_{\mathrm{rt}} < 1 r Λ rt < 1 and
sup Y ^ ∈ X R ∥ D Y ^ V ^ ( Y ^ ) ∥ ∞ ≤ Λ r t < 5. \sup_{\hat{Y}\in X_R} \lVert D_{\hat{Y}}\hat{V}(\hat{Y})\rVert_{\infty} \le\Lambda_{\mathrm{rt}} < 5. Y ^ ∈ X R sup ∥ D Y ^ V ^ ( Y ^ ) ∥ ∞ ≤ Λ rt < 5.
The Gronwall estimate in (59) , with all weights equal to one and in either time direction, gives
∥ D Y ^ Ψ ^ σ ( Y ^ ) − I ∥ ∞ ≤ r Λ r t 1 − r Λ r t , Y ^ ∈ X R ( 0 ) , ∣ σ ∣ ≤ r . (43) \left\lVert D_{\hat{Y}}\hat{\Psi}_{\sigma}(\hat{Y})-I\right\rVert_{\infty}
\le\frac{r\Lambda_{\mathrm{rt}}}{1-r\Lambda_{\mathrm{rt}}}, \qquad\hat{Y}\in X_R^{(0)},\quad|\sigma|\le r. \tag*{(43)} D Y ^ Ψ ^ σ ( Y ^ ) − I ∞ ≤ 1 − r Λ rt r Λ rt , Y ^ ∈ X R ( 0 ) , ∣ σ ∣ ≤ r . (43)
Let
P R ( σ , ω ^ ) = D ω ^ Y ^ R ( T R , 0 + σ ; ω ^ ) . P_R(\sigma,\hat{\omega})=D_{\hat{\omega}}\hat{Y}_R(T_{R,0}+\sigma;\hat{\omega}). P R ( σ , ω ^ ) = D ω ^ Y ^ R ( T R , 0 + σ ; ω ^ ) .
This is a 7 × 3 7 \times3 7 × 3 matrix. Bounds on P R ( 0 , ω ^ ) P_R(0,\hat{\omega}) P R ( 0 , ω ^ ) are obtained from the bounds on D ω ^ W D_{\hat{\omega}}W D ω ^ W in the preceding subsection by applying (39) and its chain rule. The chain rule for the time extension gives
P R ( σ , ω ^ ) = D Y ^ Ψ ^ σ ( Y ^ R ( T R , 0 ; ω ^ ) ) P R ( 0 , ω ^ ) . P_R(\sigma,\hat{\omega})=D_{\hat{Y}}\hat{\Psi}_{\sigma}\bigl(\hat{Y}_R(T_{R,0};\hat{\omega})\bigr)P_R(0,\hat{\omega}). P R ( σ , ω ^ ) = D Y ^ Ψ ^ σ ( Y ^ R ( T R , 0 ; ω ^ ) ) P R ( 0 , ω ^ ) .
Hence, using the induced infinity norm also for rectangular matrices,
∥ P R ( σ , ω ^ ) − P R ( 0 , ω ^ ) ∥ ∞ ≤ r Λ r t 1 − r Λ r t ∥ P R ( 0 , ω ^ ) ∥ ∞ . \lVert P_R(\sigma,\hat{\omega})-P_R(0,\hat{\omega})\rVert_{\infty}
\le\frac{r\Lambda_{\mathrm{rt}}}{1-r\Lambda_{\mathrm{rt}}}\lVert P_R(0,\hat{\omega})\rVert_{\infty}. ∥ P R ( σ , ω ^ ) − P R ( 0 , ω ^ ) ∥ ∞ ≤ 1 − r Λ rt r Λ rt ∥ P R ( 0 , ω ^ ) ∥ ∞ .
Let [ a i j , b i j ] [a_{ij},b_{ij}] [ a ij , b ij ] be the computed interval containing ( P R ( 0 , ω ^ ) ) i j (P_R(0,\hat{\omega}))_{ij} ( P R ( 0 , ω ^ ) ) ij for every allowed choice of the parameters ω ^ \hat{\omega} ω ^ at B + B_+ B + . Set
ε R = r Λ r t 1 − r Λ r t max 1 ≤ i ≤ 7 ∑ j = 1 3 max { ∣ a i j ∣ , ∣ b i j ∣ } . \varepsilon_R=\frac{r\Lambda_{\mathrm{rt}}}{1-r\Lambda_{\mathrm{rt}}}\max_{1\le i\le7}\sum_{j=1}^{3}\max\{|a_{ij}|,|b_{ij}|\}. ε R = 1 − r Λ rt r Λ rt 1 ≤ i ≤ 7 max j = 1 ∑ 3 max { ∣ a ij ∣ , ∣ b ij ∣ } .
The preceding estimate gives
a i j − ε R ≤ ( P R ( σ , ω ^ ) ) i j ≤ b i j + ε R , ∣ σ ∣ ≤ r . a_{ij}-\varepsilon_R\le(P_R(\sigma,\hat{\omega}))_{ij}\le b_{ij}+\varepsilon_R,\qquad|\sigma|\le r. a ij − ε R ≤ ( P R ( σ , ω ^ ) ) ij ≤ b ij + ε R , ∣ σ ∣ ≤ r .
Let P R P_R P R be the 7 × 3 7\times3 7 × 3 matrix whose ( i , j ) (i,j) ( i , j ) entry is the interval
[ a i j − ε R , b i j + ε R ] , [a_{ij}-\varepsilon_R,b_{ij}+\varepsilon_R], [ a ij − ε R , b ij + ε R ] ,
with the lower bound rounded down and the upper bound rounded up. Thus P R P_R P R bounds the derivative entries for every allowed right shooting time and every allowed choice of ω ^ \hat{\omega} ω ^ .
Let P L P_L P L record interval bounds for the derivatives of the left solution with respect to the parameters α ^ \hat{\alpha} α ^ at B − B_- B − , evaluated at T L T_L T L and expressed in the coordinates Y ^ \hat{Y} Y ^ . Thus both matrices bound derivatives in the coordinates C ( Y ) C(Y) C ( Y ) , whereas P k P_k P k in Subsection A.2 bounds derivatives of W W W . We collect bounds for each entry of the derivative of the matching map in
A = [ P L , − R P R , − R V ^ ( X R ) ] . A=[P_L,-RP_R,-R\hat{V}(X_R)]. A = [ P L , − R P R , − R V ^ ( X R )] .
Its ( i , j ) (i,j) ( i , j ) entry is an interval containing ∂ F i / ∂ p j ( p ) \partial\mathcal{F}_i/\partial p_j(p) ∂ F i / ∂ p j ( p ) for every p ∈ p 0 + [ − r , r ] 7 p\in p_0+[-r,r]^7 p ∈ p 0 + [ − r , r ] 7 . The right block includes all ∣ σ ∣ ≤ r |\sigma|\le r ∣ σ ∣ ≤ r , so D F ( p 0 + [ − r , r ] 7 ) ⊂ A D\mathcal{F}(p_0+[-r,r]^7)\subset A D F ( p 0 + [ − r , r ] 7 ) ⊂ A .
For the later curvature estimates, we also sharpen the interval bounds for the left solution data at T L T_L T L . Applying the mean-value formula in α ^ \hat{\alpha} α ^ over its prescribed ranges, we find that all values of the left family at T L T_L T L lie in the sum of the interval box for the solution data at the centre and P L [ − r , r ] 3 P_L[-r,r]^3 P L [ − r , r ] 3 .
The Newton–Kantorovich criterion.
We use the following consequence of the contraction mapping theorem, applied to the map p ↦ p − C F ( p ) p \mapsto p-CF(p) p ↦ p − C F ( p ) with C C C fixed.
Theorem 4.2 (Newton–Kantorovich criterion with a fixed inverse). Let F : U → R n F: U \to\mathbb{R}^{n} F : U → R n be continuously differentiable on an open neighbourhood of B = p 0 + [ − r , r ] n B=p_0+[-r,r]^n B = p 0 + [ − r , r ] n , where r > 0 r>0 r > 0 . Suppose a real n × n n \times n n × n matrix C C C and constants η , α ≥ 0 \eta,\alpha\ge0 η , α ≥ 0 satisfy
∥ C F ( p 0 ) ∥ ∞ ≤ η , sup p ∈ B ∥ I − C D F ( p ) ∥ ∞ ≤ α < 1 , η + α r < r . (44) \|CF(p_0)\|_{\infty} \le\eta,\qquad\sup_{p\in B}\|I-CDF(p)\|_{\infty} \le\alpha< 1,\qquad\eta+\alpha r<r. \tag*{(44)} ∥ C F ( p 0 ) ∥ ∞ ≤ η , p ∈ B sup ∥ I − C D F ( p ) ∥ ∞ ≤ α < 1 , η + α r < r . (44) Then F F F has a unique zero p ∗ p^* p ∗ in B B B , and
∥ p ∗ − p 0 ∥ ∞ ≤ η 1 − α < r . \|p^*-p_0\|_{\infty} \le\frac{\eta}{1-\alpha}<r. ∥ p ∗ − p 0 ∥ ∞ ≤ 1 − α η < r . Proof. Put N ( p ) = p − C F ( p ) \mathcal{N}(p)=p-CF(p) N ( p ) = p − C F ( p ) . Integration of its derivative along segments in the convex box gives
∥ N ( p ) − N ( q ) ∥ ∞ ≤ α ∥ p − q ∥ ∞ , ∥ N ( p ) − p 0 ∥ ∞ ≤ η + α r < r . \|\mathcal{N}(p)-\mathcal{N}(q)\|_{\infty}\le\alpha\|p-q\|_{\infty},\qquad\|\mathcal{N}(p)-p_0\|_{\infty}\le\eta+\alpha r<r. ∥ N ( p ) − N ( q ) ∥ ∞ ≤ α ∥ p − q ∥ ∞ , ∥ N ( p ) − p 0 ∥ ∞ ≤ η + α r < r . Thus N \mathcal{N} N is a contraction of B B B into its interior and has a unique fixed point p ∗ p^* p ∗ . Since ∥ I − C D F ( p 0 ) ∥ ∞ < 1 \|I-CDF(p_0)\|_{\infty}<1 ∥ I − C D F ( p 0 ) ∥ ∞ < 1 , the Neumann series shows that C D F ( p 0 ) CDF(p_0) C D F ( p 0 ) , and therefore C C C , is invertible. Fixed points of N \mathcal{N} N are consequently exactly the zeros of F F F . Finally,
∥ p ∗ − p 0 ∥ ∞ ≤ α ∥ p ∗ − p 0 ∥ ∞ + η , \|p^*-p_0\|_{\infty}\le\alpha\|p^*-p_0\|_{\infty}+\eta, ∥ p ∗ − p 0 ∥ ∞ ≤ α ∥ p ∗ − p 0 ∥ ∞ + η , which gives the stated distance bound. □
Let r 0 ⊂ R n r_0\subset\mathbb{R}^{n} r 0 ⊂ R n be a box containing F ( p 0 ) F(p_0) F ( p 0 ) , and let A ⊂ R n × n A\subset\mathbb{R}^{n\times n} A ⊂ R n × n be a box containing D F ( p ) DF(p) D F ( p ) for every p ∈ B p\in B p ∈ B . Recall that N ( p ) = p − C F ( p ) \mathcal{N}(p)=p-CF(p) N ( p ) = p − C F ( p ) is the correction map. Using interval arithmetic, define
E = I − C A , K = − C r 0 + E [ − r , r ] n . (45) \mathcal{E}=I-CA,\qquad\mathcal{K}=-Cr_0+\mathcal{E}[-r,r]^n. \tag*{(45)} E = I − C A , K = − C r 0 + E [ − r , r ] n . (45)
The entries of E \mathcal{E} E bound the corresponding derivatives of the correction map, since
D N ( p ) = I − C D F ( p ) ∈ E ( p ∈ B ) . D\mathcal{N}(p)=I-CDF(p)\in\mathcal{E}\qquad(p\in B). D N ( p ) = I − C D F ( p ) ∈ E ( p ∈ B ) .
The box K ⊂ R n \mathcal{K}\subset\mathbb{R}^{n} K ⊂ R n bounds the corrected points relative to p 0 p_0 p 0 :
N ( B ) ⊂ p 0 + K . \mathcal{N}(B)\subset p_0+\mathcal{K}. N ( B ) ⊂ p 0 + K .
Consequently, K ⊂ ( − r , r ) n \mathcal{K}\subset(-r,r)^n K ⊂ ( − r , r ) n ensures that N \mathcal{N} N sends B B B into its interior.
For an interval, its absolute value denotes the largest absolute value it contains. We obtain bounds for the correction at p 0 p_0 p 0 and the derivative throughout B B B by taking
η = max i ∣ ( − C r 0 ) i ∣ , α = max i ∑ j ∣ E i j ∣ . \eta=\max_i |(-Cr_0)_i|,\qquad\alpha=\max_i\sum_j|\mathcal{E}_{ij}|. η = i max ∣ ( − C r 0 ) i ∣ , α = i max j ∑ ∣ E ij ∣.
These satisfy
∥ N ( p 0 ) − p 0 ∥ ∞ ≤ η , sup p ∈ B ∥ D N ( p ) ∥ ∞ ≤ α . \|\mathcal{N}(p_0)-p_0\|_{\infty}\le\eta,\qquad\sup_{p\in B}\|D\mathcal{N}(p)\|_{\infty}\le\alpha. ∥ N ( p 0 ) − p 0 ∥ ∞ ≤ η , p ∈ B sup ∥ D N ( p ) ∥ ∞ ≤ α .
Finally, 0 ∈ E [ − r , r ] n 0\in\mathcal{E}[-r,r]^n 0 ∈ E [ − r , r ] n implies − C r 0 ⊂ K -Cr_0\subset\mathcal{K} − C r 0 ⊂ K , and hence
η ≤ max i ∣ K i ∣ . \eta\le\max_i|\mathcal{K}_i|. η ≤ i max ∣ K i ∣.
Verification of the matching inequalities.
Proposition 4.3. For every p ∈ p 0 + [ − r , r ] 7 p\in p_0+[-r,r]^7 p ∈ p 0 + [ − r , r ] 7 , both solutions exist and are regular with f 0 , f 1 , f 2 > 0 f_0,f_1,f_2>0 f 0 , f 1 , f 2 > 0 at positive inward times through their respective matching times. There exist an interval box r 0 r_0 r 0 containing F ( p 0 ) F(p_0) F ( p 0 ) and a matrix A A A of intervals containing D F ( p 0 + [ − r , r ] 7 ) DF(p_0+[-r,r]^7) D F ( p 0 + [ − r , r ] 7 ) such that, for the fixed matrix C C C , exact rational evaluation of (45) with n = 7 n=7 n = 7 gives
max i ∑ j ∣ E i j ∣ ≤ α ∗ < 1 10 , max i ∣ K i ∣ r ≤ β ∗ < 1 10 . (46) \max_i\sum_j|\mathcal{E}_{ij}|\le\alpha_*<\frac{1}{10},\qquad\max_i\frac{|\mathcal{K}_i|}{r}\le\beta_*<\frac{1}{10}. \tag*{(46)} i max j ∑ ∣ E ij ∣ ≤ α ∗ < 10 1 , i max r ∣ K i ∣ ≤ β ∗ < 10 1 . (46) where the exact rational constants listed in Appendix A satisfy
α ∗ < 0.096 , β ∗ < 0.097. \alpha_*<0.096,\qquad\beta_*<0.097. α ∗ < 0.096 , β ∗ < 0.097.
Thus the correction map N ( p ) = p − C F ( p ) \mathcal{N}(\mathbf{p})=\mathbf{p}-C\mathcal{F}(\mathbf{p}) N ( p ) = p − C F ( p ) has contraction factor less than one tenth and satisfies
N ( p 0 + [ − r , r ] 7 ) ⊂ p 0 + [ − r / 10 , r / 10 ] 7 . \mathcal{N}\left(\mathbf{p}_0+[-r,r]^7\right)\subset\mathbf{p}_0+[-r/10,r/10]^7. N ( p 0 + [ − r , r ] 7 ) ⊂ p 0 + [ − r /10 , r /10 ] 7 .
At t = T L t=T_L t = T L , the left solution values for all allowed data at B − B_- B − lie in the rational intervals in Table 4 .
Table 4. Bounds at t = T L t = T_L t = T L for all allowed data at the left singular orbit. Divide each number in the last two columns by 10 7 10^7 1 0 7 .
Proof. The verification program verify.py, described in Appendix A, first checks that the prescribed parameter ranges at both singular orbits lie within the domains used in Section 3. Theorem 3.3, the Taylor recurrence in (33) and the remainder estimates then supply intervals containing the exact solutions and their parameter derivatives at t 0 = 1 / 32 t_0=1/32 t 0 = 1/32 . These starting bounds hold for every allowed choice of data at either singular orbit.
At each integration step, the program verifies (56), positivity of f ~ 0 \tilde{f}_0 f ~ 0 , f 1 2 f_1^2 f 1 2 , f 2 2 f_2^2 f 2 2 , and h Λ k < 7 / 20 h\Lambda_k<7/20 h Λ k < 7/20 . The validated Taylor estimates in Subsection A.2 guarantee that the exact solutions exist throughout the step and give valid solution and derivative bounds at the next step. After 231 left steps and 162 right steps, this gives bounds at T L T_L T L and T R , 0 T_{R,0} T R , 0 , with sharper bounds for the solutions at p 0 \mathbf{p}_0 p 0 . The program checks in exact rational arithmetic that each step starts where the preceding one ends and that the final steps end exactly at T L T_L T L and T R , 0 T_{R,0} T R , 0 ; thus no interval of integration is skipped.
At these terminal times, (39) and the chain rule convert the bounds on W W W and its parameter derivatives into bounds in the coordinates C ( Y ) C(Y) C ( Y ) . The estimates allowing T R T_R T R to vary then cover every right shooting time in p 0 + [ − r , r ] 7 \mathbf{p}_0+[-r,r]^7 p 0 + [ − r , r ] 7 . Thus both solutions exist through their required matching times, with f 0 , f 1 , f 2 > 0 f_0,f_1,f_2>0 f 0 , f 1 , f 2 > 0 at positive inward times, so p 0 + [ − r , r ] 7 ⊂ D \mathbf{p}_0+[-r,r]^7\subset\mathcal{D} p 0 + [ − r , r ] 7 ⊂ D . Substituting the bounds at the matching times into the matching map in (41) and its derivative gives
F ( p 0 ) ∈ r 0 , D F ( p ) ∈ A for every p ∈ p 0 + [ − r , r ] 7 . \mathcal{F}(\mathbf{p}_0)\in\mathbf{r}_0,\qquad D\mathcal{F}(\mathbf{p})\in\mathcal{A}\quad\text{for every }\mathbf{p}\in\mathbf{p}_0+[-r,r]^7. F ( p 0 ) ∈ r 0 , D F ( p ) ∈ A for every p ∈ p 0 + [ − r , r ] 7 .
It remains to check that these intervals satisfy the inequalities required by the contraction criterion.
For this final calculation, verify.py treats the lower and upper bounds of each interval as exact fractions and uses integer numerators and denominators. It computes E \mathcal{E} E and K \mathcal{K} K from r 0 \mathbf{r}_0 r 0 , A \mathcal{A} A and C \mathcal{C} C using (45), and verifies (46). It also checks that the reported intervals contain these exact rational results. Since E \mathcal{E} E bounds D N D\mathcal{N} D N on p 0 + [ − r , r ] 7 \mathbf{p}_0+[-r,r]^7 p 0 + [ − r , r ] 7 and K \mathcal{K} K contains N ( p 0 + [ − r , r ] 7 ) − p 0 \mathcal{N}(\mathbf{p}_0+[-r,r]^7)-\mathbf{p}_0 N ( p 0 + [ − r , r ] 7 ) − p 0 , these inequalities prove the asserted contraction and image bounds.
Finally, the mean-value formula and the computed left derivative bounds give intervals containing the solution values at t = T L t=T_L t = T L for all allowed data at the left singular orbit. Rounding the lower bounds of these intervals down and their upper bounds up gives Table 4 .
Corollary 4.4. The matching map F \mathcal{F} F has a unique zero in p 0 + [ − r , r ] 7 \mathbf{p}_0+[-r,r]^7 p 0 + [ − r , r ] 7 .
Proof. The initial data at t 0 t_0 t 0 depend analytically on the parameters. The boxes containing the solutions throughout the integration, including the extension in T R T_R T R , lie inside the regular ODE domain. Openness of the existence domain in initial data and time therefore makes F \mathcal{F} F analytic on an open neighbourhood of p 0 + [ − r , r ] 7 \mathbf{p}_0+[-r,r]^7 p 0 + [ − r , r ] 7 . By Theorem 4.3 and (45),
∥ C F ( p 0 ) ∥ ∞ ≤ max ∣ K i ∣ ≤ β ∗ r , sup p ∈ p 0 + [ − r , r ] 7 ∥ I − C D F ( p ) ∥ ∞ ≤ α ∗ . \lVert C\mathcal{F}(\mathbf{p}_0)\rVert_\infty\leq\max\lvert\mathcal{K}_i\rvert\leq\beta_\ast r,\qquad\sup_{\mathbf{p}\in\mathbf{p}_0+[-r,r]^7}\lVert I-CD\mathcal{F}(\mathbf{p})\rVert_\infty\leq\alpha_\ast. ∥ C F ( p 0 ) ∥ ∞ ≤ max ∣ K i ∣ ≤ β ∗ r , p ∈ p 0 + [ − r , r ] 7 sup ∥ I − C D F ( p ) ∥ ∞ ≤ α ∗ . Taking η = β ∗ r \eta=\beta_\ast r η = β ∗ r and α = α ∗ \alpha=\alpha_\ast α = α ∗ gives α < 1 \alpha<1 α < 1 and
η + α r = ( β ∗ + α ∗ ) r < r / 5 < r . \eta+\alpha r=(\beta_\ast+\alpha_\ast)r<r/5<r. η + α r = ( β ∗ + α ∗ ) r < r /5 < r . Theorem 4.2 therefore gives the unique zero p ∗ \mathbf{p}^\ast p ∗ . Since p ∗ = N ( p ∗ ) ∈ p 0 + K \mathbf{p}^\ast=\mathcal{N}(\mathbf{p}^\ast)\in\mathbf{p}_0+\mathcal{K} p ∗ = N ( p ∗ ) ∈ p 0 + K , we also have
∥ p ∗ − p 0 ∥ ∞ ≤ β ∗ r < r / 10. \lVert\mathbf{p}^\ast-\mathbf{p}_0\rVert_\infty\leq\beta_\ast r<r/10. ∥ p ∗ − p 0 ∥ ∞ ≤ β ∗ r < r /10.
Remark 4.5. The uniqueness in Theorem 4.4 holds within p 0 + [ − r , r ] 7 \mathbf{p}_0+[-r,r]^7 p 0 + [ − r , r ] 7 , with the left shooting time T L T_L T L fixed.
The bounds at t = T L t=T_L t = T L in Table 4 apply in particular to the matching solution and will be used for the curvature estimates in Subsection 5.2.
The manifold and the geometry of the soliton
Theorem 4.4 and Theorem 4.1 give a smooth compact gradient shrinking Ricci soliton on the disc-bundle manifold defined in (5). We identify this manifold and use the curvature of the soliton to prove that it is non-Einstein, nowhere locally Kähler and locally irreducible.
Identification of the manifold
Both singular orbits are equivariantly diffeomorphic to the six-dimensional full flag manifold
B = G / K ≅ SU ( 3 ) / T 2 . B = G/K \cong\operatorname{SU}(3)/T^2. B = G / K ≅ SU ( 3 ) / T 2 .
Define the associated complex line bundle
L = SU ( 3 ) × χ C ⟶ B , ( g , w ) ∼ ( g ϑ , χ ( ϑ ) − 1 w ) . (47) \mathcal{L} = \operatorname{SU}(3) \times_{\chi} \mathbb{C} \longrightarrow B,\qquad(g,w) \sim(g\vartheta,\chi(\vartheta)^{-1}w). \tag*{(47)} L = SU ( 3 ) × χ C ⟶ B , ( g , w ) ∼ ( g ϑ , χ ( ϑ ) − 1 w ) . (47)
Here χ : T 2 → S 1 \chi:T^2 \to S^1 χ : T 2 → S 1 is the homomorphism in (7), which specifies how the torus rotates the complex fibre, and L R \mathcal{L}_{\mathbb{R}} L R is the same bundle with each complex fibre viewed as a real plane. We write S ( E ) S(E) S ( E ) for the unit sphere bundle of a real vector bundle and P ( E ) P(E) P ( E ) for the bundle of complex lines in a complex vector bundle. The symbols R \mathbb{R} R and C \mathbb{C} C in (48) denote trivial line bundles over B B B . The identifications below are fibrewise: gluing the two normal discs gives a sphere S 2 S^2 S 2 , which is then identified with C P 1 \mathbb{CP}^1 CP 1 .
Proposition 5.1. The manifold defined in (5) is equivariantly diffeomorphic to
M ≅ S ( R ⊕ L R ) ≅ P ( C ⊕ L ) ≅ SU ( 3 ) × T 2 C P 1 . (48) M \cong S(\mathbb{R} \oplus\mathcal{L}_{\mathbb{R}}) \cong P(\mathbb{C} \oplus\mathcal{L}) \cong\operatorname{SU}(3) \times_{T^2} \mathbb{CP}^1. \tag*{(48)} M ≅ S ( R ⊕ L R ) ≅ P ( C ⊕ L ) ≅ SU ( 3 ) × T 2 CP 1 . (48) The torus acts on C P 1 \mathbb{CP}^1 CP 1 by ϑ ⋅ [ a : b ] = [ a : χ ( ϑ ) b ] \vartheta\cdot[a:b] = [a:\chi(\vartheta)b] ϑ ⋅ [ a : b ] = [ a : χ ( ϑ ) b ] , and the G G G -action in this model is
( g 0 , z ) ⋅ [ g , [ a : b ] ] = [ g 0 g , [ a : z b ] ] . (49) (g_0,z)\cdot[g,[a:b]]=[g_0g,[a:zb]]. \tag*{(49)} ( g 0 , z ) ⋅ [ g , [ a : b ]] = [ g 0 g , [ a : z b ]] . (49)
Proof. The action of K K K on each normal disc is k ⋅ w = ρ ( k ) w k\cdot w=\rho(k)w k ⋅ w = ρ ( k ) w in a complex coordinate w w w , where
ρ : K ⟶ S 1 , ρ ( ϑ , z ) = χ ( ϑ ) z , ker ρ = H . \rho:K\longrightarrow S^1,\qquad\rho(\vartheta,z)=\chi(\vartheta)z,\qquad\ker\rho=H. ρ : K ⟶ S 1 , ρ ( ϑ , z ) = χ ( ϑ ) z , ker ρ = H .
In each fibre, the boundary identification in (5) glues two copies of the unit disc in C \mathbb{C} C , each carrying the action w ↦ ρ ( k ) w w\mapsto\rho(k)w w ↦ ρ ( k ) w for k ∈ K k\in K k ∈ K , along their boundary circles. The resulting fibre is the unit sphere in R ⊕ C \mathbb{R}\oplus\mathbb{C} R ⊕ C , with K K K acting trivially on R \mathbb{R} R and through ρ \rho ρ on C \mathbb{C} C . Explicitly, the two maps from the unit discs are
r η ⟼ ( ± cos ( π r / 2 ) , sin ( π r / 2 ) η ) , 0 ≤ r ≤ 1 , ∣ η ∣ = 1. r\eta\longmapsto\left(\pm\cos(\pi r/2),\sin(\pi r/2)\eta\right),\qquad0\le r\le1,\quad|\eta|=1. r η ⟼ ( ± cos ( π r /2 ) , sin ( π r /2 ) η ) , 0 ≤ r ≤ 1 , ∣ η ∣ = 1.
They agree at r = 1 r=1 r = 1 . Set s = 1 − r s=1-r s = 1 − r on the first disc and s = r − 1 s=r-1 s = r − 1 on the second. Both maps then take the form
( s , η ) ⟼ ( sin ( π s / 2 ) , cos ( π s / 2 ) η ) , (s,\eta)\longmapsto\left(\sin(\pi s/2),\cos(\pi s/2)\eta\right), ( s , η ) ⟼ ( sin ( π s /2 ) , cos ( π s /2 ) η ) ,
which is smooth across s = 0 s=0 s = 0 . At r = 0 r=0 r = 0 , the quotient sin ( π r / 2 ) / r \sin(\pi r/2)/r sin ( π r /2 ) / r extends smoothly as a function of r 2 r^2 r 2 , with nonzero value. Thus these maps give a smooth equivariant identification
M ≅ G × K S ( R ⊕ C ) . M\cong G\times_K S(\mathbb{R}\oplus\mathbb{C}). M ≅ G × K S ( R ⊕ C ) .
To remove the extra S 1 S^1 S 1 -factor from the group coordinate, write an element of G G G as ( g , z ) (g,z) ( g , z ) and use the map
[ ( g , z ) , ( r 0 , v ) ] ⟼ [ g , ( r 0 , z v ) ] [(g,z),(r_0,v)]\longmapsto[g,(r_0,zv)] [( g , z ) , ( r 0 , v )] ⟼ [ g , ( r 0 , z v )]
to identify this associated sphere bundle with S U ( 3 ) × T 2 S ( R ⊕ C ) = S ( R ⊕ L R ) \mathrm{SU}(3) \times_{T^2} S(\mathbb{R} \oplus\mathbb{C}) = S(\mathbb{R} \oplus L_{\mathbb{R}}) SU ( 3 ) × T 2 S ( R ⊕ C ) = S ( R ⊕ L R ) . The formula for ρ \rho ρ shows that this map respects the associated-bundle equivalence relations.
For ( r 0 , v ) ∈ S ( R ⊕ C ) (r_0,v) \in S(\mathbb{R} \oplus\mathbb{C}) ( r 0 , v ) ∈ S ( R ⊕ C ) , the image of the rank-one Hermitian projection
1 2 ( 1 + r 0 v ‾ v 1 − r 0 ) (50) \frac{1}{2}\begin{pmatrix} 1+r_0 & \overline{v} \\ v & 1-r_0 \end{pmatrix} \tag*{(50)} 2 1 ( 1 + r 0 v v 1 − r 0 ) (50)
is a complex line in C 2 \mathbb{C}^2 C 2 . Conversely, the orthogonal projection onto any complex line has this form, giving the inverse map C P 1 → S 2 \mathbb{CP}^1 \to S^2 CP 1 → S 2 . Rotation v ↦ z v v \mapsto zv v ↦ z v corresponds to [ a : b ] ↦ [ a : z b ] [a:b] \mapsto[a:zb] [ a : b ] ↦ [ a : z b ] , so it descends to the last two identifications in (48) and gives (49) . □
The singular orbits are the two sections of M → B M \to B M → B given in each projective fibre by [ 1 : 0 ] [1:0] [ 1 : 0 ] and [ 0 : 1 ] [0:1] [ 0 : 1 ] . Near these sections, use the fibre coordinates q = b / a q=b/a q = b / a and q ~ = a / b \widetilde{q}=a/b q = a / b , respectively. An element k ∈ K k \in K k ∈ K acts by q ↦ ρ ( k ) q q \mapsto\rho(k)q q ↦ ρ ( k ) q and q ~ ↦ ρ ( k ) − 1 q ~ \widetilde{q} \mapsto\rho(k)^{-1}\widetilde{q} q ↦ ρ ( k ) − 1 q . Complex conjugation identifies these as real representations, consistently with the two identical real discs used in Section 2. The principal orbit is the unit circle bundle S ( L ) S(\mathcal{L}) S ( L ) through ( g , z ) H ↦ [ g , z ] (g,z)H \mapsto[g,z] ( g , z ) H ↦ [ g , z ] .
Proposition 5.2. The manifold M M M is simply-connected. The kernel of its G G G -action is Z 3 × { 1 } \mathbb{Z}_3 \times\{1\} Z 3 × { 1 } , so the effective group is P S U ( 3 ) × S 1 \mathrm{PSU}(3) \times S^1 PSU ( 3 ) × S 1 .
Proof. Since S U ( 3 ) \mathrm{SU}(3) SU ( 3 ) is simply-connected and T 2 T^2 T 2 is connected, the homotopy sequence of T 2 → S U ( 3 ) → B T^2 \to\mathrm{SU}(3) \to B T 2 → SU ( 3 ) → B gives π 1 ( B ) = 0 \pi_1(B)=0 π 1 ( B ) = 0 . The identification in (48) exhibits M M M as a C P 1 \mathbb{CP}^1 CP 1 -bundle over B B B . Both base and fibre are simply-connected, so its homotopy sequence gives π 1 ( M ) = 0 \pi_1(M)=0 π 1 ( M ) = 0 .
For the action kernel, an element acting trivially on M M M must act trivially on G / K G/K G / K . The kernel of the latter action is
⋂ g ∈ G g K g − 1 = Z 3 × S 1 . \bigcap_{g\in G} gKg^{-1} = \mathbb{Z}_3 \times S^1. g ∈ G ⋂ g K g − 1 = Z 3 × S 1 .
Indeed, the intersection of all conjugates of the maximal torus in S U ( 3 ) \mathrm{SU}(3) SU ( 3 ) is its centre. On the slice, an element ( η I d , z ) (\eta\,\mathrm{Id},z) ( η Id , z ) of this subgroup acts as multiplication by z z z , since χ ( η I d ) = 1 \chi(\eta\,\mathrm{Id})=1 χ ( η Id ) = 1 . Hence it acts trivially on M M M exactly when z = 1 z=1 z = 1 , proving the assertion. □
Remark 5.3 (The complex structure on the flag base and bundle). We describe a complex structure on M M M and a necessary condition for g g g to be Kähler with respect to it. Choose the invariant complex structure J 0 J_0 J 0 on B B B defined by J 0 X i j = Y i j J_0X_{ij}=Y_{ij} J 0 X ij = Y ij for i < j i<j i < j . The positive-root spaces form a subalgebra, so J 0 J_0 J 0 is integrable. Let Ω i \Omega_i Ω i be the invariant two-form equal to Q ( J 0 ⋅ , ⋅ ) Q(J_0\cdot,\cdot) Q ( J 0 ⋅ , ⋅ ) on m i \mathfrak{m}_i m i and zero whenever either argument lies in another isotropy summand.
The circle-invariant form θ \theta θ from (11) satisfies θ ( Z ) = 1 \theta(Z)=1 θ ( Z ) = 1 , so it is a connection one-form on P → B P \to B P → B . With the Ω i \Omega_i Ω i pulled back from B B B , (9) gives
d θ = 1 3 Ω 1 + 1 6 Ω 2 + 1 6 Ω 3 . d\theta= \frac{1}{3}\Omega_1+\frac{1}{6}\Omega_2+\frac{1}{6}\Omega_3. d θ = 3 1 Ω 1 + 6 1 Ω 2 + 6 1 Ω 3 .
The induced unitary connection on L \mathcal{L} L has curvature i d θ i\,d\theta i d θ of type ( 1 , 1 ) (1,1) ( 1 , 1 ) , so its ( 0 , 1 ) (0,1) ( 0 , 1 ) part squares to zero and defines a homogeneous holomorphic structure on L \mathcal{L} L . The torus acts on the three complex tangent lines of B B B by z ‾ 1 / z ‾ 2 \overline{z}_1/\overline{z}_2 z 1 / z 2 , z ‾ 2 / z ‾ 3 \overline{z}_2/\overline{z}_3 z 2 / z 3 , z ‾ 1 / z ‾ 3 \overline{z}_1/\overline{z}_3 z 1 / z 3 . Their product is χ − 2 \chi^{-2} χ − 2 , so the dual determinant bundle K B K_B K B has multiplier χ 2 \chi^2 χ 2 , the same as L 2 \mathcal{L}^2 L 2 . Thus, with the connection in (47) ,
K B ≅ L 2 , c 1 ( L ) = − 1 2 c 1 ( B ) = − [ d θ / ( 2 π ) ] , K_B \cong\mathcal{L}^2,\qquad c_1(\mathcal{L})=-\frac{1}{2}c_1(B)=-[d\theta/(2\pi)], K B ≅ L 2 , c 1 ( L ) = − 2 1 c 1 ( B ) = − [ d θ / ( 2 π )] ,
where K B K_B K B is the canonical bundle and c 1 ( B ) = c 1 ( T 1 , 0 B ) c_1(B)=c_1(T^{1,0}B) c 1 ( B ) = c 1 ( T 1 , 0 B ) . In particular, L \mathcal{L} L is a holomorphic square root of K B K_B K B , and P ( C ⊕ L ) \mathbb{P}(\mathbb{C}\oplus\mathcal{L}) P ( C ⊕ L ) supports a complex structure on M M M .
On the regular set, lift J 0 J_0 J 0 to ker θ \ker\theta ker θ and set
J ∂ t = f 0 − 1 Z , J Z = − f 0 ∂ t . J\partial_t=f_0^{-1}Z,\qquad JZ=-f_0\partial_t. J ∂ t = f 0 − 1 Z , J Z = − f 0 ∂ t .
This defines an integrable complex structure for which g g g is Hermitian, as in [10 §4] . Its fundamental form is
g ( J ⋅ , ⋅ ) = f 0 d t ∧ θ + f 1 2 Ω 1 + f 2 2 Ω 2 + f 3 2 Ω 3 . g(J\cdot,\cdot)=f_0\,dt\wedge\theta+f_1^2\Omega_1+f_2^2\Omega_2+f_3^2\Omega_3. g ( J ⋅ , ⋅ ) = f 0 d t ∧ θ + f 1 2 Ω 1 + f 2 2 Ω 2 + f 3 2 Ω 3 .
At fixed t t t , the root brackets give
d B ( f 1 2 Ω 1 + f 2 2 Ω 2 + f 3 2 Ω 3 ) ( X 12 , X 23 , Y 13 ) = f 2 2 + f 3 2 − f 1 2 12 , (51) d_B(f_1^2\Omega_1+f_2^2\Omega_2+f_3^2\Omega_3)(X_{12},X_{23},Y_{13})=\frac{f_2^2+f_3^2-f_1^2}{\sqrt{12}}, \tag*{(51)} d B ( f 1 2 Ω 1 + f 2 2 Ω 2 + f 3 2 Ω 3 ) ( X 12 , X 23 , Y 13 ) = 12 f 2 2 + f 3 2 − f 1 2 , (51)
where d B d_B d B is exterior differentiation on B B B . For g g g to be Kähler with respect to this J J J , its fundamental form must be closed, including on each principal orbit. The preceding calculation therefore requires f 1 2 = f 2 2 + f 3 2 f_1^2=f_2^2+f_3^2 f 1 2 = f 2 2 + f 3 2 , or f 1 2 = 2 f 2 2 f_1^2=2f_2^2 f 1 2 = 2 f 2 2 in the family f 2 = f 3 f_2=f_3 f 2 = f 3 . This condition concerns the chosen J J J ; the curvature argument below applies to all compatible parallel complex structures.
Curvature of the matching solution
Let ( g , u ) (g,u) ( g , u ) be the soliton constructed in Theorem 4.1 from the zero p ∗ p^* p ∗ supplied by Theorem 4.4. We calculate its curvature at the regular orbit t = T L t=T_L t = T L . Table 4 gives bounds at this time for every solution obtained from the allowed data at the left singular orbit, including the matching solution. For the curvature calculation we enlarge the box defined by these intervals to the following rational box:
f 0 ∈ [ 8128 , 8129 ] / 10 4 , f 1 2 ∈ [ 2870 , 2871 ] / 10 4 , f 2 2 ∈ [ 4875 , 4876 ] / 10 4 , L 0 ∈ [ 1154 , 1155 ] / 10 4 , L 1 ∈ [ 2406 , 2408 ] / 10 4 , L 2 ∈ [ 3207 , 3209 ] / 10 4 , ξ ∈ [ − 1 , 1 ] / 10 4 . (52) \begin{aligned}
f_0 &\in[8128,8129]/10^4,\quad f_1^2 \in[2870,2871]/10^4,\quad f_2^2 \in[4875,4876]/10^4,\\
L_0 &\in[1154,1155]/10^4,\quad L_1 \in[2406,2408]/10^4,\quad L_2 \in[3207,3209]/10^4,\quad\xi\in[-1,1]/10^4. \tag*{(52)}
\end{aligned} f 0 L 0 ∈ [ 8128 , 8129 ] /1 0 4 , f 1 2 ∈ [ 2870 , 2871 ] /1 0 4 , f 2 2 ∈ [ 4875 , 4876 ] /1 0 4 , ∈ [ 1154 , 1155 ] /1 0 4 , L 1 ∈ [ 2406 , 2408 ] /1 0 4 , L 2 ∈ [ 3207 , 3209 ] /1 0 4 , ξ ∈ [ − 1 , 1 ] /1 0 4 . (52)
Recall that
u ′ = L 0 + 2 L 1 + 4 L 2 − ξ , tr ( L 2 ) = L 0 2 + 2 L 1 2 + 4 L 2 2 . u'=L_0+2L_1+4L_2-\xi,\qquad\operatorname{tr}(L^2)=L_0^2+2L_1^2+4L_2^2. u ′ = L 0 + 2 L 1 + 4 L 2 − ξ , tr ( L 2 ) = L 0 2 + 2 L 1 2 + 4 L 2 2 .
The orbit Ricci eigenvalues R 0 , R 1 , R 2 R_0,R_1,R_2 R 0 , R 1 , R 2 are given by (18). We distinguish these from the Ricci eigenvalues of the full eight-dimensional metric, denoted by ρ ⊥ , ρ 0 , ρ 1 , ρ 2 \rho_\perp,\rho_0,\rho_1,\rho_2 ρ ⊥ , ρ 0 , ρ 1 , ρ 2 in the normal, m 0 , m 1 \mathfrak{m}_0,\mathfrak{m}_1 m 0 , m 1 , and m 2 ⊕ m 3 \mathfrak{m}_2\oplus\mathfrak{m}_3 m 2 ⊕ m 3 directions, respectively. They are
ρ ⊥ = 7 − ( R 0 + 2 R 1 + 4 R 2 ) − tr ( L 2 ) + ξ tr L , ρ i = 1 − u ′ L i , i = 0 , 1 , 2 , (53) \begin{aligned}
\rho_\perp&= 7-(R_0+2R_1+4R_2)-\operatorname{tr}(L^2)+\xi\operatorname{tr}L,\\
\rho_i &= 1-u'L_i,\qquad i=0,1,2, \tag*{(53)}
\end{aligned} ρ ⊥ ρ i = 7 − ( R 0 + 2 R 1 + 4 R 2 ) − tr ( L 2 ) + ξ tr L , = 1 − u ′ L i , i = 0 , 1 , 2 , (53)
on mutually orthogonal blocks of dimensions 1, 1, 2, 4, respectively. Distinct blocks may have equal eigenvalues away from the certified orbit. To obtain the normal Ricci eigenvalue, take the trace of the shape evolution equation:
( tr L ) ′ = − ξ tr L + ( R 0 + 2 R 1 + 4 R 2 ) − 7. (\operatorname{tr}L)'=-\xi\operatorname{tr}L+(R_0+2R_1+4R_2)-7. ( tr L ) ′ = − ξ tr L + ( R 0 + 2 R 1 + 4 R 2 ) − 7.
Substitution into
ρ ⊥ = Ric g ( ∂ t , ∂ t ) = − ( tr L ) ′ − tr ( L 2 ) \rho_\perp=\operatorname{Ric}_g(\partial_t,\partial_t)=-(\operatorname{tr}L)'-\operatorname{tr}(L^2) ρ ⊥ = Ric g ( ∂ t , ∂ t ) = − ( tr L ) ′ − tr ( L 2 )
gives the first expression in (52). The tangential expressions follow from Hess g u ( V , W ) = u ′ h ( t ) ( L V , W ) \operatorname{Hess}_g u(V,W)=u'h(t)(LV,W) Hess g u ( V , W ) = u ′ h ( t ) ( L V , W ) for V , W ∈ T P V,W\in TP V , W ∈ T P .
With the convention V ↦ R ( V , N ) N V\mapsto R(V,N)N V ↦ R ( V , N ) N , the normal Jacobi endomorphism is − ( L ′ + L 2 ) -(L'+L^2) − ( L ′ + L 2 ) . Its three tangential block coefficients are
κ i = 1 + ξ L i − R i − L i 2 , i = 0 , 1 , 2. (54) \kappa_i=1+\xi L_i-R_i-L_i^2,\qquad i=0,1,2. \tag*{(54)} κ i = 1 + ξ L i − R i − L i 2 , i = 0 , 1 , 2. (54)
All denominators in these formulas are positive on the box in (51).
Lemma 5.4. On the box in (51), the quantities in Table 5 satisfy the stated bounds.
Table 5. Exact rational curvature bounds on the box in (51). Every row has denominator 10 6 10^6 1 0 6 .
Proof. Exact rational interval evaluation of (18), (52), and (53) on the box in (51) gives the displayed bounds. The implementation in verify.py combines repeated terms before evaluation and checks that each interval denominator excludes zero. The arithmetic and source are described in Appendix A.
Geometric consequences.
Lemma 5.5. Let ( N 0 , h ) (N_0,h) ( N 0 , h ) be a Riemannian manifold and p ∈ N 0 p \in N_0 p ∈ N 0 . Suppose a unit vector E ∈ T p N 0 E \in T_pN_0 E ∈ T p N 0 spans a simple eigenspace of the Ricci endomorphism and the Jacobi endomorphism
span { E } ⊥ ⟶ span { E } ⊥ , X ⟼ R ( X , E ) E \operatorname{span}\{E\}^{\perp} \longrightarrow\operatorname{span}\{E\}^{\perp}, \qquad X \longmapsto R(X,E)E span { E } ⊥ ⟶ span { E } ⊥ , X ⟼ R ( X , E ) E is invertible. No neighbourhood of p p p admits a compatible parallel complex structure or a nontrivial Riemannian product decomposition.
Proof. A compatible parallel complex structure commutes with the Ricci endomorphism and hence preserves each of its eigenspaces. Its square is − Id -\operatorname{Id} − Id , which is impossible on a real line.
Suppose a neighbourhood of p p p is isometric to a nontrivial Riemannian product. The orthogonal projection pr 1 \operatorname{pr}_1 pr 1 onto the tangent bundle of the first factor is parallel and satisfies 0 < rank ( pr 1 ) < dim N 0 0 < \operatorname{rank}(\operatorname{pr}_1) < \dim N_0 0 < rank ( pr 1 ) < dim N 0 . The Ricci endomorphism is block diagonal for this product, so pr 1 \operatorname{pr}_1 pr 1 commutes with it. Simplicity of the eigenspace spanned by E E E and pr 1 2 = pr 1 \operatorname{pr}_1^2=\operatorname{pr}_1 pr 1 2 = pr 1 give pr 1 E = a E \operatorname{pr}_1 E=aE pr 1 E = a E for a ∈ { 0 , 1 } a \in\{0,1\} a ∈ { 0 , 1 } . Since pr 1 \operatorname{pr}_1 pr 1 is parallel, it also commutes with every curvature endomorphism. Consequently, for every X ∈ span { E } ⊥ X \in\operatorname{span}\{E\}^{\perp} X ∈ span { E } ⊥ ,
pr 1 ( R ( X , E ) E ) = R ( X , E ) pr 1 ( E ) = a R ( X , E ) E . \operatorname{pr}_1(R(X,E)E)=R(X,E)\operatorname{pr}_1(E)=aR(X,E)E. pr 1 ( R ( X , E ) E ) = R ( X , E ) pr 1 ( E ) = a R ( X , E ) E . The assumed invertibility implies that these vectors R ( X , E ) E R(X,E)E R ( X , E ) E span span { E } ⊥ \operatorname{span}\{E\}^{\perp} span { E } ⊥ . Hence pr 1 = a Id \operatorname{pr}_1=a\operatorname{Id} pr 1 = a Id at p p p , contradicting the rank condition.
Proposition 5.6. The compact soliton determined by Theorem 4.4 has a nonconstant potential, is non-Einstein, and has no nonempty open Kähler or nontrivial product subset.
Proof. At t = T L t=T_L t = T L , Table 5 gives u ′ ( T L ) ≥ 1879300 / 10 6 > 0 u'(T_L) \ge1879300/10^6 > 0 u ′ ( T L ) ≥ 1879300/1 0 6 > 0 and
min μ ∈ { ρ 0 , ρ 1 , ρ 2 } ∣ ρ ⊥ − μ ∣ ≥ 22435 10 6 > 0. \min_{\mu\in\{\rho_0,\rho_1,\rho_2\}}|\rho_{\perp}-\mu| \ge\frac{22435}{10^6}>0. μ ∈ { ρ 0 , ρ 1 , ρ 2 } min ∣ ρ ⊥ − μ ∣ ≥ 1 0 6 22435 > 0. Thus the potential is nonconstant and the metric is non-Einstein. Moreover, for every vector V V V tangent to this orbit,
g ( R ( V , ∂ t ) ∂ t , V ) ≥ 10537 10 6 g ( V , V ) . g(R(V,\partial_t)\partial_t,V) \ge\frac{10537}{10^6}g(V,V). g ( R ( V , ∂ t ) ∂ t , V ) ≥ 1 0 6 10537 g ( V , V ) . Thus the normal Ricci eigenspace is simple and the normal Jacobi endomorphism is positive definite. On the connected regular interval ( 0 , T ) (0,T) ( 0 , T ) , the solution of (20) is analytic. The six functions
ρ ⊥ − ρ 0 , ρ ⊥ − ρ 1 , ρ ⊥ − ρ 2 , κ 0 , κ 1 , κ 2 \rho_{\perp}-\rho_0,\quad\rho_{\perp}-\rho_1,\quad\rho_{\perp}-\rho_2,\quad\kappa_0,\quad\kappa_1,\quad\kappa_2 ρ ⊥ − ρ 0 , ρ ⊥ − ρ 1 , ρ ⊥ − ρ 2 , κ 0 , κ 1 , κ 2 are analytic there. Their product is nonzero at T L T_L T L by Table 5 , so its zero set is discrete in ( 0 , T ) (0,T) ( 0 , T ) . At every time outside that set, ∂ t \partial_t ∂ t spans a simple Ricci eigenspace, and the normal Jacobi endomorphism is invertible: its three block values are κ 0 , κ 1 , κ 2 \kappa_0,\kappa_1,\kappa_2 κ 0 , κ 1 , κ 2 . Theorem 5.5 applies at every point of such an orbit.
Let U ⊂ M U \subset M U ⊂ M be any nonempty open subset. It meets the regular set, since the two singular orbits have codimension two. Arc length is a submersion on the regular set, so the image of this intersection contains an open interval. That interval contains a time outside the discrete exceptional set, and U U U therefore contains a point to which Theorem 5.5 applies. A Kähler structure or product decomposition on U U U would restrict to a neighbourhood of that point, which is impossible.
Proof of Main Theorem. Theorem 4.4 supplies the matching solution, and Theorem 4.1 constructs the smooth compact metric and potential on the prescribed double disc bundle. Theorem 5.2 proves that M M M is simply-connected. Theorem 5.1 identifies the total space and its two singular orbits. Theorem 5.6 proves that the metric is non-Einstein, nowhere locally Kähler and locally irreducible.
Appendix A. Reproducibility and the finite certificate
Arithmetic and implementation.
The recorded replay used Python 3.12.10, NumPy 2.3.5 and Nodo 22.23.4 on 64-bit AMD64 Windows. Interval bounds use binary64 arithmetic with round-to-nearest, gradual underflow, and correctly rounded addition, subtraction, multiplication and division. Each elementary result is enlarged by one nextafter step towards each relevant infinity. Products use all four endpoint products, reciprocals use [ 1 / b , 1 / a ] [1/b,1/a] [ 1/ b , 1/ a ] when 0 ∉ [ a , b ] 0 \notin[a,b] 0 ∈ / [ a , b ] , and matrix products use explicit interval sums. Rational coefficients are enclosed by outward-rounded division of exactly represented integers; the root data and preconditioner entries are the exact dyadic rationals printed in the manuscript. Taylor coefficients are computed by standard power-series arithmetic, and derivatives by the product and quotient rules applied to the underlying functions. The remainder estimates in § 3 and Subsection A.2 give bounds for the solutions and their parameter derivatives. Every proposed integration step must pass the interval tests in Subsection A.2; overflow, invalid operations and division by zero cause rejection.
For the final checks, Fraction.from_float converts interval endpoints to exact dyadic rationals. The checker recomputes the matrix products and fixed-point image using rational arithmetic, checks the stated containments and inequalities, and verifies that consecutive integration steps meet exactly. Its constants in Theorem 4.3 are
α ∗ = 247976859883368044686315079332527 2 111 , (55) \alpha_* = \frac{247976859883368044686315079332527}{2^{111}}, \tag*{(55)} α ∗ = 2 111 247976859883368044686315079332527 , (55)
β ∗ = 236696180595019440104046978422827173381464177735 944473296573929 ⋅ 2 111 . (56) \beta_* = \frac{236696180595019440104046978422827173381464177735}{944473296573929 \cdot2^{111}}. \tag*{(56)} β ∗ = 944473296573929 ⋅ 2 111 236696180595019440104046978422827173381464177735 . (56)
They satisfy α ∗ < 0.096 \alpha_* < 0.096 α ∗ < 0.096 and β ∗ < 0.097 \beta_* < 0.097 β ∗ < 0.097 .
Validated continuation
We record the implementation of the standard validated Taylor estimates used in Subsection 4.3; see [19 ] and [27 Lemma 1 and Sections 2.4–2.6] . The nonautonomous equation is covered by adjoining the time variable with derivative 1. Below, derivatives with respect to W W W keep the initial time fixed.
At t = t k t=t_k t = t k , let c k c_k c k approximate the solution at p 0 p_0 p 0 . The vectors r k r_k r k and r k ( 0 ) r_k^{(0)} r k ( 0 ) bound, componentwise, the differences from c k c_k c k for all allowed data at the relevant singular orbit and for the data at p 0 p_0 p 0 , respectively. The 7 × 3 7 \times3 7 × 3 matrix P k P_k P k records an interval bound for each parameter derivative. For a step of length h = t k + 1 − t k h=t_{k+1}-t_k h = t k + 1 − t k , the program finds a compact box X k X_k X k , with f ~ 0 , f 1 2 , f 2 2 > 0 \tilde f_0,f_1^2,f_2^2>0 f ~ 0 , f 1 2 , f 2 2 > 0 , and checks
c k + [ − r k , r k ] + [ 0 , h ] G ( [ t , t + h ] , X k ) ⊂ int X k . (57) c_k + [-r_k,r_k] + [0,h]G([t,t+h],X_k) \subset\operatorname{int} X_k. \tag*{(57)} c k + [ − r k , r k ] + [ 0 , h ] G ([ t , t + h ] , X k ) ⊂ int X k . (57)
It uses the weighted maximum norm and its induced matrix norm,
w = ( 1 , 1 / 10 , 1 / 10 , 10 , 1 , 1 , 10 ) , ∥ v ∥ w = max i ∣ v i ∣ w i , ∥ A ∥ w = max i ∑ j ∣ A i j ∣ w j w i , (58) w=(1,1/10,1/10,10,1,1,10), \qquad\lVert v\rVert_w=\max_i\frac{|v_i|}{w_i}, \qquad\lVert A\rVert_w=\max_i\sum_j |A_{ij}|\frac{w_j}{w_i}, \tag*{(58)} w = ( 1 , 1/10 , 1/10 , 10 , 1 , 1 , 10 ) , ∥ v ∥ w = i max w i ∣ v i ∣ , ∥ A ∥ w = i max j ∑ ∣ A ij ∣ w i w j , (58)
and bounds ∥ D W G ( τ , W ) ∥ w ≤ Λ k \lVert D_WG(\tau,W)\rVert_w\leq\Lambda_k ∥ D W G ( τ , W ) ∥ w ≤ Λ k for τ ∈ [ t , t + h ] \tau\in[t,t+h] τ ∈ [ t , t + h ] and W ∈ X k W\in X_k W ∈ X k , requiring h Λ k < 7 / 20 h\Lambda_k<7/20 h Λ k < 7/20 . Since t > 0 t>0 t > 0 and the first three coordinates are positive, G G G is analytic near this box. The strict inclusion in (57) guarantees continuation inside X k X_k X k . Gronwall’s inequality for the first variational equation gives, for W ∗ ∈ c k + [ − r k , r k ] W_*\in c_k+[-r_k,r_k] W ∗ ∈ c k + [ − r k , r k ] and 0 ≤ σ ≤ h 0\leq\sigma\leq h 0 ≤ σ ≤ h ,
∥ D W Φ t , σ ( W ∗ ) − I ∥ w ≤ e h Λ k − 1 ≤ h Λ k 1 − h Λ k . (59) \lVert D_W\Phi_{t,\sigma}(W_*)-I\rVert_w\leq e^{h\Lambda_k}-1\leq\frac{h\Lambda_k}{1-h\Lambda_k}. \tag*{(59)} ∥ D W Φ t , σ ( W ∗ ) − I ∥ w ≤ e h Λ k − 1 ≤ 1 − h Λ k h Λ k . (59)
Thus the matrix whose ( i , j ) (i,j) ( i , j ) entry is
Σ i j = I i j + h Λ k 1 − h Λ k w i w j [ − 1 , 1 ] \Sigma_{ij}=I_{ij}+\frac{h\Lambda_k}{1-h\Lambda_k}\frac{w_i}{w_j}[-1,1] Σ ij = I ij + 1 − h Λ k h Λ k w j w i [ − 1 , 1 ]
bounds the flow derivative throughout the step.
Let T n ( t , W ) T_n(t,W) T n ( t , W ) be the coefficient of v n v^n v n in Φ t , v ( W ) \Phi_{t,v}(W) Φ t , v ( W ) . The program computes these coefficients and their derivatives by interval evaluation of the recurrence
T 0 ( t , W ) = W , T n + 1 ( t , W ) = 1 n + 1 [ v n ] G ( t + v , ∑ j = 0 n T j ( t , W ) v j ) . T_0(t,W)=W,\qquad T_{n+1}(t,W)=\frac{1}{n+1}[v^n]G\left(t+v,\sum_{j=0}^{n}T_j(t,W)v^j\right). T 0 ( t , W ) = W , T n + 1 ( t , W ) = n + 1 1 [ v n ] G ( t + v , j = 0 ∑ n T j ( t , W ) v j ) .
Here [ v n ] [v^n] [ v n ] extracts the coefficient of v n v^n v n in each coordinate. The finite Taylor estimates of [27 Lemma 1] give
Φ t , h ( c k ) ∈ ∑ n = 0 14 h n T n ( t , c k ) + h 15 T 15 ( [ t , t + h ] , X k ) , \Phi_{t,h}(c_k)\in\sum_{n=0}^{14}h^nT_n(t,c_k)+h^{15}T_{15}([t,t+h],X_k), Φ t , h ( c k ) ∈ n = 0 ∑ 14 h n T n ( t , c k ) + h 15 T 15 ([ t , t + h ] , X k ) ,
J k = ∑ n = 0 8 h n D W T n ( t , c k + [ − r k , r k ] ) + h 9 R 9 , \mathcal{J}_k=\sum_{n=0}^{8}h^nD_WT_n(t,c_k+[-r_k,r_k])+h^9\mathcal{R}_9, J k = n = 0 ∑ 8 h n D W T n ( t , c k + [ − r k , r k ]) + h 9 R 9 ,
where J k \mathcal{J}_k J k bounds each entry of D W Φ t , h ( c k + [ − r k , r k ] ) D_W\Phi_{t,h}(c_k + [-r_k,r_k]) D W Φ t , h ( c k + [ − r k , r k ]) . The remainder R 9 \mathcal{R}_9 R 9 bounds D W T 9 ( τ , W ) Σ D_WT_9(\tau,W)\boldsymbol{\Sigma} D W T 9 ( τ , W ) Σ for τ ∈ [ t , t + h ] \tau\in[t,t+h] τ ∈ [ t , t + h ] and W ∈ X k W \in X_k W ∈ X k . This bound is computed by applying the product and quotient rules to the same recurrence, with the initial perturbation matrix at T 0 = W T_0=W T 0 = W set to Σ \boldsymbol{\Sigma} Σ .
Choose c k + 1 c_{k+1} c k + 1 as the midpoint of the computed interval for Φ t , h ( c k ) \Phi_{t,h}(c_k) Φ t , h ( c k ) , and let q k q_k q k bound its componentwise radius about c k + 1 c_{k+1} c k + 1 , including rounding error. The mean-value formula and the chain rule justify the updates
r k + 1 ≥ q k + ∣ J k ∣ r k , r k + 1 ( 0 ) ≥ q k + ∣ J k ∣ r k ( 0 ) , P k + 1 ⊃ J k P k . r_{k+1} \ge q_k + |\mathcal{J}_k|r_k,\qquad r_{k+1}^{(0)} \ge q_k + |\mathcal{J}_k|r_k^{(0)},\qquad P_{k+1} \supset\mathcal{J}_kP_k. r k + 1 ≥ q k + ∣ J k ∣ r k , r k + 1 ( 0 ) ≥ q k + ∣ J k ∣ r k ( 0 ) , P k + 1 ⊃ J k P k .
Here ∣ J k ∣ |\mathcal{J}_k| ∣ J k ∣ denotes the componentwise supremum of the absolute values of its entries. All bounds are rounded outwards, and 0 ≤ r k ( 0 ) ≤ r k 0 \le r_k^{(0)} \le r_k 0 ≤ r k ( 0 ) ≤ r k is checked at every step. The mesh is common to all allowed data at a given singular orbit and is held fixed when differentiating. Stored times and step sizes are interpreted as exact dyadic rationals, and each equality t k + h = t k + 1 t_k+h=t_{k+1} t k + h = t k + 1 is checked in exact fraction arithmetic, including the final step to the prescribed matching time.
Numerical exploration and reproduction
The historical search described in Subsection 3.5 is recorded in the embedded numerical-discovery.json, which contains the initial and final parameters, outcomes and source fingerprints. The historical record is not an input to the rigorous matching test. The optional command
python -B verify.py --numerics --output numerical-evidence
reproduces the refinement, convergence tables and figure data using 60-digit decimal arithmetic. Starting from the six-decimal rounding of p 0 p_0 p 0 , the fixed-inverse iteration uses order 24 and a maximum step size of 1 / 32 1/32 1/32 , stopping when the residual falls below 10 − 28 10^{-28} 1 0 − 28 . A separate order-30 evaluation with a maximum step size of 1 / 64 1/64 1/64 gives a residual of approximately 4.32 ⋅ 10 − 31 4.32 \cdot10^{-31} 4.32 ⋅ 1 0 − 31 . The convergence table uses the same exact verification inputs at every order. This mode requires Python and NumPy and reports numerical evidence only; full interval verification is described below.
Reproduction of the computation
The sole computational attachment, verify.py, contains all numerical inputs, readable source blocks and the historical search record, and checks each block’s hash before extraction. Its SHA256, using UTF-8 with LF line endings, is
4D9C89E9945DE6FEC40C268EDC9EA344AE4A1909115DA68AD3F5231C234BFB65.
The source uses documented ASCII versions of the manuscript’s notation. States, radii, flow weights and flow Jacobians follow ( W 1 , … , W 7 ) (W_1,\ldots,W_7) ( W 1 , … , W 7 ) ; solution data follow C ( Y ) C(Y) C ( Y ) , and parameters follow (40) . The tuple p_hat is α ^ \hat{\alpha} α ^ or ω ^ \hat{\omega} ω ^ , denoted by c ^ \hat{c} c ^ in the derivative estimates. The Nomizu calculation uses tangent-block order ( m 0 , m 1 , m 2 , m 3 ) (m_0,m_1,m_2,m_3) ( m 0 , m 1 , m 2 , m 3 ) and polynomial-variable order ( f 0 2 , f 1 2 , f 2 2 , f 3 2 ) (f_0^2,f_1^2,f_2^2,f_3^2) ( f 0 2 , f 1 2 , f 2 2 , f 3 2 ) . The source blocks have the following roles:
Table 7.
With the runtime described above and Node on the executable search path, run
python -B verify.py --output replay
The directory must be new or empty; it receives the extracted sources, step traces, certificates and a summary identifying the executed verifier by its SHA256. All flow bounds are recomputed, and a successful full run verifies (54)–(55) and reports FULL COMPUTATIONAL PROOF REPLAY PASS. The wrapper rejects Python options (-O, -OO), since assertions perform mathematical checks. The optional --quick mode checks source integrity, local estimates and exact curvature, but omits the integrations and does not verify the matching zero. Omitting --output uses a temporary directory; --node specifies the Node executable by path.
The exact preconditioner
The table specifies every entry of the fixed 7 × 7 7 \times7 7 × 7 matrix C C C in Theorem 4.3. The column order is the order of the solution data ( f 0 , f 1 2 , f 2 2 , L 0 , L 1 , L 2 , ξ ) (f_0,f_1^2,f_2^2,L_0,L_1,L_2,\xi) ( f 0 , f 1 2 , f 2 2 , L 0 , L 1 , L 2 , ξ ) and the row order is the coordinate order of (40) . All entries are exact dyadic rationals. The matrix was obtained by numerically inverting a central-difference Jacobian at the candidate. Its entries are then treated as fixed exact inputs. Its adequacy is established by the bound on I − C A I-CA I − C A in (46) .
Table 6. Exact entries of C C C .
References