_id
stringlengths
36
36
text
stringlengths
5
665k
marker
stringlengths
3
6
marker_offsets
sequence
label
stringlengths
28
32
a381057b-9259-4cd1-b352-224bd76a4a64
The following theorem was proved by Agol [1]} and Wise [2]}, [3]} in the hyperbolic case. It was proved by Liu [4]} and Przytycki-Wise [5]} for graph manifolds with boundary and it was proved by Przytycki-Wise [6]} for manifolds with a non-trivial Jaco-Shalen-Johannson decomposition and at least one hyperbolic piece in the JSJ decomposition.
[3]
[ [ 61, 64 ] ]
https://openalex.org/W3080612394
587be717-9224-49ab-8539-b9c25832bd4d
By the Sphere Theorem [1]}, an irreducible 3-manifold is aspherical, i.e., all its higher homotopy groups vanish, if and only if it is a 3-disk or has infinite fundamental group. If \(M\) and \(N\) are two aspherical closed 3-manifolds, then they are homeomorphic if and only if their fundamental groups are isomorphic. Actually, every isomorphism between their fundamental groups is induced by a homeomorphism. More generally, every 3-manifold \(N\) with torsionfree fundamental group group is topologically rigid in the sense that any homotopy equivalence of closed 3-manifolds with \(N\) as target is homotopic to a homeomorphism. This follows from results of Waldhausen, see Hempel [1]} and Turaev [3]}, as explained for instance [4]}.
[1]
[ [ 22, 25 ], [ 690, 693 ] ]
https://openalex.org/W4238266252
1e522c5f-462d-46b5-9016-5c90478ebdb3
The fundamental group of a closed manifold is finitely presented. Fix a natural number \(d \ge 4\) . Then a group \(G\) is finitely presented if and only if it occurs as fundamental group of a closed orientable \(d\) -dimensional manifold. This is not true in dimension 3. A detailed exposition about the problem, which finitely presented groups occur as fundamental groups of closed 3-manifolds, can be found in [1]}. For us it will be important that the fundamental group of any 3-manifold is residually finite, This follows from [2]} and the proof of the Geometrization Conjecture. More information about fundamental groups of 3-manifolds can be found for instance in [1]}.
[1]
[ [ 414, 417 ], [ 672, 675 ] ]
https://openalex.org/W1956890014
c2adccc7-e3bb-4608-b3de-91e206db7880
Let \(M\) be a compact oriented 3-manifold. Recall the definition in [1]} of the Thurston norm \(x_M(\phi )\) of a 3-manifold \(M\) and an element \(\phi \in H^1(M;{\mathbb {Z}})=\operatorname{Hom}(\pi _1(M),{\mathbb {Z}})\) : \(x(\phi )–:=\min \lbrace \chi _-(F)\, | \, F \subset M \textup { properly embedded surface dualto }\phi \rbrace ,\)
[1]
[ [ 70, 73 ] ]
https://openalex.org/W122881644
32ad9e2a-6075-4b90-b5f0-678387026b5b
Thurston [1]} showed that this defines a seminorm on \(H^1(M;\mathbb {Z})\) which can be extended to a seminorm on \(H^1(M;\mathbb {R} )\) which we also denote by \(x_M\) . In particular we get for \(r \in {\mathbb {R}}\) and \(\phi \in H^1(M;{\mathbb {R}})\) \(x_M(r \cdot \phi )& = &|r| \cdot x_M(\phi ).\)
[1]
[ [ 9, 12 ] ]
https://openalex.org/W122881644
750f1f73-f4f8-40c6-9719-047ca607ce77
If \(p \colon \widetilde{M} \rightarrow M\) is a finite covering with \(n\) sheets, then Gabai [1]} showed that \(x_{\widetilde{M}}(p^*\phi )& = &n \cdot x_M(\phi ).\)
[1]
[ [ 97, 100 ] ]
https://openalex.org/W1560527130
5bc115ef-6925-49a1-b72e-df616ccd4106
If \(F \rightarrow M \xrightarrow{} S^1\) is a fiber bundle for a 3-manifold \(M\) and compact surface \(F\) , and \(\phi \in H^1(M;{\mathbb {Z}})\) is given by \(H_1(p) \colon H_1(M) \rightarrow H_1(S^1)={\mathbb {Z}}\) , then by [1]} we have \(x_M(\phi ) & = &{\left\lbrace \begin{array}{ll}- \chi (F), & \text{if} \;\chi (F) \le 0;\\0, & \text{if} \;\chi (F) \ge 0.\end{array}\right.}\)
[1]
[ [ 234, 237 ] ]
https://openalex.org/W122881644
ccc790ab-de69-44fa-b74b-85020a6594b5
Thurston [1]} has shown that \(T(M)^*\) is an integral polytope, i.e, the convex hull of finitely many points in the integral lattice \(H_1(M;{\mathbb {Z}})/\mbox{torsion} \subseteq H_1(M;{\mathbb {R}})\) .
[1]
[ [ 9, 12 ] ]
https://openalex.org/W122881644
1bd55364-3241-4748-93e9-0d41a58403b8
A marking for a polytope is a (possibly empty) subset of the set of its vertices. We conclude from Thurston [1]} that we can equip \(T(M)^*\) with a marking such that \(\phi \in H^1(M;{\mathbb {R}})\) is fibered if and only if it pairs maximally with a marked vertex, i.e., there exists a marked vertex \(v\) of \(T(M)^*\) , such that \(\phi (v) >\phi (w)\) for any vertex \(w\ne v\) .
[1]
[ [ 108, 111 ] ]
https://openalex.org/W122881644
87f14232-ecdd-4835-8a91-f3fa7021248f
For some information about the proof and in particular of references in the literature we refer to [1]} except for assertion (REF ) which is due to Jaikin-Zapirain and Lopez-Alvarez [2]}. A group is called locally indicable if every non-trivial finitely generated subgroup admits an epimorphism onto \({\mathbb {Z}}\) . Examples are torsionfree one-relator groups.
[2]
[ [ 182, 185 ] ]
https://openalex.org/W3105807034
ce6a6f60-9b6e-4a56-a8f1-fef54d3a1312
There is a program of Linnell [1]} to prove the Atyiah Conjecture which is discussed in details for instance in [2]} and [3]}. This shows that one has at least some ideas why the Atyiah Conjecture is true and that the Atiyah Conjecture is related to some deep ring theory and to algebraic \(K\) -theory, notably to projective class groups. This connection to ring theory has been explained and exploited for instance in [4]}, [5]}, where the division closure is replaced by the \(\ast \) -regular closure.
[4]
[ [ 420, 423 ] ]
https://openalex.org/W2922526935
6f39928a-8a0a-44d3-9703-153fe86eefb0
There is a program of Linnell [1]} to prove the Atyiah Conjecture which is discussed in details for instance in [2]} and [3]}. This shows that one has at least some ideas why the Atyiah Conjecture is true and that the Atiyah Conjecture is related to some deep ring theory and to algebraic \(K\) -theory, notably to projective class groups. This connection to ring theory has been explained and exploited for instance in [4]}, [5]}, where the division closure is replaced by the \(\ast \) -regular closure.
[5]
[ [ 426, 429 ] ]
https://openalex.org/W3105807034
6f3afbfc-b3de-41a3-a6c9-296861418d36
The class of sofic groups is very large. It is closed under direct and free products, taking subgroups, taking inverse and direct limits over directed index sets, and is closed under extensions with amenable groups as quotients and a sofic group as kernel. In particular it contains all residually amenable groups and fundamental groups of 3-manifolds. One expects that there exists non-sofic groups but no example is known. More information about sofic groups can be found for instance in [1]} and [2]}.
[2]
[ [ 499, 502 ] ]
https://openalex.org/W2153827714
5870d41a-7f0f-4e88-88f5-b21e7148bcf6
Remark 3.22 The conjectures above imply a positive answer to [1]} and [2]}. They also would settle [3]} and [4]}. One may wonder whether it is related to the Volume Conjecture due to Kashaev [5]} and H. and J. Murakami [6]}.
[6]
[ [ 220, 223 ] ]
https://openalex.org/W2023409861
81b5e8ee-528b-43b0-ab4d-6b34bed4a8c0
The proof of the following result can be found in [1]}. It reduces in the weakly acyclic case Conjecture REF to Conjecture REF .
[1]
[ [ 50, 53 ] ]
https://openalex.org/W2557229284
10374974-cf0b-4fde-ad4b-978fd1e92dae
It is conceivable that Theorem REF remains true if we drop the assumption that \(b_p^{(2)}(\overline{M};{\mathcal {N}}(G))\) vanishes for all \(p \ge 0\) , but our present proof works only under this assumption, see [1]}.
[1]
[ [ 218, 221 ] ]
https://openalex.org/W2557229284
59ce72cd-9857-4534-adad-155820dd8081
More information about the conjectures above can be found in [1]}.
[1]
[ [ 61, 64 ] ]
https://openalex.org/W2557229284
db7a7163-c10b-40a4-a69f-edd42dcd2ca3
Conjecture REF is attributed to Bergeron-Venkatesh [1]}. They allow only locally symmetric spaces for \(M\) . They also consider the case of twisting with a finite-dimensional integral representation. Further discussions about this conjecture can be found for instance in [2]}, [3]}, and [4]}.
[1]
[ [ 52, 55 ] ]
https://openalex.org/W2963260686
1d34ca4b-d33c-4234-8aac-653aec519c19
Conjecture REF is attributed to Bergeron-Venkatesh [1]}. They allow only locally symmetric spaces for \(M\) . They also consider the case of twisting with a finite-dimensional integral representation. Further discussions about this conjecture can be found for instance in [2]}, [3]}, and [4]}.
[2]
[ [ 273, 276 ] ]
https://openalex.org/W1956890014
a6fae5ed-6757-4798-9eea-c76bb8bfb1e9
Conjecture REF is attributed to Bergeron-Venkatesh [1]}. They allow only locally symmetric spaces for \(M\) . They also consider the case of twisting with a finite-dimensional integral representation. Further discussions about this conjecture can be found for instance in [2]}, [3]}, and [4]}.
[3]
[ [ 279, 282 ] ]
https://openalex.org/W3100738199
9c3e93e9-1bb1-485d-b8bd-0e50ff61508b
Conjecture REF is attributed to Bergeron-Venkatesh [1]}. They allow only locally symmetric spaces for \(M\) . They also consider the case of twisting with a finite-dimensional integral representation. Further discussions about this conjecture can be found for instance in [2]}, [3]}, and [4]}.
[4]
[ [ 289, 292 ] ]
https://openalex.org/W3101583560
f309edfa-e2ab-48b0-9062-8d30537770e6
The relation between Conjecture REF and Conjecture REF is discussed in [1]}.
[1]
[ [ 73, 76 ] ]
https://openalex.org/W2557229284
9ddc271a-2143-4d91-8d02-09aacedf22ef
The chain complex version Conjecture REF is stated in [1]}. We at least explain what it says for 1-dimensional chain complexes, or, equivalently, matrices. Here it is important to work over the integral group ring.
[1]
[ [ 55, 58 ] ]
https://openalex.org/W2557229284
0b67926f-8992-4d2d-8150-bd6dfe404a7c
Notice that \(||y||_1\) defines only a seminorm on \(H_p^{\operatorname{sing}}(X;{\mathbb {R}})\) , it is possible that \(||y||_1 = 0\) but \(y \ne 0\) . The next definition is taken from [1]}.
[1]
[ [ 191, 194 ] ]
https://openalex.org/W1570262040
df1a3754-3fb5-4f0c-99bc-34fd608b5d21
Bergeron-Sengun-Venkatesh [1]} consider the equality above for arithmetic hyperbolic 3-manifolds and relate it to a conjecture about classes in the second integral homology.
[1]
[ [ 26, 29 ] ]
https://openalex.org/W3100738199
f91cff73-a969-403d-a2ff-9eec9569e33d
Define the positive real number \(v_3\) to be the supremum of the volumes of all \(n\) -dimensional geodesic simplices, i.e., the convex hull of \((n+1)\) points in general position, in the \(n\) -dimensional hyperbolic space \({\mathbb {H}}^3\) . If \(M\) is an admissible 3-manifold, then one gets from [1]}, [2]}, and [3]}, see [4]} \(||M|| = \frac{- 6\pi }{v_3} \cdot \rho ^{(2)}(\widetilde{M}).\)
[2]
[ [ 314, 317 ] ]
https://openalex.org/W2019578387
144e1559-62b5-4af9-837b-b9b70a0bf0c7
There are variants of the simplicial volume, namely, the notion of the integral foliated simplicial volume, see [1]}, [2]}, or [3]}, and of the stable integral simplicial volume, see [3]}. The integral foliated simplicial volume gives an upper bound on the torsion growth for an oriented closed manifold, i.e, an upper bound on \(\limsup _{i \rightarrow \infty } \;\frac{\ln \big (\bigl |\operatorname{tors}(H_n(M[i];{\mathbb {Z}}))\bigr |\bigr )}{[G:G_i]}\) in the situation of Conjecture REF , see [3]}. There are the open questions whether for an aspherical oriented closed manifold the simplicial volume and the integral foliated simplicial volume agree and whether for an aspherical oriented closed manifold with residually finite fundamental group the integral foliated simplicial volume and the stable integral simplicial volume agree, see [3]}. The stable integral simplicial volume and the simplicial volume agree for aspherical oriented closed 3-manifolds, see [7]}.
[1]
[ [ 112, 115 ] ]
https://openalex.org/W1483679902
f6504ac4-067d-4234-a224-4fb888ac676e
If \(\eta ^{(2)}_{V_u,B_u} (C_*)\) has a gap at the spectrum at zero, then obviously the \(L^2\) -torsion of \(\eta ^{(2)}_{V_u,B_u} (C_*)\) is well-defined. Moreover the function sending \(v \in R(G,\operatorname{GL}_n({\mathbb {C}}))\) to the \(L^2\) -torsion of \(\eta ^{(2)}_{V_u,B_u} (C_*)\) is well-defined and continuous in neighborhood of \(u\) . This follows form the continuity of the Fulgede-Kadison determinant for invertible matrices over the group von Neumann algebra with respect to the norm topology, see [1]}, [2]}, or, [3]}. This is studied in more detail for a hyperbolic 3-manifold \(M\) with empty or incompressible torus boundary and the canonical holonomy representation \(h \colon \pi _1(M) \rightarrow \operatorname{SL}_2({\mathbb {C}})\) by Bénard-Raimbault [4]}. They actually show that this function is real analytic near \(h\) .
[2]
[ [ 531, 534 ] ]
https://openalex.org/W2323746695
b2dd72d2-40d6-4e1c-bc27-0e48ca6bede1
Remark 7.2 (Assumption REF ) The reader does not need to know what the \(K\) -theoretic Farrell-Jones Conjecture for \({\mathbb {Z}}G\) is, it can be used as a black box. The reader should have in mind that it is known for a large class of groups, e.g., hyperbolic groups, CAT(0)-groups, solvable groups, lattices in almost connected Lie groups, fundamental groups of 3-manifolds and passes to subgroups, finite direct products, free products, and colimits of directed systems of groups (with arbitrary structure maps). For more information we refer for instance to [1]}, [2]}, [3]}, [4]}, [5]}, [6]}, [7]}.
[4]
[ [ 586, 589 ] ]
https://openalex.org/W2094789018
2818f70b-e6da-4105-a71e-74cbbafd66cf
Notice that for polytopes \(P_0\) , \(P_1\) and \(Q\) in a finite-dimensional real vector space we have the implication \(P_0 + Q = P_1 + Q \Longrightarrow P_0 = P_1\) , see [1]}. Hence elements in \(\mathcal {P}_{{\mathbb {Z}}}(H)\) are given by formal differences \([P] - [Q]\) for integral polytopes \(P\) and \(Q\) in \({\mathbb {R}}\otimes _{{\mathbb {Z}}} H\) and we have \([P_0] - [Q_0] = [P_1] - [Q_1] \Longleftrightarrow P_0 + Q_1 = P_1 + Q_0\) .
[1]
[ [ 176, 179 ] ]
https://openalex.org/W4245290017
1d2fe87e-a508-4bab-a98c-098fc16ecc9c
It makes no difference whether \(\widehat{K} \cong \widehat{G}\) means abstract isomorphism of groups or topological group isomorphism, see Nikolov and Segal [1]}.
[1]
[ [ 159, 162 ] ]
https://openalex.org/W1997331766
ba73689d-76ea-4974-949a-0de902e8a57e
To the author's knowledge profinite rigidity of fundamental groups of hyperbolic closed 3-manifolds, even among themselves, is an open question. Examples of hyperbolic closed 3-manifolds, whose fundamental groups are profinite rigid in the absolute sense, are constructed in [1]}. A weaker but still open problem is the following which is equivalent to [2]}.
[1]
[ [ 275, 278 ] ]
https://openalex.org/W3102237430
61bec780-e31a-4747-b586-f782bf2e72cb
Our model may be seen as modelling another version of `clumping' as discussed by [1]} which, like the models they investigate, allows for more variability in outcomes than standard homogeneously mixing models. The model is also closely related to the `epidemic among giants' of [2]} and the discussion at end of section 4.3 of that paper; but that model considers only Reed-Frost epidemic dynamics (see Section REF , 3rd paragraph) and here we provide much more detailed and complete results. The stochastic multi-type model with weaker transmission between types than within types goes back at least to [3]}, who cites related models in publications from the late 1950s. The idea of a population of communities with relatively strong within-community links and weaker between-community links is similar in spirit to some motivations for the Stochastic Block Model or planted partition model (see e.g. [4]} in the probabilistic literature or [5]} in the networks and community detection literature); though in that context the strength of between- and within-community connections are usually, but not always, assumed to scale with population size in the same way as each other (as is the case in the usual multi-type epidemic model).
[5]
[ [ 942, 945 ] ]
https://openalex.org/W2559839022
d4cef450-b432-4ae9-be44-20d6deae2312
Thanks to the above references on the Cauchy problem for (REF ), solving (REF ) will not be an issue here because we know from [1]}, [2]}, [3]}, [4]} that (REF ) is globally well-posed in \(H^s(\) , \(s\ge 1\) (and even locally well-posed for \(s\ge \frac{1}{2}\) ). In fact, for \(M<\infty \) an easier argument can be applied by using that for frequencies \(\le M\) the equation (REF ) becomes an ODE (with a Lipschitz vector field) for which global well-posedness holds thanks to the \(L^2\) conservation law, while for frequencies \(>M\) the equation (REF ) becomes a linear equation. However, if one wishes to have \(H^s(\) bounds uniformly with respect to \(M\) , the analysis of [1]}, [2]}, [3]}, [4]} cannot be avoided. Let us denote by \(\Phi _M(t)\) the flow of (REF ). For \(M=\infty \) we simply write \(\Phi _{\infty }(t)=\Phi (t)\) . We shall note specify the dependence on \(p\) in these notations.
[3]
[ [ 139, 142 ], [ 704, 707 ] ]
https://openalex.org/W2075677712
4ec1e11d-e13d-475b-89e2-193034831fd4
The proof of Theorem REF relies on a key improvement on our previous papers [1]} and [2]}, together with bounds resulting from dispersive estimates such as Bourgain's \(L^6\) Strichartz inequality for (linear) KdV. Let us now briefly explain how we improve: following our argument in [2]} we would get on the r.h.s. of (REF ) a power larger than two of the \(H^{(k-\frac{1}{2})^-}\) norm of the initial datum, while in (REF ) we get a power less than two. This is the key to proving, beyond quasi-invariance, \(L^p\) regularity for the density of the transported Gaussian measure (see Theorem REF below). Similarly, one should also compare (REF ) with the estimate obtained in [1]} which allows to get on the r.h.s. of (REF ) a power of the \(H^{k}\) norm of the initial datum that is worse than the one that we get in this paper. The improvement on the growth of the Sobolev norms that we get in Theorem REF below relies in a fundamental way on improving this power. For an overview of the results in [1]} and [2]} we refer to [7]}. The aforementioned improvements on both exponents come from a refinement of the energies, compared with the ones used in the previous papers. In particular, once we compute its time derivative along solutions we get a multilinear expression of densities in which for every single term at least five factors involve one nontrivial derivative. This key property of distribution of derivatives on several factors was out of reach with our previous constructions of modified energies. For more details see Section . Then the dispersive effect, through the \(L^6\) Strichartz bound, allows us to transform the aforementioned distribution of derivative in terms of powers of Sobolev norms of the initial datum, as discussed above.
[1]
[ [ 77, 80 ], [ 682, 685 ], [ 1009, 1012 ] ]
https://openalex.org/W3102750408
1d6059da-a2d5-42df-a744-819ee6f33811
Finally, we should point out that, in the context of gKdV, modified energies already appeared, at the level of \(k=2\) , in [1]}, where they are used in connection with \(N\) -soliton asymptotics.
[1]
[ [ 124, 127 ] ]
https://openalex.org/W2043098860
2f40ae57-f577-4627-87ba-1588d2e9772f
The line of research leading to results as the one in Theorem REF was initiated in [1]}. We improve results obtained in [2]} (where the growth had exponent \(2k\) ) and [3]} (where the growth was lowered to \(k-1+\varepsilon \) .) For details on how (REF ), (REF ) in Theorem REF imply Theorem REF we refer to [4]}.
[4]
[ [ 313, 316 ] ]
https://openalex.org/W3102750408
379d6f7a-052c-4565-9374-8bc5af5774da
It would be very interesting to construct solutions of the defocusing gKdV such that the \(H^k\) , \(k>1\) norms do not remain bounded in time. Unfortunately such results are rare in the context of canonical dispersive models (with the notable exception of [1]}).
[1]
[ [ 258, 261 ] ]
https://openalex.org/W2963594978
83adb9ec-7252-4f9f-864e-be907267ed63
Theorem REF fits in the line of research aiming to describe macroscopical (statistical dynamics) properties of Hamiltonian PDE's. The earliest references we are aware of is [1]}, followed by [2]}, [3]}, [4]}, [5]}. Inspired by the work on invariant measures for the Benjamin-Ono equation [6]}, [7]}, [8]}, [9]}, quasi-invariance of Gaussian measure for several dispersive models was obtained in recent years, see [10]}, [11]}, [12]}, [13]}, [14]}, [15]}, [16]}, [17]}, [18]}, [19]}, [20]}, [21]}, [22]}, [23]}. The method to identify the densities in Theorem REF is inspired by recent works [10]}, [25]}. In Theorem REF , we provide much more information on the densities when compared to [21]}, which used modified energies on the nonlinear Schrödinger equation. It should be underlined that a key novelty in the proof of Theorem REF with respect to [25]} and [21]} is that we crucially use dispersive estimates in the analysis.
[4]
[ [ 204, 207 ] ]
https://openalex.org/W1990089063
bdc7bcb7-8991-4797-bcf2-402309b0bd50
Theorem REF fits in the line of research aiming to describe macroscopical (statistical dynamics) properties of Hamiltonian PDE's. The earliest references we are aware of is [1]}, followed by [2]}, [3]}, [4]}, [5]}. Inspired by the work on invariant measures for the Benjamin-Ono equation [6]}, [7]}, [8]}, [9]}, quasi-invariance of Gaussian measure for several dispersive models was obtained in recent years, see [10]}, [11]}, [12]}, [13]}, [14]}, [15]}, [16]}, [17]}, [18]}, [19]}, [20]}, [21]}, [22]}, [23]}. The method to identify the densities in Theorem REF is inspired by recent works [10]}, [25]}. In Theorem REF , we provide much more information on the densities when compared to [21]}, which used modified energies on the nonlinear Schrödinger equation. It should be underlined that a key novelty in the proof of Theorem REF with respect to [25]} and [21]} is that we crucially use dispersive estimates in the analysis.
[9]
[ [ 307, 310 ] ]
https://openalex.org/W2255092212
41f0118c-7ace-4aac-8490-337dca5fa82f
Theorem REF fits in the line of research aiming to describe macroscopical (statistical dynamics) properties of Hamiltonian PDE's. The earliest references we are aware of is [1]}, followed by [2]}, [3]}, [4]}, [5]}. Inspired by the work on invariant measures for the Benjamin-Ono equation [6]}, [7]}, [8]}, [9]}, quasi-invariance of Gaussian measure for several dispersive models was obtained in recent years, see [10]}, [11]}, [12]}, [13]}, [14]}, [15]}, [16]}, [17]}, [18]}, [19]}, [20]}, [21]}, [22]}, [23]}. The method to identify the densities in Theorem REF is inspired by recent works [10]}, [25]}. In Theorem REF , we provide much more information on the densities when compared to [21]}, which used modified energies on the nonlinear Schrödinger equation. It should be underlined that a key novelty in the proof of Theorem REF with respect to [25]} and [21]} is that we crucially use dispersive estimates in the analysis.
[15]
[ [ 449, 453 ] ]
https://openalex.org/W3183210163
c8f741ef-16b2-4e0c-a493-20654b4812ca
Theorem REF fits in the line of research aiming to describe macroscopical (statistical dynamics) properties of Hamiltonian PDE's. The earliest references we are aware of is [1]}, followed by [2]}, [3]}, [4]}, [5]}. Inspired by the work on invariant measures for the Benjamin-Ono equation [6]}, [7]}, [8]}, [9]}, quasi-invariance of Gaussian measure for several dispersive models was obtained in recent years, see [10]}, [11]}, [12]}, [13]}, [14]}, [15]}, [16]}, [17]}, [18]}, [19]}, [20]}, [21]}, [22]}, [23]}. The method to identify the densities in Theorem REF is inspired by recent works [10]}, [25]}. In Theorem REF , we provide much more information on the densities when compared to [21]}, which used modified energies on the nonlinear Schrödinger equation. It should be underlined that a key novelty in the proof of Theorem REF with respect to [25]} and [21]} is that we crucially use dispersive estimates in the analysis.
[17]
[ [ 463, 467 ] ]
https://openalex.org/W2964230361
9f93712b-7c09-40a0-ae12-4e6cc5d102c9
The results of this paper and previous works of the second and third authors [1]}, [2]}, [3]}, [4]} can be summarized as follows. In the case of integrable models, exact conservation laws for all Sobolev regularities imply existence of invariant measures; the modified energies we construct in the context of non integrable models imply existence of quasi-invariant measures. Concerning the deterministic behavior of the solutions, exact conservation laws imply uniform bounds on Sobolev norms of solutions while the modified energies we construct imply polynomial bounds on Sobolev norms of solutions.
[1]
[ [ 77, 80 ] ]
https://openalex.org/W3102008150
d207e293-9735-44da-863b-c95b2c76ad6e
The results of this paper and previous works of the second and third authors [1]}, [2]}, [3]}, [4]} can be summarized as follows. In the case of integrable models, exact conservation laws for all Sobolev regularities imply existence of invariant measures; the modified energies we construct in the context of non integrable models imply existence of quasi-invariant measures. Concerning the deterministic behavior of the solutions, exact conservation laws imply uniform bounds on Sobolev norms of solutions while the modified energies we construct imply polynomial bounds on Sobolev norms of solutions.
[4]
[ [ 95, 98 ] ]
https://openalex.org/W2255092212
965e27be-d17c-4b96-bc62-b367f8857557
Acknowledgement. The third author is grateful to Yvan Martel for pointing out the reference [1]} and for interesting discussions about gKdV.
[1]
[ [ 92, 95 ] ]
https://openalex.org/W2043098860
70036992-5101-4e32-958c-b1b86bf33c20
The aim of this section is to collect useful results on the flows associated with (REF ) and (REF ). Firstly, global existence and uniqueness of solutions for the truncated flows follow by a straightforward O.D.E. argument, along with conservation of \(L^2\) mass. From now on we assume without further comment existence and uniqueness of global flows \(\Phi _M(t)\) for \(M\in {\mathbb {N}}\) . The Cauchy problem associated with (REF ) is much more involved. In particular we quote [1]}, [2]}, [3]}, [4]} whose analysis implies that for every \(s\ge 1\) there exists a unique global solution associated with the initial datum \(\varphi \in H^s\) ; moreover we have continuous dependence on the initial datum. The analysis in [3]} allows to treat local Cauchy theory down to low regularity \(H^\frac{1}{2}\) .
[3]
[ [ 498, 501 ], [ 730, 733 ] ]
https://openalex.org/W2075677712
be585cbe-58e8-4d88-8f0e-b41f9f08174d
It will later be important to have a series of uniform bounds with respect to \(M\) (in particular suitable \(L^6\) bounds), as well as some delicate convergences in suitable topologies of the finite dimensional flows to the infinite dimensional one. To the best of our knowledge, those properties do not follow in a straightforward way from the aforementioned works and their proofs require some further arguments. Indeed in our analysis we shall borrow many ideas from references above (in particular [1]}), that in conjunction with new ingredients will imply several properties for the flows \(\Phi _M(t)\) with \(M\in {\mathbb {N}}\cup \lbrace \infty \rbrace \) .
[1]
[ [ 505, 508 ] ]
https://openalex.org/W2075677712
e957e8d7-dd78-4142-b6a4-241aca205af3
We now present the gauge transform following [1]}. Set \(u_M(t,x)=\pi _M (\Phi _M(t)\varphi ) \) and introduce a change of unknown, \(v_M(t,x)=u_M\big (t,x+(p+1)\int _{0}^t \int _ṵ_M^p dxdt\big )\,,\)
[1]
[ [ 45, 48 ] ]
https://openalex.org/W2075677712
d6f847df-917a-4a7c-b430-28f306d4b87e
Denote by \(S(t)\) the linear group associated with linear KdV equation, namely \(S(t)=e^{t\partial _x^3}\) . Then () rewrites, in integral form, \(v_M(t)=S(t)(\pi _M \varphi )+(p+1)\int _{0}^t S(t-\tau ) \pi _M\Pi ( \partial _x v_M(\tau ) \Pi v_M^p(\tau ))d\tau \,.\) The analysis of [1]}, pages 183-186 and pages 197-200 may be used to obtain that for \(s\ge 1\) , \(\Big \Vert \int _{0}^t S(t-\tau ) \pi _M\Pi ( \partial _x w(\tau ) \Pi w^p(\tau ))d\tau \Big \Vert _{Y^s_T}\le CT^{\kappa }\Vert w\Vert _{Y^1_T}^p \Vert w\Vert _{Y^s_T},\) where \(\kappa >0\) and \(T\in (0,1)\) . We refer to the appendix for the proof of (REF ). Notice that (REF ) is a slightly modified version compared with the one available in the literature: we gain a power of \(T\) , which is very important later. By a similar argument one proves a multi-linear estimate for \(s\ge 1\) : \(\Big \Vert \int _{0}^t S(t-\tau ) \pi _M\Pi ( \partial _x w_{p+1}(\tau )\Pi (w_1(\tau )\times \dots \times w_p(\tau )))d\tau \Big \Vert _{Y^s_T}\le CT^{\kappa }\sum _{i=1}^{p+1}\big ( \Vert w_{i}\Vert _{Y^s_T} \prod _{\begin{array}{c}j=1,\dots , p+1\\j\ne i\end{array}} \Vert w_{j}\Vert _{Y^1_T}\big )\) and existence and uniqueness follows by a classical fixed point argument in the space \(Y_T^s\) . Applying (REF ) with \(s=1\) , \( w=v_N\) and recalling (REF ), we obtain that \(\Vert v_M\Vert _{Y^1_T}\le C\Vert \varphi \Vert _{H^1}\) provided \(T\) is small enough depending only on a bound for \(\varphi \) in \(H^1\) . Applying once again (REF ), we get \(\Vert v_M\Vert _{Y^s_T}\le C\Vert \varphi \Vert _{H^s}+CT^\kappa (C\Vert \varphi \Vert _{H^1})^p \Vert v_M\Vert _{Y^s_T}\) which implies \(\Vert v_M\Vert _{Y^s_T}\le C\Vert \varphi \Vert _{H^s}\) by possibly taking \(T\) smaller but still depending only on an \(H^1\) bound for \(\varphi \) . By the embedding \(Y_T^s\subset L^\infty ([0,T];H^s)\) , (REF ) follows and we also get \(\Vert v_M\Vert _{X^{s,\frac{1}{2}}_T}\le C\Vert \varphi \Vert _{H^s}.\) Now we invoke the Strichartz estimate \((8.37)\) of [2]} : \(\Vert S(t)g\Vert _{L^6((0,T); L^6)}\le C \Vert g\Vert _{H^{\sigma }},\quad \sigma >0\) which together with the transfer principle from [3]} yields \(\Vert w\Vert _{L^6((0,T); L^{6}) }\le C \Vert w\Vert _{ X^{\sigma ,b}_T}, \quad b>\frac{1}{2}.\) Next let \(w\in X^{\frac{1}{3}, \frac{1}{3}}_T\) , then we may assume without loss of generality that \(w\) is a global space time function such that \( \Vert w\Vert _{X^{\frac{1}{3}, \frac{1}{3}}}\le 2 \Vert w\Vert _{X^{\frac{1}{3}, \frac{1}{3}}_T}\) . By Sobolev embedding \(H^\frac{1}{3}\subset L^6\) and \(S(t)\) being an isometry on \(H^s\) , \(\Vert w\Vert _{L^6({\mathbb {R}};L^6()}\le C \Vert S(-t) w(t,.)\Vert _{L^6({\mathbb {R}};H^\frac{1}{3}()}\le C \Vert \langle D \rangle _x^\frac{1}{3} (S(-t) w(t,.))\Vert _{L^6({\mathbb {R}};L^2()}\) and by Minkowski inequality and Sobolev embedding (that we now exploit w.r.t. the time variable) \(\dots \le C \Vert \langle D \rangle _x^{\frac{1}{3}} (S(-t) w(t,.))\Vert _{L^2( L^6({\mathbb {R}}))}\le C \Vert \langle D \rangle _x^{\frac{1}{3}} S(-t) w(t,.)\Vert _{L^2( H^\frac{1}{3} ({\mathbb {R}}))}\\=C \Vert \langle D \rangle _t^{\frac{1}{3}}\langle D \rangle _x^\frac{1}{3} (S(-t) w(t,.))\Vert _{L^2({\mathbb {R}}\times } =C\Vert w\Vert _{X^{\frac{1}{3},\frac{1}{3}}}\le 2C \Vert w\Vert _{X_T^{\frac{1}{3},\frac{1}{3}}}\) so that \(\Vert w\Vert _{L^6((0,T); L^{6}) }\le C \Vert w\Vert _{X^{\frac{1}{3},\frac{1}{3}}_T}\) . Interpolation with (REF ) yields \(\forall \,\varepsilon >0\,,\quad \Vert w\Vert _{L^6((0,T); L^6)}\le C \Vert w\Vert _{X^{\varepsilon ,\frac{1}{2}}_T}\,.\) By choosing \(w=v_M\) and recalling (REF ) where we replace \(s\) by \(s+\varepsilon \) , \(\Vert v_M\Vert _{L^6((0,T); W^{s,6})}\le C \Vert v_M\Vert _{X^{s+\varepsilon ,\frac{1}{2}}_T}\le C \Vert \varphi \Vert _{H^{s+\varepsilon }}, \quad \forall \varepsilon >0,\) and we get (REF ). The proof of (REF ) follows by (REF ) by considering the difference of two solutions. Finally, \(\pi _M\Pi (\partial _x v_M \Pi v_M^p))-\Pi (\partial _x v \Pi v^p )=\\\pi _M \Pi (\partial _x v_M\Pi (v_M^p-v^p))+ (\partial _x v_M-\partial _x v)\Pi v^p)-(1-\pi _M)\Pi ( \partial _x v \Pi v^p)\,,\) where \(v_M, v\) are solutions to () and (). Therefore using (REF ), where we choose \(p\) factors \(w_i\) equal to either \(v_M, v\) and one factor equal to \(v-v_M\) , writing the fixed point equation solved by \(v-v_M\) , and recalling (REF ), we get (see e.g. [4]} for details), with \(\mathcal {K}\) being a compact in \(H^s\) , \(\sup _{\varphi \in {\mathcal {K}}} \Vert \pi _M \Phi _M^{\mathcal {G}}(t)\varphi -\Phi ^{\mathcal {G}} (t)\varphi \Vert _{Y^s_T}\overset{M\rightarrow \infty }{\longrightarrow }0\,.\) Therefore we get (REF ) by using the continuous embedding \(Y^s_T\subset L^\infty ([0,T]; H^s)\) .
[1]
[ [ 287, 290 ] ]
https://openalex.org/W2075677712
03063a5d-5448-4585-870f-08072507445b
Denote by \(S(t)\) the linear group associated with linear KdV equation, namely \(S(t)=e^{t\partial _x^3}\) . Then () rewrites, in integral form, \(v_M(t)=S(t)(\pi _M \varphi )+(p+1)\int _{0}^t S(t-\tau ) \pi _M\Pi ( \partial _x v_M(\tau ) \Pi v_M^p(\tau ))d\tau \,.\) The analysis of [1]}, pages 183-186 and pages 197-200 may be used to obtain that for \(s\ge 1\) , \(\Big \Vert \int _{0}^t S(t-\tau ) \pi _M\Pi ( \partial _x w(\tau ) \Pi w^p(\tau ))d\tau \Big \Vert _{Y^s_T}\le CT^{\kappa }\Vert w\Vert _{Y^1_T}^p \Vert w\Vert _{Y^s_T},\) where \(\kappa >0\) and \(T\in (0,1)\) . We refer to the appendix for the proof of (REF ). Notice that (REF ) is a slightly modified version compared with the one available in the literature: we gain a power of \(T\) , which is very important later. By a similar argument one proves a multi-linear estimate for \(s\ge 1\) : \(\Big \Vert \int _{0}^t S(t-\tau ) \pi _M\Pi ( \partial _x w_{p+1}(\tau )\Pi (w_1(\tau )\times \dots \times w_p(\tau )))d\tau \Big \Vert _{Y^s_T}\le CT^{\kappa }\sum _{i=1}^{p+1}\big ( \Vert w_{i}\Vert _{Y^s_T} \prod _{\begin{array}{c}j=1,\dots , p+1\\j\ne i\end{array}} \Vert w_{j}\Vert _{Y^1_T}\big )\) and existence and uniqueness follows by a classical fixed point argument in the space \(Y_T^s\) . Applying (REF ) with \(s=1\) , \( w=v_N\) and recalling (REF ), we obtain that \(\Vert v_M\Vert _{Y^1_T}\le C\Vert \varphi \Vert _{H^1}\) provided \(T\) is small enough depending only on a bound for \(\varphi \) in \(H^1\) . Applying once again (REF ), we get \(\Vert v_M\Vert _{Y^s_T}\le C\Vert \varphi \Vert _{H^s}+CT^\kappa (C\Vert \varphi \Vert _{H^1})^p \Vert v_M\Vert _{Y^s_T}\) which implies \(\Vert v_M\Vert _{Y^s_T}\le C\Vert \varphi \Vert _{H^s}\) by possibly taking \(T\) smaller but still depending only on an \(H^1\) bound for \(\varphi \) . By the embedding \(Y_T^s\subset L^\infty ([0,T];H^s)\) , (REF ) follows and we also get \(\Vert v_M\Vert _{X^{s,\frac{1}{2}}_T}\le C\Vert \varphi \Vert _{H^s}.\) Now we invoke the Strichartz estimate \((8.37)\) of [2]} : \(\Vert S(t)g\Vert _{L^6((0,T); L^6)}\le C \Vert g\Vert _{H^{\sigma }},\quad \sigma >0\) which together with the transfer principle from [3]} yields \(\Vert w\Vert _{L^6((0,T); L^{6}) }\le C \Vert w\Vert _{ X^{\sigma ,b}_T}, \quad b>\frac{1}{2}.\) Next let \(w\in X^{\frac{1}{3}, \frac{1}{3}}_T\) , then we may assume without loss of generality that \(w\) is a global space time function such that \( \Vert w\Vert _{X^{\frac{1}{3}, \frac{1}{3}}}\le 2 \Vert w\Vert _{X^{\frac{1}{3}, \frac{1}{3}}_T}\) . By Sobolev embedding \(H^\frac{1}{3}\subset L^6\) and \(S(t)\) being an isometry on \(H^s\) , \(\Vert w\Vert _{L^6({\mathbb {R}};L^6()}\le C \Vert S(-t) w(t,.)\Vert _{L^6({\mathbb {R}};H^\frac{1}{3}()}\le C \Vert \langle D \rangle _x^\frac{1}{3} (S(-t) w(t,.))\Vert _{L^6({\mathbb {R}};L^2()}\) and by Minkowski inequality and Sobolev embedding (that we now exploit w.r.t. the time variable) \(\dots \le C \Vert \langle D \rangle _x^{\frac{1}{3}} (S(-t) w(t,.))\Vert _{L^2( L^6({\mathbb {R}}))}\le C \Vert \langle D \rangle _x^{\frac{1}{3}} S(-t) w(t,.)\Vert _{L^2( H^\frac{1}{3} ({\mathbb {R}}))}\\=C \Vert \langle D \rangle _t^{\frac{1}{3}}\langle D \rangle _x^\frac{1}{3} (S(-t) w(t,.))\Vert _{L^2({\mathbb {R}}\times } =C\Vert w\Vert _{X^{\frac{1}{3},\frac{1}{3}}}\le 2C \Vert w\Vert _{X_T^{\frac{1}{3},\frac{1}{3}}}\) so that \(\Vert w\Vert _{L^6((0,T); L^{6}) }\le C \Vert w\Vert _{X^{\frac{1}{3},\frac{1}{3}}_T}\) . Interpolation with (REF ) yields \(\forall \,\varepsilon >0\,,\quad \Vert w\Vert _{L^6((0,T); L^6)}\le C \Vert w\Vert _{X^{\varepsilon ,\frac{1}{2}}_T}\,.\) By choosing \(w=v_M\) and recalling (REF ) where we replace \(s\) by \(s+\varepsilon \) , \(\Vert v_M\Vert _{L^6((0,T); W^{s,6})}\le C \Vert v_M\Vert _{X^{s+\varepsilon ,\frac{1}{2}}_T}\le C \Vert \varphi \Vert _{H^{s+\varepsilon }}, \quad \forall \varepsilon >0,\) and we get (REF ). The proof of (REF ) follows by (REF ) by considering the difference of two solutions. Finally, \(\pi _M\Pi (\partial _x v_M \Pi v_M^p))-\Pi (\partial _x v \Pi v^p )=\\\pi _M \Pi (\partial _x v_M\Pi (v_M^p-v^p))+ (\partial _x v_M-\partial _x v)\Pi v^p)-(1-\pi _M)\Pi ( \partial _x v \Pi v^p)\,,\) where \(v_M, v\) are solutions to () and (). Therefore using (REF ), where we choose \(p\) factors \(w_i\) equal to either \(v_M, v\) and one factor equal to \(v-v_M\) , writing the fixed point equation solved by \(v-v_M\) , and recalling (REF ), we get (see e.g. [4]} for details), with \(\mathcal {K}\) being a compact in \(H^s\) , \(\sup _{\varphi \in {\mathcal {K}}} \Vert \pi _M \Phi _M^{\mathcal {G}}(t)\varphi -\Phi ^{\mathcal {G}} (t)\varphi \Vert _{Y^s_T}\overset{M\rightarrow \infty }{\longrightarrow }0\,.\) Therefore we get (REF ) by using the continuous embedding \(Y^s_T\subset L^\infty ([0,T]; H^s)\) .
[3]
[ [ 2198, 2201 ] ]
https://openalex.org/W1964699420
930a7a23-15b2-4a9a-8fbd-b2d9edf9b147
for \(T\le 1\) and \(C>0\) independent of \(M\) and \(\kappa \) . The arguments that we will perform below are standard. Our only goal is to provide a complete argument for a reader unfamiliar with the \(X^{s,b}\) machinery, as well as proving how to gain the positive power of \(T\) at the r.h.s. (which was of importance for our analysis in Section ). At the best of our knowledge the estimate above written in this form is not readily available in the literature, even if we closely follow [1]}. As \(\pi _M\) is bounded on \(Y^s\) , it suffices to prove \(\Big \Vert \int _{0}^t S(t-\tau ) \Pi (\partial _x v(\tau ) \Pi v^p(\tau ))d\tau \Big \Vert _{Y^s_T}\le CT^{\kappa }\Vert v\Vert _{Y^1_T}^p \Vert v\Vert _{Y^s_T},\quad \kappa >0\,.\)
[1]
[ [ 498, 501 ] ]
https://openalex.org/W2075677712
f260d0b6-2e4e-4678-a661-7d55d0b5fd9b
where \(\psi \in C^\infty _0({\mathbb {R}})\) is such that \(\psi \equiv 1\) on \([-1,1]\) . Using [1]}, we obtain that \(\Big \Vert \psi (t) \int _{0}^t S(t-\tau ) \Pi (\partial _x v(\tau ) \Pi v^p(\tau ))d\tau \Big \Vert _{Y^s}\le C\big \Vert \Pi (\partial _x v \Pi v^p )\big \Vert _{Z^s}\, ,\)
[1]
[ [ 101, 104 ] ]
https://openalex.org/W2075677712
b6835926-a582-4ea0-bb59-14308c35cc85
will be enough. Its proof follows by combining the following propositions. The next statement is a slightly modified version of [1]}.
[1]
[ [ 128, 131 ] ]
https://openalex.org/W2075677712
a13d87ee-b58f-4eca-9e68-37b861d055de
Write \({\mathcal {F}}(u^p)(\tau ,n)= \int _{\tau =\tau _1+\cdots +\tau _p}\,\,\,\sum _{n=n_1+\cdots +n_p}\,\,\prod _{k=1}^p \hat{u}(\tau _k,n_k)\,,\) where \({\mathcal {F}}\) and \(\hat{u}\) denote the space time Fourier transform (continuous in time and discrete in space). \(\Vert u^p\Vert _{X^{s-1,\frac{1}{2}}}^2=\int _{{\mathbb {R}}}\sum _{n\in {\mathbb {Z}}}\langle n\rangle ^{2(s-1)} \langle \tau +n^3\rangle \, |{\mathcal {F}}(u^p)(\tau ,n)|^2\, d\tau \,.\) Notice that the r.h.s. in (REF ) may be bounded with \(\int _{{\mathbb {R}}}\sum _{n\in {\mathbb {Z}}}\langle n\rangle ^{2(s-1)} \langle \tau +n^3\rangle \big (\int _{\tau =\tau _1+\cdots +\tau _p}\,\,\sum _{n=n_1+\cdots +n_p}\,\,\prod _{k=1}^p |\widehat{u}(\tau _k,n_k)|\big )^2\, d\tau .\) Hence if we define \(w(t,x)\) by \(\hat{w}(\tau ,n)=|\hat{u}(\tau ,n)|\) we get \(\Vert u\Vert _{X^{s,b}}=\Vert w\Vert _{X^{s,b}}\) , \(\Vert u\Vert _{Y^s}=\Vert w\Vert _{Y^s}\) , and we are reduced to estimate \(\int _{{\mathbb {R}}}\sum _{n\in {\mathbb {Z}}}\langle n\rangle ^{2(s-1)} \langle \tau +n^3\rangle \,\big (\int _{\tau =\tau _1+\cdots +\tau _p}\,\,\sum _{n=n_1+\cdots +n_p}\,\,\prod _{k=1}^p \widehat{w}(\tau _k,n_k)\big )^2 d\tau \,\,\,.\) Next we split the domain of integration and we consider first the contribution to (REF ) in the region \(|\tau +n^3|\le 10p |\tau _1+n_1^3|.\) If we define \(w_1\) by \(\widehat{w_1}(\tau ,n)=\langle \tau +n^3\rangle ^{\frac{1}{2}}\, \widehat{w}(\tau ,n)\) , then the contribution to (REF ) in the region (REF ) can be controlled in the physical space variables as follows \(C\Vert w_1 w^{p-1}\Vert _{L^2({\mathbb {R}};H^{s-1})}^2\le & C\big (\Vert w_1\Vert _{L^2({\mathbb {R}}; H^{s-1})}^2 \Vert w^{p-1}\Vert _{L^\infty ({\mathbb {R}}; L^\infty )}^2+\Vert w_1\Vert _{L^2({\mathbb {R}}; L^\infty )}^2\Vert w^{p-1}\Vert _{L^\infty ({\mathbb {R}}; H^{s-1})}^2\big )\\\le & C \big (\Vert w\Vert _{X^{s-1, \frac{1}{2}}}^2 \Vert w\Vert _{L^\infty ({\mathbb {R}}; H^1)}^{2(p-1)} + \Vert w_1\Vert _{L^2({\mathbb {R}}; H^1)}^2\Vert w\Vert _{L^\infty ({\mathbb {R}}; H^{s-1})}^2\Vert w\Vert _{L^\infty ({\mathbb {R}}; H^{1})}^{2(p-2)}\big )\) where we have used standard product rules and Sobolev embedding \(H^1\subset L^\infty \) . We proceed with \((\dots ) \le C \big (\Vert w\Vert _{X^{s-1, \frac{1}{2}}}^2 \Vert w\Vert _{Y^1}^{2(p-1)} + \Vert w_1\Vert _{X^{1,\frac{1}{2}}}^2\Vert w\Vert _{Y^{s-1}}^2\Vert w\Vert _{Y^1}^{2(p-2)} \big )\) where we used \(Y^1\subset L^\infty ({\mathbb {R}}; H^1)\) , \(Y^{s-1}\subset L^\infty ({\mathbb {R}}; H^{s-1})\) . Notice that we have a better estimate, when compared with (REF ), in the region (REF ). Similarly, we can evaluate the contributions to (REF ) of the regions \(| \tau +n^3|\le 10 p| \tau _k+n_k^3|,\quad 2\le k\le p\,.\) Therefore, we may assume that the summation and the integration in (REF ) is performed in the region \(\max _{1\le k\le p}|\tau _k+n_k^3|\le \frac{1}{10p} |\tau +n^3|\,.\) Write \((\tau +n^3)-\sum _{k=1}^p(\tau _k+n_k^3)=\Big (\sum _{k=1}^p n_k\Big )^3-\sum _{k=1}^p n_k^3\,,\) therefore in the region (REF ) we have \(\Big |\Big (\sum _{k=1}^p n_k\Big )^3-\sum _{k=1}^p n_k^3\Big |\ge |\tau +n^3|-\sum _{k=1}^p |\tau _k+n_k^3|\ge \frac{9}{10}|\tau +n^3|\) hence \(\langle \tau +n^3\rangle \le C\Big |\Big (\sum _{k=1}^p n_k\Big )^3-\sum _{k=1}^p n_k^3\Big |\,.\) By symmetry we can assume \(|n_1|\ge |n_2|\ge \cdots \ge |n_k|\) and by using [1]}, we obtain that \(\Big |\Big (\sum _{k=1}^p n_k\Big )^3-\sum _{k=1}^p n_k^3\Big |\le C |n_1|^2 |n_2|.\) Consequently in the region (REF ) we get \(\langle \tau +n^3\rangle \le C \langle n_1\rangle ^2 \langle n_2\rangle \) , and the corresponding contribution to (REF ) can be estimated as \(C\, \int _{{\mathbb {R}}}\sum _{n\in {\mathbb {Z}}}\,\big (\int _{\tau =\tau _1+\cdots +\tau _p}\,\,\sum _{n=n_1+\cdots +n_p}\,\langle n_1\rangle ^{s} \langle n_2\rangle ^\frac{1}{2} \,\prod _{k=1}^p (\widehat{w}(\tau _k,n_k)\big )^2 \, d\tau \) If we define \(w_1\) , \(w_2\) by \(\widehat{w_1}(\tau ,n)=\langle n\rangle ^{s} \widehat{w}(\tau ,n)\) , \(\widehat{w_2}(\tau ,n)=\langle n\rangle ^{\frac{1}{2}} \widehat{w}(\tau ,n)\) , going back to physical space variables, we estimate (REF ) as \(C\Vert w_1 w_2 w^{p-2}\Vert _{L^2({\mathbb {R}}; L^2)}^2 \le &C\Vert w_1\Vert _{L^\infty ({\mathbb {R}}; L^2)}^2\Vert w_2\Vert _{L^4({\mathbb {R}}; L^\infty )}^2\Vert w\Vert _{L^4({\mathbb {R}}; L^\infty )}^2\Vert w\Vert _{L^\infty ({\mathbb {R}}; L^\infty )}^{2(p-3)}\\\le & C \Vert w\Vert _{L^\infty ({\mathbb {R}}; H^s)}^2\Vert w_2\Vert _{L^4({\mathbb {R}}; W^{\frac{1}{2},4})}^2\Vert w\Vert _{L^4({\mathbb {R}};W^{1,4})}^2\Vert w\Vert _{L^\infty ({\mathbb {R}}; H^1)}^{2(p-3)}.\) Hence by using \(Y^1\subset L^\infty ({\mathbb {R}};H^1)\) and \(Y^s\subset L^\infty ({\mathbb {R}};H^s)\) , along with the estimate \(\Vert u\Vert _{L^4({\mathbb {R}}; L^4)}\le C\Vert u\Vert _{X^{0,\frac{1}{3}}}\) established in the fundamental work [2]}, we proceed with \((\dots ) \le C \Vert w\Vert _{Y^s}^2\Vert w\Vert _{X^{1,\frac{1}{3}}}^2\Vert w\Vert _{X^{1,\frac{1}{3}}}^2\Vert w\Vert _{Y^1}^{2(p-3)}\) and this concludes the proof.
[1]
[ [ 3444, 3447 ] ]
https://openalex.org/W2075677712
916f41fe-69e8-45bf-98a5-d2d017293cf1
On the other hand, it has turned out that there is another particular class of STIT tessellations, called Mondrian tessellations, for which a second-order description is desirable, since such tessellations have found numerous applications in machine learning. Reminiscent of the famous paintings of the Dutch modernist painter Piet Mondrian, the eponymous tessellations are a version of STIT tessellations with only axis-parrallel cutting directions. Originally established by Roy and Teh [1]}, Mondrian tessellations have been shown to have multiple applications in random forest learning [2]}, [3]} and kernel methods [4]}. Both random forest learners and random kernel approximations based on the Mondrian process have shown significant results, especially as they are substantially more adapted to online-learning (i.e., the ability to incorporate new data into an existing model without having to completely retrain it) than many other of their tessellation-based counterparts. This is due to the self-similarity of Mondrian tessellations, which stems from their defining characteristic of being iteration stable (see [5]}), and allows to obtain explicit representations for many conditional distributions of Mondrian tessellations. This property allows a tessellation-based learner to be re-trained on new data without having to newly start the training process and is thus considerably more efficient on large data sets. These methods have recently been carried over back to their origin in stochastic geometry, i.e., to general STIT tessellations [6]}, [7]}.
[2]
[ [ 590, 593 ] ]
https://openalex.org/W2159187228
11d619e5-a96a-4ea6-a345-2cf9e77c987d
Let \([\mathbb {R}^2]\) be the space of lines in \(\mathbb {R}^2\) . Equipped with the Fell topology, \([\mathbb {R}^2]\) carries a natural Borel \(\sigma \) -field \(\mathfrak {B}([\mathbb {R}^2])\) , see [1]}. Further, define \([\mathbb {R}^2]_0\) to be the space of all lines in \(\mathbb {R}^2\) passing through the origin. For a line \(L\in [\mathbb {R}^2]\) , we write \(L^+\) and \(L^-\) for the positive and negative half-spaces of \(L\) , respectively, and \(L^\perp \) for its orthogonal line passing through the origin. For a compact set \(A\subset \mathbb {R}^2\) define \([A]:= \lbrace L \in [\mathbb {R}^2] : L \cap A\ne \emptyset \rbrace \in \mathfrak {B}([\mathbb {R}^2])\)
[1]
[ [ 208, 211 ] ]
https://openalex.org/W2767091268
36fe2d06-fdc3-4c0d-829c-1173674b2cbd
for any non-negative measurable function \(g: [\mathbb {R}^2] \rightarrow \mathbb {R}\) , see [1]}. Sufficient normalization is usually applied to \(\mathcal {R}\) in order to gain a probability distribution, which is then referred to as the directional distribution.
[1]
[ [ 94, 97 ] ]
https://openalex.org/W565214016
13a3874b-343a-4f32-a91c-2bec1e3bde4d
Remark 2.4 It is instructive to compare this result to the corresponding asymptotic formulas for isotropic STIT tessellations in the plane and the rectangular Poisson line process. We therefore denote by \(\Lambda _{\rm iso }\) the isometry invariant measure on the space of lines in the plane normalized in such a way that \(\Lambda _{\rm iso}([[0,1]^2])={4\over \pi }\) (this is the same normalization as the one used in [1]}).
[1]
[ [ 425, 428 ] ]
https://openalex.org/W565214016
cf2c6d00-28bb-4c6d-a6bf-af5b29aed283
In [1]} an explicit description of the pair-correlation function of the vertex point process of an isotropic planar STIT tessellation has been derived, while such a description for the random edge length measure can be found in [2]}. Also the so-called cross-correlation function between the vertex process and the random length measure was computed in [1]}. In the present paper we develop similar results for planar Mondrian tessellations. To define the necessary concepts, we suitably adapt the notions used in the isotropic case. We let \(Y_t\) be a weighted Mondrian tessellation of \(\mathbb {R}^2\) with weight \(p\in (0,1)\) and time parameter \(t>0\) , define \(R_p:=[0,1-p]\times [0,p]\) and let \(R_{r,p}:=rR_p\) be the rescaled rectangle with side lengths \(r(1-p)\) and \(rp\) . In the spirit of Ripley's K-function widely used in spatial statistics [4]}, we let \(t^2K_{\cal E}(r)\) be the total edge length of \(Y_t\) in \(R_{r,p}\) when \(Y_t\) is regarded under the Palm distribution with respect to the random edge length measure \(\mathcal {E}_t\) concentrated on the edge skeleton. On an intuitive level the latter means that we condition on the origin being a typical point of the edge skeleton, see [4]}. The classic version of Ripley's K-function considers a ball of radius \(r>0\) , but since our driving measure is non-isotropic, we account for that by considering \(R_{r,p}\) instead. Similarly, we let \(\lambda K_{\cal V}(r)\) be the total number of vertices of \(Y_t\) in \(R_{r,p}\) , where \(\lambda =t^2p(1-p)\) stands for the vertex intensity of \(Y_t\) and where we again condition on the origin being a typical vertex of the tessellation (in the sense of the Palm distribution with respect to the random vertex point process). While these functions still have a complicated form (which we will determine in the course of our arguments below), we consider their normalized derivatives – provided these derivatives are well defined as it is the case for us. In the isotropic case, these are known as the pair-correlation functions of the random edge length measure or the vertex point process, respectively. In our case, the following normalization turns out to be most suitable: \(g_\mathcal {E}(r)= \frac{1}{2p(1-p)r} \frac{\textup {d}K_\mathcal {E}(r)}{\textup {d}r} \qquad \text{and}\qquad g_\mathcal {V}(r)= \frac{1}{2p(1-p)r} \frac{\textup {d}K_\mathcal {V}(r)}{\textup {d}r},\)
[4]
[ [ 869, 872 ], [ 1232, 1235 ] ]
https://openalex.org/W2118166339
c4563541-8877-40c1-a769-aef390e0b460
(i) In the isotropic case Schreiber and Thäle showed in [1]} that the pair-correlation function of the random edge length measure \(\mathcal {E}_{t}\) has the form \(g_\mathcal {E}(r) = 1 + \frac{1}{2t^2 r^2}\Big (1-e^{-\frac{2}{\pi } tr}\Big ).\) In [2]} the same authors showed that the pair-correlation function of the vertex point process \(\mathcal {V}_t\) and the cross-correlation function of the random edge length measure and the vertex point process are given by \(g_{\mathcal {E},\mathcal {V}}(r) = 1+ \frac{1}{t^2 r^2}-\frac{\pi }{4t^3r^3}- \frac{e^{-\frac{2}{\pi }tr}}{2t^2r^2}\Big (1-\frac{\pi }{2tr}\Big )\) and \(g_\mathcal {V}(r) = 1 + \frac{2}{t^2r^2} - \frac{\pi }{t^3r^3} + \frac{\pi ^2}{4t^4r^4} - \frac{e^{-\frac{2}{\pi }tr}}{2t^2r^2} \Big ( 1 - \frac{\pi }{tr} + \frac{\pi ^2}{2t^2r^2} \Big ).\) (ii) For the rectangular Poisson line process as given in Remark REF one can use the theorem of Slivnyak-Mecke (see for example [3]}) to deduce that the corresponding analogues of the cross- and pair-correlation functions are given by \(g_\mathcal {E}(r) = 1 + \frac{1}{tr},\qquad g_{\mathcal {E},\mathcal {V}}(r) = 1 + \frac{1}{4trp(1-p)}\qquad \text{and}\qquad g_\mathcal {V}(r)=1 + \frac{1}{2tr p^2(1-p)^2 }.\)
[3]
[ [ 953, 956 ] ]
https://openalex.org/W2118166339
7c08224e-9351-453d-b10d-8bb0fd03dfc9
   + 12 p2(1-p) RRR1(() ) 1(() )   ( )  I1 (s2 e-sp|-|;t)  d d d. Having established the covariance measure of the edge process \(\mathcal {E}_{t}\) , we now aim at giving the corresponding pair-correlation function \(g_\mathcal {E}(r)\) . In a first step towards this we need to establish the reduced covariance measure \(\widehat{\operatorname{Cov}}(\mathcal {E}_{t})\) defined by the relation Cov(Et)(AB)= A B-xCov(Et)(dy) 2(dx) for a measurable product \(A\times B\subset \mathbb {R}^2\times \mathbb {R}^2\) (cf. [1]}). We now examine the first of the two integral summands in (). Using Lemma REF (ii) we see that    12 p(1-p)2 RRR1(A ())   1(B ())   I1 (s2 e-s(1-p)|-|;t)  d d d
[1]
[ [ 522, 525 ] ]
https://openalex.org/W2525528836
e9dab924-aa79-4b23-ac0c-d3b007959d7d
Proceeding analogously with the second summand in () and using the diagonal shift argument from [1]}, we get the reduced covariance measure \( \widehat{\operatorname{Cov}}(\mathcal {E}_{t})\) on \(\mathbb {R}^2\) : Cov(Et)(     ) = 12 p(1-p)2 R(0 z)0 (0 z)0 y-x()   dx   dy   I1 (s2 e-s(1-p)|z|;t)   dz
[1]
[ [ 96, 99 ] ]
https://openalex.org/W2525528836
7e3080bd-66d9-4529-af63-01db061f333c
   + 12 p2(1-p) R0(0 z) 0(0 z) y-x()   dx   dy   I1 (s2 e-s(1-p)|z|;t)   dz. Noting that the intensity of the random measure \(\mathcal {E}_{t}\) is just \(t\) , see [1]}, we apply Equation (8.1.6) in [2]} to see that the corresponding reduced second moment measure \(\widehat{\mathcal {K}}(\mathcal {E}_{t})\) is K(Et)( ) = Cov(Et)(     ) + t2 2(). While the classical Ripley's K-function would be \(t^{-2}\) times the \(\widehat{\mathcal {K}}(\mathcal {E}_{t})\) -measure of a disc of radius \(r>0\) , we define our Mondrian analogue as KE(r):=1t2   K(Et)(Rr,p), where \(R_{r,p}:=rR_p\) with \(R_p:=[0,1-p] \times [0, p]\) as before. Calculating \(K_{\mathcal {E}}(r)\) explicitly via Lemma REF yields KE(r) = r2p(1-p) +p(1-p)22t2 R(0 z)0 (0 z)0 y-x(Rr,p )   dx   dy   I1 (s2 e-s(1-p)|z|;t)   dz
[2]
[ [ 203, 206 ] ]
https://openalex.org/W2525528836
58e94361-c53b-406f-9ebb-6b930b85d193
   =p2 (1-p)R2 w(A)R  1(B-w0(0 z))) I1(s2 (-s p|z| );t)   dz   dw. The second summand in each of the terms in () can be dealt with using Lemma REF (ii), see also Equation (). As in the previous section, we want to proceed by giving the reduced covariance measure via the diagonal-shift argument in the sense of [1]}. Plugging the terms we just deduced into (), we end up with the covariance measure Cov( Vt,Et)(AB)= A B-xCovV,E(dy) 2(dx), where the reduced cross-covariance measure is given by CovV,E(     )= p(1-p)( (1-p)R 1( (0z)0)  I1(s2 (-s(1-p)z ;t)  dz
[1]
[ [ 313, 316 ] ]
https://openalex.org/W2525528836
2fc2d0a4-7370-49a5-9905-fdb9ab11015b
We now aim at giving the corresponding Mondrian analogue of the pair-correlation function of the vertex point process. As in the previous sections, we do so by giving the reduced covariance measure via a diagonal-shift argument in the sense of [1]}. Consider the first integral term in (REF ) without its coefficient for the Borel set \( A\times B \subset \mathbb {R}^2\times \mathbb {R}^2\) . After multiplying the Dirac measures, we only consider the first two summands that integrate over \(\delta _{(\tau ,\sigma )}(A)\delta _{(\tau ,\sigma )}(B)\) and \(\delta _{(\tau ,\sigma )}(A)\delta _{(\vartheta , \sigma )}(B)\) , respectively, as the other two can be handled in the same fashion. Using Lemma REF (iii) yields \(&&\int _{\mathbb {R}} \int _{\mathbb {R}} \int _{\mathbb {R}} \, \delta _{(\tau ,\sigma )}(A)\delta _{(\tau ,\sigma )}(B) \, \mathcal {I}^1\big (s^2 \exp (-s(1-p)| \tau -\vartheta | ) ;t\big ) \, \textup {d}\vartheta \, \textup {d}\tau \, \textup {d}\sigma \\\\&=& \int _{\mathbb {R}^2} \delta _{w}(A)\, \int _{\mathbb {R}} \delta _{\mathbf {0}}(B - w) \mathcal {I}^1\big (s^2 \exp (-s(1-p)|-z_1| ) ;t\big ) \, \textup {d}z \, \textup {d}w\)
[1]
[ [ 244, 247 ] ]
https://openalex.org/W2525528836
bf1936d9-5d5e-4788-a77f-e587166900aa
We again define a function in the spirit of Ripley's K-function via the reduced second moment measure \(\widehat{\mathcal {K}}( \mathcal {V}_{t})(R_{r,p})\) of \(R_{r,p}\) , \(r>0\) , and the corresponding normalized derivative as KV(r)=1V2   K( Vt)(Rr,p)=1(t2p(1-p))2   K( Vt)(Rr,p). Combining the considerations above with the diagonal shift argument from [1]} we obtain that the reduced covariance measure \( \widehat{\operatorname{Cov}}( \mathcal {V}_{t})\) with Cov( Vt)(AB)= A B-x Cov( Vt)(dy) 2(dx), is given by \( \nonumber && \widehat{\operatorname{Cov}} ( \mathcal {V}_{t})( \, \cdot \, )\nonumber \\\nonumber && =p(1-p)\Bigg [(1-p) \bigg (\int _{\mathbb {R}} \delta _{\mathbf {0}}(\cdot ) \mathcal {I}^1\big (s^2 \exp (-s(1-p)|z| ) ;t\big ) \, \textup {d}z\\\nonumber && \qquad \qquad \qquad \qquad + \int _{\mathbb {R}} \delta _{(z, 0)}(\cdot ) \, \mathcal {I}^1\big (s^2 \exp (-s(1-p)|z|) ;t\big ) \, \textup {d}z \bigg )\\\nonumber && \qquad + p \bigg ( \int _{\mathbb {R}} \delta _{\mathbf {0}}(\cdot ) \mathcal {I}^1\big (s^2 \exp (-sp|z| ) ;t\big ) \, \textup {d}z+ \, \int _{\mathbb {R}} \delta _{(0, z)}(\cdot ) \, \mathcal {I}^1\big (s^2 \exp (-sp|z|) ;t\big ) \, \textup {d}z \, \bigg )\\\nonumber \\\nonumber && \qquad + 4\Bigg ( (1-p)^2 \int _{\mathbb {R}} \,\ell _1( \cdot \cap \overline{(0 z)}_0) \, \, \mathcal {I}^2\big (s^2 \exp (-s(1-p)|z|);t\big )\, \textup {d}z\\\nonumber \\\nonumber && \qquad \qquad \quad + p^2 \int _{\mathbb {R}} \,\ell _1( \cdot \cap {_0}\overline{(0 z)}) \, \, \mathcal {I}^2 \big (s^2 \exp (-sp|z|);t\big )\, \textup {d}z\\\nonumber && \qquad \qquad \quad + (1-p)^3 \int _{\mathbb {R}} \Big (\int _{ \overline{(0z)}_0} \int _{ \overline{(0z)}_0} \delta _{y-x}(\cdot )\, \textup {d}x \, \textup {d}y\Big )\,\mathcal {I}^3\big (s^2 \exp (-s(1-p)\vert z\vert ) ;t\big )\, \textup {d}z\\\nonumber \\&& \qquad \qquad \quad + p^3 \int _{\mathbb {R}} \Big (\int _{ \overline{{_0}(0z)}} \int _{ {_0}\overline{(0z)}} \delta _{y-x}(\cdot )\, \textup {d}x \, \textup {d}y\Big )\,\mathcal {I}^3\big (s^2 \exp (-sp\vert z\vert ) ;t\big )\, \textup {d}z\Bigg ) \Bigg ].\)
[1]
[ [ 361, 364 ] ]
https://openalex.org/W2525528836
1b041648-1c6d-4b06-998d-6893942698e6
The relationship [1]} now yields the reduced second moment measure \(\widehat{\mathcal {K}}( \mathcal {V}_{t})\) \( \widehat{\mathcal {K}}( \mathcal {V}_{t})( \, \cdot \, )&=& \widehat{\operatorname{Cov}}( \mathcal {V}_{t})( \, \cdot \, )+ (t^2p(1-p))^2 \ell _2( \, \cdot \, ).\)
[1]
[ [ 17, 20 ] ]
https://openalex.org/W2525528836
10a21af1-6f48-4f05-9b6a-685575dee5a9
This is a sample theorem. The run-in heading is set in bold, while the following text appears in italics. Definitions, lemmas, propositions, and corollaries are styled the same way. Proofs, examples, and remarks have the initial word in italics, while the following text appears in normal font. For citations of references, we prefer the use of square brackets and consecutive numbers. Citations using labels or the author/year convention are also acceptable. The following bibliography provides a sample reference list with entries for journal articles [1]}, an LNCS chapter [2]}, a book [3]}, proceedings without editors [4]}, and a homepage [5]}. Multiple citations are grouped [1]}, [2]}, [3]}, [1]}, [3]}, [4]}, [5]}.
[1]
[ [ 557, 560 ], [ 684, 687 ], [ 702, 705 ] ]
https://openalex.org/W4232554906
af0c3339-2c30-476d-a15d-03eaf6812949
Regarding the above, it would be very useful to see how relevant is the role that each fluid, represented by their respective energy-momentum tensor \(T^{i}_{\mu \nu }\) in (REF ), plays on a self-gravitating system, as well as how these gravitational sources interact with each other. This would allow, for instance, detecting which source dominates over the others, and consequently rule out any equation of state incompatible with the dominant source. Conceptually, achieving this in general relativity should be extremely difficult, given the nonlinear nature of the theory. However, since the Gravitational Decoupling approach (GD) [1]}, [2]} is precisely designed for coupling/decoupling gravitational sources in general relativity, we will see that, indeed, it is possible to elucidate the role played by each gravitational source, without resorting to any numerical protocol or perturbation scheme, as explained in the next paragraph.
[1]
[ [ 638, 641 ] ]
https://openalex.org/W2608205523
1c8948bc-3e0a-49a0-a465-bbb24c5ad842
Regarding the above, it would be very useful to see how relevant is the role that each fluid, represented by their respective energy-momentum tensor \(T^{i}_{\mu \nu }\) in (REF ), plays on a self-gravitating system, as well as how these gravitational sources interact with each other. This would allow, for instance, detecting which source dominates over the others, and consequently rule out any equation of state incompatible with the dominant source. Conceptually, achieving this in general relativity should be extremely difficult, given the nonlinear nature of the theory. However, since the Gravitational Decoupling approach (GD) [1]}, [2]} is precisely designed for coupling/decoupling gravitational sources in general relativity, we will see that, indeed, it is possible to elucidate the role played by each gravitational source, without resorting to any numerical protocol or perturbation scheme, as explained in the next paragraph.
[2]
[ [ 644, 647 ] ]
https://openalex.org/W2901000943
5b6d07ab-e323-4d49-8604-5f57e092aeb5
In this Section, we briefly review the GD for spherically symmetric gravitational systems described in detail in Ref. [1]}. For the axially symmetric case, see Ref. [2]}. The gravitational decoupling approach and its simplest version [3]}, based in the Minimal Geometric Deformation (MGD) [4]}, [5]}, [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}, [14]}, [15]}, [16]}, [17]}, [18]}, [19]}, [20]}, [21]}, [22]}, [23]}, [24]}, [25]}, [26]}, [27]}, are attractive for many reasons (for an incomplete list of references, see [28]}, [29]}, [30]}, [31]}, [32]}, [33]}, [34]}, [35]}, [36]}, [37]}, [38]}, [39]}, [40]}, [41]}, [42]}, [43]}, [43]}, [45]}, [46]}, [47]}, [48]}, [49]}, [50]}, [51]}, [49]}, [53]}, [54]}, [55]}, [56]}, [57]}, [58]}, [59]}, [60]}, [61]}, [62]}, [63]}, [64]}, [65]}, [66]}, [67]}, [68]}, [69]}, [70]}, [71]}, [72]}, [73]}, [74]}, [75]}, [76]}, [77]}, [78]}, [79]}, [80]}. Among them we can mention i) the coupling of gravitational sources, which allows for extending known solutions of the Einstein field equations into more complex domains; ii) the decoupling of gravitational sources, which is used to systematically reduce (decouple) a complex energy-momentum tensor \(T_{\mu \nu }\) into simpler components; iii) to find solutions in gravitational theories beyond Einstein's; iv) to generate rotating hairy black hole solutions, among many others applications.
[1]
[ [ 118, 121 ] ]
https://openalex.org/W2901000943
bd2b6133-5a08-4eaa-9da2-7c74f4c8ea75
In this Section, we briefly review the GD for spherically symmetric gravitational systems described in detail in Ref. [1]}. For the axially symmetric case, see Ref. [2]}. The gravitational decoupling approach and its simplest version [3]}, based in the Minimal Geometric Deformation (MGD) [4]}, [5]}, [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}, [14]}, [15]}, [16]}, [17]}, [18]}, [19]}, [20]}, [21]}, [22]}, [23]}, [24]}, [25]}, [26]}, [27]}, are attractive for many reasons (for an incomplete list of references, see [28]}, [29]}, [30]}, [31]}, [32]}, [33]}, [34]}, [35]}, [36]}, [37]}, [38]}, [39]}, [40]}, [41]}, [42]}, [43]}, [43]}, [45]}, [46]}, [47]}, [48]}, [49]}, [50]}, [51]}, [49]}, [53]}, [54]}, [55]}, [56]}, [57]}, [58]}, [59]}, [60]}, [61]}, [62]}, [63]}, [64]}, [65]}, [66]}, [67]}, [68]}, [69]}, [70]}, [71]}, [72]}, [73]}, [74]}, [75]}, [76]}, [77]}, [78]}, [79]}, [80]}. Among them we can mention i) the coupling of gravitational sources, which allows for extending known solutions of the Einstein field equations into more complex domains; ii) the decoupling of gravitational sources, which is used to systematically reduce (decouple) a complex energy-momentum tensor \(T_{\mu \nu }\) into simpler components; iii) to find solutions in gravitational theories beyond Einstein's; iv) to generate rotating hairy black hole solutions, among many others applications.
[2]
[ [ 165, 168 ] ]
https://openalex.org/W3121669597
a7a72f39-1169-405d-acab-90fddfad8384
In this Section, we briefly review the GD for spherically symmetric gravitational systems described in detail in Ref. [1]}. For the axially symmetric case, see Ref. [2]}. The gravitational decoupling approach and its simplest version [3]}, based in the Minimal Geometric Deformation (MGD) [4]}, [5]}, [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}, [14]}, [15]}, [16]}, [17]}, [18]}, [19]}, [20]}, [21]}, [22]}, [23]}, [24]}, [25]}, [26]}, [27]}, are attractive for many reasons (for an incomplete list of references, see [28]}, [29]}, [30]}, [31]}, [32]}, [33]}, [34]}, [35]}, [36]}, [37]}, [38]}, [39]}, [40]}, [41]}, [42]}, [43]}, [43]}, [45]}, [46]}, [47]}, [48]}, [49]}, [50]}, [51]}, [49]}, [53]}, [54]}, [55]}, [56]}, [57]}, [58]}, [59]}, [60]}, [61]}, [62]}, [63]}, [64]}, [65]}, [66]}, [67]}, [68]}, [69]}, [70]}, [71]}, [72]}, [73]}, [74]}, [75]}, [76]}, [77]}, [78]}, [79]}, [80]}. Among them we can mention i) the coupling of gravitational sources, which allows for extending known solutions of the Einstein field equations into more complex domains; ii) the decoupling of gravitational sources, which is used to systematically reduce (decouple) a complex energy-momentum tensor \(T_{\mu \nu }\) into simpler components; iii) to find solutions in gravitational theories beyond Einstein's; iv) to generate rotating hairy black hole solutions, among many others applications.
[3]
[ [ 234, 237 ] ]
https://openalex.org/W2608205523
6dfe4bcd-0bd2-4433-8607-bd874d267ec2
Of course the tensor \(\theta _{\mu \nu }\) vanishes when the deformations vanish (\(f=g=0\) ). We see that for the particular case \(g=0\) , Eqs. (REF )-() reduce to the simpler “quasi-Einstein" system of the MGD of Ref. [1]}, in which \(f\) is only determined by \(\theta _{\mu \nu }\) and the undeformed metric (REF ). Also, notice that the set (REF )-() contains \(\lbrace \xi ,\,\mu \rbrace \) , and therefore is not independent of (REF )-(). This of course makes sense since both systems represent a simplified version of a more complex whole, described by Eqs. (REF )-().
[1]
[ [ 223, 226 ] ]
https://openalex.org/W2608205523
1e16177d-a205-473b-98a4-d96d690e2e3d
The vision field is undergoing two revolutionary trends since about two years ago. The first trend is self-supervised visual representation learning pioneered by MoCo [1]}, which for the first time demonstrated superior transferring performance on seven downstream tasks over the previous standard supervised methods by ImageNet-1K classification. The second is the Transformer-based backbone architecture [2]}, [3]}, [4]}, which has strong potential to replace the previous standard convolutional neural networks such as ResNet [5]}. The pioneer work is ViT [2]}, which demonstrated strong performance on image classification by directly applying the standard Transformer encoder [7]} in NLP on non-overlapping image patches. The follow-up work, DeiT [3]}, tuned several training strategies to make ViT work well on ImageNet-1K image classification. While ViT/DeiT are designed for the image classification task and has not been well tamed for downstream tasks requiring dense prediction, Swin Transformer [4]} is proposed to serve as a general-purpose vision backbone by introducing useful inductive biases of locality, hierarchy and translation invariance.
[1]
[ [ 167, 170 ] ]
https://openalex.org/W3035524453
23dc155a-a24f-40a1-9854-c72ab2669f0f
The vision field is undergoing two revolutionary trends since about two years ago. The first trend is self-supervised visual representation learning pioneered by MoCo [1]}, which for the first time demonstrated superior transferring performance on seven downstream tasks over the previous standard supervised methods by ImageNet-1K classification. The second is the Transformer-based backbone architecture [2]}, [3]}, [4]}, which has strong potential to replace the previous standard convolutional neural networks such as ResNet [5]}. The pioneer work is ViT [2]}, which demonstrated strong performance on image classification by directly applying the standard Transformer encoder [7]} in NLP on non-overlapping image patches. The follow-up work, DeiT [3]}, tuned several training strategies to make ViT work well on ImageNet-1K image classification. While ViT/DeiT are designed for the image classification task and has not been well tamed for downstream tasks requiring dense prediction, Swin Transformer [4]} is proposed to serve as a general-purpose vision backbone by introducing useful inductive biases of locality, hierarchy and translation invariance.
[2]
[ [ 406, 409 ], [ 559, 562 ] ]
https://openalex.org/W3094502228
a2f2ab67-cb75-4de3-bdb6-51c94deddaa5
The vision field is undergoing two revolutionary trends since about two years ago. The first trend is self-supervised visual representation learning pioneered by MoCo [1]}, which for the first time demonstrated superior transferring performance on seven downstream tasks over the previous standard supervised methods by ImageNet-1K classification. The second is the Transformer-based backbone architecture [2]}, [3]}, [4]}, which has strong potential to replace the previous standard convolutional neural networks such as ResNet [5]}. The pioneer work is ViT [2]}, which demonstrated strong performance on image classification by directly applying the standard Transformer encoder [7]} in NLP on non-overlapping image patches. The follow-up work, DeiT [3]}, tuned several training strategies to make ViT work well on ImageNet-1K image classification. While ViT/DeiT are designed for the image classification task and has not been well tamed for downstream tasks requiring dense prediction, Swin Transformer [4]} is proposed to serve as a general-purpose vision backbone by introducing useful inductive biases of locality, hierarchy and translation invariance.
[3]
[ [ 412, 415 ], [ 752, 755 ] ]
https://openalex.org/W3116489684
e0fa8f84-b6c5-4183-bf85-882580528e66
The vision field is undergoing two revolutionary trends since about two years ago. The first trend is self-supervised visual representation learning pioneered by MoCo [1]}, which for the first time demonstrated superior transferring performance on seven downstream tasks over the previous standard supervised methods by ImageNet-1K classification. The second is the Transformer-based backbone architecture [2]}, [3]}, [4]}, which has strong potential to replace the previous standard convolutional neural networks such as ResNet [5]}. The pioneer work is ViT [2]}, which demonstrated strong performance on image classification by directly applying the standard Transformer encoder [7]} in NLP on non-overlapping image patches. The follow-up work, DeiT [3]}, tuned several training strategies to make ViT work well on ImageNet-1K image classification. While ViT/DeiT are designed for the image classification task and has not been well tamed for downstream tasks requiring dense prediction, Swin Transformer [4]} is proposed to serve as a general-purpose vision backbone by introducing useful inductive biases of locality, hierarchy and translation invariance.
[4]
[ [ 418, 421 ], [ 1007, 1010 ] ]
https://openalex.org/W3138516171
46ae90da-bb29-4431-9981-19c3a51eb5ea
The vision field is undergoing two revolutionary trends since about two years ago. The first trend is self-supervised visual representation learning pioneered by MoCo [1]}, which for the first time demonstrated superior transferring performance on seven downstream tasks over the previous standard supervised methods by ImageNet-1K classification. The second is the Transformer-based backbone architecture [2]}, [3]}, [4]}, which has strong potential to replace the previous standard convolutional neural networks such as ResNet [5]}. The pioneer work is ViT [2]}, which demonstrated strong performance on image classification by directly applying the standard Transformer encoder [7]} in NLP on non-overlapping image patches. The follow-up work, DeiT [3]}, tuned several training strategies to make ViT work well on ImageNet-1K image classification. While ViT/DeiT are designed for the image classification task and has not been well tamed for downstream tasks requiring dense prediction, Swin Transformer [4]} is proposed to serve as a general-purpose vision backbone by introducing useful inductive biases of locality, hierarchy and translation invariance.
[5]
[ [ 529, 532 ] ]
https://openalex.org/W2194775991
b60c8d28-2e3d-49bc-ab30-15d759695a68
The vision field is undergoing two revolutionary trends since about two years ago. The first trend is self-supervised visual representation learning pioneered by MoCo [1]}, which for the first time demonstrated superior transferring performance on seven downstream tasks over the previous standard supervised methods by ImageNet-1K classification. The second is the Transformer-based backbone architecture [2]}, [3]}, [4]}, which has strong potential to replace the previous standard convolutional neural networks such as ResNet [5]}. The pioneer work is ViT [2]}, which demonstrated strong performance on image classification by directly applying the standard Transformer encoder [7]} in NLP on non-overlapping image patches. The follow-up work, DeiT [3]}, tuned several training strategies to make ViT work well on ImageNet-1K image classification. While ViT/DeiT are designed for the image classification task and has not been well tamed for downstream tasks requiring dense prediction, Swin Transformer [4]} is proposed to serve as a general-purpose vision backbone by introducing useful inductive biases of locality, hierarchy and translation invariance.
[7]
[ [ 681, 684 ] ]
https://openalex.org/W2963403868
02a1be7a-4940-4a82-a280-b356dc267b3b
While the two revolutionary waves appeared independently, the community is curious about what kind of adaptation is needed and what it will behave when they meet each other. Nevertheless, until very recently, a few works started to explore this space: MoCo v3 [1]} presents a training recipe to let ViT perform reasonably well on ImageNet-1K linear evaluation; DINO [2]} presents a new self-supervised learning method which shows good synergy with the Transformer architecture.
[1]
[ [ 260, 263 ] ]
https://openalex.org/W3145450063
e655a890-7073-4ef4-a906-a901eff76866
While the two revolutionary waves appeared independently, the community is curious about what kind of adaptation is needed and what it will behave when they meet each other. Nevertheless, until very recently, a few works started to explore this space: MoCo v3 [1]} presents a training recipe to let ViT perform reasonably well on ImageNet-1K linear evaluation; DINO [2]} presents a new self-supervised learning method which shows good synergy with the Transformer architecture.
[2]
[ [ 366, 369 ] ]
https://openalex.org/W3159481202
c15f115f-fbe0-46bf-b3d4-7658dd80b595
In addition to this backbone architecture change, we also present a self-supervised learning approach by combining MoCo v2 [1]} and BYOL [2]}, named MaroonMoBY (by picking the first two letters of each). We tune a training recipe to make the approach performing reasonably high on ImageNet-1K linear evaluation: 72.8% top-1 accuracy using DeiT-S with 300-epoch training which is slightly better than that in MoCo v3 and DINO but with lighter tricks. Using Swin-T architecture instead of DeiT-S, it achieves 75.0% top-1 accuracy with 300-epoch training, which is 2.2% higher than that using DeiT-S. Initial study shows that some tricks in MoCo v3 and DINO are also useful for MoBY, e.g. replacing the LayerNorm layers before the MLP blocks by BatchNorm like that in MoCo v3 bring additional +1.1% gains using 100 epoch training, indicating the strong potential of MoBY.
[1]
[ [ 123, 126 ] ]
https://openalex.org/W3009561768
ec990692-8193-4885-b85b-1af46ab42e85
In addition to this backbone architecture change, we also present a self-supervised learning approach by combining MoCo v2 [1]} and BYOL [2]}, named MaroonMoBY (by picking the first two letters of each). We tune a training recipe to make the approach performing reasonably high on ImageNet-1K linear evaluation: 72.8% top-1 accuracy using DeiT-S with 300-epoch training which is slightly better than that in MoCo v3 and DINO but with lighter tricks. Using Swin-T architecture instead of DeiT-S, it achieves 75.0% top-1 accuracy with 300-epoch training, which is 2.2% higher than that using DeiT-S. Initial study shows that some tricks in MoCo v3 and DINO are also useful for MoBY, e.g. replacing the LayerNorm layers before the MLP blocks by BatchNorm like that in MoCo v3 bring additional +1.1% gains using 100 epoch training, indicating the strong potential of MoBY.
[2]
[ [ 137, 140 ] ]
https://openalex.org/W3035060554
487d1fca-23a8-4c20-9d89-e8dab88d1fcd
When transferred to downstream tasks of COCO object detection and ADE20K semantic segmentation, the representations learnt by this self-supervised learning approach achieves on par performance compared to the supervised method. Noting self-supervised learning with ResNet architectures has shown significantly stronger transferring performance on downstream tasks than supervised methods [1]}, [2]}, [3]}, the results indicate large space to improve for self-supervised learning with Transformers.
[1]
[ [ 388, 391 ] ]
https://openalex.org/W3035524453
599e6a47-bb14-479c-ae02-bc8a9f048a6b
When transferred to downstream tasks of COCO object detection and ADE20K semantic segmentation, the representations learnt by this self-supervised learning approach achieves on par performance compared to the supervised method. Noting self-supervised learning with ResNet architectures has shown significantly stronger transferring performance on downstream tasks than supervised methods [1]}, [2]}, [3]}, the results indicate large space to improve for self-supervised learning with Transformers.
[2]
[ [ 394, 397 ] ]
https://openalex.org/W3172615411
96f38a14-992b-4be6-b4ff-1e7a875ce18f
When transferred to downstream tasks of COCO object detection and ADE20K semantic segmentation, the representations learnt by this self-supervised learning approach achieves on par performance compared to the supervised method. Noting self-supervised learning with ResNet architectures has shown significantly stronger transferring performance on downstream tasks than supervised methods [1]}, [2]}, [3]}, the results indicate large space to improve for self-supervised learning with Transformers.
[3]
[ [ 400, 403 ] ]
https://openalex.org/W3135958856
a78f0cae-5cf1-4311-a3f5-fe13ac6fc8e3
MoBY is a combination of two popular self-supervised learning approaches: MoCo v2 [1]} and BYOL [2]}. It inherits the momentum design, the key queue, and the contrastive loss used in MoCo v2, and inherits the asymmetric encoders, asymmetric data augmentations and the momentum scheduler in BYOL. We name it MaroonMoBY by picking the first two letters of each method.
[1]
[ [ 82, 85 ] ]
https://openalex.org/W3009561768
ce0fffa5-803e-482c-88ef-616feb7b3b4a
MoBY is a combination of two popular self-supervised learning approaches: MoCo v2 [1]} and BYOL [2]}. It inherits the momentum design, the key queue, and the contrastive loss used in MoCo v2, and inherits the asymmetric encoders, asymmetric data augmentations and the momentum scheduler in BYOL. We name it MaroonMoBY by picking the first two letters of each method.
[2]
[ [ 96, 99 ] ]
https://openalex.org/W3035060554
8285f7d9-de3f-4005-ae38-8f38b2dfabcf
In training, like most Transformer-based methods, we also adopt the AdamW [1]}, [2]} optimizer, in contrast to previous self-supervised learning approaches built on ResNet backbone where usually SGD [3]}, [4]} or LARS [5]}, [6]}, [7]} is used. We also introduce a regularization method of asymmetric drop path which proves crucial for the final performance.
[1]
[ [ 74, 77 ] ]
https://openalex.org/W2964121744
34d1adcb-d252-4a69-bc0d-8e8e7f027049
In training, like most Transformer-based methods, we also adopt the AdamW [1]}, [2]} optimizer, in contrast to previous self-supervised learning approaches built on ResNet backbone where usually SGD [3]}, [4]} or LARS [5]}, [6]}, [7]} is used. We also introduce a regularization method of asymmetric drop path which proves crucial for the final performance.
[2]
[ [ 80, 83 ] ]
https://openalex.org/W2908510526
7332fbe5-51d5-469e-90ae-d3b9daa0b281
In training, like most Transformer-based methods, we also adopt the AdamW [1]}, [2]} optimizer, in contrast to previous self-supervised learning approaches built on ResNet backbone where usually SGD [3]}, [4]} or LARS [5]}, [6]}, [7]} is used. We also introduce a regularization method of asymmetric drop path which proves crucial for the final performance.
[3]
[ [ 199, 202 ] ]
https://openalex.org/W3035524453
46a815be-b146-4481-9122-fee0a42aba01
In training, like most Transformer-based methods, we also adopt the AdamW [1]}, [2]} optimizer, in contrast to previous self-supervised learning approaches built on ResNet backbone where usually SGD [3]}, [4]} or LARS [5]}, [6]}, [7]} is used. We also introduce a regularization method of asymmetric drop path which proves crucial for the final performance.
[4]
[ [ 205, 208 ] ]
https://openalex.org/W3106005682
2ce253ee-c6fa-4fed-8e61-06454cc41b6e
In training, like most Transformer-based methods, we also adopt the AdamW [1]}, [2]} optimizer, in contrast to previous self-supervised learning approaches built on ResNet backbone where usually SGD [3]}, [4]} or LARS [5]}, [6]}, [7]} is used. We also introduce a regularization method of asymmetric drop path which proves crucial for the final performance.
[5]
[ [ 218, 221 ] ]
https://openalex.org/W3005680577
9be25d9e-8c8d-4c20-b078-db7d7240f49b
In training, like most Transformer-based methods, we also adopt the AdamW [1]}, [2]} optimizer, in contrast to previous self-supervised learning approaches built on ResNet backbone where usually SGD [3]}, [4]} or LARS [5]}, [6]}, [7]} is used. We also introduce a regularization method of asymmetric drop path which proves crucial for the final performance.
[6]
[ [ 224, 227 ] ]
https://openalex.org/W3035060554
63712964-378f-4474-9e50-3ff8ec15f49e
In training, like most Transformer-based methods, we also adopt the AdamW [1]}, [2]} optimizer, in contrast to previous self-supervised learning approaches built on ResNet backbone where usually SGD [3]}, [4]} or LARS [5]}, [6]}, [7]} is used. We also introduce a regularization method of asymmetric drop path which proves crucial for the final performance.
[7]
[ [ 230, 233 ] ]
https://openalex.org/W3172615411
05a29663-9167-48fb-bebc-f49dbb09c782
In this work, we adopt the tiny version of Swin Transformer (Swin-T) as our default backbone, such that the transferring performance on downstream tasks of object detection and semantic segmentation can be also evaluated. The Swin-T has similar complexity with ResNet-50 and DeiT-S. The details of specific architecture design and hyper-parameters can be found in [1]}.
[1]
[ [ 364, 367 ] ]
https://openalex.org/W3138516171
12b6988c-bf0d-467f-ab0a-dcb1e56e1f82
Linear evaluation on ImageNet-1K dataset is a common evaluation protocol to assess the quality of learnt representations [1]}. In this protocol, a linear classifier is applied on the backbone, with the backbone weights frozen and only the linear classifier trained. After training this linear classifier, the top-1 accuracy using center crop is reported on the validation set.
[1]
[ [ 121, 124 ] ]
https://openalex.org/W3035524453
278c7baf-86dd-411a-bf35-5ce93188af3c
During training, we follow [1]} to use random resize cropping with scale from \([0.08, 1]\) and horizontal flipping as the data augmentation. 100-epoch training with a 5-epoch linear warm-up stage is conducted. The weight decay is set as 0. The learning rate is set as the optimal one of \(\lbrace 0.5, 0.75, 1.0, 1.25\rbrace \) through grid search for each pre-trained model.
[1]
[ [ 27, 30 ] ]
https://openalex.org/W3035524453
c370a077-c02f-4188-8499-04d48ae507bb
Regarding previous methods such as MoCo v3 [1]} and DINO [2]} adopt ViT/DeiT as their backbone architecture, we first report results of MoBY using DeiT-S [3]} for fair comparison with them. Under 300-epoch training, MoBY achieves 72.8% top-1 accuracy, which is slightly better than MoCo v3 and DINO (without the multi-crop trick), as shown in Table REF .
[1]
[ [ 43, 46 ] ]
https://openalex.org/W3145450063