Arxiv:1804.10425V1 [Math.DG]

arXiv:1804.10425v1 [math.DG] 27 Apr 2018 neaayi fteudryn on-e fadmi n the and domain a of point-set underlying theor the a of provide analysis invariants ance Integral diff of available. instead is integrals points employing of whose of scale benefits at the estimators of shape vantage and spac feature in as surfaces serve Learni way and this Manifold curves perform like submanifolds, to dimensional tools theoretical as literature eeecs...... Spheres...... over ...... Monomials ...... of . . . . . References Integration ...... A...... Appendix ...... Conclusions ...... 7. . Descriptors . . . . Curvature ...... Analysis . . . 6. Covariance . . . . Spherical ...... Analysis . Submanifold 5. . Covariance . Riemannian . Cylindrical . a . . of . . . . Form . . 4. Fundamental . Submanifolds. . Third . Riemannian . . . of . Invariants . 3. . Integral . . PCA ...... 2. . Introduction 1. Date oa nerlivrat ae nPicplCmoetAna Component Principal on based invariants integral Local pi 0 2018. 30, April : obidu siaoso h eodfnaetlfr,adth and form, fundamental second the submanifolds. of whi estimators directions, hypersurf up principal of build case and to the curvatures In co principal tensor. that the this scale ers of with traces the expansions by asymptotic encoded have decom domains eigenvalue the the that of prove and o submanifolds generalization general ex a to asymptotic propose we have Moreover, space, int ambient curvatures. scalar the the sho in by We balls determined and curvature. codimension, cylinders intrinsic general of and submanifold extrinsic covarianc a local the the relate to with order set in manifold Riemannian bedded Abstract. NERLIVRAT RMCVRAC NLSSOF ANALYSIS COVARIANCE FROM INVARIANTS INTEGRAL AIRÁVRZVZS,MCALKRY N HI PETERSON CHRIS AND KIRBY, MICHAEL ÁLVAREZ-VIZOSO, JAVIER rnia opnn nlsscnb efre vrsaldo small over performed be can Analysis Component Principal MEDDREANA MANIFOLDS RIEMANNIAN EMBEDDED 1. Introduction Contents 1 27 ...... tclln ewe h ttsia covari- statistical the between link etical hcnb sda ecitr tscale at descriptors as used be can ch h lsia hr udmna form fundamental third classical the f 1 ...... 25 ...... 22 ...... cs hscvrac nlssrecov- analysis covariance this aces, gadGoer rcsigo low- of Processing Geometry and ng .Crauedsrposotie in obtained descriptors Curvature e. 10 ...... asosi em ftema and mean the of terms in pansions 17 ...... rnil hnol iceesample discrete a only when erentials oiino h oainematrices covariance the of position httevlm fdmison domains of volume the that w nlsso h neligpoint underlying the of analysis e reto ihhigher-dimensional with ersection steReantno,o general of tensor, Riemann the us ueia mlmnaintksad- takes implementation numerical ti h uvtr information curvature the ntain ieeta-emti nainsat invariants differential-geometric yi aebe nrdcdi the in introduced been have lysis 26 ...... 3 ...... 5 ...... an fa em- an of mains 2 J.ÁLVAREZ-VIZOSO,M.KIRBY,ANDC.PETERSON a point of the domain inside the manifold. In particular, intersecting the submanifold with a ball in the ambient space cuts out a subdomain whose covariance matrix has an eigenvalue decomposition that asymptotically expands with the scale of the ball. The geometric interpretation of this analysis lies in the fact that the first and second terms of the eigenvalue series encode the curvature information of the submanifold at the center of the ball. Integral invariants have been introduced and used in Computer Graphics and Geometry Pro- cessing by [11], [6, 7], [9, 10], [21, 22]. The integral invariant viewpoint via Principal Component Analysis has been introduced and studied theoretically and numerically [1], [4], [9,10], [15], [22], [35], with a focus on curves and surface, in order to process discrete samples of points to determine features and detect shapes at scale, and study stability with respect to noise [19], [28,29]. Voronoi- based covariance matrices have been also been of interest, [23,24]. The eigenvalue decomposition of covariance matrices of spherical intersection domains was also introduced by [5], [31, 32], in order to obtain local adaptive Galerkin bases for the invariant manifold of large-dimensional dynamical systems. For curves, the Frenet-Serret frame is recovered in the scale limit, and ratios of the covariance matrix eigenvalues provide descriptors at scale of the generalized curvatures [2], but the tools needed to study the curve case are significantly different due to the fact that one-dimensional submanifolds have only extrinsic curvature. In the present work we generalize to embedded Riemannian manifolds of general codimension the recent study of PCA integral invariants of hypersurfaces [3], that followed the theoretical study of surfaces in [29], with the purpose of obtaining analogous asymptotic formulas between eigenvalues of covariance matrices and curvature, as it was found for curves in [2]. We shall also introduce a generalization to general codimension of the third fundamental form in order to encapsulate all the curvature information hidden in the covariance analysis. Our main result shows how the eigenvalue decomposition of the covariance of cylindrical and spherical intersection domains has the first two orders of the asymptotic expansion given in terms of the dimension, and the extrinsic and intrinsic curvature encoded in the traces of the third fundamental form, with limit eigenvectors playing the role of generalized principal directions. Geodesic balls inside manifolds have asymptotic series for their intrinsic volume given as corrections to the Euclidean ball completely determined by intrinsic scalar curvature invariants [14]. In our case, the domains of integration depend on the embedding of the submanifold so the extrinsic curvature will play a crucial role in the volume corrections, as in [16]. Normal coordinates via the exponential map are naturally used to do geometric measurements needed for probability and statistics from an intrinsic perspective inside Riemannian manifolds, e.g. [26,27]. The generalized definition of integral invariants makes use of the exponential map in the ambient manifold to make measurements over the underlying point-set of a submanifold. The structure of the paper is as follows: in section §2 we propose a general definition of integral invariants in the context of general Riemannian submanifolds by use of the exponential map, along with the two types of kernel domains on which we will perform the PCA. In section §3, the study of the geometry of submanifolds via the second fundamental form is briefly reviewed and the classical third fundamental form is generalized to submanifolds of general codimension. In section §4 we compute the volume, barycenter and covariance matrix of a cylindrical domain inside an embedded submanifold; in particular, we show that the scaling of the eigenvalues of the covariance matrix singles out the tangent and normal spaces of the manifold at the point INTEGRAL INVARIANTS FROM COVARIANCE ANALYSIS OF EMBEDDED RIEMANNIAN MANIFOLDS 3 by the span of the corresponding limit eigenvectors, and how the next-to-leading order term in the asymptotic series of the eigenvalues is determined by the eigenvalues of the tangent and normal traces of the third fundamental form. In section §5 an analogous analysis is carried out for the domain determined by the intersection of a ball in ambient space with the manifold, which introduces considerable correction terms with respect to the previous case. This leads to an eigenvalue decomposition of the covariance matrix with tangent part given in terms of the Weingarten operator corresponding to the mean curvature vector. Finally, in section §6 we obtain the limit ratios of the eigenvalues in terms of this curvature information, and invert the asymptotic series to get descriptors at scale for the case of hypersurfaces, where the second and third fundamental forms are completely given by the principal curvatures and principal directions. These results show how Principal Component Analysis can be carried out on a general embedded Riemannian submanifold to probe its local geometry. It establishes the relationship between the statistical covariance analysis of the underlying point-set of the manifold and the classical differential-geometric curvature via the third fundamental form. Applying the integral invariant approach to hypersurfaces provides a method to build multi-scale descriptors of curvature also for the case of general codimension.

2. PCA Integral Invariants of Riemannian Submanifolds In our context, integral invariants are local integrals over domains of a submanifold determined by intersection with objects in the ambient space, like spheres. Two such integrals are the volume of the domain and the point in the ambient manifold that represents the center of mass of the region. A more interesting object is the covariance matrix obtained by integrating the relative covariance of the degrees of freedom of the points in the domain, i.e., the products of the coordinates of the points with respect to a chosen frame. In order to get a frame independent integral invariant, one takes the eigenvalue decomposition of the covariance matrix. Since the kernel domains have a natural scale, e.g., the radius of the sphere, it is useful to think of them as a matrix-valued function of scale at every point. Therefore, these integral invariants correspond to eigenvalues and eigenvectors that can be interpreted respectively as a set of scalar and frame- valued functions of scale at every point. The study of covariance matrices in order to obtain adapted frames of general submanifols was studied for example in [5] and [31, 32], whereas the integral invariant approach was developed in detail to extract the curvature information of surfaces in space in [29]. In order to do this type of Principal Component Analysis on a general Riemannian submanifold and generalize local integral invariants, deﬁnitions using Cartesian coordinates must naturally be promoted to Riemann normal coordinates [8], [25]. If the n-dimensional submanifold Mpnq sits inside an ambient Riemannian manifold N pn`kq, g , the curves in N that generalize the axis used n`k p q in R are the geodesic curves γv t and these always exist and are unique locally at any point p q p N and direction v. Given an orthonormal frame in TpM NpM, the geodesics tangent to P ‘ each of the vectors will trace out generalized coordinate axis in N that, through the exponential map will uniquely specify any point in a local neighborhood around p. Assuming N is geodesically complete to simplify the exposition, the exponential map collects all geodesics starting at p by 4 J.ÁLVAREZ-VIZOSO,M.KIRBY,ANDC.PETERSON

n`k mapping straight lines through the origin in TpN R to geodesics through p: – exp : TpM N such that exp tv γtv 1 γv t . p Ñ pp q“ p q“ p q

At any point p there is a neighborhood U of 0 in TpN where exp is a diffeomorphism onto a neighborhood U of p in N . From this, for star-shaped U, there is also a unique geodesic γ t p q connecting p and any other point q U rsuch that the tangent γ1 0 exp´1 q . Moreover, P p q “ p p q the arclength of γ between the two points, i.e. the distancer d p,q between them determined p q by the metric g, is the length of the tangent vector representation through this map, d p,q ´1 p q “ expp q . These normal neighborhoods allow the parametrization of points using the geodesic } p q} n`k distances tangent to a given frame eµ at p. The injectivity radius rp is the radius of the t uµ“1 largest ball B0 ε in TpN where exp is a diffeomorphism, so Bp rp exp B0 rp is the largest p q p q“ pp p qq ball in N created by radial geodesics of the same length around p where normal coordinates are well-defined. In fact rp 0 always. Since our main theorems 4.5 and 5.5 are asymptotic results ą with scale, in a general Riemannian manifold one could always use normal coordinates to study domains of submanifolds small enough so that they can be mapped to Euclidean space, thus, we propose the following general definition of PCA integral invariants in a general Riemannian manifold.

Deﬁnition 2.1. Let D be a measurable domain in a Riemannian manifold N , g such that p q D Bp rp for some point p N , The integral invariants associated to the moments of order 0, Ă p q P 1 and 2 of the geodesic coordinate functions of the points of D with respect to p are: the volume V D 1 dVol, (2.1) p q“ żD the barycenter 1 s D exp´1 q dVol, (2.2) p q“ V D r p p qs p q żD and the eigenvalue decomposition of the covariance matrix:

C D exp´1 q exp´1 q dVol. (2.3) p q“ r p p qs b r p p qs żD Here dVol is the measure on D, restriction of the measure of N induced by the metric g, and the tensor product is to be understood as the outer product of the components of the exp´1 map in a chosen orthonormal basis of TpN . The reference point of the covariance matrix is often chosen to be the barycenter exp s instead of p. pp q The two types of domains that we shall study are regions in a submanifold M N deter- Ă mined by the intersection with a ball and a cylinder. Using the exponential map one can deﬁne such intersections by mapping Euclidean balls and higher-dimensional cylinders in TpN to their geodesic generalizations in the ambient manifold N .

Deﬁnition 2.2. The spherical component of radius ε rp, at a point p of a submanifold M of ď a Riemannian manifold N is the domain given by:

´1 Dp ε : M q N : exp q ε rp . (2.4) p q “ X t P } p p q} ď ď u INTEGRAL INVARIANTS FROM COVARIANCE ANALYSIS OF EMBEDDED RIEMANNIAN MANIFOLDS 5

An element V in the Grassmannian Gr m,n k is an m-dimensional linear subspace of Rn`k. p ` q Fixing a point and m-dimensional ball inside V, the standard three dimensional cylinder over the xy-plane can be generalized to an V-cylinder by taking all points in the ambient space that project down onto the ball inside V.

Deﬁnition 2.3. The cylindrical component of radius ε rp, at a point p of a submanifold M of ď a Riemannian manifold N over the m-plane V Gr m,n k , is the V-cylinder intersection: P p ` q ´1 Cyl ε, V : M q N : projV exp q ε rp , (2.5) pp q “ X t P } p p p qq} ď ď u where projV is the orthogonal projection onto V as a linear subspace of TpN . We shall write p¨q Cyl ε when V TpM is assumed. pp q “ We will compute these integral invariants for embedded submanifolds in Euclidean ambient space, N Rn`k, where exp´1 q q p as vectors and the tensor product recovers the “ p p q “ ´ common deﬁnition of PCA integral invariants studied in the literature. The points q D are P then parametrized by a vector X such that the barycenter is the center of mass 1 s D X dVol, (2.6) p q“ V D p q żD and the the covariance matrix can be interpreted as analogous to a moment of inertia matrix, which for the cylindrical component shall be taken with respect to the center p, following the convention and motivation of [32],

C Cyl ε X p X p dVol, (2.7) p pp qq “ p ´ q b p ´ q żCylppεq whereas for the spherical component the covariance matrix shall be taken with respect to the barycenter following [29],

C Dp ε X s Dp ε X s Dp ε dVol. (2.8) p p qq “ p ´ p p qqq b p ´ p p qqq żDppεq Without loss of generality, these deﬁnitions could have been normalized by the volume of the domain to make the integral measure become a probability density and thus make the matrices actual statistical covariances.

3. Third Fundamental Form of a Riemannian Submanifold For a complete analysis of the the geometry of Riemannian submanifolds see [8], [17], [25], [33]. Let M, g be an n-dimensional manifold isometrically embedded in an n k -dimensional p q p ` q Riemannian manifold N , g , and let ∇, ∇ be the respective Levi-Civita connections. We shall p q write g , , , classically called the first fundamental form of M in N . Then, at any point p¨ ¨q “ x ¨ ÿ p M and for any vector y TpM, and vector field X Γ T M , the metric connection of M P P P p Jq J is the projection of the metric connection of N : ∇yX ∇yX , where : TpN TpM. “ p q p ¨ q Ñ The second fundamental form II of M in N is defined to be the normal projection of the ambient K covariant derivative when acting on vectors fields tangent to M, i.e., denoting : TpN p ¨ q Ñ NpM, K II x, y ∇yX , i.e., ∇yX ∇yX II x, y , (3.1) p q “ p q “ ` p q 6 J.ÁLVAREZ-VIZOSO,M.KIRBY,ANDC.PETERSON for all x, y TpM, and X Γ T M such that X p x. It is a symmetric bilinear form on the P P p q | “ tangent space at every point taking values in the normal space, II : TpM TpM NpM. Fixing b Ñ a normal vector n NpM, the scalar-valued bilinear form II x, y , n has a corresponding self- P x p q y adjoint map Sn End TpM , called the Weingarten map at n, such that: P p q II x, y , n Sn x, y x, Sn y . (3.2) p x p q y“x y“x y n k Fixing orthonormal bases eµ of TpM, and nj of NpM, the components of the second t uµ“1 p t uj“1 p fundamental form at point p are:

k k k j II eµ, eν II eµ, eν nj II eµ, eν , nj nj Sj eµ, eν nj. (3.3) p q“ j“1 p q “ j“1x p q y “ j“1x y ÿ ÿ ÿ The geometric meaning of II lies in the fact that the Weingarten map measuresp the tangential rate of change of normal vectors to M when moving in tangent directions, cf. [8, Eq. II.2.4]:

J Sn x ∇xN , “ ´p q for any N Γ NM such that N p n. From this, [25, Ch. 4, Cor. 9, 10], II x, x is to be P p q | “p p q interpreted as the curve acceleration in N of a geodesic inside M at p with tangent velocity x. Therefore, II naturally measures the extrinsic curvature of the embedding since it represents the forced curving of the straightest lines in M due to the curving of M itself in N . The inverse function theorem and [17, Ch. VII, Ex. 3.3] establish the following lemma, of fundamental importance for the computations in the proofs of the present work.

Lemma 3.1. Let M be an n-dimensional submanifold of an n k -dimensional Riemannian p ` q manifold N , g , with the induced metric g M. For any point p M and orthonormal basis n p Mq | 1 P n`k N eµ µ“1 of Tp , it is possible to choose normal coordinates y ,...,y in such that the t u 1 n p n q n`1 n`k coordinate tangent vectors at the origin Y ,..., Y coincide with eµ µ“1, and Y ,..., Y k t u are an orthonormal basis nj of NpM. Moreover, M is locally given by a graph manifold t uj“1 y1 x1,...,yn xn,yn`1 f 1 x ,...,yn`k f k x , such that the components of the second “ “ “ p q “ p q fundamental form at p can be written as:

k 2f j II eµ, eν B 0 nj. (3.4) p q“ xµ xν p q j“1 ÿ „B B  n The invariance of the trace of II for any orthonormal tangent frame eµ leads to the t uµ“1 deﬁnition of the mean curvature vector: n k n j j j H II eµ, eµ H nj, where H II eµ, eµ . (3.5) “ µ“1 p q“ j“1 “ µ“1 p q ÿ ÿ ÿ The study of the intrinsic geometry of M, g depends only on the metric and is given in terms p q of the Riemann curvature tensor:

R x, y z ∇x∇y ∇y∇x ∇ Z, p q “ p ´ ´ rx,ysq for any x, y, z TpM and Z Γ T M such that Z p z. This fundamental tensor equivalently P P p q | “ measures the integrability of parallel transport, geodesic deviation and local ﬂatness. Its traces INTEGRAL INVARIANTS FROM COVARIANCE ANALYSIS OF EMBEDDED RIEMANNIAN MANIFOLDS 7 yield the Ricci tensor n Ric x, y R eµ, x y, eµ R x, y , p q“ µ“1x p q y“x y ÿ p and the scalar curvature, R Ric eµ, eµ . Here, R End TpM is the Ricci operator “ µ p q P p q associated to the Ricci tensor with respect to the metric. ř Gauß Theorema Egregium establishes that the intrinsicp curvature of surfaces is a particular combination of products of the components of the second fundamental form. This generalizes to higher dimension to

Theorem 3.2 (Gauß equation). The Riemann curvature tensor of a submanifold M is related to the curvature R of the ambient manifold N via

R x, y z, w R x, y z, w II x, w , II y, z II x, z , II y, w (3.6) x p q y“x p q y`x p q p qy´x p q p qy for all x, y, z, w TpM. P In classical differential geometry, [13], [34], the third fundamental form is the natural object to construct out of scalar products after the first fundamental form, I x, y x, y , and the p q“x y second fundamental form II x, y Sx, y , so it is defined for hypersurfaces, e.g. [20], as p q“x y 2 III x, yp Sx, Sy S x, y . p q“x y“x y However, it does not provide new informationp sincep it is compp letely determined by Gauß equation 3.2, e.g., in Euclidean space [17]:

2 S x, y H Sx, y Ric x, y , (3.7) x y“ x y´ p q 2 or, in terms of the Ricci operator,p S HSp R. For a manifold M of higher codimension “ ´ k, there are k linearly independent normal vectors at every point and, as mentioned before, the generalized second fundamental formp takes valuesp p in the normal bundle precisely to reﬂect this structure in terms of the corresponding Weingarten operators at every normal vector. Therefore, the natural generalization of Sx, Sy to this context is x y Deﬁnition 3.3. The third fundamentalp p form of a Riemannian submanifold M N is the fourth- ˚ 2 ˚ Ă rank tensor III TpM NpM NpM, given at every point p M by P p q b b P

III x, y n, m : Sm x , Sn y . (3.8) x p q y “x y for any x, y TpM, and n, m NpM. P P p p At any speciﬁc point, and because the Weingarten maps are self-adjoint, the linear operator III x, y End NpM is written as the following linear combination, when a particular p q P p k q j orthonormal basis nj of the normal space is ﬁxed and η g , nj is the dual basis: t uj“1 “ p¨ q k i III x, y Si Sj x , y η nj. (3.9) p q“ x y b i,j“1 ÿ p p 8 J.ÁLVAREZ-VIZOSO,M.KIRBY,ANDC.PETERSON

j This is due to the linearlity of the map n Sn : NpM End TpM ; if n n nj then ÞÑ Ñ p q “ j k k ř pj j Sn x, y II x, y , n n II x, y , nj n Sj x, y , x y“x p q y“ x p q y“x y j“1 ˜j“1 ¸ ÿ ÿ for all x, y pTpM. p P ˚ 2 ˚ Let us deﬁne the tangent trace of a tensor A TpM NpM NpM as the operator sum Pn p q b b of the evaluations at an orthonormal basis eµ of TpM: t uµ“1 n tr kA : A eµ, eµ End NpM , (3.10) “ µ“1 p q P p q ÿ And let the normal trace of such a tensor be k ˚ 2 tr KA : III , nj, nj TpM , (3.11) “ j“1x p¨ ¨q y P p q ÿ k for any orthonormal basis nj of NpM. These tensors are well-deﬁned since the sums are t uj“1 independent of the orthonormal basis chosen.

Lemma 3.4. At any point p M, for any x, y TpM, and n, m NpM, the normal trace of P P P the third fundamental form is k 2 tr III x, y S x, y SH R R x, y , (3.12) K p q“ x j y“xp ´ ` q y j“1 ÿ where R and R are the Ricci operatorsp of M and Np respectively.p In particular, the sum of k squares of the Weingarten operators Sj, for an orthonormal basis nj of NpM, is independent t uj“1 of the basis.p The tangent trace of the third fundamental form is a linear operator on NpM whose components with respect to thep metric are the Frobenius inner products of the corresponding Weingarten operators:

tr III n, m tr SnSm . (3.13) x p k q y“ p q The total trace is p p tr III tr tr III H 2 R R. (3.14) “ K k “} } ´ ` Proof. The normal trace biliniar form has components k k n tr KIII eµ, eν Sj eµ, Sj eν Sjeα, eµ Sjeα, eν p q“ j“1x y“ j“1 α“1x yx y ÿ ÿ ÿ k np p p n p j j II eα, eµ II eα, eν II eα, eµ , II eα, eν , (3.15) “ j“1 α“1 p q p q“ α“1x p q p qy ÿ ÿ ÿ that using Gauß equation lead to the corresponding linear operator with respect to the metric: n n n tr KIII eµ, eν II eα, eα , II eν, eµ R eα, eν eµ, eα R eα, eν eµ, eα p q“ α“1x p q p qy` α“1x p q y´ α“1x p q y ÿ ÿ ÿ II eµ, eν , H Ric eµ, eν Ric eµ, eν “x p q y` p q´ p q SH eµ, eν R eµ, eν R eµ, eν . “x y`x y´x y p p INTEGRAL INVARIANTS FROM COVARIANCE ANALYSIS OF EMBEDDED RIEMANNIAN MANIFOLDS 9

This is the generalization of the operator of the classical third fundamental form, equation 3.7:

k 2 S SH R R. j “ ´ ` j“1 ÿ p p p The tangent trace is trivial by deﬁnition of trace of a linear operator with respect to the metric and the self-adjointness of the Weingarten operators:

n III tr k n, m SmSn eµ, eµ Sm, Sn F . x p q y“ µ“1x y “ p q ÿ p p p p In a ﬁxed orthonormal basis this tensor is the linear combination

k n k III i i tr k Si Sj eµ , eµ η nj tr Si Sj η nj, “ i,j“1 µ“1x y b “ i,j“1 p q b ÿ ÿ ÿ p p p p whose components can be expressed in terms of the second fundamental form as

n n i j tr SiSj Si eµ, eν Sj eµ, eν II eµ, eν II eµ, eν . (3.16) p q“ µ, ν“1x yx y“ µ, ν“1 p q p q ÿ ÿ p p p p Taking the total trace of III is analogous to the complete contraction of the Riemann curvature tensor indices to obtain the scalar curvature:

n III III tr tr ktr K SH R R eµ, eµ tr SH tr R tr R “ “ µ“1x p ´ ` q y“ ´ ` ÿ n p p n p p 2 tr KIII eµ, eµ II eα, eβ , (3.17) “ µ“1 p q“ α, β } p q} ÿ ÿ n 2 where tr SH II eµ, eµ , H H , and the traces of the Ricci operators are by “ µ“1x p q y“} } deﬁnition the scalar curvatures. p ř Equations 3.16 and 3.15 will be recognized inside the elements of the tangent and normal matrix blocks in our covariance matrices to express its eigenvalues in terms of the third fundamental form. The asymmetry of the components of the third fundamental form operator III x, y encodes p q K the curvature information of the connection deﬁned on the normal bundle NM by ∇xN , for p q any x TpM, N Γ NM , where an analog to Gauß equation holds. P P p q

Lemma 3.5 (Ricci equation). The Riemann curvature of the induced normal connection, RK, satisﬁes:

R x, y n, m R x, y n, m III x, y n, m III x, y m, n , (3.18) x Kp q y“x p q y`x p q y´x p q y for all x, y TpM, and n, m NpM, at any point p M. P P P 10 J.ÁLVAREZ-VIZOSO,M.KIRBY,ANDC.PETERSON

Proof. Writing the classical equation [8, Ex. II.11] in terms of Weingarten maps leads to n R x, y n, m RK x, y n, m II eµ, x , n II eµ, y , m x p q y´x p q y“ µ“1 rx p q yx p q y` ÿ II eµ, y , n II eµ, x , m ´x p q yx p q y s n Sn x, eµ Sm y, eµ Sn y, eµ Sm x, eµ Sn x, Sm y Sn y, Sm x . “ µ“1x yx y´x yx y“x y´x y ÿ p p n p p p p p p for any orthonormal basis eµ of TpM. t uµ“1 4. Cylindrical Covariance Analysis In this section we compute the integral invariants of the cylindrical domain around a point on an n-dimensional submanifold M of Rn`k. In the case the cylinder is not normal to the manifold at the point, we can only establish the leading order terms, but that is sufficient in the generic case to be able to detect the tangent space of the manifold by the scaling behaviour of the eigenvalues of the covariance matrix. Once the cylinder is fixed to be normal to this tangent space, the integral invariants can be computed to next-to-leading order to see how they encode the geometric information of the third fundamental form. We shall always work in a neighborhood U Rn`k of p M, sufficiently small so that U M Ă P X is given by a graph representation x1,...,xn,f 1 x ,...,f k x T over its tangent space, i.e., 0 1 n T r j p q p qs represents p, x x ,...,x TpM, and ∇f 0 0, so that the manifold is approximated “ r s P p q “ at p by its osculating paraboloids.

Lemma 4.1. The first fundamental form components of a graph manifold M Rn`k, parametrized 1 n 1 k T n`k Ă by x ,...,x ,f x ,...,f x TpM NpM R , are: r p q p qs P ‘ – k f j f j gµν x δµν B B . (4.1) p q“ ` xµ xν j“1 ÿ B B The induced measure on M in these coordinates is given by 2 1 k n n 2f j dVol det g x dnx 1 B 0 xβ O x3 dnx. (4.2) “ p q “ ¨ ` 2 » xα xβ p q fi ` p q˛ j“1 α“1 β“1 ˆ ˙ a ÿ ÿ ÿ B B ˝ – fl ‚ Proof. The tangent space in these coordinates is spanned by the vectors 1 k 1 n 1 k T f f T Xµ B x ,...,x ,f x ,...,f x 0,..., 1,..., 0, B ,..., B , “ xµ r p q p qs “ r xµ xµ s B B B for µ 1,...,n, which yields the canonical orthonormal basis at p since ∇f j 0 0. The “ p q “ induced metric tensor is then k f j f j g x X , X δ B B . µν µ ν µν xµ xν p q“x y“ ` j“1 ÿ B B From this, recalling that the f j x have Taylor expansions starting at order 2 in these coordinates, p q the matrix of the metric components is of the form g Idn h , where the correction matrix r s“ ` r s INTEGRAL INVARIANTS FROM COVARIANCE ANALYSIS OF EMBEDDED RIEMANNIAN MANIFOLDS 11

j j j h µf ν f is small because we are in a neighborhood of 0 with ∇f 0 0. Let r s “ r j B B s p q“ ř 1 n 2f j f j x B 0 xαxβ O x3 , p q“ 2 xα xβ p q ` p q α,β“1 ÿ ˆB B ˙ for every j 1,...,k, then “ f j n 2f j B B 0 xβ O x2 . xµ “ xβ xµ p q ` p q β“1 B ÿ ˆB B ˙ The natural volume form of a Riemannian manifold is given by ?det g dx1 dxn, [25, Ch. ^¨¨¨^ 7, Lem. 19], whose lowest order approximation is det g 1 tr h, so ?det g 1 1 tr h, i.e., « ` « ` 2 2 1 n k f j 2 1 n k n 2f j det g x 1 B ... 1 B 0 xβ O x3 . p q“ ` 2 xα ` “ ` 2 » xβ xµ p q ﬁ ` p q α“1 j“1 ˆ ˙ α“1 j“1 β“1 ˆ ˙ a ÿ ÿ B ÿ ÿ ÿ B B – ﬂ

In the rest of this paper we shall abbreviate second derivatives at the origin by 2f j κj κj : B 0 , αβ “ βα “ xα xβ p q B B motivated by the notation of hypersurface principal curvatures, which are the eigenvalues of the local Hessian of the deﬁning function. We can now compute the Taylor expansion of the integral invariants in the chosen coordinates, and then relate the terms to the curvature diﬀerential invariants which are always combinations of second derivatives.

Theorem 4.2. The n-dimensional volume of the cylindrical component for a generic V Gr n,n k , K P p ` q such that V TpM 0 , is to leading order the volume of the ellipsoid of intersection between X “ t u the V-cylinder and TpM: n V n`1 V Cylp ε, Vn 1 ℓµ O ε , (4.3) p p qq “ p q µ“1 ` p q ź where ℓµ are the the principal semi-axes of the ellipsoid. When V TpM, the volume is “ 2 ε 4 V Cyl ε Vn ε 1 tr III O ε (4.4) p pp qq “ p q ` 2 n 2 ` p q „ p ` q  where tr III H 2 R. “} } ´ Proof. To compute the leading term of V Cyl ε, V we can approximate M near p by its tangent p pp qq space, such that, ﬁxing local coordinates with a basis for TpM NpM, a point is speciﬁed by T K ‘ K n`k X x, 0 , with x TpM, 0 NpM. Since V TpM 0 , we have TpM V R , and “ r s K Pn`k P n X “ t u ‘ n“ k of course V V R . Let eµ be an orthornomal basis of TpM, and uα vj ‘ “ t uµ“1 t uα“1 Y t uj“1 an orthonormal basis of V VK, then the elements of the former are a linear combination of the ‘ latter, so there are matrices A, B such that: n k α j eµ A uα B vj. “ µ ` µ α“1 j“1 ÿ ÿ 12 J.ÁLVAREZ-VIZOSO,M.KIRBY,ANDC.PETERSON

µ We need to ﬁnd the region projV X ε, and since X x eµ, when X TpM, the } p q} ď “ µ P projection is n n n ř µ α projV X X, uα uα x Aµuα, p q“ α“1x y “ α“1 µ“1 ÿ ÿ ÿ hence, the domain of integration in x in this approximation is

n n 2 2 µ α 2 projV X x Aµ ε . } p q} “ α“1 ˜µ“1 ¸ ď ÿ ÿ This is a quadratic equation that can be written as

n n µ α α ν T T T 2 2 x AµAν x x A A x y y y ε , µ, ν «α“1 ff “ r ¨ s “ ¨ “} } ď ÿ ÿ where y AT x. The matrix A AT is positive definite since it is clearly nonnegative, and if “T r ¨ s K x ker A for nonzero x, then projV X 0, thus X V , which contradicts X TpM under P K p q“ P P our assumption V TpM 0 . Therefore, the cylindrical domain is an n-dimensional ellipsoid X “ t u in the tangent space at p, whose volume is given in terms of its principal semi-axes: πn{2 n V Cyl ε, V ℓ O εn`1 . p Γ n 1 µ p p qq “ 2 µ“1 ` p q p ` q ź When V TpM, the local graph approximation of M over TpM yields “ proj X proj x,f 1 x ,...,f k x T x ε, TpMp q“} TpMpr p q p qs q}“} }ď pnq thus, we are integrating det g x over the ball Bp ε TpM, which can be computed using p q p q Ă the integrals in the appendix:a 2 ε 1 k n n S n´1 i β 3 V Cylp ε d ρ 1 καβρ x O x dρ p p qq “ Sn´1 ¨ ` 2 » fi ` p q˛ 0 i“1 α“1 β“1 ż ż ÿ ÿ ÿ n`2 ˝k n n – fl ‚ ε i i β γ S n`4 Vn ε καβκαγ x x d O ε “ p q` 2 n 2 Sn´1 ` p q i“1 α“1 β,γ p ` q ÿ ÿ ÿ ż C εn`2 k n V ε 2 κi 2 O εn`4 n 2 n 2 αβ “ p q` i“1 α,β p q ` p q p ` q ÿ ÿ 2 n Vn ε ε V ε p q II e , e , II e , e O εn`4 . n 2 n 2 α β α β “ p q` α,β x p q p qy` p q p ` q ÿ Here the spherical integral is only nonzero when β γ, and the last term is the component “ expression of equation 3.14.

Proposition 4.3. The barycenter of the cylindrical component, for V as in the previous theorem, is s Cyl ε, V 0 O ε2 . (4.5) p pp qq “ ` p q INTEGRAL INVARIANTS FROM COVARIANCE ANALYSIS OF EMBEDDED RIEMANNIAN MANIFOLDS 13

In the case V TpM, the barycenter is: “ ε2 s Cyl ε 0, H T O ε4 . (4.6) p pp qq “ r 2 n 2 s ` p q p ` q Proof. For generic V, approximating the manifold again by its tangent space, X x, 0 O ε2 T , “ r ` p qs the normal component does not contribute until order two and the tangent component also vanishes at order 1 in ε. When V TpM, we saw that the integration domain reduces to a ball. “ The integrals of the tangent components xµ weighed by ?det g are of order O εn`4 , since the p q ﬁrst terms in the expansion have odd powers in the coordinates. On the other hand the normal components integrate as:

ε ε 1 n j S j n´1 S n´1 j 2 α β O 3 V s d f det gρ dρ d ρ καβρ x x x dρ r s “ Sn´1 “ Sn´1 ¨2 ` p q˛ 0 0 α, β“1 ż ż a ż ż ÿ εn`2 n C˝εn`2 ‚ j α β S O n`4 2 j O n`4 καβ x x d ε H ε , “ 2 n 2 Sn´1 ` p q“ 2 n 2 ` p q α,β“1 p ` q ÿ ż p ` q n Dividing by V V Cyl ε cancels C2ε Vn ε to leading order. “ p pp qq “ p q In order to study the eigenvalue decomposition of the covariance matrix we need to establish how to determine the limit eigenvectors and the ﬁrst two terms of the series expansion of the eigenvalues, so that computing the integrals in an arbitrary orthonormal basis produces blocks identiﬁable in terms of the coordinate expressions of the second and third fundamental forms in that basis. An analogous result to the matrix expansion in [3] generalizes to higher codimension.

Lemma 4.4. Let C ε be an n k n k real symmetric matrix depending on a real parameter p q p ` qˆp ` q ε with convergent series expansion in a neighborhood of 0 such that:

a Idn 0nˆk Anˆn Bnˆk C ε ε2 ε4 O ε5 , p q“ ˜0kˆn 0kˆk¸ ` ˜Bkˆn Γkˆk ¸ ` p q where a 0, and the blocks A, B, Γ are not completely zero. Let V , V denote the ﬁrst ‰ r sJ r sK n and last k components of a vector in Rn`k. Then the series of eigenvectors of C ε form n`k p q an orthonormal basis of R that converges for ε 0. The ﬁrst n eigenvalues are λµ ε 2 p4q 4 5 p4q Ñ p0q n p q “ aε λµ ε O ε , where λµ and the corresponding limit eigenvectors V µ satisfy the ` ` p q t uµ“1 eigenvalue decomposition of A:

p4q p0q p0q λ Idn A V 0n 1, V 0k 1. p µ ´ q r µ sJ “ ˆ r µ sK “ ˆ p4q 4 5 p4q The last k eigenvalues are λj ε λ ε O ε , where λ and the corresponding limit eigen- p q“ j ` p q j vectors V p0q n`k satisfy the eigenvalue decomposition of Γ: t j uj“n`1 p4q p0q p0q λ Idk Γ V 0n 1, V 0n 1. p j ´ q r j sK “ ˆ r j sJ “ ˆ Therefore, the fourth-order term of the eigenvalues is given by the eigenvalues of the blocks A and Γ, with the respective eigenvectors as the limit eigenvectors of C ε for ε 0. p q Ñ 14 J.ÁLVAREZ-VIZOSO,M.KIRBY,ANDC.PETERSON

Proof. The eigenvalue decomposition C ε V ε λ ε V ε can be written as a convergent series p q p q“ p q p q expansion in ε within a neighborhood of 0 for all Hermitian matrices of converging power series elements [30]:

a Idn 0nˆk Anˆn Bnˆk ε2 ε4 O ε5 V p0q V p1qε V p2qε2 . . . r ˜0kˆn 0kˆk¸ ` ˜Bkˆn Γkˆk ¸ ` p qs¨r ` ` ` s“ λp1qε1 λp2qε2 λp3qε3 λp4qε4 . . . V p0q V p1qε V p2qε2 . . . . “ p ` ` ` ` qr ` ` ` s The zero matrix C 0 is the limit when ε 0, with λ 0 λp0q 0 as a totally degenerate p q Ñ p q “ “ eigenvalue of multiplicity n k . By [30, ch. I, Th. 1], for ε 0, this eigenvalue branches out p ` q ą into n k eigenvalues λi ε with n k orthonormal eigenvectors V i ε , all convergent in a p ` q p q p ` p0qq p q neighborhood of 0. Thus, the vectors V limε 0 V i ε are a unique orthonormal basis of i “ Ñ p q Rn`k that is completely determined by the perturbation matrix. The eigenvalue difference between C ε and its full diagonalization is bounded by the matrix p q norm difference between them, which implies λp1q λp3q 0, and also λp2q a, for i 1,...,n, “ “ i “ “ and λp2q 0, for i n 1,...,n k, since C ε is already diagonal up to that order. One can i “ “ ` ` p q obtain the relations satisfied by λp4q and V p0q equating order by order. At second order, λp2q a i “ is nonzero for i 1,...,n, hence “

a Idn 0nˆk p2q p0q 0nˆn 0nˆk p0q λi Idn`k V i V i 0 r ˜0k n 0k k¸ ´ s “ ˜0k n a Idk¸ “ ˆ ˆ ˆ ´ p0q implies that V µ 0k 1, for the limit of the ﬁrst n eigenvectors. At fourth order we have r sK “ ˆ

p4q Anˆn Bnˆk p0q a Idn 0nˆk p2q p2q λi Idn`k V i λi Idn`k V i , r ´ ˜Bkˆn Γkˆk ¸ s “ r ˜0kˆn 0kˆk¸ ´ s which in the present case, i 1,...,n, makes the right-hand side become 0 for the ﬁrst n rows. p0q “ On the other hand, V 0k 1 makes B not contribute in the left-hand side, hence the ﬁrst r i sK “ ˆ n rows lead to the equation: p4q p0q λ Idn A V 0n 1. p i ´ q r i sJ “ ˆ p2q p0q When i n 1,...,n k, an analogous argument using λ 0, leads to V 0n 1, and “ ` ` i “ r i sJ “ ˆ in turn to: p4q p0q λ Idn Γ V 0k 1. p i ´ q r i sK “ ˆ Since the limit eigenvectors are an orthonormal basis they cannot be zero and, therefore, the previous equations establish λp4q and the nonzero components of V p0q as the eigenvalue decom- i r i s position of A and Γ, which always has a solution due to being symmetric matrices.

The previous lemma is a fundamental step to establish the main theorem of this and the next section.

K Theorem 4.5. For V Gr n,n k such that V TpM 0 , i.e. for non-normal transver- P p ` q X “ t u sality, and when Cyl ε, V is ﬁnite, the covariance matrix Cp ε, V has as limit eigenvectors pp q p q INTEGRAL INVARIANTS FROM COVARIANCE ANALYSIS OF EMBEDDED RIEMANNIAN MANIFOLDS 15

2 spanning TpM those corresponding to the ﬁrst n eigenvalues, which scale as ε . The other k eigenvalues scaling at higher order have limit eigenvectors that span NpM:

2 n ε 2 n`3 λµ Cyl ε, V ℓ Vn 1 ℓα O ε , µ 1,...,n, (4.7) p pp qq “ n 2 µ p q ` p q “ ` α“1 n`3 ź λj Cyl ε, V 0 O ε , j n 1,...,n k, (4.8) p pp qq “ ` p q “ ` ` where ℓµ are the principal lengths of the ellipsoid in 4.2. When V TpM, let λl denote taking “ r¨s the l-th eigenvalue of a linear operator at p, or of its associated bilinar form with respect to the metric. Then the eigenvalues of the covariance matrix of the cylindrical component are:

2 4 ε ε 6 λµ Cyl ε Vn ε λµ tr III Idn 2 tr III O ε (4.9) p pp qq “ p q n 2 ` 2 n 2 n 4 r p q ` K s` p q „  ` 4 p ` qp ` q ε 6 λj Cyl ε Vn ε λj H H 2 tr III O ε (4.10) p pp qq “ p q 4 n 2 n 4 r b ` k s` p q „ p ` qp ` q  for all µ 1,...,n, and j n 1,...,n k. Moreover, the corresponding first n eigenvectors “ “ ` ` converge to the principal directions of the operator tr KIII SH R, and the last k eigenvectors III “ ´ to those of H H 2 tr k . b ` p p Proof. For generic V the manifold is again approximated by its tangent space as X x, 0 T , “ r s which produces no contribution to the normal block at leading order O εn`2 . Choosing the p q tangent orthonormal basis to be aligned with the principal axis of the ellipsoid, and changing µ µ variables so that x y ℓµ, the tangent block becomes an integration over a ball: “ n V µν µ ν n µ ν n C Cylp ε, x x d x y y ℓµℓν ℓα d y r p p qqs “ T T 2 “ 2 x AÄ xďε µ yµď1 α“1 ż ż ź εn`2 n ř δ ℓ ℓ V 1 ℓ O εn`3 . µν n 2 µ ν n α “ p q α“1 ` p q ` ź Thus, the covariance matrix leading term is proportional to diag ℓ2,...,ℓ2 , 0,..., 0 , which has p 1 n q limit eigenvectors corresponding to the first n eigenvalues spanning TpM, and the other k eigen- 2 vectors spanning NpM, by an straightforward extension to lemma 4.4 at order ε . µ ν n i j k For V TpM, we shall compute the integrals of the matrix blocks x x , and f f , “ r sµ,ν“1 r si,j“1 so the next-to-leading order elements of those blocks will suffice to obtain the eigenvalues and limit eigenvectors by the results of the previous lemma. The tangent block is:

µν µ ν n C Cylp ε x x det g x d x r p p qqs “ pnq p q żB pεq a 2 ε 1 k n n S n`1 µ ν j β O 3 d ρ x x 1 καβρ x x dρ “ Sn´1 ¨ ` 2 » ﬁ ` p q˛ 0 i“1 α“1 β“1 ż ż ÿ ÿ ÿ n`2 ˝ n`4 k n– n ﬂ ‚ ε µ ν S ε i i µ ν β γ S n`6 x x d καβκαγ x x x x d O ε , n 2 Sn´1 2 n 4 Sn´1 “ ` i“1 α“1 β,γ ` p q ` ż p ` q ÿ ÿ ÿ ż 16 J.ÁLVAREZ-VIZOSO,M.KIRBY,ANDC.PETERSON and the last integral is only nonzero for the following combination of indices using the notation in the appendix

µ ν β γ x x x x d S C4 µνβγ C22 µνβγ µνβγ µνβγ . (4.11) Sn´1 “ p q` p q` p q` p q ż ” ı This simpliﬁes the sums using the relationship between C4,C22 and C2, and writing 1 δµν to p ´ q enforce µ ν in the last two terms of C22: ‰

n`2 n`4 k n n n δµν C2ε C2ε i 2 i 2 i i »3δµν κ δµν κ 2 1 δµν κ κ fi ... n 2 ` 2 n 2 n 4 p αµq ` p αβ q ` p ´ q αµ αν ` ` p ` qp ` q i“1 α“1 α, β α“1 ÿ — ÿ βÿ‰µ ÿ ffi — ffi – fl 2 4 k n k n ε Vn ε ε i 2 i i n`6 Vn ε δµν p q δµν κ 2 κ κ O ε “ p qn 2 ` 2 n 2 n 4 » p αβq ` αµ αν fi ` p q i“1 α,β i“1 α“1 ` p ` qp ` q ÿ ÿ ÿ ÿ – fl 2 4 n n Vn ε ε Vn ε ε 2 n`6 p q δµν p q δµν II eα, e 2 II eα, eµ , II eα, eν O ε . n 2 2 n 2 n 4 ¨ β ˛ “ ` α, β } p q} ` α“1x p q p qy ` p q ` p ` qp ` q ÿ ÿ ˝ ‚ The component expression of equations 3.15 and 3.17 identify this block matrix at order O εn`4 p q as the matrix elements of the operator tr tr III Idn 2tr III in our chosen orthonormal rp k K q ` K s basis, whose eigenvalues are then by lemma 4.4 the next-to-leading order contribution to the first n eigenvalues of C Cyl ε , and whose eigenvectors are the limit eigenvectors of C Cyl ε . p pp qq p pp qq We perform now the integration of the normal block, which truncated to leading order yields: ε ρn`3 n n ij i j n S i j α β γ δ O n`6 C Cylp ε f x f x d x ... d dρ καβκγδx x x x ε pnq Sn´1 4 r p p qqs “ B pεq p q p q ` “ 0 α,β γ, δ ` p q ż ż ż ÿ ÿ where the angular integral is only nonzero in the same cases as in equation 4.11 above, but with the indices relabeled accordingly. This again simplifies every summation by matching the combination of indices and using the relations among the constants:

n`4 n n n ij ε i j i j i j n`6 C Cyl ε »C4 κ κ C22 ¨ κ κ 2 κ κ ˛fi O ε r p pp qqs “ 4 n 4 αα αα ` αα γγ ` αβ αβ ` p q p ` q α“1 α, γ α, β — ÿ ˚αÿ‰γ αÿ‰β ‹ffi — ˚ ‹ffi – ˝ ‚fl n`4 n n C2 ε i j i j 3 II eα, eα II eα, eα II eα, eα II eγ, eγ “ 4 n 2 n 4 » p q p q` p q p q` p ` qp ` q α“1 α, γ — ÿ αÿ‰γ – n i j n`6 2 II eα, eβ II eα, eβ fi O ε , ` α, β p q p q ` p q αÿ‰β ffi ffi in which the first sum precisely completes the elements missing from the other twofl

2 n n n Vn ε ε i j i j n`6 p q II eα, eα II eγ, eγ 2 II eα, e II eα, e O ε . 4 n 2 n 4 » β β ﬁ “ ˜α“1 p q¸˜γ“1 p q¸ ` α,β p q p q ` p q p ` qp ` q ÿ ÿ ÿ – ﬂ INTEGRAL INVARIANTS FROM COVARIANCE ANALYSIS OF EMBEDDED RIEMANNIAN MANIFOLDS 17

In this last expression we clearly identify the components H H ij, and those of 2 tr III using r b s k the deﬁnition of H and equation 3.16.

We shall see below that the spherical covariance matrix has the same normal eigenvalues, to leading order, as the cylindrical case above. In [31,32] these were expressed as an average of the squares of the curvatures of curves inside the manifold M. Therefore, our previous computation provides an explicit formula for this interpretation of the normal eigenvalues.

Corollary 4.6. Let M be an n-dimensional submanifold of Euclidean space Rn`k, then the first generalized curvatures κ γ, x, nj of curves γ M passing through p with tangent vector x and p q Ă principal normal vectors any of the eigenvectors nj, j 1,...,k, of H H 2 tr III , integrate “ r b ` k s to: 4 1 2 n ε III κ γ, x, nj d x λj H H 2 tr k . (4.12) Vn ε pnq p q “ n 2 n 4 r b ` s p q żB pεq p ` qp ` q In particular: k 2 R 1 2 n 3 H 2 4 κ γ, x, nj d x } } ´ ε . (4.13) V ε pnq p q “ n 2 n 4 j“1 n B pεq ÿ p q ż p ` qp ` q 5. Spherical Covariance Analysis The difference between the cylindrical and spherical intersection domains for a graph manifold lies in the irregular projection onto the tangent space: by definition the cylinder is the extension pnq in the normal directions of the ball Bp ε TpM, so the points of the graph manifold satisfy p q Ă proj x, f x T x ε, and thus the integration region is a perfect ball. However, in } TpMpr p qs q} “ } }ď the spherical case the domain of integration is x 2 f x 2 ε2, which is nontrivial and in } } `} p q} ď general cannot be parametrized exactly. One can nevertheless apply the same procedure as done originally in [29] and [3] to find the leading order corrections to the ball domain.

Lemma 5.1. For ε 0 small enough so that M is a graph manifold over TpM, using cylindrical ą coordinates, the radial parametric equation of a point X ρx1, . . . , ρxn,f 1 ρx ,...,f k ρx T n “ r p q p qs in Dp ε M S ε , is B p q“ X p p q 2 K x 3 4 r x : ρ x1,..., xn ε p q ε O ε , (5.1) p q “ p q“ ´ 8 ` p q n´1 where x S TpM, and P Ă k n n K x 2 : II x, x 2 κi κi xαxβxγ xδ (5.2) p q “} p q} “ αβ γδ i“1 α,β γ,δ ÿ ÿ ÿ is the square of the ambient space acceleration of a geodesic curve of M with tangent x at p.

Proof. A point of the spherical boundary satisﬁes x 2 k f i x 2 ε2. Since x 2 ρ2, } } ` i“1p p qq “ } } “ and f i x 1 n κi xαxβ O x3 , it is immediate that p q“ 2 α,β αβ ` p q ř ř 2 1 k n ρ2 ρ4 κi xαxβ ε2 O ρ5 . 4 ¨ αβ ˛ ` i“1 α,β ´ “ p q ÿ ÿ ˝ ‚ 18 J.ÁLVAREZ-VIZOSO,M.KIRBY,ANDC.PETERSON

4 Defining K x 2 as the coefficient of ρ , we can solve the equation to order four to get p q 4 2 1 ρ2 1 1 K x 2ε2 ε2 K x 2ε4 O ε6 , “ K x 2 ´ ` ` p q “ ´ 4 p q ` p q p q ´ a ¯ whose square root yields the result. Note that the actual error may be of order four because this could contribute at order fie upon squaring the expression, which is the order ne- glected in the original equation. In our chosen orthonormal basis at p, we have that II x, x p q “ k n i α β M i“1 α,β καβx x ni, and this is precisely the ambient space acceleration of a geodesic of , cf.ř [25,´ř ch. 4, Cor. 10].¯ Proposition 5.2. The n-dimensional volume of the spherical component is 2 ε 2 3 V Dp ε Vn ε 1 2 tr III H O ε (5.3) p p qq “ p q ` 8 n 2 p ´} } q` p q „ p ` q  where 2 tr III H 2 H 2 2R. ´} } “} } ´ Proof. In contrast to the proof of the cylindrical domain, the radial integration introduces new angular corrections due to r x : p q rpxq n´1 V Dp ε d S ρ det g ρx dρ p p qq “ Sn´1 p q ż ż0 n a n`2 k n r x S r x i i β γ S n`3 p q d p q καβκαγ x x d O ε , “ Sn´1 n ` Sn´1 2 n 2 ` p q i“1 α,β,γ ż ż p ` q ÿ ÿ the second integral is the same to leading order as in the cylindrical case, hence n 2 2 ε K x Vn ε ε d S 1 n p q ε2 O ε3 p q tr III O εn`3 “ Sn´1 n ´ 8 ` p q ` 2 n 2 ` p q ż „  p ` q εn`2 k n n V ε ε2 i i α β γ δ S n III O n`3 Vn ε καβκγδ x x x x d p q tr ε , “ p q´ 8 Sn´1 ` 2 n 2 ` p q i“1 α,β γ,δ ÿ ÿ ÿ ż p ` q where the integral is only nonzero as in equation 4.11, so

n`2 k n n n 2 C2 ε i 2 i i i 2 Vn ε ε n`3 Vn ε »3 κ κ κ 2 κ fi p q tr III O ε “ p q´ 8 n 2 p ααq ` αα γγ ` p αβq ` 2 n 2 ` p q p ` q i“1 a“1 α,γ α,β p ` q ÿ — ÿ αÿ‰γ αÿ‰β ffi — ffi – fl 2 k n k n ε i i i 2 3 Vn ε 1 4 tr III κ κ 2 κ O ε » 8 n 2 ¨ αα γγ αβ ˛ fi “ p q ` ´ i“1 α,γ ´ i“1 α,βp q ` p q p ` q ÿ ÿ ÿ ÿ – ˝ ‚ fl 2 Now, the first set of sums in the braces is II eα, eα , II eγ, eγ H , and the second x α p q γ p qy“} } set is tr III. ř ř Remark 5.3. Notice that it is not known the dependence of the error generated by the irregular radius r x , O εn`3 in the previous proof, and whether it cancels at that order upon spherical p q p q integration, so the spherical component invariants may have error terms at lower order than the cylindrical ones. INTEGRAL INVARIANTS FROM COVARIANCE ANALYSIS OF EMBEDDED RIEMANNIAN MANIFOLDS 19

Proposition 5.4. The barycenter of the spherical component is to leading order the same as for the cylindrical component:

2 ε T 4 s Dp ε 0, H O ε . (5.4) p p qq “ r 2 n 2 s ` p q p ` q Proof. The new contributions from r x to the cylindrical computations are at least of the same p q order, O ε4 , as the overall error. p q The covariance integral invariants for the spherical domain were obtained for hypersurfaces in [3] by performing the computations in the basis of principal and normal directions. In arbitrary codimension, the diﬀerent osculating paraboloids of f i x , i 1,...,k, cannot be diagonalized p q “ simultaneously to a common basis in general. The amount of terms and simpliﬁcations needed in this general case is of much higher complexity than for hypersurfaces but, nevertheless, an analogous result for the eigenvalue decomposition obtains.

Theorem 5.5. Let λl denote taking the l-th eigenvalue of a linear operator at p, or of its r¨s associated bilinar form with respect to the metric. Then the eigenvalues of the covariance matrix of the spherical component are:

2 4 ε ε 2 5 λµ Dp ε Vn ε λµ 2 tr III H Idn 4SH O ε (5.5) p p qq “ p q n 2 ` 8 n 2 n 4 r p ´} } q ´ s` p q „  ` 4 p ` qp ` q ε 1 6 p λj Dp ε Vn ε λj tr III H H O ε (5.6) p p qq “ p q 2 n 2 n 4 r k ´ n 2 b s` p q „ p ` qp ` q `  for all µ 1,...,n, and j n 1,...,n k. Moreover, the corresponding ﬁrst n eigenvectors “ “ ` ` converge to the principal directions of the Weingarten operator at H, i.e., SH, and the last k III 1 eigenvectors to those of tr k n`2 H H . r ´ b s p Proof. From lemma 4.4 again, only the tangent and normal blocks need to be computed. Now, however, the covariance matrix is taken with respect to the barycenter, so there is an extra matrix contribution from the tensor product,

C Dp ε X X dVol X s dVol, p p qq “ b ´ b żDppεq żDppεq because the other two products cancel each other upon integration. From the proof of the barycenter formula, this integral is to leading order:

n`8 n`6 O ε n n O ε n k p q ˆ p q ˆ X s dVol V Dp ε s s 4 n`6 Vnpεqε Dppεq b “ p p qq b “ ¨O ε k n 2 H H˛ ż p q ˆ 4pn`2q b ˝ ‚ There is no diﬀerence in the normal block computations of this covariance matrix and the cylindrical case proved before, since the corrections coming from r x are O εn`6 . Thus, subtracting p q p q the barycenter contribution:

4 4 4 Vn ε ε Vn ε ε Vn ε ε 1 p q H H 2 tr III p q H H p q tr III H H . 4 n 2 n 4 p b ` k q´ 4 n 2 2 b “ 2 n 2 n 4 p k ´ n 2 b q p ` qp ` q p ` q p ` qp ` q ` 20 J.ÁLVAREZ-VIZOSO,M.KIRBY,ANDC.PETERSON

For the tangent block, the number of correction terms due to the spherical domain irregularities with respect to the cylindrical case makes a substantial contribution at O εn`4 : p q

2 rpxq k n n µν S n`1 µ ν 1 i β 3 C Dp ε d ρ x x 1 καβρ x O x dρ r p p qqs “ Sn´1 ¨ ` 2 » fi ` p q˛ 0 i“1 α“1 β“1 ż ż ÿ ÿ ÿ n`2 ˝ 2 2 – fl ‚ ε µ ν K x ε 3 δµν C2 n 2 x x p q d S O ε “ n 2 ´ p ` q Sn´1 8 ` p q ` „ ż  n`4 k n ε i i µ ν β γ S καβκαγ x x x x d ... 2 n 4 Sn´1 ` i“1 α,β,γ ` p ` q ÿ ÿ ż 2 n`4 k n n n n Vn ε ε ε i i n 4 i i n`5 δµν p q κ κ C ` κ κ C O ε , “ n 2 ` 2 n 4 » αβ αγ pµνβγq ´ 4 αβ γδ pµναβγδqfi ` p q i“1 α“1 β,γ α,β γ,δ ` p ` q ÿ ÿ ÿ ÿ ÿ – fl n´1 where we have made use of equation 5.2, and written Cpαβ... q for the integral over S of the monomial product xαxβ . . . , (notice here the indices are not exponents but coordinate components). The first summation simplifies again with equation 4.11 to yield the cylindrical tangent block, but the other set of sums comprises the 31 spherical integrals of all possible monomials of degree six:

µ ν α β γ δ S Cpµναβγδq x x x x x x d C6 µναβγδ “ Sn´1 “ p q` ż

C24 µναβγδ µναβγδ µναβγδ µναβγδ µναβγδ µναβγδ µναβγδ ` p q` p q` p q` p q` p q` p q` p q` ” µναβγδ µναβγδ µναβγδ µναβγδ µναβγδ µναβγδ µναβγδ µναβγδ p q` p q` p q` p q` p q` p q` p q` p q ı

C222 µναβγδ µναβγδ µναβγδ µναβγδ µναβγδ µναβγδ µναβγδ ` p q` p q` p q` p q` p q` p q` p q` ” µναβγδ µναβγδ µναβγδ µναβγδ µναβγδ µναβγδ µναβγδ µναβγδ p q` p q` p q` p q` p q` p q` p q` p q 

Each of these contractions are only nonzero when the connected indices are equal, and at the same time diﬀerent from the indices of the other connected groups, for instance:

n n n n i i i i καβκγβ µναβγδ δµν καακγγ . α,β γ,δ p q“ α‰µ γ‰µ ÿ ÿ ÿ γÿ‰α

Matching all the indices in this way for each of the terms just found, and taking into account C2 the relation of C6,C24 and C222 to C2 in the appendix, we take out a common factor 4pn`2q , and INTEGRAL INVARIANTS FROM COVARIANCE ANALYSIS OF EMBEDDED RIEMANNIAN MANIFOLDS 21 abbreviate the sum notation to produce all the terms of order O εn`4 : p q

2 n`4 µν δµν Vn ε ε C2 ε i 2 i i i 2 C Dp ε p q »4δµν κ 8δ✁µν κ κ 12δµν κ r p p qqs “ n 2 ` 8 n 2 n 4 p αβq ` αµ αν ` p αν q ` p ` qp ` q i α, β α α ÿ — βÿ‰µ ÿ ÿ — – i 2 i 2 i i i i i i i i i i i 15δµν κνν 3 δµν καα δ✁µν κµν κνν κνµκνν κνν κµν κνν κνµ κνµκµµ κµν κµµ ´ p q ´ # α‰µp q ` p ` ` ` ` ` ÿ

i i i i i i i 2 i 2 i 2 i 2 κ κ κ κ δµν κ κ κ κ κ κ ` µµ νµ ` µµ µν q` ¨ αα νν ` p αν q ` p αν q ` p νβq ` p νβq ` α‰µ α‰µ α‰µ β‰µ β‰µ ÿ ÿ ÿ ÿ ÿ ˝ i i i i i 2 i 2 i i κ κ δµν κ κ κ κ δ✁µν κ κ γγ νν ´ ¨ αα γγ ` p αβ q ` p αβq ˛ ´ µν γγ γ‰µ ¸+ α‰µ γ‰µ,α α‰µ β‰µ,α α‰µ β‰µ,α #γ‰µ,ν ÿ ÿ ÿ ÿ ÿ ÿ ÿ ÿ κi κi ˝κi κi κi κi κi κi κi κi ‚ κi κi ` µβ νβ ` µβ βν ` νµ γγ ` αµ να ` αµ αν ` νβ µβ β‰µ,ν β‰µ,ν γ‰µ,ν α‰µ,ν α‰µ,ν β‰µ,ν ÿ ÿ ÿ ÿ ÿ ÿ

κi κi κi κi κi κi κi κi κi κi O εn`5 ` αν µα ` αα µν ` νβ βµ ` αν αµ ` αα νµ,ﬁ ` p q α‰µ,ν α‰µ,ν β‰µ,ν α‰µ,ν α‰µ,ν ÿ ÿ ÿ ÿ ÿ . ﬂ Many of the resulting summations are the same after relabeling and using κi- κi , so they αβ “ βα can be gathered into common factors:

2 4 µν Vn ε ε Vn ε ε i 2 i i i 2 C Dp ε δµν p q p q 4δµν κ 8 κ κ 15δµν κ n 2 8 n 2 n 4 » αβ αµ αν νν r p p qqs “ ` i α,βp q ` α ´ p q ` p ` qp ` q ÿ ÿ ÿ i 2 i i i – i i i 2 3δµν καα 12 1 δµν κµν κµµ κνν 6δµν καακνν 12δµν καν ´ α‰µp q ´ p ´ q p ` q´ α‰µ ´ α‰µp q ÿ ÿ ÿ

i i i 2 i i i i δµν κ κ 2δµν κ 1 δµν 4κ κ 8 κ κ ´ αα γγ ´ p αβq ´ p ´ q µν αα ` αµ να fi α‰µ γ‰α,µ α‰µ β‰α,µ ˜ α‰µ,ν α‰µ,ν ¸ ÿ ÿ ÿ ÿ ÿ ÿ fl for which regrouping terms and completing some sums will clarify the simplifications below,

2 4 Vn ε ε Vn ε ε δ p q p q 8 κi κi 12κi κi κi 4κi κi µν n 2 8 n 2 n 4 αµ αν µν µµ νν µν αα “ ` i « α ´ p ` q´ α‰µ,ν ` p ` qp ` q ÿ ÿ ÿ

8 κi κi δ 4 κi 2 3 κi 2 21 κi 2 2κi κi 12 κi 2 αµ να µν $ αβ αα µµ µµ αα αµ ´ α‰µ,ν ` α, βp q ´ α‰µp q ` p q ´ α‰µ ´ α p q ÿ & ÿ ÿ ÿ ÿ % κi κi 2 κi 2 8 κi 2 O εn`5 . ´ αα γγ ´ p αβq ` p αµq ,ﬁ ` p q α‰µ γ‰α,µ α‰µ β‰α,µ α‰µ ÿ ÿ ÿ ÿ ÿ . ﬂ Some terms inside the curly braces complement the missing- elements of other summations: i 2 i i i 2 i 2 i 2 i i i 2 21 κµµ 2κµµ καα 12 καµ 8 καµ 15 κµµ 2κµµ καα 4 καµ , p q ´ α‰µ ´ α p q ` α‰µp q “ p q ´ α ´ α p q ÿ ÿ ÿ ÿ ÿ 22 J.ÁLVAREZ-VIZOSO,M.KIRBY,ANDC.PETERSON and

3 κi 2 κi κi 2 κi 2 κi κi 2 κi 2. ´ p ααq ´ αα γγ ´ p αβq “´ αα γγ ´ p αβq α‰µ α‰µ γ‰α,µ α‰µ β‰α,µ α,γ‰µ α,β‰µ ÿ ÿ ÿ ÿ ÿ ÿ ÿ Now, notice that this last type of double sum decomposes as follows

αγ αγ αγ αγ µµ, ´ α,γ‰µr ¨ s “´ α, γr ¨ s ` γ r ¨ s ` α r ¨ s ´r¨s ÿ ÿ αÿ“µ γÿ“µ therefore, the right hand side of the previous two equations complement each other:

2 4 δµν Vn ε ε Vn ε ε C D ε µν p q p q 8 κi κi 12κi κi κi 4κi κi p n 2 8 n 2 n 4 αµ αν µν µµ νν µν αα r p p qqs “ ` i « α ´ p ` q´ α‰µ,ν ` p ` qp ` q ÿ ÿ ÿ

i i i 2 i 2 i i i 2 n`5 8 κ κ δµν 4 κ 12 κ κ κ 2 κ O ε . αµ να $ αβ µµ αα γγ αβ ,ﬁ ´ α‰µ,ν ` α, βp q ` p q ´ α, γ ´ α, βp q ` p q ÿ & ÿ ÿ ÿ . i 2 ﬂ To simplify further, use 12 κ%µµ to complete the remaining sums and cancel terms:- p q i i i i i i i i 2 8 καµκνα 8κµν κµµ κνν 8 καµκνα 8 κµµ δµν 0, α ´ p ` q´ α‰µ,ν ` p q “ ÿ ÿ and i i i i i i 2 i i 4κµν κµµ κνν 4κµν καα 4 κµµ δµν 4κµν καα. ´ p ` q´ α‰µ,ν ` p q “´ α ÿ ÿ Finally, all these computations lead us to the simple expression:

2 4 µν δµν Vn ε ε Vn ε ε i 2 i 2 i i n`5 C Dp ε p q p q δµν 2 κ H 4κ H O ε n 2 8 n 2 n 4 » $ αβ , µν ﬁ r p p qqs “ ` i α, βp q ´ p q ´ ` p q ` p ` qp ` q ÿ & ÿ . – ﬂ where % - i i II κµν H eµ, eν , H SH eµ, eν , i “x p q y“x y ÿ and p i 2 i 2 III 2 2 καβ H 2 tr H , i p α, βp q ´ p q q“ ´} } ÿ ÿ identify the covariance tangent block to be the matrix of the Weingarten operator at the mean curvature, plus a constant, in the orthonormal basis chosen.

6. Curvature Descriptors Curvature descriptors in terms of the covariance eigenvalues were introduced in [29] for surfaces and in [3] for hypersurfaces. A limit formula for the ratio of the eigenvalues was found for curves [2] to establish a direct relationship between the local covariance analysis of a domain containing the point p and the Frenet-Serret curvature information at p, which in the case of curves completely determines the curve locally up to rigid motion [18, Th. 2.13]. The two main theorems of the present work generalize this type of result to general submanifolds by directly taking the limits of the covariance matrix eigenvalues. INTEGRAL INVARIANTS FROM COVARIANCE ANALYSIS OF EMBEDDED RIEMANNIAN MANIFOLDS 23

Corollary 6.1. Writing λµ p, ε for the tangent eigenvalues of the cylindrical covariance matrix p q C Cyl ε , they satisfy the asymptotic ratio p pp qq λµ p, ε λν p, ε n 2 lim Vn ε p q´ p q ` λµ tr KIII λν tr KIII , (6.1) εÑ0 p q λµ p, ε λν p, ε “ n 4 p r s´ r s q p q p q ` and the normal eigenvalues satisfy

n`k Vn ε n 2 2 lim p q λj p, ε ` H 2 tr III , (6.2) εÑ0 λ p, ε λ p, ε p q“ 4 n 4 } } ` µ ν j“n`1 p q p q ÿ p ` q ` ˘ for any µ,ν 1,...,n. Let λµ p, ε denote the eigenvalues in the case of the spherical domain “ p q covariance matrix, Cp Dp ε , then the corresponding limits are p p qqr λµ p, ε λν p, ε n 2 lim Vn ε p q´ p q ` λν SH λµ SH , (6.3) εÑ0 p q λµ p, ε λν p, ε “ 2 n 4 r s´ r s r p q rp q p ` q ´ ¯ and r p r p r r n`k Vn ε n 2 1 2 lim p q λj p, ε ` tr III H . (6.4) εÑ0 2 n 4 n 2 λµ p, ε λν p, ε j“n`1 p q“ ´ } } p q p q ÿ p ` q ˆ ` ˙ r Now we focus onr smoothr hypersurfaces in Rn`1. Theorems 4.5 and 5.5 provide formulas to extract curvature estimators at scale from the eigenvalues of the covariance matrices. Doing this analysis on a hypersurface furnishes descriptors at scale of the principal curvatures, and the principal and normal directions. As explained in [3], for an embedded Riemannian manifold M Rn`k, of general codimension k, it can always be projected down locally to k hypersurfaces Ă by choosing k linearly independent orthogonal directions nj of its normal space, and project the points to the linear subspace TpM nj . Approximations of the principal curvatures and ‘x y directions of these hypersurfaces are suﬃcient to build an estimator of the second fundamental form of the original manifold and, by Gauß equation 3.2, get in turn a descriptor of its Riemann curvature tensor.

Example 6.2. For a smooth hypersurface S, there is only one unit normal vector n at every n point p S, up to orientation. Choosing eµ as the orthonormal basis of the tangent space P t uµ“1 given by the principal directions at p, the components of the third fundamental form are:

2 2 III eµ, eν n, n S eµ, S eν S eµ, eν κ δµν tr III eµ, eν . (6.5) x p q y“x y“x y“ µ “ K p q The tangent trace components arep p p n III 2 2 2 tr k n, n tr S κµ H R, (6.6) x y“ p q“ µ“1 “ ´ ÿ p that coincides with the total trace, tr III n κ2 H2 R. “ µ“1 µ “ ´ The limit eigenvectors of either C Cylp ε or C Dp ε yield a local adapted orthonormal n`1 p přqq p p qq frame e1,..., en n of R that precisely singles out the tangent and normal spaces at x y‘x y every generic point. If the principal curvatures at p are of diﬀerent absolute value, this basis 24 J.ÁLVAREZ-VIZOSO,M.KIRBY,ANDC.PETERSON exactly points in the principal and normal directions. The tangent eigenvalues of the cylindrical covariance matrix C Cyl ε satisfy p pp qq λµ p, ε λν p, ε n 2 2 2 lim Vn ε p q´ p q ` κµ p κν p , (6.7) εÑ0 p q λµ p, ε λν p, ε “ n 4p p q´ p q q p q p q ` and the normal eigenvalue has

λn`1 p, ε n 2 2 lim Vn ε p q ` 3H p 2R p , (6.8) εÑ0 p qλµ p, ε λν p, ε “ 4 n 4 p q´ p q p q p q p ` q ` ˘ for any µ,ν 1,...,n. For the spherical covariance matrix C Dp ε , the limits are “ p p qq

λµ p, ε λν p, ε n 2 lim Vn ε p q´ p q ` κν p κµ p H p , (6.9) εÑ0 p q λµ p, ε λν p, ε “ 2 n 4 r p q´ p qs p q r p q rp q p ` q and r r λn`1 p, ε n 2 n 1 2 lim Vn ε p q ` ` H p R p . (6.10) εÑ0 p qλµ p, ε λν p, ε “ 2 n 4 n 2 p q´ p q pr q p q p ` q „ `  The known terms of the series expansion of the eigenvalue decomposition of the covariance r r matrices can be inverted to extract the curvature descriptors upon truncations of the series. In the spherical case, one recovers the results and descriptors already obtained in [3].

Corollary 6.3. Let us write λ p, ε λ Dp ε ,Vp ε V Dp ε for the integral invariants p q ” p p qq p q ” p p qq of a spherical domain on a hypersurface S, then the corresponding curvature descriptors at scale ε 0 and point p S, for any µ 1,...,n, are: ą P “ λ p, ε 8 n 1 n 2 V ε R ` 2 n`1 p Dp ε 2 n 2 n 4 4 p q p ` qp2 ` q p q 1 (6.11) p p qq “ p ` q p ` q n ε Vn ε ´ n ε Vn ε ´ p q ˆ p q ˙

2 ` 2 λn`1 p, ε 8 n 2 Vp ε H Dp ε 4 n 2 n 4 4 p q p `2 q 1 p q , (6.12) p p qq “ p˘q p ` q p ` q n ε Vn ε ` n ε ´ Vn ε d p q ˆ p q˙ 2 n 2 V ε n 4 ε2 λ p, ε κ D` ε p µ 1 , µ p 2 p `` q p q `2 p q (6.13) p p qq “ ε H Dp ε Vn ε ` ε n 2 ´ Vn ε ´ p p qq „ p q ˆ ` p q ˙  where the overall sign can be chosen by ﬁxing a normal orientation from

sgn en 1 Dp ε , s Dp ε . p˘q “ x ` p p qq p p qq y The eigenvectors eµ Dp ε and en 1 Dp ε are descriptors of the principal and normal direc- p p qq ` p p qq tions respectively. The errors are:

Corollary 6.4. Denote λ p, ε λ Cyl ε ,Vp ε V Cyl ε the integral invariants of a p q ” p pp qq p q ” p pp qq cylindrical domain on a hypersurface S, then the corresponding curvature descriptors at scale INTEGRAL INVARIANTS FROM COVARIANCE ANALYSIS OF EMBEDDED RIEMANNIAN MANIFOLDS 25

ε 0 and point p S, for any µ 1,...,n, are: ą P “ 2 n 2 2 n 4 λ p, ε V ε R n`1 p Cylp ε p `2 q p `2 q p q 3 1 p q (6.14) p p qq “ ε ε Vn ε ` ´ Vn ε „ p q ˆ p q ˙

2 n 2 2 n 4 λn`1 p, ε Vp ε H Cylp ε p `2 q p `2 q p q 2 1 p q , (6.15) p p qq “ p˘q ε ε Vn ε ` ´ Vn ε d „ p q ˆ p q˙ 2 2 n 2 n 4 λµ p, ε ε Vp ε κµ Cylp ε `2 `2 p q p q 1 , (6.16) p p qq “ ε ε Vn ε ´ n 2 ´ Vn ε ` „ ˆ p q ` ˙ p q  where the overall sign can be chosen by ﬁxing a normal orientation from

sgn en 1 Cyl ε , s Cyl ε . p˘q “ x ` p pp qq p pp qq y The eigenvectors eµ Cyl ε and en 1 Cyl ε are descriptors of the principal and normal di- p pp qq ` p pp qq rections respectively. The truncation errors are: H2 p H2 Cyl ε O ε2 , R p R Cyl ε O ε2 , κ2 p κ2 Cyl ε O ε2 . | p q´ p pp qq| ď p q | p q´ p pp qq| ď p q | µp q´ µp pp qq| ď p q Proof. Solving for the next-to-leading order term in the volume formula 4.4, and for the normal eigenvalue in equation 4.10, we get a system of two equations H2 R A ε , 3H2 2R B ε , ´ “ p q ´ “ p q whose solution is H2 B 2A and R B 3A, where “ ´ “ ´ 2 n 2 V ε 4 n 2 n 4 λ p, ε p O 2 n`1 O 2 A ε p `2 q p q 1 ε , B ε p ` qp4 ` q p q ε . p q“ ε Vn ε ´ ` p q p q“ ε Vn ε ` p q ˆ p q ˙ p q Finally, solving for κ2 from the tangent eigenvalue equation 4.9, and using A ε κ2 , the µ p q “ α α last formula obtains. ř The spherical descriptors can be used to determine the relative signs of the principal curvatures, and the cylindrical descriptors can be used to estimate with higher precision the absolute value of the principal curvatures.

7. Conclusions We have used the exponential map to propose a generalization of the multi-scale integral invariants determined by performing Principal Component Analysis in small regions of n-dimensional submanifolds inside a general n k -dimensional Riemannian manifold. The kernel domains p ` q studied for Riemannian manifolds embedded in Euclidean space were determined by the manifold intersection with higher-dimensional cylinders and balls in the ambient space. The volume of these regions expands with scale as the volume of the n-dimensional ball plus second order corrections proportional to the mean curvature and scalar curvature of the submanifold at the center point. We have also introduced a generalization of the classical third fundamental form to any codimension and showed how it relates to the Weingarten and Ricci operators and the Ricci equation. Then, the covariance analysis of the region point-set was found to have eigenvalues encoding curvature in terms of the third fundamental form; in particular, the ﬁrst n eigenvalues are related to those of the normal trace of the third fundamental form operator and the corresponding eigenvectors converge to its principal directions, whereas the last k eigenvalues and eigenvectors are related to the tangent trace of this tensor. In the case of the spherical domain 26 J.ÁLVAREZ-VIZOSO,M.KIRBY,ANDC.PETERSON the tangent eigenvalues and eigenvectors of the covariance matrix are related to the Weingarten operator at the mean curvature vector. For hypersurfaces, these eigenvalues provide a method to estimate the principal curvatures and principal directions, furnishing descriptors for general submanifolds via the analysis of their independent hypersurface projections. These results show how local integral invariants relate to the same geometric information traditionally characterized by diﬀerential-geometric invariants.

Appendix A. Integration of Monomials over Spheres Let x x1,...,xn T Rn, and denote unit the sphere and ball of radius ε in Rn by: “ r s P Sn´1 x Rn : x 1 , Bn ε x Rn : x ε . “ t P } }“ u p q “ t P } }ď u µ µ n´1 General spherical coordinates r, φ1,...,φn 1 are given by r x , where x : x r S . p ´ q “} } “ { P Deﬁnition A.1. For any integers p1,...,pn 0, 1, 2,... , the integrals of the monomials P t u x1 p1 xn pn over the unit sphere and the ball of radius ε are denoted by: p q ¨¨¨p q pnq 1 p1 n pn Sn´1 pnq 1 p1 n pn n Cp1...pn x x d , Dp1...pn x x d B. (A.1) “ Sn´1 p q ¨¨¨p q “ n p q ¨¨¨p q ż żB pεq where d Sn´1 is the Euclidean measure on the sphere and dnB dx1 dxn rn´1dr d Sn´1. “ ¨ ¨ ¨ “ The following formula is crucial to the computations of the present paper, cf. [12].

1 Theorem A.2. Let bi pi 1 , then the values of the integrals A.1 over spheres are “ 2 p ` q

0, if some pi is odd, Cpnq (A.2) p1...pn $ Γ b1 Γ b2 Γ bn “ 2 p q p q¨¨¨ p q , if all pi are even, &’ Γ b1 b2 bn p ` `¨¨¨` q and the integrals over balls become %’ n`p1`¨¨¨`pn pnq ε pnq Dp1...pn Cp1...pn . (A.3) “ n p1 pn ` `¨¨¨` Example A.3. We shall need the relations among integrals of monomials of even powers: Γ 3 Γ 1 n´1 n{2 1 2 S 2 2 π C2 x d 2 p 3q p nq´1 n , “ Sn´1 p q “ Γ “ Γ 1 ż p 2 ` 2 q p 2 ` q 1 2 2 2 1 C22 x x d S C2, “ Sn´1 p q p q “ n 2 ż ` 1 4 3 C4 x d S C2 3 C22, “ Sn´1 p q “ n 2 “ ż ` 1 2 2 2 3 2 1 C222 x x x d S C2, “ Sn´1 p q p q p q “ n 2 n 4 ż p ` qp ` q 1 2 2 4 3 C24 x x d S C2 3 C222, “ Sn´1 p q p q “ n 2 n 4 “ ż p ` qp ` q 1 6 15 C6 x d S C2 15 C222. “ Sn´1 p q “ n 2 n 4 “ ż p ` qp ` q INTEGRAL INVARIANTS FROM COVARIANCE ANALYSIS OF EMBEDDED RIEMANNIAN MANIFOLDS 27

The volume of a ball of radius ε, and the area of the unit sphere satisfy: n n n´1 Vn ε Vol B ε ε C2, Sn 1 Area S nC2. p q“ p p qq “ ´ “ p q“ The integral of a general combination of coordinates depends on the superindices involved, which must not be confused with exponents. For instance

µ ν β γ x x x x d S C4 µνβγ C22 µνβγ µνβγ µνβγ Sn´1 “ p q` p q` p q` p q ż ” ı is the general value of the integral of any product of 4 coordinates, that can be all equal to produce C4, or be a couple of diﬀerent pairs to result in C22. We introduce the following notation:

µνβγ δµν δβγ δ✁µβ, p q“ so that the symbol is 1 only when the connected superindices are equal and the nonconnected superindices are diﬀerent, and 0 otherwise, and where δ✁µβ : 1 δµβ is the negation of the “ p ´ q Kronecker delta, i.e., nonzero only if µ β. An example of order 6 is ‰

µναβγδ δµγ δνδ δαβ δ✁µν δ✁µα δ✁να. p q“ Acknowledgments

We would like to thank Louis Scharf for very helpful discussions. J.Á.V. would like to thank Miguel Dovale Álvarez for many useful discussions during the writing of this paper. References

1. P. Alliez, D. Cohen-Steiner, Y. Tong, and M. Desbrun, Voronoi-based variational reconstruction of unori- ented point sets, Proceedings of the Fifth Eurographics Symposium on Geometry Processing (Aire-la-Ville, Switzerland), SGP ’07, Eurographics Association, 2007, pp. 39–48. 2. J. Álvarez-Vizoso, R. Arn, M. Kirby, C. Peterson, and B. Draper, Geometry of curves in Rn, singular value decomposition, and Hankel determinants, arXiv:1511.05008v2. 3. J. Álvarez-Vizoso, M. Kirby, and C. Peterson, Manifold Curvature Descriptors from Hypersurface Integral Invariants, preprint arXiv:1804.04808 (2018). 4. J. Berkmann and T. Caelli, Computation of surface geometry and segmentation using covariance techniques, IEEE Transactions on Pattern Analysis and Machine Intelligence 16 (1994), 1114–1116. 5. D.S. Broomhead, R. Indik, A.C. Newell, and D.A. Rand, Local adaptive Galerkin bases for large-dimensional dynamical systems, Nonlinearity 4 (1991), no. 2, 159. 6. F. Cazals and M. Pouget, Estimating differential quantities using polynomial fitting of osculating jets, Symp. Geometry Processing. Eurographics (2003), 177–178. 7. F. Cazals and M. pp. 351-360 Pouget, Molecular shape analysis based upon the Morse-Smale complex and the Connolly function, Proc. Symp. Comp. Geometry (2003). 8. I. Chavel, Riemannian Geometry, 2nd ed., Cambridge University Press, 2006. 9. U. Clarenz, M. Griebel, M. Rumpf, M.A. Schweitzer, and A. Telea, Feature sensitive multiscale editing on surfaces, The Visual Computer 20 (2004), no. 5, 329–343. 10. U. Clarenz, M. Rumpf, and A. Telea, Robust feature detection and local classification for surfaces based on moment analysis, IEEE Transactions on Visualization and Computer Graphics 10 (2003), 516–524. 11. M. Connolly, Measurement of protein surface shape by solid angles, J. Mol. Graphics 4 (1986), no. 1, 3–6. 12. G.B. Folland, How to integrate a polynomial over a sphere, The American Mathematical Monthly 108 (2001), no. 5, 446–448. 13. A. Gray, E. Abbena, and S. Salamon, Modern Differential Geometry of Curves and Surfaces with Mathematica, 3rd ed., Chapman and Hall/CRC, 2006. 28 J.ÁLVAREZ-VIZOSO,M.KIRBY,ANDC.PETERSON

14. A. Gray and L. Vanhecke, Riemannian geometry as determined by the volumes of small geodesic balls, Acta Math. 142 (1979), 157–198. 15. H. Hoppe, T. DeRose, T. Duchamp, J. McDonald, and W. Stuetzle, Surface reconstruction from unorganized points, SIGGRAPH Comput. Graph. 26 (1992), no. 2, 71–78. 16. D. Hulin and M. Troyanov, Mean curvature and asymptotic volume of small balls, The American Mathematical Monthly 110 (2003). 17. S. Kobayashi and K. Nomizu, Foundations of differential geometry, John Wiley & Sons, Inc., 1969. 18. W. Kühnel, Differential Geometry: curves-surfaces-manifolds, vol. 16, American Mathematical Soc., 2006. 19. Y.-K. Lai, S.-M. Hu, and T. Fang, Robust principal curvatures using feature adapted integral invariants, Proceedings - SPM 2009: SIAM/ACM Joint Conference on Geometric and Physical Modeling (2009), 325– 330. 20. H.L. Liu, U. Simon, L. Verstraelen, and C.P. Wang, The third fundamental form metric for hypersurfaces in nonflat space forms, J. geom. 65 (1999), 130–142. 21. S. Manay, D. Cremers, B.-W. Hong, and A. J. Yezzi, Integral invariants for shape matching, Pattern Analysis and Machine Intelligence, IEEE Transactions on 28 (2006), no. 10, 1602–1618. 22. S. Manay, B.-W. Hong, A. J. Yezzi, and S. Soatto, Integral invariant signatures, Computer Vision - ECCV 2004 (Berlin, Heidelberg) (Tomás Pajdla and Jiří Matas, eds.), Springer Berlin Heidelberg, 2004, pp. 87–99. 23. Q. Mérigot, Geometric structure detection in point clouds, PhD. Thesis, Université Nice Sophia Antipolis, 2009. 24. Q. Mérigot, M. Ovsjanikov, and L. Guibas, Voronoi-based curvature and feature estimation from point clouds, Visualization and Computer Graphics, IEEE Transactions on 17 (2011), 743 – 756. 25. B. O’Neill, Semi-Riemannian Geometry, Academic Press, 1983. 26. X. Pennec, Probabilities and statistics on Riemannian manifolds: basic tools for geometric measurements, Int. Workshop on Nonlinear Signal and Image Processing (NSIP’99), 1999. 27. , Intrinsic statistics on Riemannian manifolds: basic tools for geometric measurements, J. Math. Imag- ing Vis. 25 (2006), 127–154. 28. H. Pottmann, J. Wallner, Q.-X. Huang, and Y.-L. Yang, Integral invariants for robust geometry processing, CAGD 26 1 (2008). 29. H. Pottmann, J. Wallner, Y.-L. Yang, Y.-K. Lai, and S.-M. Hu, Principal curvatures from the integral invariant viewpoint, Computer Aided Geometric Design 24 (2007), 428–442. 30. F. Rellich, Perturbation Theory of Eigenvalue Problems, Gordon and Breach, 1969. 31. F.J. Solis, Geometric aspects of local adaptive Galerkin bases, PhD. Thesis, University of Arizona, 1993. 32. , Geometry of local adaptive Galerkin bases, Applied Mathematics & Optimization (2000), no. 41, 331–342. 33. M. Spivak, A Comprehensive Introduction to Differential Geometry, 3rd ed., Publish or Perish, Inc., 1999. 34. V.A. Toponogov, Differential Geometry of Curves and Surfaces. A Concise Guide, Birkhäuser Boston, 2006. 35. Y.-L. Yang, Y.-K.-Kun Lai, S.-M. Hu, and H. Pottmann, Robust principal curvatures on multiple scales, Proc. Eurographics Symposium on Geometry Processing (2006), 223–226.

Department of Mathematics, Colorado State University, Fort Collins, CO, USA E-mail address: [email protected], [email protected], [email protected]