8

The basis vectors on a manifold are defined as partial derivative operators of any function that can be locally mapped around a point to $\mathbb R^n.$ The Christoffel symbols come about when a vector field on a such a manifold is differentiated, and the product rule calls for differentiating not just the components, but also these basis vectors, since they change from point to point.

Is there any sense, therefore, in which the Christoffel symbols can be understood as a second derivative?

JAP
  • 553
  • Your best bet is to focus on concrete examples rather than in abstract. Consider the polar coordinates in $\mathbb R^3$. The Christoffel symbols are computer, for example, here. Do you think they look like a second derivative? I would say no, but the question is rather subjective. – Giuseppe Negro Mar 26 '22 at 21:59
  • @GiuseppeNegro Yes, that is exactly my issue - the two topics, i.e. covariant basis and Christoffel symbols - are explained typically in two different chapters or classes, and when it comes to Christoffel symbols a less abstract example like the one you mention is used. So it makes the calculations possible, but abandons a bit the abstract idea. – JAP Mar 26 '22 at 22:03

4 Answers4

10

No, they are not second derivatives of any kind. They're the "expansion coefficients" for differentiating one basis vector field along another. The expression $\nabla_XY$ means roughly to "differentiate the vector field $Y$ along the direction $X$". Now, one can consider any local frame $\{\xi_1,\dots, \xi_n\}$ of the tangent bundle, and in this regard, consider $\nabla_{\xi_i}(\xi_j)$. This is yet again another local vector field, so can be written as a linear combination $\nabla_{\xi_i}(\xi_j)=\Gamma^k_{ij}\xi_k$ for some $\Gamma$'s. So $\Gamma^{k}_{ij}$ tells us the $\xi_k$ component of the change in $\xi_j$ along $\xi_i$; but this 'change' arises due to a 'first derivative'. Indeed, the fact that $\nabla$ satisfies the product rule is very much characteristic of first-derivatives.

Do not be confused by the fact that the coordinate-induced basis vector fields can be written as $\frac{\partial}{\partial x^i}$. This merely has to do with one of the many equivalent ways one defines the tangent space to a manifold.


Edit In Response to Comments:

Ok my first sentence was too harsh (see Deane's excellent and concise answer for one interpretation, and below for a different special case interpretation), but still, I maintain that as far as it pertains to derivatives of vector field/sections of bundles (which is mainly why covariant derivatives are introduced), you shouldn't think of second derivatives.

Suppose $M$ is a smooth $m$-dimensional embedded submanifold of $\Bbb{R}^n$, let $\phi:A\to \phi[A]\subset M\subset\Bbb{R}^n$ be a local parametrization of $M$, where $A\subset\Bbb{R}^m$ is an open set (so $\phi$ is smooth, a homeomorphism onto its image, and has injective derivative at every point, all in the usual multivariable calculus sense). Now, for each $p\in \phi[A]\subset M$, and each $i\in\{1,\dots, m\}$ let us define \begin{align} \xi_i(p):=(\partial_i\phi)_{\phi^{-1}(p)}\equiv \frac{\partial \phi}{\partial u^i}\bigg|_{\phi^{-1}(p)}\in\Bbb{R}^n. \end{align} This is a vector tangent to $M$ at the point $p$; see this answer just in case you're not comfortable with relating the abstract definitions with concrete stuff in the special $\Bbb{R}^n$ case. Let $\langle\cdot,\cdot\rangle$ denote the standard inner product in $\Bbb{R}^n$, and define for each $a,b\in\{1,\dots, m\}$, \begin{align} g_{ab}(p):=\langle \xi_a(p),\xi_b(p)\rangle, \end{align} and let $[g^{ab}]$ denote the inverse matrix to $[g_{ab}]$. This $g$ is actually none other than the metric tensor on $M$ induced via pull-back of the standard RIemannian metric on $\Bbb{R}^n$. Now, by using the Levi-Civita connection on $M$, we can easily work out that the Christoffel symbols are \begin{align} \Gamma^k_{ij}(p)&=\frac{1}{2}g^{ks}(p)\left(\partial_i(g_{js}\circ \phi)_{\phi^{-1}(p)}+ \partial_j(g_{si}\circ \phi)_{\phi^{-1}(p)}- \partial_s(g_{ij}\circ \phi)_{\phi^{-1}(p)}\right)\\ &=g^{ks}(p)\cdot \langle(\partial_i\partial_j\phi)_{\phi^{-1}(p)}, (\partial_i\phi)_{\phi^{-1}(p)}\rangle\\ &\equiv g^{ks}(p)\cdot \left\langle\frac{\partial^2\phi}{\partial u^i\partial u^j}\bigg|_{\phi^{-1}(p)},\frac{\partial\phi}{\partial u^s}\bigg|_{\phi^{-1}(p)}\right\rangle \end{align} (the last equality is just a change to more classical Leibniz notation). So, ok the Christoffel symbols corresponding to the Levi-Civita connection for an embedded submanifold, and corresponding to a local parametrization, does depend on second derivatives of the parametrization.

Having said this, here's why I'd strongly urge you to not think that just because of this, the Christoffel symbols capture second derivatives

  • This relies on having $M$ be embedded in some $\Bbb{R}^n$, and we exploited heavily that this is an inner-product vector space (we needed the inner product of $\Bbb{R}^n$ and we needed the vector space structure to differentiate $\Bbb{R}^n$-valued maps, namely the parametrization $\phi$). Yes, there are several embedding theorems (even isometric ones), so in a sense, this is no loss of generality. However typically one needs much higher dimensional ambient space (and the larger the ambient space, the more seemingly superfluous information there is, which automatically makes it difficult to recognize intrinsic properties). We're also talking only about the Levi-Civita connection.

  • Elaborating on the previous comment, a general manifold is not given to us already embedded in $\Bbb{R}^n$, and as such it doesn't have "position vectors", or as I used above, $\Bbb{R}^n$-valued parametrizations $\phi$. These parametrizations aren't the fundamental quantities. i.e in basic math and physics courses, we learn about distances first, then we learn about position/displacement, and then we learn about speed and velocity and then about acceleration. However, in differential geometry, the fundamental object is the manifold, and its various tangent spaces (i.e velocities are primary object). "position vectors" are a meaningless concept unless you're already embedded. "speed" is an extra concept which can only be made sense once you have a metric tensor, and finally distances/lengths are obtained by integrating the speed of curves. So, our logical development is very much reversed (this is more general, and more useful). So, as much as possible, we should try to avoid these parametrizations when formulating the theory (for concrete calculations, we of course need the charts/parametrizations, but when developing the theory, they for many purposes obscure the picture).

  • Just because second derivatives of $\phi$ appear in this formula, it doesn't mean the $\Gamma$'s are "second order effects". Given any smooth function $f$, say $\Bbb{R}\to \Bbb{R}$, I can find a function $F_1$ such that $F_1'=f$. Likewise, I can find another function $F_2$ such that $F_2''=f$. FOr any positive integer, I can find a function which if differentiated that many times, yields $f$. So does that mean $f$ should be thought of as a "$0^{th}$ derivative" (since $f^{(0)}=f$)? Or should I think of $f$ as a "first derivative" (since $F_1'=f$)? Should I think of it as a "second derivative" (since $F_2''=f$)? Same thing with the $\Gamma$'s. (The $\nabla$ on the other hand satisfies the product rule, which is very much characteristic of first derivatives, and the $\Gamma$'s appear simply due to linear algebra, as the "expansion coefficients" relative to a basis).

  • The connections we're talking about here are more properly called connections on $TM$ (and the induced connections on the various tensor bundles $T^r_s(TM)$). A slightly more general situation is that you have a "vector bundle" $(E,\pi, M)$. VERY roughly speaking, this means we have two smooth manifolds $E,M$ and a surjective mapping $\pi:E\to M$. THe idea is that this is a "smoothly varying family of vector spaces". So for each $p\in M$, we have a vector space $E_p:=\pi^{-1}(\{p\})$. These vector spaces make up $E$, i.e $E=\bigcup_{p\in M}E_p$. Now, we can consider an "$E$-vector field", or more technically a smooth section of the vector bundle. This means we consider a map $\xi:M\to E$ such that $\pi\circ \xi=\text{id}_M$. Unwinding the definitions, this just means at each point $p\in M$ we have a vector $\xi(p)\in E_p$ in this particular vector space. Now, we can ask about what it means to take directional derivatives of such a mapping where the target space consists of not just a single vector space like $\Bbb{R}^n$, but many of them. Note that these vector spaces $E_p$ need not be the tangent spaces $T_pM$ (so the $\xi$'s need not have anything to do with $\frac{\partial}{\partial x^i}$). They could be something else entirely. There's also no $\Bbb{R}^n$-valued parametrizations lying around, so no second derivatives of parametrizations to be even considered. Yet, the question of such differentiation seems quite fundamental; we can define $\nabla$ in a straighforward way (Koszul's definition for a covariant derivative, or we can use the much more geometric definition of Ehresmann to eventually arrive at $\nabla$). And even in this general circumstance, the $\Gamma$'s can be viewed simply as a linear-algebraic consequence: they're just certain coefficients in the expansion of relative choices of bases.

The main point I wish to drive home is that sometimes, the way certain quantities look like when we generalize the theory tells us how we should interpret it. We have already encountered so many bumps in the road when trying to impart this second-derivative interpretation to the $\Gamma$'s. So, that should hopefully be strong enough motivation that it isn't a fruitful way to think about it.

peek-a-boo
  • 55,725
  • 2
  • 45
  • 89
  • Very helpful. Thanks. The last paragraph is the the problem - I am trying to think through your first example but with the local basis expressed as ${\frac \partial{\partial x^i}}.$ – JAP Mar 26 '22 at 22:17
  • This is a different question entirely, but now I'm a bit confused myself. In Frederic Schuller's lecture, he introduces the covariant derivative as an axiomatic construction not depending on anything else but in Pavel Grinfeld's book, he does it by just showing vwhat happens when we take derivative along gridlines.. how are these two related? – tryst with freedom Mar 26 '22 at 22:39
  • @Buraian Note that a manifold doesn't have "position vectors" so trying to use that (intuitive in $\Bbb{R}^3$, but highly misleading) notion isn't really helpful, and your confusion perfectly illustrates why it is so misleading. – peek-a-boo Mar 26 '22 at 22:39
  • In the book, he does find that it is actually generalizes to higher dimension somehow. I am not sure where he said that.. let me find it actually – tryst with freedom Mar 26 '22 at 22:42
  • 2
    the most intuitive notion I can give you is to consider an embedded $m$-dimensional submanifold $M\subset\Bbb{R}^n$. Suppose you have vector fields $X,Y$ defined on an open neighborhood $U$ of $M$ in $\Bbb{R}^n$; so these can be viewed as usual smooth functions $X,Y:U\to\Bbb{R}^n$. Then, you can define $\nabla_XY(p)$ by taking the usual directional derivative $D_XY(p):=\frac{d}{dt}\bigg|_{t=0}Y(p+tX(p))$, and orthogonally projecting it to $T_pM$. See Lee's book for a clearer discussion – peek-a-boo Mar 26 '22 at 22:45
  • I just went through the material again, and I think I sort of got the thing going on here. We are actually saying a lot when we give the exact parameterization of the position vector to give points on the surface. The connection coefficients will come out naturally as result of derivative. When doing the axiomatic wya, we are saying the least possible amount of stop. We don't talk about coordinates, position vector etc, we just hand out the christoffels in the beginning itself and say "use this to take the derivatives". – tryst with freedom Mar 26 '22 at 23:02
  • Do you agree with my conclusion? – tryst with freedom Mar 26 '22 at 23:05
  • 1
    @Buraian right, and in particular if you have a metric tensor field $g$ on the manifold, then there's a "standard" way of "handing out" these Christoffel symbols. This gives rise to the "Levi-Civita connection" (the unique metric-compatible, torsion-free connection). – peek-a-boo Mar 26 '22 at 23:08
  • Very coarsely the Christoffel symbols are the components of the differentiation of the vector field formed by the a set of basis vectors along those same basis vectors? – JAP Mar 27 '22 at 01:04
  • @Patti "are the components...", yes and very slightly more precisely, they're the components with respect to that same set of basis vectors. – peek-a-boo Mar 27 '22 at 01:40
  • One final question - do you see any contradiction between you answer and the less radical sounding ("No, they are not second derivatives of any kind.) edit to Buraians answer? That sounds great, but do I have to choose, or they are the same idea? – JAP Mar 27 '22 at 18:20
  • 1
    @Patti ok perhaps I've been to harsh in my first sentence "not second derivatives of any kind". I should have really said that in special situations, one can interpret them as comprising of second derivatives (thereby answering your question in the affirmative), and then I should have followed that up (as I have done in my edit now) with why thinking in this approach isn't really fruitful. – peek-a-boo Mar 28 '22 at 00:39
  • @Patti lol even in my previous comment I should have started with "ok I've definitely been too harsh..." because as has been pointed out in Deane's excellent answer, you can definitely say $\Gamma^k_{ij}=-\nabla^2_{ij}x^k$. My point was mainly that the definition of Christoffel symbols is introduced mainly from a linear algebra perspective: certain expansion coefficients, and that it isn't fruitful to force a second-derivative interpretation starting from this definition. So yes my apologies, I should have phrased my topic sentence more clearly. – peek-a-boo Mar 28 '22 at 01:04
4

The Christoffel symbols arise naturally when you want to differentiate a scalar function $f$ twice and want the resulting Hessian to be a $2$-tensor. When you work out the details, you discover that with respect to local coordinates, the Hessian of $f$ is given by $$ \nabla^2_{ij}f = \partial^2_{ij}f - \Gamma^k_{ij}\partial_kf. $$ In particular, if you set $f(x) = x^k$, you get $$ -\nabla^2_{ij}x^k = \Gamma^k_{ij}. $$

Also, notice that the Hessian of $f$ will always be a symmetric $2$-tensor if $\Gamma^k_{ij} = \Gamma^k_{ji}$, which is equivalent to saying that the connection is torsion-free. That's one of the reasons you want the Levi-Civita connection to be torsion-free.

Deane
  • 7,582
  • Good answer, but isn't this essentially the exact same thing I wrote o_o – tryst with freedom Mar 26 '22 at 22:59
  • @Buraian, could be, but I don't see any mention of the Hessian of a scalar function in your answer. Any chance you want to elaborate? – Deane Mar 26 '22 at 23:01
  • I had mentioned the idea of taking a derivative of vector in a direction. When we take a derivative of scalar in a direction, we take a vector and dot it in direction we want. When we take derivative of vector in a direction, we take a matrix and multiply it with the vector we want. In my answer, I've mentioned there is a matrix somehow but not the details, in yours, the only difference is you've said it is a hessian only – tryst with freedom Mar 26 '22 at 23:03
  • @Buraian, the Christoffel symbols arise from taking only one derivative of a vector field. But for a scalar function, they arise only after you have differentiated twice, roughly speaking because one derivative of a function is a vector fields (actually, a $1$-form). So the only one way I know to make a Christoffel symbol a second derivative of something is to make that something a scalar function. On the other hand, in your answer I see only derivatives of vector fields. – Deane Mar 27 '22 at 15:57
  • 1
    just a small issue (which doesn't really affect the rest of your answer) but I think there should be a minus sign $\nabla^2_{ij}f=\partial^2_{ij}f-\Gamma^k_{ij}\partial_kf$ (because $\nabla(dx^k)=-\Gamma^k_{ij},dx^i\otimes dx^j$) – peek-a-boo Mar 27 '22 at 19:25
  • @peek-a-boo, yes. You are correct. I will fix that. Thanks! – Deane Mar 27 '22 at 19:53
3

Suppose we have a vector field depending on what point we are in space $v= v^i e_i$, the Christtofel appear naturally when we take the derivative of vector with the parameterization of the coordinate. They're sort of like a 'spatial derivative' of the basis vectors.

What do I mean by that? Suppose you move from one point in the surface to another, then the vector field on the surface can change due two reasons:

  1. Change in vector component
  2. Change in basis

The change in basis part is what the Christoffel's try to capture.

Suppose I move from point to $A$ to point $B$, along a coordinate gridline:

$$e_i(B) - e_i(A) \approx \Gamma_{ik}^l e_l d \lambda$$

Note that I am tkaing that $B$ is the point at parameter value $\lambda+ d \lambda$ and $A$ is the point at parameter value $\lambda$ on the gridlines. For a particular value of $i$, we can think of RHS as a matrix multiplying on the unit vectors to say what the change in that basis is as we move along the gridline.


Also, I guess they are sort of second derivatives? In Pavel Grinfeld's Tensor analysis book, he talks about how given a parametrization of a surface we automatically know the tangent vector of the surface by differentiation of the position vector.

For example, let's consider a sphere:

$$ R(a,\theta,\phi)= a \sin \theta \cos \phi \hat{i} + a \sin \phi sin \theta \hat{j}+ a \cos \theta \hat{k}$$

If you differentiate the above vector w.r.t. the coordinates, we can get two tangents vector at a point i.e: $e_{\theta} =\frac{\partial R}{\partial \theta}$ and $e_{\phi} = \frac{\partial R}{\partial \phi}$. The Christoffel would then be related to the second derivative of position vector (going by previous eq which I introduced the symbols with).

$$ e_r= \frac{\partial R}{\partial r} = ( \sin \theta \cos \phi, \sin \phi \sin \theta, \cos \theta)$$

$$ e_{\phi} = a \sin \theta( - \sin \phi, \cos \phi)$$

$$ \frac{\partial e_r}{\partial \phi} = \sin \theta (-\sin \phi , \cos \phi)= \frac{e_{\phi} }{a}$$

This above equation is telling us about the Christoffel $\Gamma_{r \phi}^i$ for $ \{r , \theta ,\phi \}$ it tells us that $ r$ and $\phi$ entries are zero and $\theta $ entry is $\frac{1}{a}$

  • On the point of taking derivative of vector in directions : https://math.stackexchange.com/a/4359170/688539 – tryst with freedom Mar 26 '22 at 22:12
  • If you've gone through this answer and want to see the connection of with the axiomatic deftn of the covariant derivative pls see the exchange under peek-a-boo's answer – tryst with freedom Mar 26 '22 at 23:09
  • The edit touched directly on the question. Ty. Is it somehow in contradiction to the other answer by peekaboo? How to reconcile both answers? – JAP Mar 27 '22 at 15:22
  • For the sake of the course by Pavel, what my answer says will be enough information to complete the course till the end. It's like the naive notion of calculus which we get in HS physics. For the sake of rigorous differential Geometry, we take less as given and we need to be more careful. @Patti – tryst with freedom Mar 27 '22 at 15:25
  • Hi @JAP I added a concrete calculation which may help you see what I mean and also for future peeps – tryst with freedom May 26 '22 at 12:34
1

I guess the main relationship with the second derivative is the geodesic equation $\ddot{x}(t) + \Gamma(\dot{x}(t), \dot{x}(t))= 0$, or $\nabla_{\dot{x(t)}}\dot{x}(t) = 0$. I think the most intuitive explanation for the role of $\Gamma$ is for any curve $x(t)$, the acceleration $\ddot{x}$ is not tangent, and the term $\Gamma(\dot{x}(t), \dot{x}(t))$ is required to make it tangent. It is easy to see this in the example of the sphere. Consider a particle moving on the unit sphere, $x^Tx = 1$. Then $x(t)^T\dot{x}(t) = 0$, hence $$x(t)^T\ddot{x}(t) + \dot{x}(t)^T\dot{x}(t) = 0. $$ Thus, $x(t)^T\ddot{x}(t)\neq 0$ generally, or $\ddot{x}(t)$ is not tangent, as well-known from highschool physics. But the above equation implies $$x(t)^T\{\ddot{x}(t) + x(t)\dot{x}(t)^T\dot{x}(t)\} = 0 $$ or $\ddot{x}(t) + x(t)\dot{x}(t)^T\dot{x}(t)$ is always tangent. If we work with a vector field $b(t)$ along the curve $x(t)$, ie $x(t)^Tb(t) =0$ but not necessarily $b(t) = \dot{x}^T(t)$, we have $\dot{x}(t)^Tb(t) + x(t)^T\dot{b}(t) =0$, or $$x(t)^T\{\dot{b}(t) + x(t)\dot{x}(t)^Tb(t) \}=0 $$ hence $\dot{b}(t) + x(t)\dot{x}(t)^Tb(t)$ is always tangent. The term $$\Gamma(\dot{x}(t), b(t)) = x(t)\dot{x}(t)^T b(t)$$ gives us a connection, it is a Christoffel function. This $\Gamma$ is not the only adjustment to make $\dot{b}$ tangent, once we have $\Gamma$, we can add another bilinear term tangent to the sphere. Among all the possible adjustments $\Gamma$ to make $\dot{b} + \Gamma(x, b)$ tangent, there is a unique one compatible with the metric, the Levi-Civita connection. For the embedded sphere, the simplest adjustment $x(t)\dot{x}(t)^T b(t)$ is actually the Levi-Civita connection.

  • Thank you for your answer. Here are a couple of follow-up questions that shows my lack of foundation on the topic: 1. When you say "the acceleration $\ddot{x}$ is not tangent" you imply that it's understood the particle is traveling along a geodesic, and in such trajectory the only acceleration is, by definition, perpendicular to the surface. Is this correct? – JAP Jul 20 '23 at 22:55
  • You are very welcome! Actually, the equation $x(t)^T\ddot{x}(t) + \dot{x}^T\dot{x} = 0$ is valid for any curve on the sphere, not necessarily geodesic. Since $\dot{x}^T\dot{x}\neq 0$ if there is any movement, $x(t)^T\ddot{x}(t)\neq 0$, so $\ddot{x}(t)$ is not tangent. For any curve, $\ddot{x}(t) + x\dot{x}^T(t)\dot{x}(t) = \nabla_{\dot{x}}\dot{x}(t)$ is tangent, but need not be zero. For geodesic, this is zero. – tensor_and_manifold Jul 20 '23 at 23:14
  • Can you clarify what is now a 'barrier-of-entry' in the mathematical expressions in "...$x^Tx = 1$. Then $x(t)^T\dot{x}(t) = 0$" I presume $x$ is a positional vector of length $1$ on the sphere. As for the second expression, it means that the velocity is perpendicular to the positional vector? Then the product rule for the derivative, etc. And you get to $\dot b(t).$ Now this $\dot b(t)$ is not tangent to the curve, $b(t)$ is. And the difference between $\dot b(t)$ and the tangent is the Christoffel symbol?
  • – JAP Jul 20 '23 at 23:14
  • 1
    $\newcommand{\R}{\mathbb{R}}$ Yes, $x$ is a (column) vector in $\R^n$, and on the unit sphere $|x|^2=1$ or $x(t)^Tx(t) = 1$ is just the statement that the radius is $1$. Differentiate, we get $2 x(t)^T\dot{x}(t)=0$ (vector calculus), or $x(t)^T\dot{x}(t)=0$, which is the statement the tangent vector is perpendicular to the the radius. For your last question, I would like to clarify, adding the Christoffel term to $\dot{b}(t)$ does not give you back $b(t)$, but "the tangent" is a different tangent vector - the covariant derivative along the curve, typically denoted by $\nabla_{\dot{x}}b(t)$. – tensor_and_manifold Jul 20 '23 at 23:33