A geometric interpretation of the Levi-Civita connection? Let $M$ be a Riemannian manifold. There exists a unique torsion-free connection in the (co)tangent bundle of $M$ such that the metric of $M$ is covariantly constant. This connection is called the Levi-Civita connection and its existence and uniqueness are usually proven by a direct calculation in coordinates. See e.g. Milnor, Morse theory, chapter 2, \S 8. This is short and easy but not very illuminating.
According to C. Ehresmann, a connection in a fiber bundle $p:E\to B$ (where $E$ and $B$ are smooth manifolds and $p$ is a smooth fibration) is just a complementary subbundle of the vertical bundle $\ker dp$ in $T^*E$. If $G$ is the structure group of the bundle and $P\to B$ is the corresponding $G$-principal bundle, then to give a connection whose holonomy takes values in $G$ is the same as to give a $G$-equivariant connection on $P$.
If $p:E\to B$ is a rank $r$ vector bundle with a metric, then one can assume that the structure group is $O(r)$; the corresponding principal bundle $P\to B$ will in fact be the bundle of all orthogonal $r$-frames in $E$. One can then construct an $O(r)$-equivariant connection by taking any metric on $P$, averaging so as to get an $O(r)$-equivariant metric and then taking the orthogonal complement of the vertical bundle.
Notice that in general one can have several $O(r)$-equivariant connections: take $P$ to be the total space constant $U(1)$-bundle on the circle; $P$ is a 2-torus and every rational foliation of $P$ that is non-constant in the "circle" direction gives a $U(1)$-equivariant connection. (All these connections are gauge equivalent but different.)
So I would like to ask: given a Riemannian manifold $M$, is there a way to interpret the Levi-Civita connection as a subbundle of the frame bundle of the tangent bundle of $M$ so that its existence and uniqueness become clear without any calculations in coordinates?
 A: This avoids use of Christoffel symbols and index argument by re-characterizing the torsion free property.  The following is one definition for a torsion-free connection.  Let $\tau:TM \to M$ be the tangent bundle projection.  A torsion free connection is a splitting of $T(TM) = H(TM) \oplus V(TM)$ where $V(TM) = {\rm kernel}(T \tau) \equiv TM \oplus TM$ (see Kolar,Michor,Slovak (1993)).  The covariant derivative associated to this splitting is given by $\frac{Dv}{Dt} = {\rm pr_2} \left( {\rm ver}( \frac{dv}{dt} ) \right)$ where ${\rm ver}:TM \to VM$ is the projection onto $VE \equiv TM \oplus TM$ and $\rm pr_2$ is the projection onto the second factor of $TM \oplus TM$.  The torsion-free property manifests from the fact that given any surface embedding $m: (s,t) \mapsto m(s,t) \in M$ we observe $\frac{D}{Dt} ( \partial_s m) = \frac{D}{Ds}( \partial_t m)$ by the equality of mixed partials.  Consider the map $h^{\uparrow}: TM \to T^{(2)}M$ given by $h^{\uparrow}(v_0) = \left.\frac{d}{dt} \right|_{t=0}(v(t))$ where $v(t)$ is the geodesic with initial condition $v_0 \in TM$.  Then (I think) the Levi-Cevita is given by $H(TM) = h^{\uparrow}(TM)$.  Clearly, this is a horizontal space, and thus a well-defined connection.
A: The Levi-Civita connection is locally described by the Christoffel symbols.   How does one obtain these in a natural fashion: write the Euler-Lagrange  for the length functional. The  extremals of this functional are the geodesics and  once you write the Euler-Lagrange equations you obtain the   Christoffel  symbols.  For details see  Example  5.1.8  from my book.
A: I think that a good way to understand the Levi-Civita connection is to say that is is the Ehresmann connection in $TTM$ obtained from the linearization of the geodesic flow by a natural geometric construction.
I described this construction in my answer to this MO question, but I'll do so again with some improvements.
Dynamic construction.
Let $c(t)$ be an orbit of the geodesic flow in $TM$, consider the vertical subspaces $V(t)$ in $TTM$ along $c(t)$ and bring them back to the tangent space of the cotangent bundle over the point $c(0)$ by using the differential of the flow. You get a family of (Lagrangian) subspaces $l(t) := D\phi_{-t}(V(t))$ in the symplectic vector space $T_{c(0)}TM$. 
Now forget you ever had a geodesic flow: all that you need is the curve of subspaces. A bit of differential projective geometry---described below---shows that you also get a second curve $h(t)$ of (Lagrangian subspaces) in $T_{c(0)}(T^*M)$ that is transversal to $l(t)$. The subspace $h(0)$ is the horizontal subspace of the connection and  $T_{c(0)}(T^*M) = l(0) \oplus h(0)$ is the decomposition into vertical and horizontal subspaces.
Projective construction.
Now I'll describe as succintly as possible the projective-geometric construction
that underlies both the Levi-Civita connection and the Schwartzian derivative.
For the detais of what follows see this paper What's new in the description here is that I explicitly use the Springer resolution (Duran and I used implicitly in the paper).
First we need two remarks on the geometry of the Grassmannian $G_n(\mathbb{R}^{2n})$ of $n$-dimensional subspaces in $\mathbb{R}^{2n}$ 
1. The tangent  space  of $G_n(\mathbb{R}^{2n})$ at a subspace $\ell$
is canonically identified with the space of linear maps from $\ell$ to $\mathbb{R}^{2n}/\ell$ or, equivalently, with the space $(\mathbb{R}^{2n}/\ell) \otimes \ell^*$. Since $\mathbb{R}^{2n}/\ell$ and $\ell$ have the same dimension, we may distinguish a class of differentiable curves $\gamma$ on the Grassmannian  by requiring that at each instant $t$ their velocities are invertible linear maps from $\gamma(t)$ to  $\mathbb{R}^{2n}/\gamma(t)$. These curves are called fanning or regular.
Using that the cotangent space of $G_n(\mathbb{R}^{2n})$ at a subspace $\ell$
is canonically isomorphic to $\ell \otimes (\mathbb{R}^{2n}/\ell)^*$, we can lift
every fanning curve $\gamma(t)$ to  a  curve on the cotangent bundle of the Grassmannian by $t \mapsto (\dot{\gamma}(t))^{-1}$. 
2. Consider the action of the linear group $GL(2n;\mathbb{R})$ on the Grassmannian $G_n(\mathbb{R}^{2n})$ and lift it to an action on its cotangent bundle. The moment map of this action takes values on the set of nilpotent matrices. 
Now consider a fanning curve $\gamma(t)$ on the Grassmannian $G_n(\mathbb{R}^{2n})$ and lift it to  the curve $(\dot{\gamma}(t))^{-1}$ on its cotangent bundle. Use the moment map to obtain a curve $F(t)$ of nilpotent matrices. Note that everything we have done is $GL(2n,\mathbb{R})$-equivariant. 
Finally we come to the little miracle: the time derivative of $F(t)$ is a curve of reflections $\dot{F}(t)$ (i.e., $\dot{F}(t)^2 = I$) whose -1 eigenspace is the curve of subspaces $\gamma(t)$ and whose $1$-eigenspace defines a "horizontal curve" $h(t)$ equivariantly attached to $\gamma(t)$. This is the construction that yields the Levi-Civita connection (and what is behind the formalisms of Grifone and Foulon for connections of second order ODE's on manifolds).
Differentiate $F(t)$ a second time to find the Schwartzian derivative. Geometrically, it just describes how the curve $h(t)$ moves with respect to $\gamma(t)$. For comparison, recall that the curvature of a connection is obtained by differentiating (i.e., bracketing) horizonal vector fields and projecting onto the vertical bundle.
A: To understand the existence and uniqueness of the LC connection, it is not possible to sidestep some algebra, namely the fact (with a 1-line proof) that a tensor $a_{ijk}$ symmetric in $i,j$ and skew in $j,k$ is necessarily zero. The geometrical interpretation is this: once one has the $O(n)$ subbundle $P$ of the frame bundle $F$ defined by the metric, there exists (at each point) a unique subspace transverse to the fibre that is tangent both to $P$ and to a coordinate-induced section $\{\partial/\partial x_1,\ldots,\partial/\partial x_n\}$ of $F$.
A: Is the following description correct?
The metric determines the geodesics: pull a string tight enough and it will be a geodesic.
These in turn determine a class of connections, determined up to torsion: twist the string while parallel transporting a tangent vector along it, and you are changing the connection keeping the same geodesics.
Now choose a connection, parallel transport an infinitesimal vector along a geodesic curve $\gamma$. The tip of the vector will draw a curve $\gamma '$. The zero torsion connection in the class, i.e. the Levi-Civita connection, is the one minimizing the lenght of $\gamma '$.  
BTW, this question is related:
What is torsion in differential geometry intuitively?
