Why isn't the path integral defined for non-homotopic paths? Context
In the Aharonov Bohm effect, there is  a solenoid which creates a magnetic field. Since the electron cannot be inside the solenoid, the configuration space is not simply connected.
Question
I've read in this paper, that the path integral is  defined only for  paths in the same homotopy class in the configuration space. But I don't see the reason for this. Could someone explain it or give any reference?
It seems that Laidlaw, DeWitt and Schulman have done some work, but I didn't see any proof. And Feynman & Hibbs don't seem to mention it.
Furthermore, does the same problem arise in standard variational calculus when one applies Hamilton's principle?
 A: TL;DR: The Feynman path integral/kernel/amplitude $K(x_f,x_i)$ is in general a weighted sum over ALL (not necessarily homotopic) paths, as user Heidar writes in a comment above.
In more detail: Let there be given an initial point $x_i$, a final point $x_f$, and a fiducial point $\ast$. Fix two paths $\gamma_i: x_i\to \ast$ and $\gamma_f:\ast \to x_i$. It is natural to assume that the full path integral is of the form $$K(x_f,x_i) ~=~ \sum_{\gamma\in \pi^1(X, \ast)} \chi(\gamma)~ K^{\gamma}(x_f,x_i),$$ where $\chi(\gamma)\in\mathbb{C}$ is some weight. Here the partial path integral $K^{\gamma}(x_f,x_i)$ consists of all paths $x_i\to x_f$ in the homotopy class $[\gamma_i+\gamma+\gamma_f]$. In this way we have formally counted each path exactly once.
For consistency, it turns out that $\chi$ must be a 1-dimensional unitary representation of the fundamental group $\pi^1(X, \ast)$ using well-known group and probability properties of the path integral, cf. the theorem of Ref. 1. The upshot is that we have reduced the full path integral $K(x_f,x_i)$ to a weighted sum of partial path integrals $K^{\gamma}(x_f,x_i)$ whose paths are all within the same homotopy class. But we still have to define/calculate the weighted sum $K^{\gamma}(x_f,x_i)$ of paths within each homotopy class.
References:

*

*M.G.G. Laidlaw & C.M. DeWitt,
Phys. Rev. D3 (1971) 1375.

A: The principle of the superposition of quantum states, or, as I shall
refer to it, the sum over the alternatives, holds for particles belonging
to a multiply-connected space in the same way as it holds for particles
belonging to a simply-connected one, since it is one of the fundamental
principles of quantum theory. On the other hand, what must be better
explained here is why the sum over the alternatives, or in the present
case, over the paths, in a simply-connected space can be constructed
as a single path integral, differently from multiply-connected spaces.
First, I shall begin with an intuitive argument. Let $X$ be a "nice" topological
space (we mean, for instance, that $X$ is arcwise connected or locally
simply connected), $a,b\in X$, $\Omega(a,b)$ the set of paths $[t_{a},t_{b}]\longrightarrow X$ from $a$ to $b$ and $t_{b}>t_{a}>0$. To each $x(t)\in\Omega(a,b)$, we associate an amplitude $\phi[x(t)]$. Recall that, heuristically, we write the following proportionally relation for the propagator $K=K(b,t_{b};a,t_{a})$,
$$
K\sim\sum_{x(t)\in\Omega(a,b)}\phi[x(t)].
$$
If $S$ is the action governing the dynamics of our system and if
$t_{b}-t_{a}$ is small enough, we know that $\phi[x(t)]\sim e^{iS[x(t)]}$
(where we have assumed $h=2\pi$). But there is no a priori
reason, without evoking any property of $X$, to ensure that all the paths should contribute to $K$ with
the same phase. For example, if $x(t),y(t)\in\Omega(a,b)$, why we
cannot have
$$
K\sim e^{iS[x(t)]}-e^{iS[y(t)]}+...\,\,\,?
$$
It turns out that if our topological space $X$ is simply-connected,
we can always deform the path $x(t)$ to $y(t)$ continuously, a deformation
which, in effect, should make $\phi[x(t)]$ approach $\phi[y(t)]$
continuously too. Formally,
$$
\phi[y(t)]=\lim\phi[x(t)]=e^{iS[y(t)]},\,\,\,\mbox{as }x(t)\rightarrow y(t)\mbox{ continuously.}
$$
From this, we can conclude two things:


*

*Paths in a simply connected space contribute to the total amplitude
with the same phase. So if $X$ is simply connected, we can then write
the familiar expression
$$
K\sim\sum_{x(t)\in\Omega(a,b)}e^{iS[x(t)]},
$$
which, upon introducing the appropriate measure, results in the Feynman
path integral
$$
K=\int_{\Omega(a,b)}e^{iS[x(t)]}\mathcal{D}x(t).
$$

*Paths in the same homotopy class contribute to the total amplitude
with the same phase. So, for the propagator $K^{q}$ restricted to
paths constrained in the homotopy class $q$, we can write similarly
$$
K^{q}\sim\sum_{x(t)\in q}e^{iS[x(t)]},
$$
that also becomes a path integral
$$
K^{q}=\int_{q}e^{iS[x(t)]}\mathcal{D}x(t),
$$
but whose domain of functional integration is now $q$. Each such
$K^{q}$ is called a partial amplitude.


Since the principle of the sum over the alternatives allows us to
write the propagator $K$ as the sum of the amplitudes of each homotopy
class $q$ individually (namely, the partial amplitudes), each one
contributing with a phase that will be labeled by $\xi_{q}\in\mathbb{C},|\xi_{q}|=1$,
we have that
$$
K=\sum_{q\in\pi(a,b)}\xi_{q}K^{q},
$$
where $\pi(a,b)$ is the set of all homotopy classes for the paths from $a$ to $b$. This answers the question raised by jinawee, I hope.
But now, it will be instructive if we sketch on the proof of that result
discovered first by Schulman and proved a little later by Laidlaw
and DeWitt. Namely, that the set of phases $\{\xi_{q}\}$ can be "identified"
with a scalar unitary representation of the fundamental group of $X$.
The idea is the following. Let $c\in X$ fixed and choose $C(x)$
to be any path connecting $c$ to whatever $x\in X$. Such $C(\alpha)$
is known as a homotopy mesh. To each pair $(a,b)\in X\times X$,
we can construct a mapping 
$$
f_{ab}:\pi(c)\longrightarrow\pi(a,b)
$$
by $f_{ab}(\alpha)=[C^{-1}(a)]\alpha[C(b)]$. This is an injection
between the fundamental group $\pi\equiv\pi(c)$ at $c$ and the homotopy
class $\pi(a,b)$, allowing us to label the propagator $K^{q}$ and
the phase factor $\xi_{q}$ associated to a homotopy class $q$ with
the elements of the fundamental group $\pi$, say,
$$
K^{q}\rightarrow K^{\alpha},\,\,\,\xi_{q}\rightarrow\xi(\alpha)
$$
iff $f_{ab}(\alpha)=q$. So finally, our propagator assumes the form
of a sum over the elements of a group:
$$
K=\sum_{\alpha\in\pi}\xi(\alpha)K^{\alpha}.
$$
The result then follows from this consideration: the association $\alpha\mapsto K^{\alpha}$ of a partial amplitude $K^{\alpha}$ to each element $\alpha$ of
the group $\pi$ depends on the injection $f_{ab}$, which in turn,
depends on the choice of the mesh function $C$$(x)$. The (absolute
value) of the propagator $K$, however, must be the same independently
of the adopted mesh function.
The best place to find the details of the proof is still the paper
"Feynman Functional Integrals for Systems of Indistinguishable Particles"
(1971) by Laidlaw and DeWitt.
Additionally, there is another way to motivate the formula of the propagator
$K$ as a sum over partial amplitudes associated to homotopy classes,
which is based in a covering space of $X$. This was in fact one of
the original reasonings employed by Schulman in "A Path Integral
for Spin" (1968) in order to discuss the spin of a (quantum) non-relativistic
particle using exclusively the method of path integration.
Roughly, it is like this. Let $\mathbf{X}$ be covering space of $X$
and $\mathrm{p}:\mathbf{X}\longrightarrow X$ the covering projection.
Moreover, let $\mathcal{L}$ be the Lagrangian of our system in $X$,
for which $S=\int\mathcal{L}dt$, $\mathbf{L}$ the lift of $\mathcal{L}$
to our covering space $\mathbf{X}$ induced by the projection $\mathrm{p}$
and $\mathbf{S}=\int\mathbf{L}dt$ the action on $\mathbf{X}$. To
each pair $(a,b)\in X\times X$, choose some $\mathbf{a}\in\mathrm{p}^{-1}(a)$
and let $\mathbf{b}_{\alpha}\in\mathrm{p}^{-1}(b)$ define a sequence
in $\mathbf{X}$ indexed by the elements of the fundamental group,
that is, $\alpha\in\pi$.
Since the covering space is simply-connected, to each $\alpha\in\pi$,
the propagator $K^{\alpha}$ associated to the amplitude for going
from $\mathbf{a}$ to $\mathbf{b}_{\alpha}$ in the interval $[t_{a},t_{b}]$
on $\mathbf{X}$ is given, as familiar, by the path integral
$$
K^{\alpha}=\int_{\mathbf{a}}^{\mathbf{b}_{\alpha}}e^{i\mathbf{S}[\mathbf{x}(t)]}\mathcal{D}[\mathbf{x}(t)],
$$
where in this case, the functional integral runs over the paths $\mathbf{x}(t):[t_{a},t_{b}]\longrightarrow\mathbf{X}$
connecting $\mathbf{a}$ to $\mathbf{b}_{\alpha}$.
Finally, by the principle of the sum over the alternatives, the propagator
$K$ for the amplitude of going from $a$ to $b$ in the time interval
$[t_{a},t_{b}]$ on our multiply-connected space $X$ is seem to be
the sum over the alternatives for going from $\mathbf{a}$ to $\mathbf{b}_{\alpha}$
for all $\alpha\in\pi$ in the covering $\mathbf{X}$. In effect,
we shall again obtain $K=\sum_{\alpha\in\pi}\xi(\alpha)K^{\alpha}$
for some phase factors $\xi(\alpha)\in\mathbb{C}$.
A proof of the result found by Schulman, Laidlaw and DeWitt described
above, using the latter approach of covering spaces, may be found
in the paper "Quantum mechanics and field theory on multiply connected
and on homogenous spaces" (1972) by Dowker. I believe that the best
sources for learning the subject are still the original papers cited
above (and which I may send upon request). Additionally, if you have
access to an university library, it is opportune to give a look at
the Chapter 8 of "Functional Integration: Action and Symmetries"
by Cartier and DeWitt.
