The answer is YES, we can move the four circles around (at least for a small distance).
Following is my attempt to derive a criterion for ability to move the circles around without bumping into each other. At the end is a modification of this criterion to answer two related questions.
Derivation of a moveability criterion
Before we start, let us generalize the problem a little bit.
We will identify the plane $\mathbb{R}^2$ with complex plane $\mathbb{C}$ and let
- $\Gamma$ be the collection of polyline/polygon which allows circular arcs as edges.
- $C(c,r)$ be the circle centered at $c$ with radius $r$.
- $\angle(u,v) = \arg(v\bar{u})$ be the angle one need to rotate something in direction $u$ counterclockwisely into direction $v$.
Start with a convex region $K$ whose boundary $\partial K \in \Gamma$.
We place $n$ circles $C_0,\ldots,C_{n-1}$ inside $K$ so that each $C_k$ touches $\partial K$ and two and only two others circles. If one extend indexing of the circles by periodicity and treat $C_{-1}, C_n$ as aliases of $C_{n-1}$ and $C_0$,
we can order the circles so that circle $C_k$ is touching circles $C_{k\pm 1}$. The generic problem is:
Can we move all circles simultaneously so that each $C_k$ remains inside $K$, touching $\partial K$ without bumping into other circles.
The initial condition each $C_k$ touches only two circles ensures us for small displacements, $C_k$ will not bump into circles other than $C_{k\pm 1}$. We can concentrate on the relative placement of $C_{k\pm 1}$ with respect to $C_k$.
Let $c_k, r_k$ be center and radius of $C_k$, ie. $C_k = C(c_k,r_k)$.
Since $C_k$ is touching $\partial K$, $c_k$ is constrained to lie on some curve $\gamma_k \in \Gamma$.
WOLOG, let's assume $C_k$ are ordered along $\partial K$ counterclockwisely
and all $\gamma_k$ are parameterized by arc-length with $\gamma_k(0) = c_k$ in positive orientation.
To proceed further, we need to make some technical assumptions:
[A1] None of $c_k$ is located at corners/endpoints of corresponding $\gamma_k$.
This means the tangent vectors $t_k = \gamma'_k(0)$ are well defined.
Let $\alpha_k = \angle(t_{k-1},c_k - c_{k-1})$ and $\beta_k = \angle(c_k - c_{k-1},t_k)$.
[A2] None of $|\alpha_k|$, $|\beta_k|$ equals to $\frac{\pi}{2}$.
Consider following problem of placement of $n+1$ points $p_0,\ldots,p_n$:
- [C1] $p_0,\ldots,p_n$ are initially located at corresponding $c_k$.
- [C2] $p_0,\ldots,p_n$ are constrained to move on corresponding $\gamma_k$.
- [C3] $|p_k-p_{k-1}| = \ell_k \stackrel{def}{=} r_k + r_{k-1}$ for $k = 1,\ldots, n$.
Under assumption [A1] and [A2], if we move $p_0$ along $\gamma_0$ for a small distance $s_0 = \epsilon$, it is possible to move $p_1$ along $\gamma_1$ for a small distance $s_1$ to maintain condition [C3].
Assumption [A2] ensures $s_1$ is of the order of $\epsilon$.
In fact, since $\gamma_1$ is either a straight edge or circular arc at $\epsilon \sim 0$, $s_1$ is locally a $C^\infty$ function in $\epsilon$.
Repeating this sort of argument, one find each $p_k$ will be moved along $\gamma_k$ for a distance $s_k$ of order $\epsilon$.
There are $n$ $C^\infty$ functions $s_1(\epsilon),\ldots, s_n(\epsilon)$ defined over some neighborhood of $0$ and $n$ constants
$\lambda_1,\ldots,\lambda_n$ such that:
$$p_k(\epsilon) = \gamma_k(s_k(\epsilon))\quad\text{ and }\quad s_k(\epsilon) = \lambda_k \epsilon + O(\epsilon^2)\quad\text{ for }\quad k = 1,\ldots, n$$
Let $C_k(\epsilon) = C(p_k(\epsilon),r_k)$ and focus on dependence of $s_n$ on $\epsilon$.
If $s_n(\epsilon) > \epsilon$, $p_n(\epsilon)$ will overtake $p_0(\epsilon)$ along $\gamma_0$.
$C_{n-1}(\epsilon)$ will bump into $C_0(\epsilon)$ (ie. their interior intersect) and the circles cannot move.
On the other direction, if $s_n(\epsilon) < \epsilon$, $p_n(\epsilon)$ will lag behind $p_0(\epsilon)$.
$C_{n-1}(\epsilon)$ will move away from $C_0(\epsilon)$ and the circles can move.
To see whether one can move the circles, the first thing we need is $\lambda_n$.
If $\lambda_n < 1$, the circles can move. If $\lambda_n > 1$, the circle cannot move.
For the problem at hand, $\lambda_n = 1$ by symmetry. One need the higher order $\epsilon$-dependence of $s_n(\epsilon)$.
Let $\rho_k$ be the curvature of $\gamma_k$ at $c_k = \gamma_k(0)$.
Choose a coordinate system where $c_{k-1} = 0$ and $c_{k} = \ell_k$, we find
$$p_k - p_{k-1} = \ell_k + \left(s_k + \frac{i}{2}\rho_k s_k^2\right) e^{i\beta_k} -
\left(s_{k-1} + \frac{i}{2}\rho_{k-1} s_{k-1}^2\right) e^{-i\alpha_k} + o(\epsilon^2)$$
This leads to
$$\begin{align}
\Re(p_k - p_{k-1}) &= \ell_k + \left(s_k\cos\beta_k - s_{k-1}\cos\alpha_k\right) - \frac{\epsilon^2}{2}\left(\rho_k\lambda_k^2\sin\beta_k + \rho_{k-1}\lambda_{k-1}^2\sin\alpha_k\right) + o(\epsilon^2)\\
\Im(p_k - p_{k-1}) &= \epsilon(\lambda_k\sin\beta_k + \lambda_{k-1}\sin\alpha_k) + O(\epsilon^2)
\end{align}
$$
To satisfy [C3], we need
$$s_k\cos\beta_k - s_{k-1}\cos\alpha_k = \frac{\epsilon^2}{2}\left\{
\begin{align} & \left(\rho_k\lambda_k^2\sin\beta_k + \rho_{k-1}\lambda_{k-1}^2\sin\alpha_k\right)\\
- & \frac1{\ell_k}\left[ (\lambda_k\cos\beta_k - \lambda_{k-1}\cos\alpha_k)^2 +
(\lambda_k\sin\beta_k + \lambda_{k-s}\sin\alpha_k)^2\right]\end{align}\right\} + o(\epsilon^2)$$
Comparing the $\epsilon$-term on both sides, we find $\frac{\lambda_k}{\lambda_{k-1}} = \frac{\cos\alpha_k}{\cos\beta_k}$. Multiplying these ratios for different $k$, we get:
$$\lambda_k = \prod_{j=1}^k \frac{\cos\alpha_j}{\cos\beta_j}$$
Substitute this back into above expression and simplify, we obtain
$$\frac{s_{k}}{s_{k-1}}
= \frac{\lambda_{k}}{\lambda_{k-1}}\left\{
1 + \frac12\epsilon(\lambda_k\cos\beta_k)
\underbrace{\left[\rho_k\frac{\sin\beta_k}{\cos^2\beta_k^2} + \rho_{k-1}\frac{\sin\alpha_k}{\cos^2\alpha_k} - \frac{(\tan\alpha_k + \tan\beta_k)^2}{\ell_k}\right]
}_{\text{ call this }\Delta_k}\right\}
$$
Multiply these ratios for different $k$ again, we obtain
$$s_n(\epsilon) = \epsilon\lambda_n + \frac{\epsilon^2\lambda_n}{2}
\underbrace{\sum_{k=1}^n(\lambda_k \cos\beta_k)\Delta_k}_{\text{ call this }\Delta} + o(\epsilon^2)$$
By a similar argument like the first order case, we arrive at following improved criterion:
If $\lambda_n > 1$ or $\lambda_n = 1$ and $\Delta > 0$, the circles cannot move.
If $\lambda_n < 1$ or $\lambda_n = 1$ and $\Delta < 0$, the circles can move.
Analysis of original problem
Back to the original problem where all $r_k = 1 \implies \ell_k = 2$.
Choose a coordinate systems where the circles are centered at
$$
\begin{cases}
c_0 = (1+\sqrt{2},1+2\sqrt{2})\\
c_1 = (1,1+\sqrt{2})\\
c_2 = (1+\sqrt{2},1)\\
c_3 = (1+2\sqrt{2},1+\sqrt{2})
\end{cases}
$$
For small $\epsilon$, one can treat
$\gamma_1,\gamma_2,\gamma_3$ as straight lines
and $\gamma_4$ as a circular arc with radius $$R = \sqrt{(1+\sqrt{2})^2+(1+2\sqrt{2})^2} = \sqrt{6(2+\sqrt{2})}$$
Furthermore, it is easy to see
$$\alpha_1 = \theta \stackrel{def}{=} \frac{\pi}{4} + \tan^{-1}\frac{1+\sqrt{2}}{1+2\sqrt{2}} = \tan^{-1}(3+\sqrt{2})$$
In terms of $R,\theta$, the parameters required to compute $\Delta$ is listed below
$$\begin{array}{r|cccc}
k & \lambda_k & \alpha_k & \beta_k & \rho_k\\
\hline
1 & \sqrt{2}\cos\theta & \theta & \frac{\pi}{4} & 0 \\
2 & \sqrt{2}\cos\theta & \frac{\pi}{4} & \frac{\pi}{4} & 0 \\
3 & 1 & \frac{\pi}{4} & \theta & 0\\
4 & 1 & \frac{\pi}{2} - \theta & \frac{\pi}{2} - \theta & \frac1R
\end{array}
$$
For the given choice of $R, \theta$, it satisfies an interesting relation $R\cos\theta = 1$. It is not clear whether this is a coincidence or has a deeper meaning. In any event, this simplifies evaluation of $\Delta_k$ a lot.
$$\begin{align}
\Delta_1 = \Delta_3 &= \frac{1}{R}\frac{\sin\theta}{\cos^2\theta} - \frac{(1 + \tan\theta)^2}{2}
= \tan\theta - \frac{(1+\tan\theta)^2}{2}
= -3(2+\sqrt{2})\\
\Delta_2 &= -\frac{(1+1)^2}{2} = -2\\
\Delta_4 &= \frac{2}{R}\frac{\cos\theta}{\sin^2\theta} - \frac{(\cot\theta + \cot\theta)^2}{2} = 2\cot^2\theta - 2\cot^2\theta
= 0
\end{align}$$
Since all $\lambda_k\cos\theta_k > 0$ and all $\Delta_k \le 0$ but not all $= 0$, we have
$$\Delta = \sum_{k=1}^4 (\lambda_k \cos\beta_k)\Delta_k < 0$$
Together with $\lambda_4 = 1$, the criterion we derived before tell us we can move the four circles (at least for a small distance).
Whether we can move the circles continuously to the attempted configuration
(the one where one unit circle is stuck at the corner of quarter disc) needs
more analysis.
Application to similar problems
Let us review what we have found in original problem.
The most surprising result is probably $\Delta_4 = 0$. It turns out it isn't that surprising. It vanishes because $p_3$ and $p_4$ are constrained to move on same circular arc. In general, if $p_{k}$ and $p_{k-1}$ are constrained to move on same straight edge/circular arc with fixed distance apart, they will always travel the same distance.
This means $\lambda_k = \lambda_{k-1}$ and $\Delta_k = 0$. In computation of $\lambda_n$ and $\Delta$, we can ignore existence of
such a pair of points $p_{k-1}, p_k$.
Another interesting result is $\Delta_2 < 0$. It is negative because
$\rho_1 = \rho_2 = 0$, ie. $p_1, p_2$ are moving on straight edge portions of $\gamma_1$ and $\gamma_2$.
In general, if $p_{k}, p_{k-1}$ are moving on straight edge portions of $\gamma_k$ and $\gamma_{k-1}$, we will have
$$\Delta_k = -\frac{(\tan\alpha_k + \tan\beta_k)^2}{2}$$
This term is $<0$ except at three cases.
- $\alpha_k = \beta_k = 0$ : Since the underlying $K$ is convex, this essentially force $\gamma_k$ and $\gamma_{k-1}$ to be same straight edge and reduce to the case discussed before.
- $\beta_k = -\alpha_k \ne 0$ : This is impossible when underlying $K$ is convex.
- $\beta_k + \alpha_k = \pi$ : $\gamma_k$ and $\gamma_{k-1}$ is parallel but pointing in opposite directions. This is possible when underlying $K$ contains a pair of parallel non-neighboring edges and $C_k$, $C_{k-1}$ are touch them.
Based on these, we have a alternate criterion.
Under the assumption
- [A0] $K$ is a polygon consists only of straight edges.
- [A1] None of $c_k$ is located at corners/endpoints of corresponding $\gamma_k$.
- [A2'] All $|\alpha_k|$, $|\beta_k| < \frac{\pi}{2}$ and not all of them are $0$.
We will avoid the problematic case $\alpha_k + \beta_k = \pi$. We will find all $\lambda_k\cos\theta_k > 0$, all $\Delta_k \ne 0$ but not all $\Delta_k = 0$. This means $\Delta < 0$ and as a result, arrive at following
polygonal version of criterion:
If $\lambda_n > 1$, the circles cannot move.
If $\lambda_n \le 1$, the circles can move.
If one look at configurations appear in following questions:
It is easy to check both configurations satisfy the new set of assumptions and $\lambda_n = 1$ by symmetry. As a result, the circles can move in both cases.