Is tracing out a subsystem always akin to discarding all information about it? Suppose we have some quantum system with sub-systems A and B. It could be, for example, two qubits or groups of qubits. Is it fair to say that tracing out the sub-system A is always akin discarding any information about the sub-system A? So, if our information comes only from the measurement of the sub-system B, the measured probabilities will correspond to the same reduced density matrix regardless of what happens to sub-system A: whether it has been measured by someone else without our knowledge, or not?
If this intuitive understanding is true, is there any mathematical proof you could point me to that a reduced density matrix is independent from what happens to the traced-out sub-system--for any mixed state, any combination of separable and entangled states, and any operators that could affect the traced-out sub-system? I feel this should be simple to prove, but couldn't find anything explicit.
 A: Detailed version
Let $\rho$ be a density matrix describing a state shared by Alice and Bob. We can generically write it as
$$\newcommand{\ketbra}[2]{\lvert#1\rangle\!\langle#2\rvert}\rho=\sum_{ijkl}\rho_{ijkl}\ketbra ij\otimes\ketbra kl,$$
where here $\ketbra ij$ refer to Alice's degrees of freedom, while $\ketbra kl$ to Bob's.
The reduced state $\rho_A$ of Alice is obtained by tracing out Bob's degrees of freedom, and reads
$$\rho_A\equiv\operatorname{Tr}_B\rho=\sum_{ij}\underbrace{\left(\sum_k\rho_{ijkk}\right)}_{\equiv(\rho_A)_{ij}}\ketbra ij.$$
The state that results from Bob doing something on his side can be written as
$$\rho'
\equiv(\mathbb1\otimes\mathcal E)\rho=\sum_{ijkl}\rho_{ijkl}\ketbra ij\otimes\mathcal E(\ketbra kl),$$
where $\mathcal E$ is a generic (completely positive, trace-preserving) map.
We can also get an explicit expression for the matrix elements of $\rho'$ by writing
$$\rho'=\sum_{ijkl}\rho_{ijkl}\ketbra ij\otimes\mathcal E(\ketbra kl)
=\sum_{ijmn}\underbrace{\left(\sum_{kl}\rho_{ijkl}\langle m\rvert\mathcal E(\ketbra kl)\lvert n\rangle\right)}_{\rho'_{ijmn}}\,\,\ketbra ij\otimes\ketbra mn.$$
In other words, we have
$\rho'_{ijmn}=\sum_{kl}\rho_{ijkl}\langle m\rvert\mathcal E(\ketbra kl)\lvert n\rangle.$
How is Alice's share of the state affected by Bob's acting on his? If no classical information is exchanged (that is, if Alice is not given any information about Bob's observations), then her share of the state is described again by the reduced density matrix, this time of $\rho'$: $\rho'_A\equiv\operatorname{Tr}_B\rho'$. This reads
$$\rho'_A=\sum_{ij}\left(\sum_m \rho'_{ijmm}\right)\ketbra ij,$$
where 
$$\sum_m\rho'_{ijmm}=\sum_{klm}\rho_{ijkl}\langle m\rvert\mathcal E(\ketbra kl)\lvert m\rangle=
\sum_k\rho_{ijkk},$$
where in the last step we used that any completely positive trace preserving linear map must satisfy $\operatorname{Tr}[\mathcal E(\ketbra kl)]=\delta_{kl}$, which you can easily see passing through Kraus' decomposition:
$$\operatorname{Tr}[\mathcal E(\ketbra kl)] =\operatorname{Tr}\left[\sum_\ell A^\ell\ketbra kl A^{\ell \dagger}\right]
= \operatorname{Tr}\left[\sum_\ell A^{\ell \dagger} A^\ell\ketbra kl \right]
= \operatorname{Tr}[\ketbra kl]. $$
(This also immediately follows from the property of any quantum channel to be trace-preserving).
You can therefore conclude that $\operatorname{Tr}_B(\rho)=\operatorname{Tr}_B(\rho')$, which means that no matter what Bob is doing on his side, Alice will not see any measurable difference on her side. There can of course be (classical and nonclassical) correlations between the measurement results, but these correlations are only observable by exchanging classical information between Alice and Bob.
Note that this is totally general, as any operation (including measurements and such) performed by Bob can be written as some quantum map $\mathcal E$ acting on his system.

Short version
Working directly on the matrix elements of the states and the channel, using Einstein's convention for repeated indices, and using numbers instead of latin characters for the indices, we have
$$(\rho'_A)_{12}=\{\operatorname{Tr}_B[(\mathbb1\otimes\mathcal E)\rho]\}_{12}
=[(\mathbb1\otimes\mathcal E)\rho]_{1233}
=\rho_{1245}\mathcal E_{3345}=\rho_{1233}\equiv(\rho_A)_{12},$$
where I'm defining the matrix elements of $\mathcal E$ as $\mathcal E_{1234}\equiv \langle1\rvert \mathcal E(\ketbra34)\lvert 2\rangle$.
