Let $H=H_1 \otimes H_2$ a composite Hilbert space and let $A, B$ bounded linear operators on $H$, and we can assume they are trace class. Let $A_2$ we denote the operator on $H_2$ obtained by taking the partial trace on $H_1$, i.e. $A_2 = \mathrm{tr}_1 (A) \in B(H_2)$. Assume that we know there is some $\varepsilon > 0$ so that $$ ||A-B||_{B(H)} < \varepsilon. $$ Can we deduce a similar bound for $$ ||A_2 - B_2 ||_{B(H_2)} ? $$
Notes: 1) The finite dimensional case $\dim (H_1) = n_1$, $\dim (H_2) = n_2$, $\dim (H) = n$ all finite would be fine (but if you have a hint or source for an answer in the general case I'd appreciate!).
2) (My approach) Choose the Frobenius (or Hilbert-Schmidt) norm, $$ ||A||_{B(H)} = \sqrt{\mathrm{tr}(A^\dagger A)}$$ and let $C=A-B$. We know that $ ||I_1 \otimes C_2 ||_{B(H)} = ||I_1||_{B(H_1)} ||C_2||_{B(H_2)}$ (cf. Reed Simon Vol. 1 page 299), then $$ \begin{array}{rcl} ||C_2||_{B(H_2)} &=& \frac{1}{\sqrt{n_1}} ||I_1 \otimes C_2||_{B(H)} \\ &=& \frac{1}{\sqrt{n_1}} ||C - (I_1 \otimes C_2) - C ||_{B(H)} \\ &\leq& \frac{\varepsilon}{\sqrt{n_1}} + \frac{1}{\sqrt{n_1}} ||C - (I_1 \otimes C_2)||_{B(H)} \end{array} $$ and expect to bound the term on the right by $\alpha ||C||_{B(H)}$. To see what is going on I put $$ C = C^{(1)}_1 \otimes C^{(2)}_1 + C^{(1)}_2 \otimes C^{(2)}_2 $$ (taking enough terms of these form should be fine to represent an arbitrary $C$). A calculation gives $$ ||C - (I_1 \otimes C_2)||_{B(H)} = ||(C^{(1)}_1 - (\mathrm{tr} C^{(1)}_1) I_1)\otimes C_1^{(2)} + (C^{(1)}_2 - (\mathrm{tr} C^{(1)}_2) I_1)\otimes C_2^{(2)}||_{B(H)} $$ how close this is to $||C||_{B(H)}$?
There is a bound on the norm of partial trace of operator that depends on the type of norm and the size of the Hilbert space you traced out. For the Schatten $p$-norm, which is defined for $p\ge1$ as $$\|A\|_p=\left[\text{Tr}(|A|^p)\right]^{1/p}$$ one can show a bound, due to Rastegin, where $$\|\text{Tr}_1(A)\|_p \le [\dim(H_1)]^{(p-1)/p} \|A\|_p$$ This bound can be made dimension-independent by choosing $p=1$, in which case the norm is the trace norm. But if you choose $p=2$, which gives the Frobenius norm, you'll have a dimension-dependent bound.