Can we use bootstrap samples that are smaller than original sample? I want to use bootstrapping to estimate confidence intervals for estimated parameters from a panel dataset with N=250 firms and T=50 month. The estimation of parameters is computationally expensive (few days of computation) due to use of Kalman filtering and complex nonlinear estimation. Therefore drawing (with replacement) B (in hundreds or more) samples of M=N=250 firms from original sample and estimating the parameters B times is computationally infeasible, even though this is the basic method for bootstrapping. 
So I am considering using smaller M (e.g. 10) for bootstrap samples (rather than the full size of N=250), drawn randomly with replacement from original firms, and then scale the bootstrap-estimated covariance matrix of model parameters with $\frac{1}{\frac{N}{M}}$ (in example above by 1/25) to calculate the covariance matrix for the model parameters estimated on the full sample. 
Desired confidence intervals can then be approximated based on normality assumption, or empirical ones for smaller sample scaled using a similar procedure (e.g. scaled down by a factor of $\frac{1}{\sqrt{\frac{N}{M}}}$. 
Does this workaround make sense? Are there theoretical results to justify this? Any alternatives to tackle this challenge? 
 A: This question was asked a long time ago, but I'm posting a response in case anyone discovers it in future. In short, the answer is yes: you can do this in many settings, and you are justified in correcting for the change in sample size by the $\sqrt{\frac{M}{N}}$. This approach is usually called the $M$ out of $N$ boostrap, and it works in most settings that the ``traditional''' bootstrap does, as well as some settings in which it doesn't.
The reason why is that many bootstrap consistency arguments use estimators of the form $\sqrt{N} (T_N - \mu)$, where $X_1, \ldots, X_N$ are random variables and $\mu$ is some parameter of the underlying distribution. For example, for the sample mean, $T_N = \frac{1}{N} \sum_{i=1}^N X_i$ and $\mu = \mathbb{E}(X_1)$.
Many bootstrap consistency proofs argue that, as $N \to \infty$, given some finite sample $\{x_1, \ldots, x_N\}$ and associated point estimate $\hat{\mu}_N = T_N(x_1, \ldots, x_N)$,
$$
  \sqrt{N}(T_N(X_1^*, \ldots, X_N^*) - \hat{\mu}_N) \overset{D}{\to} \sqrt{N}(T_N(X_1, \ldots, X_N) - \mu)
  \tag{1} \label{convergence}
$$
where the $X_i$ are drawn from the true underlying distribution and the $X_i^*$ are drawn with replacement from $\{x_1, \ldots, x_N\}$.
However, we could also use shorter samples of length $M < N$ and consider the estimator
$$
  \sqrt{M}(T_M(X_1^*, \ldots, X_M^*) - \hat{\mu}_N).
  \tag{2} \label{m_out_of_n}
$$
It turns out that, as $M, N \to \infty$, the estimator (\ref{m_out_of_n}) has s the same limiting distribution as above in most settings where (\ref{convergence}) holds and some where it does not. In this case, (\ref{convergence}) and (\ref{m_out_of_n}) have the same limiting distribution, motivating the correction factor $\sqrt{\frac{M}{N}}$ in e.g. the sample standard deviation.
These arguments are all asymptotic and hold only in the limit $M, N \to \infty$. For this to work, it's important not to pick $M$ too small. There's some theory (e.g. Bickel & Sakov below) as to how to pick the optimal $M$ as a function of $N$  to get the best theoretical results, but in your case computational resources may be the deciding factor.
For some intuition: in many cases, we have $\hat{\mu}_N \overset{D}{\to} \mu$ as $N \to \infty$, so that
$$
  \sqrt{N}(T_N(X_1, \ldots, X_N) - \mu),
  \tag{3} \label{m_out_of_n_intuition}
$$
can be thought of a bit like an $m$ out of $n$ bootstrap with $m=N$ and $n = \infty$ (I'm using lower case to avoid notation confusion). In this way, emulating the distribution of (\ref{m_out_of_n_intuition}) using an $M$ out of $N$ bootstrap with $M < N$ is a more ``right'' thing to do than the traditional ($N$ out of $N$) kind. An added bonus in your case is that it's less computationally expensive to evaluate.
As you mention, Politis and Romano is the main paper. I find Bickel et al (1997) below a nice overview of the $M$ out of $N$ bootstrap as well.
Sources:
PJ Bickel, F Goetze, WR van Zwet. 1997. Resampling fewer than $n$ observations: gains, losses and remedies for losses. Statistica Sinica.
PJ Bickel, A Sakov. 2008. On the choice of $m$ in the $m$ ouf of $n$ bootstrap and confidence bounds for extrema. Statistica Sinica.
A: After reading more on the topic, it seems there is established theory under "sub-sampling" allowing to do this type of confidence interval estimation. The key reference is "Politis, D. N.; Romano, J. P. (1994). Large sample confidence regions based on sub-samples under minimal assumptions. Annals of Statistics, 22, 2031-2050."
The idea is to draw samples of M < N size, "without replacement" for each sample (but with replacement across different samples of size B), from the N initial data points (series in my case), and estimate the confidence interval of parameter of interest using these samples and common bootstrap method. Then scale the confidence interval based on the rate of change in the variance of underlying distribution of parameter with changes in M. That rate is 1/M in many common settings, but could be empirically estimated if we repeat the procedure with a few different M values and look at the changes in the size of inter-percentile ranges. 
