Generally, text preprocessing impact depends on the NLP application and the language of these texts. Hence, you may find different results in the literature.

For instance, the Arabic language has a rich and complex morphology and text preprocessing is a key component for most of its NLP applications. However, its impact differs from an NLP application to another. On one hand, in the context of Information Retrieval, the state of the art light stemming algorithm increase the MAP (Mean Average Precision) by 30% in comparison to text normalization. On the other hand, light stemming shows a small improvement for Arabic text classification in comparison to text normalization.

In short the usefulness of pre-processing differs from application to application.

Q1 : So does it mean we can use any pre-processing technique which will increase the efficiency of data model OR there is a constraint about pre-processing techniques which can be used for a certain kind of application?

Q2 : The facts and results you have mentioned here, have you got those from some paper?

You could not find a paper because the concept of pre-processing is task specific. When you mention pre-processing, are you talking about stopword removal, stemming, lowercasing etc.? Are you talking about spelling correction/normalization? Or named entity recognition for question answering? What's considered as preprocessing for one task may not be preprocessing for another task. Also, for each preprocessing task, the effect will be different for tasks with differing intents. For example, lowercasing is a fairly common basic preprocessing step for many tasks (e.g., text classification). However, it is typically not useful for NER tasks.

So, in short, you can only view preprocessing and its effectiveness from task-specific perspective. You will need to search the NLP literature for the task you are interested in and then review the effectiveness of different preprocessing techniques.