My question is kinda silly, I know it. But I cannot understand why root raised cosine filter or raised cosine filter can eliminate ISI. I googled before, but those explanation is not easy for me to understand. Maybe someone can help me this with some figures? Thank you so much for you help in advance.
Pulse shaping is not done to "eliminate ISI": it does not act as some sort of equalizer to remove ISI that may already be in the waveform. The only reason for pulse shaping is to restrict the occupied bandwidth, and pulse shapes that meets the Nyquist ISI criterion are able to provide the desired bandwidth restriction without introducing new ISI.
In addition to the frequency response that AlexTP showed, I think the time domain view of the impulse response gives good insight into the Zero-ISI operation of a Nyquist filter. Below is shown the result of transmitting three symbols, than each began as a single impulse. Each plot below is the "impulse response" for each symbol. Note the blue symbol was transmitted first and is maximum at time index 0 on the graph (but began much earlier than this in time). This position where the impulse response is maximum is our ideal sampling location to determine what was transmitted. One sample after the original impulse for the blue symbol, came an impulse for the red symbol. Notice at the ideal sampling location for red (where it is maximum), the response from all the other symbols is zero (meaning zero ISI).
The pulses shown above have zero ISI at the sampling locations, but extend to infinity to achieve the maximimum conservation of bandwidth (1/(2T)). Of course, this is not feasible, so the raised cosine pulse shape applies a window to truncate the perfect Sinc function, resulting in a trade of bandwidth versus filter complexity, as well as trades with other characteristics related to performance for timing recovery.
The filter is often split between transmitter and receiver as a "root-raised cosine filter" (RRC) on each side. Note that a RRC filter does NOT have zero-ISI until it is paired with the second RRC filter to form in cascade a raised cosine filter.
Below is an example eye diagram, showing the trajectory over 1 symbol for multiple pulses (each trajectory is based on the history of all previous symbols within the depth of the RRC filter and results in a different path when going from 0 to 1 or 0 to 0 etc). At time 0 is the ideal sampling instant to decide between 0 and 1, notice with the RRC filter that there is significant ISI, which is eliminated after passing through the second RRC filter in the receiver (assuming no other multipath effects have occurred). Also notice for the filter with alpha = 0.25 that there is a significant spread in the zero crossing locations, which would add more noise in timing recovery operations that are sensitive to zero crossings. Compare this to the eye diagram with alpha= 1 and notice the significant decrease in zero-crossing jitter, as well as a wider "eye-opening" which also means less sensitivity to jitter or other errors in the sampling clock location, but this benefit comes at the expense of less spectral efficiency.
What is very interesting is the pulse shaping Nyquist filter that fred harris now refers to as the "harris-Moerder Nyquist Pulse". Using this as the pulse shaping filter results in an order of magnitude lower EVM for the same number of taps. This isn't commonly used even today (15 years later) to my knowledge since the root-raised cosine filter is baked into our specifications (I assume) but certainly is very compelling to be aware of for consideration in future radios.
Because it satisfies Nyquist ISI criterion Nyquist_ISI_criterion.
The simplest way to apply this criterion to root raised cosine filter is looking at its frequency response of the composite filter "raised cosine" : Raised-cosine_filter
You can see easily that the sum of $1/T$-shifted spectrum of raised cosine filter is constant.