On Dec 1, 2003, at 7:50 AM, Rafael Avila de Espíndola wrote:

Testing the Rob Leslie implementation of the Szu-Wei Lee algorithm I notice that his modification (multiplying the input of the DCTIV by 2 and dividing the output) generally improves the accuracy but, in some cases, may cause a overflow.

I have not found a .mp3 file that triggers a overflow, but the following data does. The present IMDCT implementation in libmad 0.15 return a valid result.

0x442D46F 0x217CC40 0x8E26853 0xF1888C6 0x921BF17 0xD282 0xF060CB8 0x3E82D68 0xBCA6124 0x1217700 0x9500F6D 0x3E9B4B6 0xC4B7675 0x6AE1645 0x56C266C 0x4463FD6 0x7EE5176 0xC408F90

Is there some evidence that no real mp3 may trigger an overflow?

Only that in order to overflow, I think some of the transformed sample values would have to be at least 12dB above full scale. For example, when the above data is transformed, many samples above full scale result, including two samples more than 14dB above (-5.200011).

Having said that, I would be very interested to learn of a legitimate Layer III frame that does cause an overflow.