Given that the "why int by default" question hasn't been answered ...
First, "default" is not really the right term (although close enough). As noted by VonC, an expression composed of ints and longs will have a long result. And an operation consisting of ints/logs and doubles will have a double result. The compiler promotes the terms of an expression to whatever type provides a greater range and/or precision in the result (floating point types are presumed to have greater range and precision than integral, although you do lose precision converting large longs to double).
One caveat is that this promotion happens only for the terms that need it. So in the following example, the subexpression 5/4 uses only integral values and is performed using integer math, even though the overall expression involves a double. The result isn't what you might expect...
(5/4) * 1000.0
OK, so why are byte and short promoted to int? Without any references to back me up, it's due to practicality: there are a limited number of bytecodes.
"Bytecode," as its name implies, uses a single byte to specify an operation. For example iadd, which adds two ints. Currently, 205 opcodes are defined, and integer math takes 18 for each type (ie, 36 total between integer and long), not counting conversion operators.
If short, and byte each got their own set of opcodes, you'd be at 241, limiting the ability of the JVM to expand. As I said, no references to back me up on this, but I suspect that Gosling et al said "how often do people actually use shorts?" On the other hand, promoting byte to int leads to this not-so-wonderful effect (the expected answer is 96, the actual is -16):
byte x = (byte)0xC0;
System.out.println(x >> 2);