Linus actually argued something slightly different. That discussion was about whether a "bad" random source can weaken a "good" source. This paper is about constructing a "good" source from two independent "bad" sources.
The problem with Linux was people complaining that mixing the Intel RNG into the existing entropy pool would lower the entropy of the entire pool, effectively letting a backdoored Intel RNG render the random system untrustworthy. Linus' response was the fairly simple/logical observation that "that's not how entropy works".
Suppose we have a random bit string with entropy X and we xor the bit string with a completely deterministic bitstring, what's the result? Clearly you have a bitstring that STILL has entropy X, because the result is completely dependent on our initial bitstring. Mixing with a deterministic bitstring doesn't make the output of our XOR any more predictable then the original bitstring, it's equally random.
So Linus' argument was that "IF the Intel RNG is completely backdoored and deterministic mixing it with the entropy pool will have no effect on the entropy in the pool. HOWEVER, if the Intel RNG is anything but completely deterministic, i.e. there is even a tiny bit of randomness in it, this will actually INCREASE the pool's entropy.
So mixing a completely backdoored RNG will have no negative impact, but mixing anything that's not 100% predictable will have a positive impact, so there's no reason to not always mix the hardware RNG with the pool.
The problem with Linux was people complaining that mixing the Intel RNG into the existing entropy pool would lower the entropy of the entire pool, effectively letting a backdoored Intel RNG render the random system untrustworthy. Linus' response was the fairly simple/logical observation that "that's not how entropy works".
Suppose we have a random bit string with entropy X and we xor the bit string with a completely deterministic bitstring, what's the result? Clearly you have a bitstring that STILL has entropy X, because the result is completely dependent on our initial bitstring. Mixing with a deterministic bitstring doesn't make the output of our XOR any more predictable then the original bitstring, it's equally random.
So Linus' argument was that "IF the Intel RNG is completely backdoored and deterministic mixing it with the entropy pool will have no effect on the entropy in the pool. HOWEVER, if the Intel RNG is anything but completely deterministic, i.e. there is even a tiny bit of randomness in it, this will actually INCREASE the pool's entropy.
So mixing a completely backdoored RNG will have no negative impact, but mixing anything that's not 100% predictable will have a positive impact, so there's no reason to not always mix the hardware RNG with the pool.