Conditional entropy between ciphertext, plaintext and the key

If we have a (possibly imperfect) cryptosystem that generates ciphertext $C$ from plaintext $P$ and key $K$, we have:

$H(C, P, K) = H(C | P, K) + H(P, K)$

where $H$ is entropy.

My question is why following is true:

$H(C | P, K) = 0$

It seems because each key and plaintext uniquely define a ciphertext but I want to prove this mathematically (theories in entropy).

It is not true for all cryptosystems. Most modern cryptographic systems support probabilistic encryption where there can be many ciphertexts associated with a single key-plaintext pair. This is particularly important where we wish schemes to have indistinguishability under chosen plaintext attacks (IND-CPA) for example.

Take the definition of the conditional entropy,

$H(C|P,K) = - \sum_{c \ \in \ C \\ m \ \in \ P \\ k \ \in \ K} p(c, m, k) \log\left(\frac{p(c,m,k)}{p(m,k)}\right)$

(maybe this is clearer on the Wikipedia, sorry for my rubbish formatting skills) then notice that, if $c$ is totally conditioned on $(m, k)$ - because, like you said, each ciphertext corresponds to a unique plaintext-key pair with a bijective mapping - then it must be that $p(c,m,k) = p(m,k)$. Consequently, the fraction inside the logarithm evaluates to $1$ always, and so the logarithm evaluates to $0$, always. Then each term in the sum is $0$ and the over all conditional entropy is $0$, as you expect for a distribution where $(m,k)$ totally defines $c$.

As mentioned above, for probabilistic encryption, where a nonce or IV is used, this wouldn't be the case, because there would be more inputs to define $c$, so the entropy of those inputs would 'filter through' to $c$ leading to non-zero conditional entropy.