Because a tangle on it's own works if everybody has the whole tangle, but in a real world scenario you need network logic. You cannot assume that everybody has the whole tangle like in the tangle whitepaper. And there is no information on this part besides:
Swarm Client
Another approach planned to enable the IOTA client running in these very resource restrained environments is to shard the core logic and database amongst different devices that then collectively run it. Similarly to swarm intelligence, this enables a cluster of devices to efficiently make transactions without being a full node, but having reduced trust requirements from SPV and light clients.
Status: Research / preliminary stage
In other words that means as long as you do not run a complete full node you have to trust others, which makes it more or completely centralized as the tangle grows.
Collectively run it
means you give your processing power and data in third party hands they compute something and then they have the current state (from that cluster). Because you cannot verify/confirm tx on your own (because you do not have the necessary information) you are completely dependent on third party, which makes it in fact centralized. And all that gets even worse as the tangle grows.