事实上，即使在没有可变状态的语言中，也很容易有一些看起来像可变状态的东西。

Consider a function with type s -> (a, s). Translating from Haskell syntax, it means a function which takes one parameter of type "s" and returns a pair of values, of types "a" and "s". If s is the type of our state, this function takes one state and returns a new state, and possibly a value (you can always return "unit" aka (), which is sort of equivalent to "void" in C/C++, as the "a" type). If you chain several calls of functions with types like this (getting the state returned from one function and passing it to the next), you have "mutable" state (in fact you are in each function creating a new state and abandoning the old one).

It might be easier to understand if you imagine the mutable state as the "space" where your program is executing, and then think of the time dimension. At instant t1, the "space" is in a certain condition (say for example some memory location has value 5). At a later instant t2, it is in a different condition (for example that memory location now has value 10). Each of these time "slices" is a state, and it is immutable (you cannot go back in time to change them). So, from this point of view, you went from the full spacetime with a time arrow (your mutable state) to a set of slices of spacetime (several immutable states), and your program is just treating each slice as a value and computing each of them as a function applied to the previous one.

好吧，也许这并不容易理解:-)

It might seem inneficient to explicitly represent the whole program state as a value, which has to be created only to be discarded the next instant (just after a new one is created). For some algorithms it might be natural, but when it is not, there is another trick. Instead of a real state, you can use a fake state which is nothing more than a marker (let's call the type of this fake state State#). This fake state exists from the point of view of the language, and is passed like any other value, but the compiler completely omits it when generating the machine code. It only serves to mark the sequence of execution.

举个例子，假设编译器给了我们以下函数:

readRef :: Ref a -> State# -> (a, State#)
writeRef :: Ref a -> a -> State# -> (a, State#)

从这些类似haskell的声明中转换，readRef接收到类似于指向“a”类型值和假状态的指针或句柄的东西，并返回由第一个形参和新的假状态指向的“a”类型值。writeRef类似，但是改变了所指向的值。

If you call readRef and then pass it the fake state returned by writeRef (perhaps with other calls to unrelated functions in the middle; these state values create a "chain" of function calls), it will return the value written. You can call writeRef again with the same pointer/handle and it will write to the same memory location — but, since conceptually it is returning a new (fake) state, the (fake) state is still imutable (a new one has been "created"). The compiler will call the functions in the order it would have to call them if there was a real state variable which had to be computed, but the only state which there is is the full (mutable) state of the real hardware.

(了解Haskell的人会注意到我简化了很多东西，并省略了一些重要的细节。对于那些想要看到更多细节的人，看看mtl中的Control.Monad.State，以及ST和IO(又名ST RealWorld)单子。

You might wonder why doing it in such a roundabout way (instead of simply having mutable state in the language). The real advantage is that you have reified your program's state. What before was implicit (your program state was global, allowing for things like action at a distance) is now explicit. Functions which do not receive and return the state cannot modify it or be influenced by it; they are "pure". Even better, you can have separate state threads, and with a bit of type magic, they can be used to embed an imperative computation within a pure one, without making it impure (the ST monad in Haskell is the one normally used for this trick; the State# I mentioned above is in fact GHC's State# s, used by its implementation of the ST and IO monads).

除了别人给出的很好的答案，想想Java中的Integer和String类。这些类的实例是不可变的，但这并不意味着仅仅因为它们的实例不可更改，这些类就毫无用处。不可变性给了你一定的安全性。您知道，如果使用String或Integer实例作为Map的键，则该键不能更改。将其与Java中的Date类进行比较:

Date date = new Date();
mymap.put(date, date.toString());
// Some time later:
date.setTime(new Date().getTime());

你已经无声地改变了地图中的一个键!使用不可变对象(如在函数式编程中)要干净得多。更容易推断会发生什么副作用——没有!这意味着程序员更容易，优化器也更容易。

For highly interactive applications such as games, Functional Reactive Programming is your friend: if you can formulate the properties of your game's world as time-varying values (and/or event streams), you are ready! These formulae will be sometimes even more natural and intent-revealing than mutating a state, e.g. for a moving ball, you can directly use the well-known law x = v * t. And what's better, the game's rules written such way compose better than object-oriented abstractions. For example, in this case, the ball's speed can be also a time-varying value, which depends on the event stream consisting of the ball's collisions. For more concrete design considerations, see Making Games in Elm.

TLDR:你可以在没有可变状态的情况下进行任何计算，但是当真正需要告诉计算机该做什么的时候，因为计算机只使用可变状态，你需要在某些时候改变一些东西。

有很多答案正确地说，没有可变状态就不能做任何有用的事情，我想用一些简单的(反)例子来支持这一点，以及一个普遍的直觉。

如果你看到任何一段被认为是“纯函数式”的代码，并且它是这样做的(不是真正的语言):

printUpToTen = map println [1..10]

这不是纯功能性的。有一个隐藏状态(stdout的状态)不仅被改变了，而且隐式地传递进来。看起来像这样的代码(同样不是真正的语言):

printUpToTen = map println stdout [1..10]

也是不纯的:即使显式地传入state (stdout)，它仍然是隐式突变的。

现在直观地说:可变状态是必要的，因为影响我们计算机的核心构建块是可变状态，这个构建块是内存。我们不能强迫计算机做任何事情，而不以某种方式操纵内存，即使我们的计算模型确实可以计算任何东西，而没有“内存”的概念。

Think of something like an old GameBoy Advance: in order to display something to the screen, you must modify the memory (there are certain addresses that are read many times a second that determine whats being put on the screen). Your computational model (pure functional programming) may not need state to operate, you may even be able to implement you model using an imperative, state manipulation model (like assembly) that abstracts the state manipulation, but at the end of the day, somewhere in you code you have to modify those addresses in memory for the device to actually display anything.

这就是命令式模型具有天然优势的地方:因为它们总是在操作状态，所以您可以非常容易地将其转换为实际修改内存。下面是一个渲染循环的例子:

while (1) {
   render(graphics_state);
}

如果你要展开循环，它看起来像这样:

render(graphics_state); // modified the memory
render(graphics_state); // modified the memory
render(graphics_state); // modified the memory
...

但在纯函数式语言中，你可能会得到这样的东西:

render_forever state = render_forever newState
    where newState = render state

展开(准确地说是压平)可以像这样可视化:

render(render(render(render(...state) // when is the memory actually changing??

// or if you want to expand it the other direction
...der(render(render(render(render(render(state) // no mutation

正如你所看到的，我们在状态上一遍又一遍地调用一个函数，状态是不断变化的，但我们从不改变内存:我们立即将它传递给下一个函数调用。即使我们的实现实际上在底层修改了一些表示状态的东西(甚至在适当的位置!)，它也不在正确的位置。在某些时候，我们需要暂停并修改内存中的正确地址，这涉及到突变。

我现在才开始讨论这个问题，但是我想为那些正在与函数式编程作斗争的人补充几点。

函数式语言维护与命令式语言完全相同的状态更新，但它们是通过将更新后的状态传递给后续的函数调用来实现的。这是一个沿着数轴移动的简单例子。您的状态是您当前的位置。

首先是命令式方式(在伪代码中)

moveTo(dest, cur):
    while (cur != dest):
         if (cur < dest):
             cur += 1
         else:
             cur -= 1
    return cur

现在是函数式的方式(在伪代码中)。我非常依赖三元运算符，因为我希望有命令式背景的人能够读懂这段代码。所以如果你不经常使用三元运算符(我总是避免它在我的命令式的日子)下面是它是如何工作的。

predicate ? if-true-expression : if-false-expression

您可以通过将一个新的三元表达式放在假表达式的位置来连接三元表达式

predicate1 ? if-true1-expression :
predicate2 ? if-true2-expression :
else-expression

考虑到这一点，下面是函数版本。

moveTo(dest, cur):
    return (
        cur == dest ? return cur :
        cur < dest ? moveTo(dest, cur + 1) : 
        moveTo(dest, cur - 1)
    )

这是一个简单的例子。如果这是在游戏世界中移动人，你就必须引入一些副作用，如在屏幕上绘制对象的当前位置，并根据对象移动的速度在每次调用中引入一些延迟。但你仍然不需要可变状态。

The lesson is that functional languages "mutate" state by calling the function with different parameters. Obviously this doesn't really mutate any variables, but that's how you get a similar effect. This means you'll have to get used to thinking recursively if you want to do functional programming. Learning to think recursively is not hard, but it does take both practice and a toolkit. That small section in that "Learn Java" book where they used recursion to calculate factorial does not cut it. You need a toolkit of skills like making iterative processes out of recursion (this is why tail recursion is essential for functional language), continuations, invariants, etc. You wouldn't do OO programming without learning about access modifiers, interfaces etc. Same thing for functional programming.

我的建议是做小Schemer(注意我说的是“做”而不是“读”)，然后做SICP的所有练习。当你完成时，你的大脑会和刚开始时不一样。

简单的回答是:你不能。

那么不变性有什么好大惊小怪的呢?

If you're well-versed in imperative language, then you know that "globals are bad". Why? Because they introduce (or have the potential to introduce) some very hard-to-untangle dependencies in your code. And dependencies are not good; you want your code to be modular. Parts of program not influence other parts as little as possible. And FP brings you to the holy grail of modularity: no side effects at all. You just have your f(x) = y. Put x in, get y out. No changes to x or anything else. FP makes you stop thinking about state, and start thinking in terms of values. All of your functions simply receive values and produce new values.

这有几个优点。

首先，没有副作用意味着程序更简单，更容易推理。不用担心引入程序的新部分会干扰并使现有的正在工作的部分崩溃。

其次，这使得程序的可并行性微不足道(有效的并行化是另一回事)。

第三，有一些可能的性能优势。假设你有一个函数:

double x = 2 * x

现在你输入一个3的值，得到一个6的值。每一次。但是在祈使句中也可以这样做，对吧?是的。但问题是，在命令式中，你可以做更多的事情。我可以:

int y = 2;
int double(x){ return x * y; }

但我也可以

int y = 2;
int double(x){ return x * (y++); }

命令式编译器不知道我是否会有副作用，这使得优化更加困难(即double 2不必每次都是4)。函数函数知道我不会——因此，它可以在每次看到“double 2”时进行优化。

现在，即使每次创建新值对于复杂类型的值在计算机内存方面看起来是难以置信的浪费，但它不必如此。因为，如果你有f(x) = y，并且x和y的值“基本相同”(例如，只有少数叶子不同的树)，那么x和y可以共享部分内存——因为它们都不会突变。

如果这个不可变的东西这么好，为什么我说没有可变状态就不能做任何有用的事情。如果没有可变性，整个程序就是一个巨大的f(x) = y函数。同样的道理也适用于程序的所有部分:只是函数，而且是“纯粹”意义上的函数。我说过，这意味着每次都是f(x) = y。因此，例如readFile("myFile.txt")每次都需要返回相同的字符串值。不是很有用。

因此，每个FP都提供了一些突变状态的方法。“纯”函数语言(例如Haskell)使用一些可怕的概念(如单子)来做到这一点，而“不纯”函数语言(例如ML)则直接允许这样做。

当然，函数式语言还带来了许多其他优点，使编程更加高效，比如一类函数等。