如果您没有插入足够多的元素导致频繁重散列(或冲突，如果您的HashSet不能调整大小)，那么HashSet当然可以为您提供常量时间访问的好处。但是对于有大量增长或收缩的集合，使用Treesets实际上可能会获得更好的性能，这取决于实现。

如果我没记错的话，平摊时间可以接近于一个功能性红黑树的O(1)。冈崎的书会有比我更好的解释。(或参阅他的出版物列表)

HashSet是O(1)来访问元素，所以这当然很重要。但是保持集合中对象的顺序是不可能的。

如果维护顺序(根据值而不是插入顺序)对您很重要，TreeSet是有用的。但是，正如您所注意到的，您正在以顺序换取访问元素的更慢时间:基本操作为O(log n)。

来自TreeSet的javadocs:

该实现为基本操作(添加、删除和包含)提供了log(n)的时间成本。

当然，HashSet实现要快得多——开销更少，因为没有排序。http://java.sun.com/docs/books/tutorial/collections/implementations/set.html提供了Java中各种Set实现的很好的分析。

这里的讨论还指出了一种有趣的“中间地带”方法来解决树与哈希的问题。Java提供了一个LinkedHashSet，它是一个HashSet，其中运行着一个“面向插入”的链表，也就是说，链表中的最后一个元素也是最近插入到哈希中的元素。这允许您避免无序散列的无序性，而不会增加TreeSet的成本。

A lot of answers have been given, based on technical considerations, especially around performance. According to me, choice between TreeSet and HashSet matters. But I would rather say the choice should be driven by conceptual considerations first. If, for the objects your need to manipulate, a natural ordering does not make sense, then do not use TreeSet. It is a sorted set, since it implements SortedSet. So it means you need to override function compareTo, which should be consistent with what returns function equals. For example if you have a set of objects of a class called Student, then I do not think a TreeSet would make sense, since there is no natural ordering between students. You can order them by their average grade, okay, but this is not a "natural ordering". Function compareTo would return 0 not only when two objects represent the same student, but also when two different students have the same grade. For the second case, equals would return false (unless you decide to make the latter return true when two different students have the same grade, which would make equals function have a misleading meaning, not to say a wrong meaning.) Please note this consistency between equals and compareTo is optional, but strongly recommended. Otherwise the contract of interface Set is broken, making your code misleading to other people, thus also possibly leading to unexpected behavior.

这个链接可能是关于这个问题的一个很好的信息来源。

基于@shevchyk在地图上可爱的视觉回答，以下是我的看法:

╔══════════════╦═════════════════════╦═══════════════════╦═════════════════════╗
║   Property   ║       HashSet       ║      TreeSet      ║     LinkedHashSet   ║
╠══════════════╬═════════════════════╬═══════════════════╬═════════════════════╣
║              ║  no guarantee order ║ sorted according  ║                     ║
║   Order      ║ will remain constant║ to the natural    ║    insertion-order  ║
║              ║      over time      ║    ordering       ║                     ║
╠══════════════╬═════════════════════╬═══════════════════╬═════════════════════╣
║ Add/remove   ║        O(1)         ║     O(log(n))     ║        O(1)         ║
╠══════════════╬═════════════════════╬═══════════════════╬═════════════════════╣
║              ║                     ║   NavigableSet    ║                     ║
║  Interfaces  ║         Set         ║       Set         ║         Set         ║
║              ║                     ║    SortedSet      ║                     ║
╠══════════════╬═════════════════════╬═══════════════════╬═════════════════════╣
║              ║                     ║    not allowed    ║                     ║
║  Null values ║       allowed       ║ 1st element only  ║      allowed        ║
║              ║                     ║     in Java 7     ║                     ║
╠══════════════╬═════════════════════╩═══════════════════╩═════════════════════╣
║              ║   Fail-fast behavior of an iterator cannot be guaranteed      ║
║   Fail-fast  ║ impossible to make any hard guarantees in the presence of     ║
║   behavior   ║           unsynchronized concurrent modification              ║
╠══════════════╬═══════════════════════════════════════════════════════════════╣
║      Is      ║                                                               ║
║ synchronized ║              implementation is not synchronized               ║
╚══════════════╩═══════════════════════════════════════════════════════════════╝

如果您没有插入足够多的元素导致频繁重散列(或冲突，如果您的HashSet不能调整大小)，那么HashSet当然可以为您提供常量时间访问的好处。但是对于有大量增长或收缩的集合，使用Treesets实际上可能会获得更好的性能，这取决于实现。

如果我没记错的话，平摊时间可以接近于一个功能性红黑树的O(1)。冈崎的书会有比我更好的解释。(或参阅他的出版物列表)