If you want to find paths with the minimum number of edges, then breadth-first search is the way to go. In other words, BFS is great for unweighted graphs.

What about weighted graphs? Let's remember what a weighted graph is. We assign a numeric weight to each edge. The weight can signify any quantity that we want, such as distance, time, cost, or penalty. The weight of a path is just the sum of the weights of the edges on the path. We typically want to find paths with minimum weight. We call a path from vertex u to vertex v whose weight is minimum over all paths from u to v a shortest path, since if weights were distances, a minimum-weight path would indeed be the shortest of all paths u to v.

As you can imagine, shortest paths come up frequently in real computer applications. In Lab Assignment 4, you will find shortest paths in a graph that represents the U.S. interstate highway system, and you'll display the shortest paths on a map. When a GPS finds a driving route, it is finding a shortest path. You can even use shortest paths to identify arbitrage opportunities in finance.

One key question is whether edge weights are allowed to be negative. The algorithm we'll see today, Dijkstra's algorithm, is guaranteed to find shortest paths only when all edge weights are nonnegative, such as when they represent distance, time, or monetary cost for driving. You might wonder when edge weights are negative. They come up less frequently than situations where all weights are nonnegative, but the algorithm for finding arbitrage opportunities relies on allowing negative edge weights. (We won't see that algorithm in CS 10, however.)

There is a situation that we have to watch out for when edge weights can be negative. Suppose that there is a cycle whose total weight is negative. Then you can keep going around and around that cycle, decreasing the cost each time, and getting a path weight of  − ∞. Shortest paths are undefined when the graph contains a negative-weight cycle.

Dijkstra's algorithm

When edge weights are required to be nonnegative, Dijkstra's algorithm is often the algorithm of choice. It's named after its inventor, Edsgar Dijkstra, who published it back in 1959. Yes, this algorithm is 55 years old! It's an oldie but a goodie. Dijkstra's algorithm generalizes BFS, but with weighted edges.

Dijkstra's algorithm finds a shortest path from a source vertex s to all other vertices. We'll compute, for each vertex v, the weight of a shortest path from the source to v, which we'll denote by v.dist. Now, this notation presumes that we have an instance variable dist for each vertex. We don't have to store shortest-path weights as instance variables. For example, we can use a map instead, mapping each vertex to its shortest-path weight. But for the purpose of understanding the algorithm, we'll use v.dist. We will also keep track of the back-pointer on each shortest path. Standard nomenclature for a back-pointer on a shortest path is a predecessor, and so we'll denote the predecessor of vertex v on a shortest path from the source by v.pred. As with the back-pointers in BFS, the predecessors in shortest paths form a directed tree, with edges pointing on paths toward the source.

I like to think of Dijkstra's algorithm as a simulation of sending out runners over the graph, starting at the source vertex s. Ideally, the simulation works as follows, though we'll see that Dijkstra's algorithm works slightly differently. It starts by sending out runners from the source vertex to all adjacent vertices. The first time a runner arrives at any vertex, runners immediately leave that vertex, headed to all of its adjacent vertices. Look at part (a) of this figure:

It shows a directed graph with source vertex s and weights next to the edges. Think of the weight of an edge as the number of minutes it would take a runner to traverse the edge.

Part (b) illustrates the start of the simulation, at time 0. At that time, shown inside vertex s, runners leave s and head toward its two adjacent vertices, t and y. The blackened vertex s indicates that we know that s.dist = 0.

Four minutes later, at time 4, the runner to vertex y arrives, shown in part (c). Because this runner is the first to arrive at y, we know that y.dist = 4, and so y is blackened in the figure. The shaded edge (s, y) indicates that the first runner to arrive at vertex y came from vertex s, so that y.pred  = s. At time 4, the runner from vertex s to vertex t is still in transit, and runners leave vertex y at time 4, headed toward vertices t, x, and z.

The next event, displayed in part (d), occurs one minute later, at time 5, when the runner from vertex y arrives at vertex t. The runner from s to t has yet to arrive. Because the first runner to arrive at vertex t arrived from vertex y at time 5, we set t.dist to 5 and t.pred to y (indicated by the shaded edge (y, t)). Runners leave vertex t, headed toward vertices x and y at this time.

The runner from vertex s finally arrives at vertex t at time 6, but the runner from vertex y had already arrived there a minute earlier, and so the effort of the runner from s to t went for naught.

At time 7, depicted in part (e), two runners arrive at their destinations. The runner from vertex t to vertex y arrives, but the runner from s to y had already arrived at time 4, and so the simulation forgets about the runner from t to y. At the same time, the runner from y arrives at vertex z. We set z.dist to 7 and z.pred to y, and runners leave vertex z, headed toward vertices s and x.

The next event occurs at time 8, as shown in part (f), when the runner from vertex t arrives at vertex x. We set x.dist to 8 and x.pred to t, and a runner leaves vertex x, heading to vertex z.

At this point, every vertex has had a runner arrive, and the simulation can stop. All runners still in transit will arrive at their destination vertices after some other runner had already arrived. Once every vertex has had a runner arrive, the distvalue for each vertex equals the weight of the shortest path from vertex s and the pred value for each vertex is the predecessor on a shortest path from s.

That was how the simulation proceeds ideally. It relied on the time for a runner to traverse an edge equaling the weight of the edge. Dijkstra's algorithm works slightly differently. It treats all edges the same, so that when it considers the edges leaving a vertex, it processes the adjacent vertices together, and in no particular order. For example, when Dijkstra's algorithm processes the edges leaving vertex s, it declares that y.dist  = 4, t.dist  = 6, and y.pred and t.pred are both sso far. When Dijkstra's algorithm later considers the edge (y, t), it decreases the weight of the shortest path to vertex t that it has found so far, so that t.dist goes from 6 to 5 and t.pred switches from s to y.

Implementing Dijkstra's algorithm

Dijkstra's algorithm maintains a min-priority queue of vertices, with their dist values as the keys. It repeatedly extracts from the min-priority queue the vertex u with the minimum dist value of all those in the queue, and then it examines all edges leaving u. If v is adjacent to u and taking the edge (u, v) can decrease v's dist value, then we put edge (u, v) into the shortest-path tree (for now, at least), and adjust v.dist and v.pred. Let's denote the weight of edge (u, v) by w(u,v). We can encapsulate what we do for each edge in a relaxation step, with the following pseudocode:

void relax(u, v) {
  if (u.dist + w(u,v) < v.dist) {
    v.dist = u.dist + w(u,v);
    v.pred = u;
  }
}

Whenever a vertex's dist value decreases, the min-priority queue must be adjusted accordingly.

Here is pseudocode for Dijkstra's algorithm, assuming that the source vertex is s:

void dijkstra(s) {
  queue = new PriorityQueue<Vertex>();
  for (each vertex v) {
    v.dist = infinity;  // can use Integer.MAX_VALUE or Double.POSITIVE_INFINITY
    queue.enqueue(v);
    v.pred = null;
  }   
  s.dist = 0;

  while (!queue.isEmpty()) {
    u = queue.extractMin();
    for (each vertex v adjacent to u)
      relax(u, v);
  }
}

In the following figure, each part shows the dist value (appearing within each vertex), the pred value (indicated by shaded edges), and the vertices in the min-priority queue (the vertices that are shaded, not blackened) just before each iteration of the while-loop:

The vertex that is newly blackened in each part of the figure is the vertex chosen as vertex u in in each iteration of the while-loop. In the simulation with runners, once a vertex receives dist and pred values, they cannot change, but here a vertex could receive dist and pred values in one iteration of the while-loop, and a later iteration of the while-loop for some other vertex u could change these values. For example, after edge (y, x) is relaxed in part (c) of the figure, the value of x.dist decreases from to 13, and x.pred becomes y. The next iteration of the while-loop in (part (d)) relaxes edge (t, x), and x.dist decreases further, to 8, and x.pred becomes t. In the next iteration (part (e)), edge (z, x) is relaxed, but this time x.dist does not change, because its value, 8, is less than z.dist + w(z,x), which equals 12.

When implementing Dijkstra's algorithm, you must take care that the min-priority queue is updated whenever a dist value decreases. For example, the implementation in the textbook uses a HeapAdaptablePriorityQueue, which has a method replaceKey. This method updates the min-priority queue's internal state to reflect the change in a vertex's key. The textbook also relies on a map from Vertex objects to Entry objects, which locate the vertex in the min-priority queue. It's a bit complicated.

You might wonder, since the min-priority queue is going to be implemented by a min-heap, why not just have the Dijkstra's algorithm code keep track of the index of each vertex's index in the array that implements the heap? The problem is that the premise violates our goal of abstraction. Who is to say that a min-heap implements the min-priority queue? In fact, we could just use an unsorted array. Or any other implementation of a min-priority queue. The Dijkstra's algorithm code should know only that it's using a min-priority queue, and not how the min-priority queue is implemented.

Correctness of Dijkstra's algorithm

How do we know that, at termination (when the min-priority queue is empty because every vertex has been dequeued), v.dist is in fact the shortest-path weight for every vertex v? We rely on a loop invariant argument. We state a property that pertains to a loop, which we call the loop invariant, and we have to show three things about it:

  1. The loop invariant is true before we enter the loop.
  2. If the loop invariant is true upon starting an iteration of the loop, it remains true upon starting the next iteration.
  3. The loop invariant, combined with the reason that we exit the loop, yields the property that we want to show.

To make our loop invariant a little simpler, let's adopt a notation for all vertices not in the min-priority queue at a given time; we'll call these vertices set S. Then here is the loop invariant:

At the start of each iteration of the while-loop, v.dist is the correct shortest-path weight for each vertex v in S.

It's easy to see that the loop invariant holds before the first iteration. At that time, all vertices are in the min-priority queue, and so set S is empty. Therefore, the loop invariant holds vacuously.

Next, let's see how the third part helps. When we exit the loop, the min-priority queue is empty, and so the set S consists of all the vertices. Therefore, every vertex has its correct shortest-path weight.

The hardest part is the second part, where we have to show that each iteration maintains the truth of the loop invariant. We'll give a simplified version of the reasoning. (A formal proof is a bit more involved.) Assume that as we enter an iteration of this loop, all vertices in set S have their correct shortest-path weights in their dist values. Then every edge leaving these vertices has been examined in some relaxation step. Consider the vertex u in the min-priority queue with the lowest dist value. Its dist value can never again decrease. Why not? Because the only edges remaining to be relaxed are edges leaving vertices in the min-priority queue and every vertex in the min-priority queue has a dist value at least as large as u.dist. Since all edge weights are nonnegative, we must have u.dist  ≤  v.dist + w(v,u) for every vertex v in the min-priority queue, and so no future relaxation will decrease u.dist. Therefore, u.dist is as low as it can go, and we can extract vertex u from the min-priority queue and relax all edges leaving u. That is, u.dist has its correct shortest-path weight and it is now in set S.

Running time of Dijkstra's algorithm

To analyze Dijkstra's algorithm, we'll let n denote the number of vertices and m denote the number of directed edges in the graph (as usual). Each vertex is enqueued once, at the start of the algorithm, and it is never enqueued again, and so there are n enqueue operations. Similarly, each vertex is dequeued once, in some iteration of the while-loop, and because it is never enqueued again, that one time is the only time a vertex is ever dequeued. Hence, there are n dequeue operations. Every edge is relaxed one time, and so there are m relaxation steps and, hence, at most m times that we need to update the min-priority queue because a key has changed. Let's use the generic term decreaseKey for updating the min-priority queue (what the textbook calls replaceKey; I like decreaseKey better, since it's for a min-priority queue). These are the relevant costs: n insert operations, n extractMin operations, and at most m decreaseKey operations. The running time of Dijkstra's algorithm depends on how these operations are implemented.

We can use an unsorted array for the min-priority queue. Each insert and decreaseKey operation takes Θ(1) time. Each extractMin operation takes time O(q), where q is the number of vertices in the min-priority queue at the time. We know that q ≤ n always, and so we can say that each extractMin operation takes O(n) time. But wait—what about when q is small? The way to think about it is to look at the time for all n extractMin operations altogether. You can do the analysis that results in an arithmetic series in n, or you can do it the way I like it. We have n operations, each taking O(n) time, and so the total time for all extractMin operations is O(n2). If we look at just the first n/2 extractMin operations, each of them has to examine at least n/2 vertices to find the one with the smallest dist value, and so just the first half of the extractMin operations take Ω(n2) time. Hence the total time for the n extractMin operations is Θ(n2), as is the total time for Dijkstra's algorithm.

What if we use a min-heap for the min-priority queue? Now each insert, extractMin, and decreaseKey operation takes O(lg n) time, and so the totel time for Dijkstra's algorithm is O((n + m) lg n). If we make the reasonable assumption that m ≥ n − 1 (otherwise, we don't even have a connected graph), then that running time is O(m lg n). That's always better than using an unsorted array, right? No. We say that a graph is dense if m = Θ(n2), that is if on average every vertex has Θ(n) neighbors. In a dense graph, Dijkstra's algorithm runs in time O(n2 lg n), which is worse than using an unsorted array.

Is there some way to implement a min-priority queue that gives us the best of both possibilities? Yes! There's a data structure called a Fibonacci heap that implements extractMin in O(lg n) amortized time (amortized over all the operations) and insert and decreaseKey in Θ(1) amortized time. With a Fibonacci heap, Dijkstra's algorithm runs in time O(n lg n + m), which is at least as good as using either an unsorted array or a min-heap. Fibonacci heaps are a little tricky to implement, and their hidden constant factors are a little worse than those for binary heaps, but they're not as hard to implement as some people seem to think. (Source: I have implemented them.)