Bocconi University— 30540: Computer Science II Handout D

Fabrizio Iozzi and Luca Trevisan February 21, 2022

Analysis of Dijkstra’s Algorithm

We provide additional details on the analysis of Dijkstra’s algorithm, to complement

the sketch given in the textbook.
At the end of each “while” loop, we call the nodes that have been removed from the
queue black, and the nodes still in the queue white. We call a path from s to v an
all-black path if all the vertices in the path, except possibly v, are black.
We will assume that all edge lengths are non-negative.
We first make the following simple observations:

• At all times, and for every vertex v, if dist[v] is not ∞ then there is an all-black
path from s to v of length dist[v] (This is true at the beginning, and it is an
invariant that is maintained every time we update dist[.] of a vertex.
• For every vertex v, the value dist[v] can only be decreased or remain the same
from step to step, and it is never increased. (This can be verified by inspecting
the code.)
• If, at some point, for a certain vertex v, the value dist[v] equals the length
of the shortest path from s to v, then the value of dist[v] is never changed
in subsequent steps. (This is an immediate consequence of the previous two
observations.)

We prove that the following invariant holds.

Lemma 1 At the end of each iteration of the “while” loop:

1. All the black nodes v have a value of dist[v] equal to the length of the shortest
path from s to v, or ∞ if no such path exists.
2. All the nodes v have a value of dist[v] equal to the length of the shortest all-black
path from s to v.

At the end of the execution of the algorithm, all nodes are black, and so either of the
two invariants implies that the values of dist[v] equal the length of the shortest path
from s to v, for all v.

1
Base Case of Induction. Consider what happens after the first iteration of the
while loop: the node s is taken out of the queue and is the only black node. We have
dist[s] = 0 and indeed the distance of s to itself is 0 (because the “empty path” is
a valid path of length 0, and there cannot be negative-length paths). This verifies
the first condition. For the second condition, dist[s] = 0 is also the length of the
shortest all-black path from s to itself; for vertices v 6= s, an all black path from s to
v must be a one-edge path s → v. If the edge (s, v) exists, then at the end of the first
iteration we indeed have dist[v] = `(s, v), which is the length of the all-black shortest
path from s to v. If the edge (s, v) does not exists, then we have dist[v] = ∞ which
is again correct because no all-black path exists from s to v.

Inductive Step Part (1). Now we argue that if both conditions are true after t
iterations, then the first condition must be true after t + 1 iterations. Let v be the
vertex that is removed from the queue during the (t + 1)-th iteration of the “while”
loop. We just need to argue that dist[v] is equal to the length of the shortest path
from s to v.
The operations that we perform during this iteration do not modify the value of
dist[v], which remains what it was after the t-th iteration. We want to argue that,
at time t, the two properties that we assume true after t iterations mean that v has
a value dist[v] equal to the actual shortest path length from s to v. Suppose, toward
a contradiction, that there is a path from s to v of length L < dist[v]. Because of
the second property, such path must use some white vertex as an intermediate step.
Let u be the first white vertex we encounter along such path. Then the length of this
path up to u is ≥ dist[u], because it is an all-black path from s to u and the second
property is telling us that dist[u] is the minimum length of such paths. Furthermore,
the length of this path up to u is ≤ L, so we conclude that

dist[u] ≤ L < dist[v]

where all the value of dist[] refer at the end of the t-th iteration. But this is a
contradiction because v was chosen to be the white vertex of minimum dist[.], for the
values of dist[.] that we got at the end of the t-th iteration.

Inductive Step Part (2). Now we argue that if both conditions are true after t
iterations, then the second condition must be true after t + 1 iterations. As before,
let us call v the vertex that is taken out of the queue during iteration t + 1. We will
distinguish between the value distt [x] of a vertex x at the end of the t-th iteration,
and the value distt+1 [x] at the end of iteration t + 1.
If x is a vertex that was black at time t, then distt [x] was the length of the shortest
path from s to v, and we must have distt+1 [x] = distt [x]. Furthermore, by the second
property at time t we have that there is an all-black path from s to x of length distt [x],

2
so there is an all-black path from s to v of length distt+1 [x] at time t + 1 (and there
cannot be a shorter all-black path).
If we consider v, we argued above that distt+1 [v] = distt [v] and that distt+1 [v] is the
length of a shortest path from s to v. By property 2 applied at time t, we have that
there is an all-black path of length distt [v] at time t, and so there is an all-black path
from s to v of length distt+1 [v] at time t + 1 (and there cannot be a shorter all-black
path).
Finally consider a vertex y that is white at time t+1. Suppose toward a contradiction
that there is an all-black path from s to y of length L < distt+1 [y]. If the path
does not contain v, then the path was already all-black at time t, and it has length
L < distt+1 [y] ≤ distt [y] which contradicts property 2 at step t. If the path contains
v as last intermediate vertex, then the path has a s → v part of length L − `(v, y) plus
the edge `(v, y). Note that we must have distt [v] ≤ L − `(v, y), because distt [v] was
already the length of a shortest path from s to v. Note, however, that the updates
that we do at iteration t + 1 imply distt+1 [y] ≤ distt [v] + `(v, y), so we have

distt [v] + `(v, y) ≤ L < distt+1 [y] ≤ distt [v] + `(v, y)

which is again a contradiction. Finally, if the path contains v as intermediate vertex,

but not as the last intermediate vertex, let us call x the last intermediate vertex, so
the path of length L is a path s → v, followed by a path v → x, followed by the edge
(x, y). Let us call t0 ≤ t the iteration in which x was taken out of the queue. First,
we note that L ≥ distt0 [x] + `(x, y), because the path of length L is made of a path
from s to x, which must have length at least dt0 [x], plus the edge (x, y). We also see
that
distt+1 [y] ≤ distt0 [y] ≤ distt0 [x] + `(x, y) ≤ L
where the first inequality follows from the fact that t0 ≤ t, the second inequality
follows from the updates that we do at step t0 , and the third inequality was observed
above. We have now reached the contradiction L < L.