MATHEMATICAL INDUCTION, POWER SUMS, AND

DISCRETE CALCULUS

PETE L. CLARK

1. Something interesting to say about uninteresting induction proofs

I am currently teaching mathematical induction in a “transitions” course for prospec-
tive math majors. Inevitably the first, and apparently least interesting, examples
of induction proofs are identities like the following:
n(n + 1)
(1) 1 + ... + n = .
2
n(n + 1)(2n + 1)
(2) 12 + . . . + n2 = .
6
 2
n(n + 1)
(3) 13 + . . . + n3 = = (1 + . . . + n)2 .
2
It is well known to both instructors and students that such induction proofs quickly
boil down to an algebraic computation. For instance, to show (1), we verify that
1 = 1(1+1)
2 , then assume that for some fixed n,
n(n + 1)
1 + ... + n = .
2
Then we add n + 1 to both sides, getting
n(n + 1) n2 + n + 2n + 2
(1 + . . . + n) + n + 1 = +n+1=
2 2
2
n + 3n (n + 1)(n + 2) (n + 1)((n + 1) + 1)
= = = .
2 2 2
Many beginning students would like to know what is “really going on here”. One
answer is that we are simply applying the principle of mathematical induction via
a procedure which will be more clear after viewing other, similar examples. This
is a perfectly good answer but not the only possible one. We might get a better
answer by asking a more precise question, e.g. “Why is mathematical induction
particularly well suited to proving closed-form identities involving finite sums?”

I have a quite different answer to that question. Here goes: the task in finding
a closed form expression for a sum is to eliminate the “dot dot dot”. This is exactly
what induction does for us. In general, suppose f and g are functions from the
positive integers to the real numbers, and our task is to prove that
f (1) + . . . + f (n) = g(n).
Now let us contemplate proving this by induction.

1
2 PETE L. CLARK

First we verify that f (1) = g(1).

Now assume the identity for a fixed positive integer n, and add f (n) to both sides:
IH
f (1) + . . . + f (n) + f (n + 1) = (1 + . . . + f (n)) + f (n + 1) = g(n) + f (n + 1).
Since we want to get g(n + 1) in the end, what remains to be shown is precisely
that g(n + 1) = g(n) + f (n + 1), or equivalently
(4) g(n + 1) − g(n) = f (n + 1).
If f and g are both simple algebraic functions, then (assuming the result is actually
true!) verifying (4) is a matter of high school algebra. For example, to prove (2),
then – after checking that 12 = 1(1+1)(2·1+1)
6 – the identity we need to verify is
(n + 1)(n + 2)(2(n + 1) + 1) n(n + 1)(2n + 1)
− = (n + 1)2 ,
6 6
and we need only expand out both sides and see that we get n2 + 2n + 1 either way.

I wish to suggest that this procedure is analogous to whatR happens in calculus when
x
we have two functions f and g and wish to verify that 1 f (t) = g(x): it suffices,
dg
by the Fundamental Theorem of Calculus, to show that g(1) = 0 and dx = f . This
analogy may well seem farfetched at the moment, so let’s leave it aside and press on.

But not just yet. First let us consider a slightly different framework: we have
two functions f, g : Z+ → R, but instead of trying to show f (1) + . . . + f (n) = g(n),
we are trying to show that
f (1) + . . . + f (n) = g(1) + . . . + g(n).
Let us write F (n) = f (1) + . . . + f (n) and G(n) = g(1) + . . . + g(n), so we want
to show that F (n) = G(n) for all n. Suppose we try to prove this by induction.
We must show that F (1) = G(1), and then we get to assume that for a given n,
F (n) = G(n), and we need to show F (n + 1) = G(n + 1). Here’s the point: given
(5) F (n) = G(n),
the desired conclusion G(n + 1) = F (n + 1) is equivalent to
(6) F (n + 1) − F (n) = G(n + 1) − G(n).
Indeed, if (5) and (6) both hold, then adding them together, we get
F (n + 1) = F (n + 1) − F (n) + F (n) = G(n + 1) − G(n) + G(n) = G(n + 1),
and similarly, if we know F (n + 1) = G(n + 1), then subtracting (5), we get (6). So
our application of induction gives that it is necessary and sufficient to show that
F (n + 1) − F (n) = G(n + 1) − G(n). Note however that
F (n + 1) − F (n) = (f (1) + . . . + f (n) + f (n + 1)) − (f (1) + . . . + f (n)) = f (n + 1),
and similarly G(n + 1) − G(n) = g(n + 1). We need to show this for all n, i.e., we
need to know that f (n) = g(n) for all n ≥ 2. Since we also needed this for n = 1,
we see – perhaps somewhat sheepishly – that what we have shown is the following.
Proposition 1. Let f, g : Z+ → R. The following are equivalent:
(i) For all n ∈ Z+ , f (n) = g(n).
(ii) For all n ∈ Z+ , f (1) + . . . + f (n) = g(1) + . . . + g(n).
 MATHEMATICAL INDUCTION, POWER SUMS, AND DISCRETE CALCULUS 3

This is not earth-shattering, but the following minor variation is somewhat inter-
esting. Namely, for any function f , define a new function ∆f , by
(∆f )(n) = f (n + 1) − f (n).
The point here is that (∆f )(1) + . . . + (∆f )(n) =
(f (2)−f (1))+(f (3)−f (2))+. . .+(f (n)−f (n−1))+(f (n+1)−f (n)) = f (n+1)−f (1).
Since Proposition 1 holds for all functions f and g, in particular it holds for ∆f
and ∆g, and we get:

Proposition 2. Let f, g : Z+ → R. The following are equivalent:

(i) For all n ∈ Z+ , (∆f )(n) = f (n + 1) − f (n) = g(n + 1) − g(n) = (∆g)(n).
(ii) For all n ∈ Z+ , f (n + 1) − f (1) = g(n + 1) − g(1).

We easily deduce the following:

Theorem 3. Let f, g : Z+ → R. Suppose :

(i) For all n, (∆f )(n) = (∆g)(n), and
(ii) f (1) = g(1).
Then f (n) = g(n) for all n.

But wait! This is directly reminiscent of the following theorem.

Theorem 4. Let f, g : R → R be differentiable functions. Suppose:

(i) For all x ∈ R, f 0 (x) = g 0 (x), and
(ii) f (1) = g(1).
Then f (x) = g(x) for all x.

Proof. We define the new function h(x) = f (x) − g(x). Our hypotheses are that
h0 (x) = (f (x) − g(x))0 = f 0 (x) − g 0 (x) = 0 for all x and that h(1) = f (1) − g(1) = 0,
and we want to show that h(x) = 0 for all x. So suppose not, i.e., there exists some
x0 with h(x0 ) 6= 0. Certainly x0 6= 1, and we may assume without loss of generality
that x0 > 1. Now we apply the Mean Value Theorem to h(x) on the interval [1, x0 ]:
there exists a real number c, 1 < c < x0 , such that
h(x0 ) − h(1)
h0 (c) = .
x0 − 1
Thus h0 (c) 6= 0, contradicting the hypothesis that h0 (x) for all x ∈ R. 

If we think instead of two functions x(t) and y(t) giving the position of a moving
particle at time t, the Theorem states the following physically plausible result: two
moving bodies with identical instanteous velocity functions and which have the
same position at time t = 1 will have the same position for all times t. Theorem 3
applies to x(t) and y(t) as follows: we look only at positive integer values of x and
y, and we can interpret (∆x)(n) as the average velocity of x(t) between times n
and n + 1. The result then says that if x(t) and y(t) start out at the same position
at time t = 1 and their average velocities on each interval [n, n + 1] agree, then x
and y have the same positions at all positive integer values.
Note that we needed a deep theorem from calculus to show Theorem 4, but for
the analogous Theorem 3 we only needed mathematical induction.
4 PETE L. CLARK

2. Another natural question

There is something else in the proofs by induction of identities (1), (2), (3) that con-
fuses many students. Namely, how do we figure out what goes on the right hand
side?!? The schematic introduced in the previous section allows us to rephrase
this more crisply, as follows: given a function f (n) like f (n) = n, f (n) = n2 or
f (n) = n3 , how do we find the function g(n) such that g(n) = f (1) + . . . + f (n)?

Again, there is a simple answer which is perfectly good and probably indeed should
be the first answer given. Namely, we should clarify our task: we were not claiming
to be able to find – and still less, asking the student to find! – the right hand side
of these identities. It is important to understand that induction is never used to
discover a result; it is only used to prove a result that one already either suspects
to be true or has been asked to show. In other words, the simple answer to the
question “How do we figure out what goes on the right hand side?” is: we don’t.
It is given to us as part of the problem.

But this is a very disappointing answer. I feel that the disappointment this answer
engenders in students is of pedagogical significance, so forgive me while I digress on
this point (or skip ahead, of course). In other words, it often happens in university
level math classes that we present certain techniques and advertise them as giving
solutions to certain problems, but we often do not discuss the limitations of these
techniques, or more positively, try to identify the range of problems to which the
techniques can be successfully applied. For instance, after learning several inte-
gration techniques, many calculus students become anxious when they realize that
they may not know which technique or combination of techniques to apply to a
given proble. They often ask for simple rules like, ”Can you tell us when to use
integration by parts?” I at least have found it tempting as an instructor to brush
off such questions, or answer them by saying that much of the point is for them to
gain enough experience with the various techniques so as to be able to figure out
(or guess) which techniques will work on a given problem. But calculus instructors
2
know something that the students don’t: many functions, like ex , simply do not
have elementary antiderivatives. It would be a terrible disservice not to point this
out to the students, as well as not to clue them into the truth: we carefully select
the integration problems we give the students so that (i) elementary antiderivatives
exist and (ii) they can indeed be found using the set of tools we have taught them.
There are real dangers that such practices will dampen or kill off students’ math-
ematical curiosity. Most students initially think they are being asked to solve a
robust class of problems – and thus, they think that they should know how to
solve these problems, and are disturbed that they don’t – but eventually they
learn that less knowledge than they thought is actually needed to solve the ex-
ercise. This is intellectually deadening. What use it it to know how to prove
12 + . . . + n2 = n(n+1)(2n+1)
6 without knowing how to figure out what should go
on the right hand side? The answer is that there is no inherent use that I can
Rb
see (other than being able to compute a x2 dx using Riemann sums); it is just an
opportunity to demonstrate a mastery of a very narrow skill, which we identified
in the last section as being able to verify that g(n + 1) − g(n) = f (n).

Of course, sometimes there are necessary reasons for not answering the natural
 MATHEMATICAL INDUCTION, POWER SUMS, AND DISCRETE CALCULUS 5

questions: the answer may be very complicated! For instance, although I tell my
2
calculus students that the reason they cannot integrate ex is that it is provably
impossible, I do not give any indication why this is true: such arguments are well
beyond the scope of the course.

But such is not the case for 12 + . . . + n2 , and we now present a simple method to
derive formulas for the power sums
Sd (n) = 1d + . . . + nd .
We begin with the sum
n
X
(i + 1)d+1 − id+1 ,


S=
i=1

which we evaluate in two different ways. First, writing out the terms gives
S = 2d+1 −1d+1 +3d+1 −2d+1 +. . .+nd+1 −(n−1)d+1 +(n+1)d+1 −nd+1 = (n+1)d+1 −1.
Second, by first expanding out the binomial (i + 1)d+1 we get
n n      
X
X d+1 d d+1
S= (i + 1)d+1 − id+1 = id+1 + i + ... + i + 1 − id−1 =
i=1 i=1
1 d
n       n   n n
X d+1 d d+1 d+1 X d d+1 X X
( i + ... + i) = i + ... + i+ 1=
i=1
1 d 1 i=1
d i=1 i=1
d   d  
X d+1 X d+1
Sj (n) = Sj (n).
j=0
d+1−j j=0
j
Equating our two expressions for S, we get
d  
X d+1
(n + 1)d+1 − 1 = Sj (n).
j=0
j

Solving this equation for Sd (n) gives

P 
d−1 d+1

(n + 1)d+1 − j=0 j Sj (n) −1
(7) Sd (n) = .
(d + 1)
This formula allows us to compute Sd (n) recursively: that is, given exact formulas
for Sj (n) for all 0 ≤ j < d, we get an exact formula for Sd (n). And getting the ball
rolling is easy: S0 (n) = 10 + . . . + n0 = 1 + . . . 1 = n.

Example (d = 1): Our formula gives

1 1 n(n + 1)
1+. . .+n = S1 (n) = ( )((n+1)2 −S0 (n)−1) = ( )(n2 +2n+1−n−1) = .
2 2 2
Example (d = 2): Our formula gives 12 + . . . + n2 = S2 (n) =
(n + 1)3 − S0 (n) − 3S1 (n) − 1 n3 + 3n2 + 3n + 1 − n − 23 n2 − 23 n − 1
= =
3 3
2n3 + 3n2 + n n(n + 1)(2n + 1)
= .
6 6
6 PETE L. CLARK

We leave it as an exercise for the reader to derive (3) using this method.

This approach is not ideally suited to rapid calculation. In particular, if we wanted

a formula for S10 (n) then our method requires us to first derive formulas for S1 (n)
through S9 (n), which would be rather time-consuming. On the other hand (7) has
theoretical applications: with it in hand we can harness induction to a much more
worthy goal, namely the proof of the following result.
Theorem 5. For every positive integer d, there exist a1 , . . . , ad ∈ Q such that for
all n ∈ Z+ we have
nd+1
1d + . . . + nd = + ad nd + . . . + a1 n.
d+1
We leave this interesting induction proof as an exercise for the reader.

Again though, let us not neglect the natural question: the method presented gives
a recursive method for evaluating the power sums Sd (n). Can we find a closed
form expression for Sd (n) in general? The answer is yes, and we will derive it as
an application of the discrete calculus, a topic to which we turn to next.

3. Discrete calculus
Let f : Z+ → R be any function. We define the new function ∆f as follows:
(∆f )(n) = f (n + 1) − f (n).
We view ∆ as being an operator on the set V = {f : Z+ → R} of all functions from
Z+ to R. Specifically, it is called the forward difference operator.

There is also a backward difference operator ∇, which takes f to

(∇f )(n) = f (n) − f (n − 1).
Something must be said about (∇f )(1) = f (1) − f (0), since by our hypothesis
f need not be defined at 0. Let us make the convention of “extension by zero”:
whenever a formula or definition calls upon us to evaluate a function at a value
outside its stated domain, we assign it the value 0. In particular this means that
(∇f )(1) = f (1) − f (0) = f (1).
d
Above we remarked upon an analogy between the ∆ operator and the operator dx
of differentiation. We will further develop the analogy here to the extent that it
will seem reasonable to refer to think of ∆ (and also ∇) as the discrete deriva-
d
tive of f . But we note one nonanalogy : the operator dx is not well-defined for
any function f : R → R; it is only defined for differentiable functions (which is in
fact sort of a circular definition, but that’s in the nature of things). Certainly ∆
and ∇ are defined for all f ∈ V : every discrete function is “discretely differentiable”.

Recall that the usual derivative satisfies the following properties:

(D1) The derivative of a constant function is the zero function.

d df dg
(D2) dx (f + g) = dx + dx .
d df
(D3) dx (αf ) = α dx .
d df dg
(D4) dx (f g) = dx g + f dx .
 MATHEMATICAL INDUCTION, POWER SUMS, AND DISCRETE CALCULUS 7

d f g df −f dg
(D5) dx g = dx g2 dx .
dxn
(D6) dx = nxn−1 .

It is easy to check that ∆ satisfies (D1) through (D3):

(D1∆) If f is constant, then ∀n, f (n + 1) = f (n), so ∀n, (∆f )(n) = 0.

(D2∆) ∆(f +g)(n) = (f +g)(n+1)−(f +g)(n) = f (n+1)+g(n+1)−f (n)−g(n) =

f (n + 1) − f (n) + g(n + 1) − g(n) = (∆f )(n) + (∆g)(n).
(D3∆) ∆(αf )(n) = αf (n + 1) − αf (n) = α(f (n + 1) − f (n)) = (α∆f )(n).

Similarly, ∇ satisfies (D2) and (D3), but not quite (D1): if f (n) = C for all
n, then for all n ≥ 2 we have (∇f )(n) = f (n) − f (n − 1) = C − C = 0, but
(∇f )(1) = f (1) − f (0) = C − 0 = C.

On the other hand, if you try to verify the direct analogue of (D4) – namely
∆(f g) = ∆(f )g +f ∆(g) – you soon see that it does not work out: the left hand side
has two terms, the right hand side has four terms, and no cancellation is possible.
Another way to see that this formula cannot be correct is as follows: if x denotes
the function n 7→ n, then as with the usual derivative we have
∆(x)(n) = n + 1 − n = 1;
i.e., the discrete derivative of the identity function x is the constant function 1. But
from this and the product rule (D4), the power rule (D6) follows by mathematical
induction. However, let us calculate ∆(x2 ):
∆(x2 )(n) = (n + 1)2 − n2 = n2 + 2n + 1 − n2 = 2n + 1,
whereas
((∆x)x + x∆x)(n) = n + n = 2n.
Something slightly different does hold:
(∆f g)(n) = (∆f )(n)g(n) + f (n + 1)(∆g)(n).
This formula looks a bit strange: the left hand side is symmetric in f and g, whereas
the right hand side is not. Thus there is another form of the product rule:
(∆f g)(n) = ∆(gf )(n) = f (n)(∆g)(n) + (∆f )(n)g(n + 1).
A more pleasant looking form of the product rule is
(8) (∆f g) = f ∆g + (∆f )g + ∆f ∆g.
d
This formulation makes clear the relationship with the usual product rule for dx :
if f : R → R is a differentiable function, then defining ∆f to be f (x + h) − f (x),
one checks that (8) remains valid. Now dividing both sides by ∆x = h, we get
(∆f g) ∆g ∆f ∆f
=f + g+ ∆g.
h ∆x ∆x ∆x
As h → 0, ∆f df ∆g dg df
∆x → dx , ∆x → dx and the last term approaches dx · 0 = 0. So the
d
product rule for dx is a simplification of the discrete product rule, an approximation
which becomes valid in the limit as h → 0. Thinking back to calculus, it becomes
clear that many identities in calculus are simplfications of corresponding discrete
8 PETE L. CLARK

identities obtained by neglecting higher-order differentials.

Moral: the conventional calculus is more analytically complex than discrete calcu-
lus: in the former, one must deal rigorously and correctly with limiting processes,
whereas in the latter no such processes exist. Conversely, discrete calculus can be
more algebraically complex than conventional calculus. More on this later.

There are corresponding, but slightly different, identities for ∇:

(∇f g)(n) = (∇f )(n)g(n − 1) + f (n)(∇g)(n) = (∇f )(n)g(n) + f (n − 1)(∇g)(n).
Similarly, in place of (D5) we have something slightly different:
(∆f )(n)g(n) − f (n)(∆g)(n)
∆(f /g)(n) = ,
g(n)g(n + 1)
assuming that for all n, g(n) 6= 0. In place of (D6) we get something significantly
more complicated, which we will discuss in detail in the next section.

3.1. Higher discrete derivatives.

Of course there is no difficulty in defining the second, and higher-order, discrete

derivatives. Namely, we put ∆2 f = ∆(∆f ), and for k ∈ Z, we define ∆k f as
∆(∆k−1 f ). Explicitly
(∆2 f )(n) = (∆f (n+1)−f (n)) = f (n+2)−f (n+1)−(f (n+1)−f (n)) = f (n+2)−2f (n+1)+f (n).
Similarly,
(∆3 f )(n) = (∆f (n + 2) − 2f (n + 1) + f (n))
= f (n+3)−f (n+2)−2f (n+2)+2f (n+1)+f (n+1)−f (n) = f (n+3)−3f (n+2)+3f (n+1)−f (n).
This suggests the following general formula for the kth discrete derivative:
Theorem 6. Let f : Z+ → R be a discrete function. For any k ∈ Z+ , we have
n  
k k
X
k
(∆ f )(n) = (−1) f (n + k − i).
i=0
i

We leave the proof to the reader as an exercise in mathematical induction.

3.2. The discrete antiderivative.

In the usual calculus, one studies the inverse process to differentiation, namely
antidifferentiation. The plausible candidate for the discrete antiderivative is just
the summation operator Σ : f → Σf defined as
(Σf )(n) = f (1) + . . . + f (n).
Let us now calculate the composite operators ∆◦Σ and Σ◦∆ applied to an arbitary
discrete function f :
(∆◦Σ)(f )(n) = ∆(n 7→ f (1)+. . .+f (n)) = f (1)+. . .+f (n+1)−(f (1)+. . .+f (n)) = f (n+1).
Similarly,
(Σ◦∆)(f )(n) = Σ(n 7→ f (n+1)−f (n)) = f (2)−f (1)+. . .+f (n+1)−f (n) = f (n+1)−f (1).
 MATHEMATICAL INDUCTION, POWER SUMS, AND DISCRETE CALCULUS 9

So Σ and ∆ are very close to being inverse operators, but there is something slightly
off with the indexing. Now the ∇ operator proves its worth: we have
(Σ◦∇)(f )(n) = Σ(n 7→ f (n)−f (n−1)) = f (1)−f (0)+. . .+f (n)−f (n−1) = f (n)−f (0) = f (n)
and
(∇◦Σ)(f )(n) = ∇(n 7→ f (1)+. . .+f (n)) = (f (1)+. . .+f (n))−(f (1)+. . .+f (n−1)) = f (n).
So indeed ∇ and Σ are inverse operators. This is even better than in the usual
calculus, where the antiderivative is only well-determined up to the addition of a
constant. In other words:
Theorem 7. (Fundamental Theorem of Discrete Calculus, v. 1) For functions
f, g : Z+ , R, the following are equivalent:
(i) For all n ∈ Z+ , f (n) = g(n) − g(n − 1) (f = ∇g).
(ii) For all n ∈ Z+ , g(n) = f (1) + . . . + f (n) (Σf = g).
Pn
In other words, if we want to find a closed form expression for i=1 f (i), it suffices
to find a function g such that ∇g = f .

3.3. The discrete definite integral.

Thinking carefully on traditional calculus reveals a certain discrepancy between

Theorem 7 and the usual fundamental theorem. Namely, in usual calculus the
fundamental theorem relates two notions of integrals: the indefinite and the defi-
nite integral. The indefinite integral of f is by definition any function F such that
dF
dx = F . In this sense, we found that Σf is an antiderivative of f : (∇Σf )(n) = f (n)
for all n ∈ Z+ . But what about the definite integral? Traditionally, the definite
Rb
integral of a (say, continuous) function f : [a, b] → R is a real number a f (x)dx
which is supposed to represent the signed area under the curve y = f (x) (and is
formalized as a limit of Riemann sums).

It is not difficult to come up with a discrete definite integral. There is a small

choice to be made, as to whether we use ∆ or ∇ as our differentiation operator.
Since in the previous section we developed a ∇-theory, to show that it doesn’t make
any essential difference, this time we set things up so as to work well with ∆.

Namely, for positive integers a ≤ b and a discrete function f : Z+ → R, we define

Sab f = f (a) + . . . + f (b − 1).1
This has the following property: we can define a new function F as a definite
integral with variable upper limit:
F (n) = S1n f = f (1) + . . . + f (n − 1).
Then we have a perfect analogue of the fundamental theorem of calculus, in two
parts. First,
(∆F )(n) = (f (1) + . . . + f (n)) − (f (1) + . . . f (n − 1)) = f (n),

1Stopping the sum at f (b − 1) rather than f (b) is the correct normalization for ∆, as we will
shortly see.
10 PETE L. CLARK

so that F is a function whose discrete derivative – this time ∆ and not ∇ – is equal
to f . Second, for 1 ≤ a ≤ b, we have
F (b) − F (a) =

(f (1) + . . . + f (b − 1)) − (f (1) + . . . + f (a − 1)) = f (a) + . . . + f (b − 1) = Sab f.

In summary:

Theorem 8. (Discrete fundamental theorem of calculus, v. 2) Let f : Z+ → R be

any function. For any two integers a ≤ b, define Σba f = f (a) + . . . + f (b − 1), and
define a function F : Z+ → R by F (n) = S1n f . Then:
a) ∆F = f .
b) Sab f = F (b) − F (a).

Remark: Of course in the traditional calculus the definite integral Σba f (x)dx has an
area interpretation. This can be given to the discrete definite integral Σba f as well.
Namely, for a discrete function f : Z+ → R, extend it to a function f : [1, ∞) → R
by f (x) = f (bxc). Thus f is the unique step function whose restriction to Z+ is f
and is left-continuous at each integer value. Then one has
Z b
f dx = Sab f,
a

i.e., the area under the step function f : [1, n] is precisely f (1) + . . . + f (n − 1).

Remark: We hope the reader has by now appreciated that the distinction be-
tween a discrete calculus based on ∆ versus one based on ∇ is very minor: in the
former case, we define the antiderivative to be f 7→ (n 7→ f (1) + . . . + f (n − 1))
(and similarly for the definite integral) Σba f and the in the latter case we define the
antiderivative to be f 7→ (n 7→ f (1) + . . . + f (n)). So we could make do by choosing
either one once and for all as the discrete derivative; on the other hand, there is no
compelling need to do so.

Remark: It would also make sense to consider functions f : Z → R. Then one

defines ∆ as before, and ∇ even more simply than before, namely (∇f )(n) =
f (n) − f (n − 1) for all n ∈ Z: no convention about f (0) is necessary. In this context
the definite integral Σba f still makes perfect sense, and it is more clear that the
notion of an indefinite integral well-defined only up to a constant. For instance, as
we did aove we can define a discrete antiderivative F of f by
F (n) = Σn−1
a f (n)
for any fixed a ∈ Z. Conversely, it is easy to see that all discrete antiderivatives
of f differ from each other by an additive constant. Equivalently, the choice of a
particular antiderivative F is determined by specifying its value at a single integer.

3.4. Discrete integration by parts.

Let f, g : Z → R be discrete functions. Summing the product rule

(∆f g)(k) = (∆f )(k)g(k) + f (k + 1)(∆g)(k)
 MATHEMATICAL INDUCTION, POWER SUMS, AND DISCRETE CALCULUS 11

from k = 1 to n yields
n
X n
X
f (n + 1)g(n + 1) − f (1)g(1) = (∆f )(k)g(k) + f (k + 1)(∆g)(k),
k=1 k=1
or
n
X n
X
(∆f )(k)g(k) = f (n + 1)g(n + 1) − f (1)g(1) − f (k + 1)(∆g)(k).
k=1 k=1

Applying this with f (n) = n gives

n
X n
X
(9) g(k) = (n + 1)g(n + 1) − g(1) (k + 1)(∆g)(k).
k=1 k=1

Taking g(n) = nd , we get

n
X n
X
Sd (n) = k d = (n + 1)d+1 − 1 − (k + 1)((k + 1)d − k d ).
k=1 k=1
...
RELATE THIS TO THE CALCULATION OF §2.

Example: Let h(n) = n2n . Evaluate H(n) = g(1) + . . . + g(n).

Solution: If instead we are asked to find the antiderivative of xex , we would use
integration by parts, writing h(x) = g(x)df (x), with g(x) = x, df (x) = ex dx. Then
Z Z
xe dx = xe − ex = xex − ex + C = (x − 1)ex + C.
x x

Let’s try the discrete analogue: put (∆f )(n) = 2n , g(n) = n, so that h(n) =
(∆f )(n)g(n). Then f (n) = 2n , ∆g = 1, so
n
X n
X
H(n) = (∆f )(k)g(k) = (n + 1)2n+1 − 2 − 2k+1 = (n + 1)2n+1 − 2 − (2n+2 − 4)
k=1 k=1
n+1
=2 (n + 1 − 2) = (n − 1)2n+1 + 2.
Thus the result is closely analogous but not precisely what one might guess: note
the 2n+1 in place of ex .
3.5. Some discrete differential equations.

Example: Let us find all discrete functions f : Z+ → R with (∆f ) = f . In

other words, for all n ≥ 1, we have f (n + 1) − f (n) = f (n), or f (n + 1) = 2f (n).
Evidently we need f (n) = 2n−1 f (1), so that the general solution is f (n) = C2n ,
for any real number C. Thus in some sense 2 is the discrete analogue of e!

Example: For α ∈ R, let us find all discrete functions f with (∆f ) = αf . For
all n ≥ 1, we have f (n + 1) − f (n) = αf (n) or f (n + 1) = (α + 1)f (n). The general
solution is then f (1) = C (arbitrary) and f (n) = (α + 1)n−1 C. Here we are using
the convention that 00 = 1, so that the general solution to ∆f = −f is given by
Cδ1 , where δ1 (1) = 1, δ1 (n) = 0 for all n > 1.

Example: The general solution to ∆2 f = f is n 7→ C1 2n + C2 δ1 (n).

12 PETE L. CLARK

Example: The general solution to ∆k f = f is the√ set of all real-valued functions in

the complex span of h(ζki + 1)n i, where ζk = e2π −1/k and 0 ≤ i < k.

4. Linear algebra of the discrete derivative

Let us reconsider the set V of all functions f : Z+ → R from the perspective of
linear algebra. As we saw above, we can naturally add two elements of V and also
multiply any element of V by any real number. Moreover these operations satisfy
all the usual algebraic properties, like commutativity and associativity of addition,
and so forth. What we are trying to say is:
Proposition 9. The set V = {f : Z+ → R} of discrete functions has the structure
of a vector space over R.
It is however a very large R-vector space. In particular it is not finite-dimensional:
for each n ∈ Z+ , define a function δn which maps n to 1 and every other positive
integer to 0. Then {δn }∞n=1 is an infinite linearly independent set.
2

Moreover, the operators ∆, ∇ and Σ are all linear operators on the vector space V ,
and ∇ and Σ are mutually inverse.

Again, that’s nice, but to compute things we would rather have a finite dimen-
sional vector space. In fact we can define an infinite family of finite dimensional
subspaces Pd ⊂ V , as follows.

First, let P ⊂ V be the set of all polynomial functions. In other words, it is

the span of the set of power functions {xd | d ∈ N}, where as before
xd : n 7→ nd .
It is easy to check that the set of power functions is linearly independent, hence a
basis for P, so that P is again infinite dimensional. However, for d ∈ N, define Pd
to be the span of 1, x, . . . , xd , i.e., the set of polynomial functions of degree at most
d: evidently Pd is a d + 1-dimensional subspace of V . Now they key fact:
Proposition 10. Let d ∈ Z+ . If P (x) is a degree d polynomial function, then its
discrete derivative ∆P : n 7→ P (n + 1) − P (n) is a degree d − 1 polynomial function.

Proof. First compute the discrete derivative of a monomial function xd : n 7→ nd :

(∆xd )(n) = (n + 1)d − nd =
       
d d + 1 d−1 n+1 d d + 1 d−1 n+1
n + n +...+ n+1−n = n +...+ n + 1.
1 n d n
This is, as claied, a polynomial of degree d − 1. From the linearity of ∆ it follows
that the discrete derivative of any polynomial function P (x) = ad xd + . . . + a1 x + a0
is a sum of polynomials of degree at most d − 1, so certainly it is a polynomial of
degree at most d − 1. What remains to be seen is that – assuming ad 6= 0, of
course – the coefficient of nd−1 in the polynomial (∆P )(n) = P (n + 1) − P (n) is

2Moreover, this set is not a basis: its span is the set of all functions which are zero for all
sufficiently large n. But this is not a key point for us.
 MATHEMATICAL INDUCTION, POWER SUMS, AND DISCRETE CALCULUS 13

nonzero. Direct calculation shows that this coefficient is (d+1)ad −ad = dad , which
is nonzero since we assumed that d > 0 and ad 6= 0. 
An immediate consequence of Proposition 10 is
∆(Pd ) ⊆ Pd−1 ⊆ Pd .
In particular, ∆ is a linear operator on the finite-dimensional vector space Pd .
Moreover its kernel is the set of constant functions. Indeed, if for any discrete
function f we have f (n + 1) − f (n) = 0 for all n, then by induction we have
f (n) = f (1) for all n. (This is the discrete analogue of the fact that a function with
identically zero derivative must be constant.) Therefore the kernel of ∆ on Pd is
one-dimensional. Recall the following fundamental fact of linear algebra: for any
linear operator L on a finite-dimensional vector space W , we have
dim(ker(L)) + dim(L(W )) = dim L.
Therefore we find that dim(∆(Pd )) = d. On the other hand, we know from the
proposition that the image ∆(Pd ) is contained in the d-dimensional subspace Pd−1 .
Therefore we must have equality:
Theorem 11. We have ∆(Pd ) = Pd−1 .
What is the significance of this? Applied to the function xd−1 ∈ Pd , we get that
there exists a degree d polynomial Pd (x) such that for all n ∈ Z+ ,
∆(Pd )(n) = Pd (n + 1) − Pd (n) = nd .
Since ∆(Pd )(n) = ∇(Pd )(n + 1), we have also
∇(Pd )(n + 1) = nd ,
and applying the summation operator Σ to both sides we get that for all n ∈ Z+ ,
Pd (n + 1) = 1d + . . . + nd .
Thus we’ve shown that there must be a nice closed form expression for the sum of
the first n dth powers: indeed, it must be a polynomial of degree d + 1.

Now we have undeniably done something worthwhile. But of course we won’t

stop here: we would like to actually compute the polynomials Pd !

But first we should address the following concern: given that it is ∇ and not
∆ which is the inverse to Σ on V , why did we work first with ∆ and only at the
end “shift variables” to get back to ∇?
The answer is that the linear map ∇ does not map the space of polynomial func-
tions to itself. It’s close, but remember that ∇ of the nonzero constant function C
is not zero: rather it is the function which is C at n = 1 and 0 for all larger values
of n, but this is not a polynomial function of any degree. (Recall that a degree d
polynomial function can be zero for at most d distinct real numbers.) So ∆ it is.

Now we have a finite dimensional vector space Pd , a linear operator ∆ on Pd ,

and a fixed basis 1, x, . . . , xd of Pd . So we can write down a matrix representing ∆.

d
Note that the usual derivative operator dx also carries Pd linearly onto Pd−1 so
gives a linear operator on Pd . In this classical case the corresponding matrix is
14 PETE L. CLARK

very simple: it is the matrix (INSERT ME).

The linear algebra of ∆ is more complicated (and more interesting!). E.g., since
∆(1) = 0,
∆(x) = 1,
∆(x ) : n 7→ (n + 1)2 − n2 = 2n + 1,
2

∆(x3 ) : n 7→ (n + 1)3 − n3 = 3n2 + 3n + 1,

∆(x4 ) : n 7→ (n + 1)4 − n4 = 4n3 + 6n2 + 4 + 1,
the matrix representation of ∆ on P4 with respect to the basis 1, x, x2 , x3 , x4 is
 
0 1 1 1 1
 0 0 2 3 4 
 
 0 0 0 3 6 
 
 0 0 0 0 4 
0 0 0 0 0
For any d, let Md be the matrix for ∆ on the basis 1, x, x2 , . . . , xd of Pd . Then

0,
if i ≥ j
(Md )ij = j
i , if i < j

To find the discrete antiderivative of the function xd−1 then, it suffices to solve the
matrix equation M [ad , ad−1 , . . . , a0 ]t = [0, . . . , 0, 1, 0]t for a0 , . . . , ad .

Moreover, because the matrix is in upper triangular form, we can easily solve the
linear system by back substitution. For example, when d = 4 we get:
a1 + a2 + a3 + a4 = 0,
+2a2 + 3a3 + 4a4 = 0,
3a3 + 6a4 = 0,
4a4 = 1.
(The last equation reads 0 = 0.) So we have a4 = 41 , and then
1 −1
a3 = (−6a4 ) = ,
3 2
1 1
a2 = (−3a3 − 4a4 ) = ,
2 4
a1 = −a2 − a3 − a4 = 0.
Note that the constant term a0 is undetermined, as it should be. It follows from
the above analysis that it doesn’t matter what constant term we take, so we may
as well take a0 = 0. Thus
1 1 1
P4 (x) = x4 − x3 + x2 ,
4 2 4
and we easily calculate
2
n4 n3 n2

n(n + 1)
P4 (n + 1) = + + = .
4 2 4 2
 MATHEMATICAL INDUCTION, POWER SUMS, AND DISCRETE CALCULUS 15

Since (Σ x3 )(n) = P4 (n + 1) for all n, we get the identity

 2
n(n + 1)
13 + . . . + n 3 = .
2

5. More linear algebra of the discrete derivative

5.1. Finding a better basis.

As we have seen, the basis 1, x, . . . , xd is a very nice one for the linear operator
d
dx on the space Pd : the matrix representation is . . ..
2 d
d
There is in fact an even better basis for dx on Pd : namely 1, x, x2 , . . . , xd , be-
cause with respect to this basis the matrix is . . ., a shift operator.

An important lesson in linear algebra is to find the best basis for the problem
at hand. More specifically, given a linear operator T on a finite-dimensional vector
space V over a field k, then under the assumption that all of the eigenvalues of
V are elements of k, there exists a Jordan basis for V with respect to T , i.e., a
basis in which V decomposes as a direct sum of T -stable subspaces Wi such that
Ti restricted to each Wi is the sum of a shift operator and a scalar.

The fact that the discrete derivative of a degree k polynomial is a degree k − 1

polynomial implies that the only eigenvalue of ∆ on Pd is zero and that the 0-
eigenspace is one-dimensional. This means that we have only one Jordan block, so
that there exists a basis p0 , . . . , pd of Pd with respect to which ∆ is a shift operator:
∆(pi ) = pi−1 for i > 0, ∆(p0 ) = 0.

The constant polynomial 1 generates the 0-eigenspace, so we can take p0 = 1.

We want p1 (x) such that ∆(p1 ) = 1; for this we can take p1 (x) = x. Next we want
p2 (x) such that ∆(p2 ) = p1 , and as we have seen, here p2 = 21 x2 is not quite right:
∆( 12 x2 )(n) = 12 (2x + 1) = x + 21 . So instead we need to take p2 (x) = 21 x2 − 12 x =
x(x−1)
2 . Similarly, one finds that we can take p3 (x) = x(x−1)(x−2)
6 . After enough
computation, the pattern becomes clear:
x(x−1)···(x−k+1)
Proposition 12. Define p0 (x) = 1, and for k ≥ 1, define pk (x) = k! .
Then ∆(p0 ) = 0 and for all k ≥ 1, ∆(pk ) = pk−1 .
Proof. Notice that pk (x) is nothing else than the binomial coefficient xk , viewed

as a polynomial in x! The result we are trying to prove is then x+1

x
x


k − k = k−1 ,
which is equivalent to the well-known binomial coefficient identity
     
x+1 x x
(10) = + ,
k k−1 k
which for instance can be verified combinatorially: it is enough to show it for
x ∈ Z+ , in which case the left hand side is the number N of k-element subsets of
{1, . . . x + 1}. Because every such subset either contains x + 1 or it doesn’t (and
not both!), we have N = N1 + N2 , where N1 is the number of k − 1 element subsets
of {1, . . . , x} and

N2 is the xnumber of k element subsets of {1, . . . , x}. That is to
x


say, N1 = k−1 and N2 = k . 
16 PETE L. CLARK

Pn−1
Corollary 13. a) 1 p0 (x) = p1 (n) − p1 (1) = p1 (x) − 1.
Pn−1
b) For all k ≥ 1, 1 pk (x) = pk+1 (n) − pk+1 (1) = pk+1 (n).
d
To stress the analogy between ∆ and dx , it is common to define the falling powers
xk = (k!)pk (x) = x(x − 1) · · · (x − k + 1).
So we have found what linear algebra tells us is the optimal basis for Pd (and, in
fact, for the infinite dimensional vector space P). What benefits do we reap?
5.2. The discrete Taylor series of a polynomial function.
Another aspect of the philosophy of the optimal basis {b1 , . . . , bn } for a vector space
V is that, upon expressing v as a linear combination of the basis vectors:
v = a1 b1 + . . . + an bn , ai ∈ R,
we expect that the coefficients ai will be natural functions of v.

Example: Suppose V is equipped with an inner product h, i and we choose an

orthonormal basis b1 , . . . , bn , then ai = hv, bi i.

xi
Example: Suppose that V = Pd and we choose the basis b0 = 1, bi = i! . We
know there are unique real numbers a0 , . . . , ad such that
x2 x3 xd
P (x) = a0 + a1 x + a2 + a3 + . . . + ad .
2 3! d!
But this is just the Taylor series for P (x). Explicitly, repeated evaluation at 0 and
k
differentiation gives ak = ddxP (0) for all 0 ≤ k ≤ d.

Now we keep V = Pd but consider instead the natural basis p0 (x), . . . , pd (x) for the
discrete derivative ∆. We get that for any P ∈ Pd unique real numbers a0 , . . . , ad
such that
x(x − 1) x(x − 1) · · · (x − d + 1)
P (x) = a0 + a1 x + a2 + . . . + an .
2 d!
Again, evaluating at x = 0 we find
a0 = P (0).
Taking ∆ of both sides gives
(∆P )(x) = a1 + a2 p1 (x) + . . . + ad pd−1 x,
and evaluating at 0 gives a1 = (∆P )(x). Continuing on in this way, we find that
for all 0 ≤ k ≤ d, ak = (∆k P )(0). So we have shown
Theorem 14. For any P (x) ∈ Pd , we have P (x) =
d d k  
X x(x − 1) · · · (x − k + 1) X X k
(∆k P )(0) = (−1)j P (k − j)pk (x).
k! j=0
j
k=0 k=0

Applying this theorem with P (x) = xd , we get:

 
d k  
X X k
P (n) = nd =  (−1)j (k − j)d  pk (n).
j=0
j
k=0
 MATHEMATICAL INDUCTION, POWER SUMS, AND DISCRETE CALCULUS 17

Pn−1
Taking 1 of both sides and applying Corollary 13, we get at last a closed form
expression for an arbitrary power sum:
 
d k  
X X k n(n − 1) · · · (n − k)
(11) 1d + . . . + (n − 1)d =  (−1)j (k − j)d  .
j=0
j (k + 1)!
k=1

5.3. Integer-valued polynomials.

First, the polynomials pk have the following interesting property:

Proposition 15. For every k ∈ N and n ∈ Z, pk (n) ∈ Z.
Note that if the polynomials pk (x) had coefficients in Z, this would be obvious. But
k
the leading coefficient of pk (x) is xk! , so this is certainly not the case for k ≥ 2. But
nevertheless, for any integer n, we have p2 (n) = n(n−1) 2 . Since either n or n − 1 is
an even integer, p2 (n) is still an integer.

It is a fairly standard exercise to prove Proposition 15 by a double induction on

n and k, using the basic binomal coefficient identity (10). But with the discrete
calculus in hand we can give a more thematic proof. First:
Lemma 16. For f : Z → R a discrete function, the following are equivalent:
(i) For all n ∈ Z, f (n) ∈ Z.
(ii) For all n ∈ Z, (∆f )(n) ∈ Z, and there exists N0 ∈ Z such that f (N0 ) ∈ Z.
The proof of Lemma 16 is immediate and is left to the reader.

Moreover, Proposition 15 swiftly follows: certainly p0 (x) = 1 maps integers to

integers. Since ∆pk = pk−1 and pk (0) = 0 for all k, we’re done by induction on k.

Remark: For those of a number-theoretic bent, we note the following restatement

of Proposition 15: the product of any k consecutive integers is divisible by k!.

In general, we say a polynomial f (x) ∈ R[x] is integer-valued if for all n ∈ Z,

f (n) ∈ Z. It is not hard to see that this implies that f (x) ∈ Q[x] (i.e., all coeffi-
cients of f are rational numbers), using e.g. the Lagrange Interpolation Formula.
But in fact a much more precise and beautiful result holds.
Theorem 17. Let f (x) ∈ R[x] be a polynomial of degree d. There are unique
a0 , . . . , ad ∈ R such that f (x) = a0 p0 (x) + . . . + ad pd (x). Moreover, TFAE:
(i) All of the coefficients a0 , . . . , ad lie in Z.
(ii) f (x) is integer-valued.
(iii) There is N0 ∈ Z such that f (N0 ), f (N0 + 1), . . . , f (N0 + d) are all integers.
Proof. The existence of unique a0 , . . . , ad such that f (x) = a0 p0 (x) + . . . ad pd (x)
merely expresses the fact that p0 , . . . , pd gives a basis for Pd .
(i) =⇒ (ii): By Proposition 15 we know that each pd (x) is integer-valued, and a
Z-linear combination of integer-valued polynomials is integer-valued.
(ii) =⇒ (iii) is immediate.
(iii) =⇒ (i): Suppose first that N0 = 0. By Theorem 14 we have that the ak ’s are
18 PETE L. CLARK

precisely the Taylor coefficients at 0:

d  
X k
ak = (∆k f )(0) = (−1)k f (k − i).
i=0
i
Since f (0), . . . , f (d) ∈ Z by hypothesis, evidently ak ∈ Z. The case of arbitary
N0 follows easily by a change of variables argument: for instance, by making the
evident generalization of Theorem 14 to Taylor series expansions about N0 . We
leave the details to the interested reader. 
5.4. Stirling numbers and a formula for Sd (n).