Suppose we want to prove Euler's Formula with induction for all positive real numbers.
At first this seems baffling, but an idea struck my mind today.

Prove: $$e^{ix}=\cos x+i\sin x \ \ \ \forall \ \ x\geq0$$

For $x=0$, we have $$1=1$$ So the equality holds.
Now let us assume that the given equality holds for some $x=k$.
$$e^{ik}=\cos k+i\sin k$$ Now, this is where I added my "own" axiom. Please answer whether this "axiom" is true or not. Now this equality must hold for $x=k+\Delta k$ also, for some infinitely small positive change $\Delta k$ (infinitesimal).
So $e^{i(k+\Delta k)}=e^{ik}.e^{i\Delta k}=(\cos k+i\sin k)(\cos\Delta k+i\sin\Delta k)$
$$=\cos k\cos\Delta k-\sin k\sin\Delta k+i\sin k\cos\Delta k+i\cos k\sin \Delta k$$ $$=\cos(k+\Delta k)+i\sin(k+\Delta k)$$ So we proved it for $x=k+\Delta k$, and hence it must hold for all $x\geq0$.
So is this approach correct, and is this a new type of induction?

Most of the comments below are indicating that the proof above is wrong. But we see the Euler's Formula does hold for all $x\geq0$, so can someone give a good counter example where this proof doesn't work. Or in other words, these statements are used to prove a wrong equality.?

Update: Okay, some comments below are suggesting that Euler's Formula is definitely true. So if we prove somehow that $$e^{i\Delta k}=\cos\Delta k+i\sin\Delta k$$ then? But how can we prove it for infinitesimal? Will the concept of limits be used? Can someone solve this mystery?

Mikhail Katz
  • 35,814
  • 3
  • 58
  • 116
Aditya Agarwal
  • 4,491
  • 2
  • 24
  • 49
  • 2
    There's such a thing as transfinite induction which is philosophically fairly similar to regular induction, but the details can be much more finicky. – Cameron Williams Dec 09 '15 at 05:30
  • So what can we deduce about this: infinitesimal induction? @CameronWilliams. – Aditya Agarwal Dec 09 '15 at 05:31
  • 43
    How do you know that $e^{i\Delta k}=\cos\Delta k+i\sin\Delta k$? – Eric Wofsey Dec 09 '15 at 05:33
  • Because of the fact that $e^{ik}=\cos k+i\sin k$. @EricWofsey. – Aditya Agarwal Dec 09 '15 at 05:38
  • 6
    @Aditya: but $k$ is not $\Delta k$... as in my answer, the way to make this rigorous is to talk about the derivative of both sides at $k$. – Qiaochu Yuan Dec 09 '15 at 05:39
  • 1
    It should be obvious that the equation holds for all k, no matter how small. So, what's the problem with what was done here? – Kaynex Dec 09 '15 at 05:51
  • 13
    His logic was to use an "inductive" technique, which means that assume it holds for some very specific $k$ and then show that it holds for another value of $k$. The trick with induction, though, is that you are assuming it holds *only* for the specified value of $k$, and nothing else. OP assumes that it also holds for the value of $\Delta k$, of which there is no reason for it to – ASKASK Dec 09 '15 at 05:55
  • When you think of a standard inductive proof, it goes "assume a statement holds for $n$, and show that it holds for $n+1$. You are trying to go about an interesting different route of "assume it holds for $k$, and show it holds for $k+\Delta k$". Notice that I said "assume it holds for $k$", and not "assume it holds for $k$ and $\Delta k$". This is because we are not (and cannot) assume it holds for $\Delta k$, only that it holds for $k$. – ASKASK Dec 09 '15 at 05:58
  • 5
    Roughly speaking, you have shown that if it holds for $k$ and $\Delta k$, then it must hold for $k+\Delta k$. Now the question is, why would this imply that it holds for all $x$? The answer is, it doesn't; in order for it to do that, you would need a much more sophisticated base case (for example show that it holds for all real numbers in a certain interval and you're all set) – ASKASK Dec 09 '15 at 06:01
  • 6
    You are trying to *prove* Euler's formula, so you cannot just assume that it holds for whatever value you want. – ASKASK Dec 09 '15 at 06:03
  • 2
    @CameronWilliams Ordinary induction is induction along a successor relation; transfinite induction is induction over successor and least upper bound or countable union. Neither of those apply in this situation, since there's no successor in this infinitesimal set-up. – David Richerby Dec 09 '15 at 08:12
  • If you say it is valid for $k$ *and* $\Delta k$, you are already saying it is valid for all $x$, without demonstrating it. – N74 Dec 09 '15 at 12:56
  • Yea, I got it now. @N74. – Aditya Agarwal Dec 09 '15 at 13:04
  • @DavidRicherby I'm aware. I was just providing OP with some information that he might find useful since it's rather obscure outside of advanced mathematical communities. – Cameron Williams Dec 09 '15 at 15:05
  • 1
    I'm finding myself thinking about this method a lot. Is it possible to show that an equation is true for an infinitesimal value? x = sin(x) comes to mind, which is true for x = 0, both have the same first and second derivative, but are not true on any infinitesimal value. Is it possible to do this then does this method of induction have any ground? – Kaynex Dec 10 '15 at 05:10
  • Say you had a more modest task to prove by usual induction over the integers that $e^{in}=\cos n+i\sin n$, for all $n=0,1,2,...$. So you verify the base case $e^{i0}=1=\cos 0+i\sin 0$, but I do not see how you then would prove that $e^{i1}=\cos 1+i\sin 1$ (that is going from $n=0$ to $n=1$ does not seem clear). Or, say you want to prove it for all $x$ that are positive integer multiples of $\frac\pi2$, that is $x=n\frac\pi2$. Base case $x=0$ which you did. Then how do you prove that $e^{i\frac\pi2}=\cos \frac\pi2+i\sin \frac\pi2$? Again just going from $n=0$ to $n=1$ seems unclear. – Mirko Dec 12 '15 at 03:19
  • This method doesn't work as is, but could work with some tweeking. If in your base case you show the statement holds for all $x$ with $|x|\leq k$ for some $k>0$, then you assume that it holds for any number $h$ with $|h| – Sean English Dec 12 '15 at 21:19
  • For induction, aren't you supposed to prove that it works for $x=0$, $x=a$, then for $a+1$? Unless this is what the entire question is about. Your question is whether or not you can do it with $a+i$, where $i$ is infinitesimally small? – Simply Beautiful Art Dec 18 '15 at 01:29
  • Doesn't assuming the equality works for $x=k$ fail you if the equality fails for some $k$ value? Your proof relies on knowing that the equality is true in the first place. – Simply Beautiful Art Dec 18 '15 at 01:40
  • There's a technique in SIA called microadditivity - IMO the single most important technique in mathematical phyics (apart from really basic ones like Pythagoras etc). That's what the OP is trying to do but I don't have time to check his working. –  Dec 20 '15 at 21:28
  • "Infinitesimal induction" (done right) is just calculus, Taylor series. – Abhimanyu Pallavi Sudhir Feb 20 '19 at 19:54

15 Answers15


Here is an example of your logic being used to prove a false statement. I am just basically going to take your exact proof, except replace the constant $e$ with the constant $2$. The logic is all the exact same, and you should notice that no where in your proof do you ever use any properties of the number $e$, so in theory any number could take its place.

Prove (The false euler equation): $$2^{ix}=\cos x+i\sin x \ \ \ \forall \ \ x\geq0$$

For $x=0$, we have $$1=1$$ So the equality holds.
Now let us assume that the given equality holds for some $x=k$.
$$2^{ik}=\cos k+i\sin k$$ Now, this is where I added my "own" axiom. Please answer whether this "axiom" is true or not. Now this equality must hold for $x=k+\Delta k$ also, for some infinitely small positive change $\Delta k$ (infinitesimal).
So $2^{i(k+\Delta k)}=2^{ik}.2^{i\Delta k}=(\cos k+i\sin k)(\cos\Delta k+i\sin\Delta k)$
$$=\cos k\cos\Delta k-\sin k\sin\Delta k+i\sin k\cos\Delta k+i\cos k\sin \Delta k$$ $$=\cos(k+\Delta k)+i\sin(k+\Delta k)$$ So we proved it for $x=k+\Delta k$, and hence it must hold for all $x\geq0$.

But the statement that we have just "proved" is very clearly false.

  • 8,711
  • 4
  • 24
  • 44
  • 14
    The statement _is_ false, but only up to linear variable substitution $k \to \ln 2\cdot k$. Which hints that the OP's idea isn't completely unusable, it just needs to be formulated properly (to consider not only the direction of infinitesimals but also their “length”). As explained by Qiaochu Yuan this can be done perfectly well as a solution to the ODE; I suppose you could write it in terms of nonstandard analysis to look even more like infinitesimal induction. – leftaroundabout Dec 09 '15 at 06:34
  • This would be even better if you added a linear term in k to your "assumed k" defintion equality (chosen such that a non-zero k exists). – bobuhito Dec 09 '15 at 10:10
  • @leftaroundabout, thanks for your suggestion. I fleshed it out in my answer, in terms of a hyperfinite induction. – Mikhail Katz Dec 10 '15 at 18:31
  • If you bend the OPs idea enough, you can prove anything. –  Dec 14 '15 at 16:48

An argument along these lines is the following: the standard theory of existence and uniqueness of solutions to ODEs implies that a solution to the ODE $y' = iy$ (where $y$ is a function $\mathbb{R} \to \mathbb{C}$) is uniquely determined by its initial value $y(0)$. Now, $e^{ix}$ and $\cos x + i \sin x$ are both solutions to this ODE, and they both have initial value $y(0) = 1$.

The way in which this comes to look like an inductive argument is when you explicitly solve this equation using Euler's method.

In general, there is a "principle of real induction," and it looks something like this: suppose $P(x)$ is a property of a nonnegative real number $x$ such that

  • $P(0)$ is true,
  • If $P(x)$ is true, then $P(y)$ is true for all $y$ in some interval $[x, x + \varepsilon)$, where $\varepsilon > 0$.
  • If $P(x)$ is true for all $x < y$, then $P(y)$ is true.

Then $P(x)$ is true for all $x$. However, it's unclear to me how easy it is to apply this to your case. (Several variations of this are possible; see, for example, this note by Pete Clark.)

Qiaochu Yuan
  • 359,788
  • 42
  • 777
  • 1,145
  • But why cant we take the equality to be true for $\Delta k$ as we took it for $k$.? And I didn't quite understand your derivative argument. Can you elaborate? – Aditya Agarwal Dec 09 '15 at 05:43
  • 2
    "Now let us assume the equality holds for some $x=k$". You are very specifically only assuming that it holds for $x=k$, and no other values of $x$. – ASKASK Dec 09 '15 at 06:02
  • @Aditya Induction principles on reals have been discussed here before, e.g. [see here.](http://math.stackexchange.com/q/4202/242) – Bill Dubuque Dec 10 '15 at 04:01

The first (not biggest) flaw in you proof is when you go from $k$ to some $x=k+\Delta k$. You are trying to prove some statement for ALL x. So you'd better proove it for ALL $x$ between $k$ and $x=k+\Delta k$. What if there are some $x$ in between for which the statement does not hold? For example, for each rational number there exists even bigger rational number, but still irrational numbers also exist!

To the bigger flaw now.

Let me "proove" much simplier statement: all real numbers are strictly less than 1.

Step 1: Statement holds for $x$ = 0 and all the smaller numbers.

Step 2: for any $x < 1$ there exists some positive change $dx$ such that the statement holds for all the numbers smaller than $x+dx$. F.e. $dx = (1 - x)/2$ will do.

Step 3. So that's it? For any number which satisfies our condition there exists even bigger number which also satisfies, so all the numbers must satisfy it?

Of course not, I know several numbers that are bigger than 1. (By the way, isn't it "Achilles and the tortoise" paradox?)

Moral of this story is that mathemetical induction isn't so obvious as it looks. It is a very important property of the set of integer numbers that you can use induction on it.

  • 2,087
  • 11
  • 13
  • 3
    I think adding an infinitesimal to a number will cover all the numbers we want. What is an infinitesimal? An infinitesimal is some quantity that is explicitly nonzero and yet smaller in absolute value than any real quantity. – Aditya Agarwal Dec 09 '15 at 07:44
  • 3
    It is not correct to say that infinitesimal is something "smaller than any real quantity". It's just a shorthand for "I can choose it small enough, so that some condition holds". – lesnik Dec 09 '15 at 08:09
  • Okay, then I guess Wolfram will have to redefine its definitions. http://mathworld.wolfram.com/Infinitesimal.html – Aditya Agarwal Dec 09 '15 at 08:10
  • 4
    Yes Wolfram would better do it. Mathematitians introduced infinitesemals and used them to for some time, but they were never quite happy about it. It was just a heuristic. Extremily usefull, but still a heuristic, and there was a demand for some proper background. Calculus provides such a background. As far as I know the concept of limit was introduced to justify the use / get rid of infinitesemals. Infinitesemals are still very handy, but it is very important to understand the background and be able to translate any satatement using "infinitesemals" into a "limits" language. – lesnik Dec 09 '15 at 08:37
  • See, I would believe Wolfram more, it is a reliable source. – Aditya Agarwal Dec 09 '15 at 08:43
  • And then what is your definition of an infinitesimal. According to me, infinitesimal is non-measurable, and any real quantity can be measured. But an infinitesimal cannot be measured – Aditya Agarwal Dec 09 '15 at 08:46
  • I am not able to give a good definition of infinitesemal. "Something so small that it can not be measured" is a very bad definition. Because it's still may be possible to say that this infinitesemal is exatly twice as big as that infinitesemal, so it's possible (sometimes) to compare and hence measure them. Situation is very similar to Dirac's delta function: it's very simple and handy heuristic, but proper definition is very complicated. But I can translate every statement using "infinitesemals" into "limits" language, and that's enough for me. – lesnik Dec 09 '15 at 09:03
  • So you are saying that "it's still possible to say that this infinitesemal is exatly twice as big as that infinitesemal, so it's possible (sometimes) to compare and hence measure them.", which is similar to saying that "infinity equals two times infinity and hence you can compare both and measure them." Which is quite evidently wrong. – Aditya Agarwal Dec 09 '15 at 09:07
  • No, it's not similar. Properties of infinitesemals are very different from properties of infinite sets. – lesnik Dec 09 '15 at 09:17
  • You are switching from mathematical rigour to casual sense in every subsequent comment. I don't think I will be able to argue with you eventually. – Aditya Agarwal Dec 09 '15 at 09:20
  • 2
    @lesnik: Wolfram is not wrong, the article is just incomplete. Non-zero infinitesimals are not part of the real numbers. They are, however, part of the hyperreal numbers, which are objects of study in the topic of non-standard analysis. – Jørgen Fogh Dec 09 '15 at 10:11
  • 2
    @Jørgen Fogh: thanks, but still I think definition on Wolfram is bad and misleading. To Aditya Agarwal: Obviously we do not understand each other, and it's a pity. But I'll try one more time. You have some statement which holds for x=0. You add infinitisemal to it (whatever it means) - statement still holds. You add inifinitesimal once again - it still holds. And again and again and again. But it does not mean that in this process you will ever reach, let's say, x=1 or any other number. There is no number less than 1, such that if you add infinitisemal to it and get a number larger than 1. – lesnik Dec 09 '15 at 12:09
  • @lesnik, again I contradict you. I agree with you till the point that if we add finite infinitesimals, then we may not achieve a real number. But what about infinite ones? Integrals are a good example. – Aditya Agarwal Dec 09 '15 at 12:35
  • Does it mean that all numbers are indeed less than 1? – lesnik Dec 09 '15 at 12:42
  • Nope, when did I imply/say that? – Aditya Agarwal Dec 09 '15 at 13:41
  • 3
    In your original post you started with x = 0, supposed the statement is true for some x = k and proved than in this case it's true for x = k + dk. Then you write "So we proved it for x=k+Δk, and hence it must hold for all x≥0." I am trying to use same approach to prove that all numbers are less than 1. My statement is also true for x = 0, if it is true for some x = k, than it is true for x=k+Δk. Why can't I just say "So we proved it for x=k+Δk, and hence it must hold for all x≥0."? – lesnik Dec 09 '15 at 14:04
  • 3
    I've added an answer partly in response to this thread, hopefully showing the problem in trying to use infinitesimals to get from 0 to 1. – Steve Jessop Dec 10 '15 at 12:27
  • 2
    @SteveJessop, you can get from 0 to 1 using infinitesimals just fine by using hyperfinite induction or hyperfinite partitions. This is a basic technique in the framework with Robinson's infinitesimals, and is explained in Keisler's textbook "Elementary calculus". – Mikhail Katz Dec 10 '15 at 18:29
  • @AdityaAgarwal I think I get what lesnik is trying to say, in the first comment you said you keep on adding infinitesimals to cover all the values but then from some finite value like 3.14 what do you add to get to $ \pi $ – Cloverr Dec 15 '15 at 18:16

In the added paragraph you ask for an example demonstrating where this kind of logic would give "a proof" of a wrong result. Here is one.

Claim. For all real numbers $x\ge0$ we have $$ e^{ix}=\cos 2x+i\sin 2x. $$

Proof. The formula holds for $x=0$. If we assume that the formula holds for $x=k$ AND that the formula holds for all small positive numbers $\Delta k$ AND that the usual formulas for $e^{x+y}$, $\sin(x+y)$ and $\cos(x+y)$ also hold, then we can do the "inductive step" $$ \begin{aligned} e^{i(k+\Delta k)}&=e^{ik}e^{i\Delta k}\\ &=(\cos 2k+i\sin 2k)(\cos2\Delta k+i\sin2\Delta k)\\ &=\cos (2[k+\Delta k])+i\sin (2[k+\Delta k]). \end{aligned} $$ QED.

Do you now see where the error in your logic is?

Jyrki Lahtonen
  • 123,594
  • 24
  • 249
  • 599

The "induction", as stated, does not quite work, as has already been pointed out. However, some rather similar arguments do work, and are quite useful.

Let me give you an example, which is a very basic model for some useful theorems in differential equations.

Lemma. Let $h \colon \mathbb{R} \to \mathbb{R}$ be differentiable with $h(0) = 0$ and $h'(x) = 0$ for each $x$. Then $h(x) = 0$ for all $x$.

Proof: Take any positive constant $c > 0$; we will prove an easier statement that $|h(x)| \leq |c x|$. If this is done, then (since $c$ was arbitrary) you can conclude that $h(x) = 0$ for any $x$.

So, let me try to apply "infinitesimal induction". Certainly, the claim holds for $x =0$. Suppose now that we have proved the claim for some $x$, and let us take some $\Delta x > 0$ (think of $\Delta x$ as being very small). Now, since $h'(x) = 0$, we certainly have $\lim_{y \to 0} \frac{h(x+y)-h(x)}{x-y} = 0$, so (assuming $\Delta x$ is sufficiently small) we can conclude that $\frac{h(x+\Delta x)-h(x)}{\Delta x} < c$, and consequently $$|h(x+\Delta x) | \leq c|\Delta x| + |h(x)| \leq c|x + \Delta x|.$$ Thus, the claim holds for $x + \Delta x$ as well.

Here is a slight subtlety: $\Delta x$ needed to be small, and how small it is might depend on $x$. For example, if I were trying to "inductively" prove the claim $x < 1$, then for each $x$ for which the claim holds, I could find $\Delta x$ such that the claim still holds for $x + \Delta x$ (e.g. $\Delta x = \frac{1- x}{10}$). However, the claim is blatantly false for $x =1$. Hence, we need some additional closure condition.

Returning to the lemma at hand, recall that we are proving the claim that $h(x) \leq c x$. Let me notice additionally the following simple fact: If the claim holds for some $x_i$ ($i =1,2,\dots$) and $\lim_{i \to \infty} x_i = x$ then also the claim holds for $x$. This is not terribly difficult: $$cx - |h(x)| = \lim_{i \to \infty} cx_i - |h(x_i)| \geq 0,$$ because $ cx_i - |h(x_i)| \geq 0$ for each $i$ by assumption.

Combining the "inductive step" (i.e. claim for $x$ implies claim for $x + \Delta x$, provided $\Delta x $ small enough) together with the "continuity" (i.e. claim for $x_i$ implies claim for $x = \lim_i x_i$), we can indeed conclude that the claim holds for all $x > 0$. $\square$

Now, with the above Lemma, you can try to make your reasoning rigorous. Let $f(x) = e^{ix}$, $g(x) = \cos x + i \sin x$ and $h(x) = |f(x) - g(x)|^2$. Assume further that you know $f'(x) = i f(x)$ (you have to know something specific to $e$.)

Clearly, $h(0) = 0$. Doing some algebraic manipulations, you will find that $h'(x) = 0$ for all $x$ (basically because $f'(x) - g'(x) = i(f(x) - g(x))$. Thus, the "infinitesimal induction" will tell you that $h(x) = 0$ always, which is just what you need.

Also, you may find the concept of connectedness to be very relevant.

Jakub Konieczny
  • 12,084
  • 1
  • 30
  • 71

My issue with your proposed proof is that you assumed your hypothesis is true for some infinitesimal value.

  • 119
  • 4

As many answers have pointed out, the method as provided does not work.

However, one could argue that strong induction does work, in the form of a carefully-established integral.

  1. It is true for $x=0$ (trivial)

  2. Assume it is true for all $y$ such that $0\leq y<x$

  3. Show that it is true for $x$.

$$ \frac{e^{ix}-1}i=\int_0^x e^{iy} dy = \int_0^x \cos y + i\sin y\ dy = \sin x-i(\cos x-1) $$ Rearranging gives $$ e^{ix}-1 = i\sin x + (\cos x-1)\\ e^{ix} = \cos x + i\sin x $$

It's not quite formal as written, and you would have to be careful with the exact form of integral being used when doing it fully formally, but I believe it demonstrates how one might do induction of this sort.

Note that, strictly speaking, the above is not sufficient to prove the claim. There is one additional requirement - prove that there can't be a "largest" value for which the claim is true - that is, show that, if it is true for $x$, then there must exist a $y>x$ such that it is true for $y$. Note that this would be satisfied if the interval for which the claim is true is open on the right.

This is necessary to allow it to push "past" a number, rather than getting stuck at a given $x$ (most notably, it is necessary to be able to get past $0$).

In this case, I believe it is sufficient to show that both $e^{ix}$ and $\cos x+i\sin x$ are analytic functions, as this implies (in conjunction with the above) that there is a non-zero region around any given point for which the two functions converge to the same value. Therefore, if it is true at $x$, then there is a point $y>x$ for which the equality holds for all values $[x,y]$. Of course, analyticity is proven using Cauchy-Riemann equations, and I won't go into that here.

Glen O
  • 11,786
  • 27
  • 38
  • I understand how this technique would work for natural numbers, but how would it work for reals? – ASKASK Dec 13 '15 at 20:11
  • @ASKASK - Where did I use natural numbers? $x$ can be any positive real number. The point is, it's what you might call "Strong Induction on the Reals". – Glen O Dec 14 '15 at 11:51
  • Yes I am saying I understand how strong induction in the naturals work, but not how it would work for the reals. In other words I don't see how if $0 – ASKASK Dec 14 '15 at 12:55
  • @ASKASK - Obviously, this is a generalisation of regular induction, and doesn't go case-by-case. But the principle remains the same - show that a predicate is true for all values larger than a given value by using the fact that it's true for values lower than any given value. There is a strict extra requirement that I left off, apparently - I'll edit an explanation into the answer. – Glen O Dec 14 '15 at 15:20
  • Ah yes that last requirement is what I was referring to. Thanks – ASKASK Dec 14 '15 at 16:15

The structure of the argument is basically correct, and can be formalized in Robinson's framework in terms of hyperfinite induction. Namely, the formula is first proved for all hyperinteger multiples of a fixed infinitesimal. Then one applies standard part to prove it for all real values of $x$ using the fact that any real $x$ is infinitely close to a suitable hyperinteger multiple of the fixed infinitesimal.

What is missing of course is the base of the induction (namely Euler's formula for the fixed infinitesimal itself).

To prove the formula by induction, it is not necessary to prove the exact formula $e^{i\Delta k}=\cos\Delta k+i\sin\Delta k$ but rather it is sufficient to prove it up to a higher order term, for example by using second-order approximations for both sides. Of course this would require knowing the beginning of Taylor series for each of the functions involved (though not the whole series). Then when you iterate the induction $\frac{1}{\Delta k}$ times, the error introduced into the calculation will be sufficiently small so that it will be erased when you apply the standard part function, resulting in Euler's formula for all real inputs. Then by the transfer principle it holds on the nose also for hyperreal inputs, and in particular for an infinitesimal input $\Delta k$ you will get the formula $e^{i\Delta k}=\cos\Delta k+i\sin\Delta k$ on the nose.

A concrete realisation of this scheme in the context of the study of small oscillations of the pendulum appears in this 2016 publication in Quantum Studies: Mathematics and Foundations.

Mikhail Katz
  • 35,814
  • 3
  • 58
  • 116

I love your cleverness! :)

From an aesthetic point of view, induction is an iterative process and iteration is countable. That works fine if you are proving properties of a set where your elements are countable, but you are applying it to the Reals which is an uncountable set.

Another way to see this, is you have shown your statement to be true for k, and k + dk, and (k+dk) + dk, and ((k+dk)+dk) + dk, and so on. But since dk is infinitesimal, so is N * dk. And so, you may not have expanded your statement from k, to cover the entire reals >= 0. Effectively, you haven't gotten any further than the assumption about k, as the game is played.

So you need to show that your method 'expands' to cover an interval around k.

Why does Yuan's method of induction work, because after all it too is an iterative process and thereby a countable process? well, because his assumptions are basically stating, "if you can expand an interval in a countably infinite way, then you can proceed with an induction like process."

By analogy, Yuan is stating "Assume k is an interval where P holds true. And assume k+dk is an expansion of that interval where P hold true. Then using an inductive process, we can countably expand the interval k to cover what we need to."

I hope that helps.

  • 189
  • 7
  • Yea, yea it does. And I aint clever, I am foolish. – Aditya Agarwal Dec 09 '15 at 13:40
  • 3
    The principle of transfinite induction -- and the generalization to Dedekind complete ordered sets discussed in my note linked to in Qiaochu's answer -- has nothing to do with countability. If anything it sounds like you have in mind an earlier version of "real induction" due to Chao which uses the Archimedean nature of the ordering on $\mathbb{R}$. But I would recommend that you think things through more rigorously: I don't think this is a matter of "aesthetics". – Pete L. Clark Dec 10 '15 at 07:17
  • @PeteL.Clark, Hyperfinite induction is more relevant, as per my answer below. – Mikhail Katz Dec 10 '15 at 18:27

There are two problems with your proof.

The first is handled by ASKASK's answer: even allowing your rule of infinitesimal induction there is a gap in your proof since you would need to prove "Euler's formula for infinitesimals" in order to get $e^{i\Delta k}=\cos\Delta k+i\sin\Delta k$. You haven't done this, so the proof doesn't work.

However, there's also a problem with your rule of infinitesimal induction. I will suppose that we have a theory of infinitesimal quantities, without going into any great detail of what that theory is. Let us suppose we have a theory in which infinitesimals "behave themselves" with respect to arithmetic, and are smaller than any positive quantity. Then infinitesimal induction, if allowed, would go like this:

To Prove: All non-negative real numbers are less than 1

Base case: $0 < 1$ (prove this either from the ordered field axioms or from your construction of $\mathbb{R}$, as you prefer).

Suppose for the inductive step that for some real number $k$, $k < 1$. That is, $1-k > 0$. Now, we have by definition of infinitesimals that they are less than all positive reals, so for any infinitesimal $\Delta k$, in particular $\Delta k < 1 - k$ and therefore $k + \Delta k < k + (1 - k) = 1$.

This completes the proof?

So, are we now going to conclude that all real numbers are less than one? Nope! The rule of inference, should we accept it, would lead to contradictions.

It's difficult to say "why" something isn't true. But without going into details of a theory of infinitesimals, my intuition as to "why" this rule of inference doesn't work is that adding an infinitesimal isn't sufficient to get from one real number to another real number. Therefore this induction on infinitesimals doesn't get us any (non-zero) distance along the real line, and certainly doesn't get us all the way along the real line.

Observe also that, never mind infinitesimals, for every $k < 1$ there is neighbourhood of $k$ with a positive radius, all of whose values are less than one. This is a bigger / stronger result than an infinitesimal neighbourhood of $k$, but it still isn't enough to crawl our way up the real line. All it gives us is an open set with an upper bound. So even adding a positive real number isn't sufficient to get us "past" a limit point, if the numbers we're adding get small enough fast enough.

Qiaochu Yuan's answer describes an effective means of crawling up the real line, but note that $\epsilon$ must be a fixed positive value. Not an infinitesimal, and not dependent on $x$. Then you're good to go. Adding a fixed constant does get from inside a set "past" a limit point of the set.

Steve Jessop
  • 4,046
  • 15
  • 21
  • 2
    The flaw with your proof that all numbers are less than 1 has nothing to do with infinitesimals not being able to "reach" any non-infinitesimal number. It is merely that in the induction step, $k$ does not have to be real--it is instead some arbitrary element of our "generalized reals" that include infinitesimals. In particular, $k<1$ does not imply $k+\Delta k<1$, since $k$ might be infinitesimally smaller than $1$. In fact, in the framework of nonstandard analysis, you _can_ make inductive arguments using infinitesimals (subject to certain constraints). – Eric Wofsey Dec 11 '15 at 08:41
  • @EricWofsey: but not using the questioner's rule of inference, which is all I was trying to illustrate. Fair enough, the intuition I describe doesn't apply to the hyperreals (and I don't know much about them), but I think it does apply to the questioner's attempt to use infinitesimals. If there's a good proof of Euler's formula by induction on the hyperreals then lets have it :-) – Steve Jessop Dec 11 '15 at 14:03

While the argument in general isn't true, as described in other answers, there is another way how to prove euality of functions. The important condition is that both $e^{ix}$ and $\cos x + i sin x$ are holomorphic functions. The Identity theorem states that

If two holomorphic functions $f$ and $g$ on a domain $D$ agree on a set $S$ which has an accumulation point in $D$ then $f = g$ on all of $D$.

So if you prove that there is a sequence of points $x_i$ such that $lim_{i\to\infty} x_i\in D$, you can use the theorem to prove the equality on the whole domain. And for this you could use the induction principle on $i$.

For example you could set $x_0=0$, $x_1=1$ and $x_n=(x_{n-1}+x_{0})/2$. Then if you prove that the equality holds for $0$, $1$ and that if it holds for $x$ and $y$ it also holds for $(x+y)/2$, the theorem gives you equality everywhere.

  • 2,277
  • 11
  • 21

For the infinitesimal argument, you could write

$$e^{i\Delta x}\approx1+i\Delta x.$$

Then taking the $k^{th}$ power,

$$e^{ik\Delta x}\approx(1+i\Delta x)^k.$$

And in the limit,

$$\lim_{k\to\infty,k\Delta x=x}e^{ix}=\lim_{k\to\infty}\left(1+i\frac xk\right)^k.$$

Developing the RHS by the binomial formula and taking the limit, you will find the Taylor series for $\cos(x)+i\sin(x)$.

  • Can we restrict the use of Taylor's formula? Because it can directly be used to prove the Euler's formula. – Aditya Agarwal Dec 11 '15 at 03:22
  • @AdityaAgarwal: how do you define $\sin$ and $\cos$ ? –  Dec 11 '15 at 08:18
  • Simply, by the unit circle. But please ain't there any other proof without the use of it? – Aditya Agarwal Dec 11 '15 at 08:20
  • "Simply, by the unit circle": can you elaborate ? –  Dec 11 '15 at 08:31
  • What do you want me to speak? $(\cos x,\sin x)$ are the coordinates on the unit circle. – Aditya Agarwal Dec 11 '15 at 08:33
  • @AdityaAgarwal: this "definition" is useless/meaningless as long as you don't establish an analytical form so that you can discuss the equivalence with $e^{ix}$. For instance, what is $\cos(1)$ ? –  Dec 11 '15 at 08:37
  • What do you want to prove? – Aditya Agarwal Dec 11 '15 at 08:44
  • 2
    The question is yours: $e^{ix}=\cos(x)+i\sin(x)$. You cannot deal with an equation like $e^{ix}=\text{the abscissa on the unit circle}+i\text{ the ordinate on the unit circle}$. You need a minimum of formalism. –  Dec 11 '15 at 09:04
  • Yea, but how can we prove it for infinitesimals without using taylor series? – Aditya Agarwal Dec 11 '15 at 09:12
  • 3
    Maybe if you provide an alternative analytical definition of the trigonometric functions. If you can't define them in analytical terms, the whole discussion is pointless. One approach is by integrating the element of arc along the unit circle (analytically of course). –  Dec 11 '15 at 09:22

This is not really an answer to this question as posted (I do not discuss the exponential function), but it points out that a certain well-known method might be interpreted as induction over the reals. The idea is that to prove something is true, for say all positive $x$, one could prove it for $x=0$ and then show that the set $T\subseteq\mathbb R$ for which it is true is both open and closed.

Compared to usual induction, when we would like to prove a statement or property $P(n)$, we prove $P(0)$ and then (induction step), show that there is no largest integer $n$ for which the statement is true. Indeed, if $n$ were largest, then $P(n)$ is true, but then $P(n+1)$ must also be true, contradicting that $n$ was largest.

Back to the case of the reals, if $T$ had a largest element $t$, then since $T$ is open we have $t+\varepsilon\in T$ for all small enough $\varepsilon>0$, contradicting that $t$ were largest. Note that $T$ must also be unbounded, since if $\sup T<\infty$ then using that $T$ is closed we would have $\sup T\in T$ so $\sup T$ would be the largest element of $T$, but we already saw that $T$ cannot have a largest element.

The above explains the idea but does not yet constitute a proof that $[0,\infty)\subseteq T$. To make it precise recall that a version of induction for the integers is the so-called complete induction, where (after starting with $P(0)$) one proves that $P(m+1)$ holds whenever $P(n)$ holds for all $n$ in the interval of integers $[0,m]$ (that is, whenever $P(n)$ holds for all $n\in\{0,1,\dots,m\}$). (Restated in a different way: One proves that $P(n+1)$ holds whenever each of $P(0),P(1),\dots,P(n)$ holds.)

For the reals, we could define $s=\sup\{x\ge0:[0,x]\subseteq T\}$, and then we would like to show that $s=\infty$. But if $s<\infty$ then, since $T$ is closed we have $s\in T$, and since $T$ is open we have $s+\varepsilon\in T$ for all small enough $\varepsilon>0$, contradicting the definition of $s$.

To illustrate the above with a real-world-example, my answer was inspired by the following MSE question by Saun Dev and more specifically by this answer by Michael Burr to (for which the OP requested an elaboration ... my answer here may provide this).

So the problem there was to prove that if $f:\mathbb R \to \mathbb R$ is a differentiable function such that $f(0)=0$ and $f'(x)>f(x),\forall x \in \mathbb R$ then $f(x)>0,\forall x>0$. The answer I referred to provided the following hint: Consider $x=0$. Since $f'(0)>f(0)=0$, we know $$ \lim_{h\rightarrow 0}\frac{f(h)-f(0)}{h}=\lim_{h\rightarrow 0}\frac{f(h)}{h}>0. $$ Therefore, for $h$ positive and sufficiently close to $0$, the numerator is positive. Continue with this idea.

So to elaborate, let $T(x)$ be the statement that either $x=0$ or $x>0$ and $f(x)>0$. (Formally $T(x)$ is $(x=0)\lor((x>0)\land(f(x)>0))$.) Let $T=\{x:T(x)\}$. What the hint shows is that $T$ is (relatively) open at $0$ in the interval $[0,\infty)$, that is there is $\varepsilon>0$ such that $T(h)$ holds, whenever $h\in[0,\varepsilon)$. Namely $T(h)$ holds for $h\in(0,\varepsilon)$ since $f(h)>0$ (by the hint), and $T(0)$ holds (trivially).

The induction step here could be split into two parts. First, $T$ is open, if $x\in T$ then we could push it further to the right, $T$ is open, there is $\varepsilon>0$ such that $T(h)$ holds, whenever $h\in[x,x+\varepsilon)$. This is essentially what the hint shows. Using that $f'(x)>f(x)\ge0$ (for $x\ge0$) and $$ \lim_{h\rightarrow x}\frac{f(x+h)-f(x)}{h}=f'(x)>0. $$ we conclude that $f(x+h)>f(x)\ge0$ for $h\in[x,x+\varepsilon)$ (or even for $h\in(x-\varepsilon,x+\varepsilon)$, if $x>0$) for all sufficiently small $\varepsilon>0$.
Second part of the induction step (which resembles complete induction for the integers), if $x>0$ and $T(y)$ holds for all $y\in[0,x)$ then $T(x)$ holds too. Indeed in the example here, since $f'(y)>f(y)>0$ for all $y\in(0,x)$, then $f$ must be strictly increasing on $(0,x)$ and by continuity $f(x)>0$ (and, with more detail, $f(x)>f(y)>0$ for each $y\in(0,x)$).

  • 13,040
  • 16
  • 39
  • A lot of answers to this question, I had overlooked that the ideas discussed in my answer here were already explained in @JakubKonieczny's answer, where in the comments AkivaWeinberger provides a link to a paper by Pete L. Clark, and notes that it has been already observed by many. – Mirko Dec 15 '15 at 18:09

As Eric Wofsey suggested in the comments, this part is not included in your extended induction assumptions!

$$e^{i \Delta k} = \cos \Delta k + i \sin \Delta k$$

Hence, your proof collapses when you use this!

To get more enlightened, you should note that we always start an induction to be true for a specific value not any value. Suppose that you wanted to prove the Euler formula for non-negative integers

$$e^{in} = \cos n + i \sin n, \qquad \qquad n=0,1,2,...$$

Then you would start by saying that will hold for $n=0$ since $1=1$. Next, if being true for $n=k$ implies that it is also true for $n=k+1$, you could conclude that it is valid for all $n$. You should note that the starting value for $k$ is $0$ not any other value.

But in your proof for positive reals, you are saying that we assume it is true for $x=k$ and $x=k+\Delta k$ where $k$ is a specific real number and $\Delta k$ is some arbitrary real number! Then you prove that it is valid for any real number but it didn't need any proof as you took it granted by your assumptions that the equality holds for any real number!

Hosein Rahnama
  • 13,627
  • 13
  • 47
  • 82
  • This does not provide an answer to the question. To critique or request clarification from an author, leave a comment below their post. - [From Review](/review/low-quality-posts/545030) – Tom-Tom Dec 16 '15 at 11:22
  • 1
    @Tom-Tom: This is the part that OP has made the mistake and in fact this little thing is the answer to the question in my opinion! :) Thanks for the review. :) – Hosein Rahnama Dec 16 '15 at 11:27

I think this is what you're trying to do.

First use linear approximation: $e^{i\Delta x} = 1+\left(\frac{d}{dx}e^{i\Delta x}\right)|_0 \Delta x=1+i\Delta x$. $\Delta x$ is infinitesimal so higher-order terms disappear.

Doing linear approximations with sine and cosine too, we get $\cos(\Delta x)=1+ 0\Delta x$ and $\sin(\Delta x) = 0+ \Delta x$.

Therefore $e^{i\Delta x}=\cos(\Delta x)+i\sin(\Delta x)$ holds with $\Delta x$ infinitesimal.

Now to move to the finite case, use DeMoivre's formula $(\cos(k)+i\sin(k))^n=\cos(nk)+i\sin(nk)$ ($n$ an integer). This can be proved with the angle addition formulas by induction on $n$.

\begin{align} e^{ix}&=\lim_{\Delta x\rightarrow 0} (e^{i\Delta x})^{x/\Delta x}\\&=\lim_{\Delta x\rightarrow 0} (\cos(i\Delta x)+i\sin(i\Delta x))^{x/\Delta x}\\&=\lim_{\Delta x\rightarrow 0} \cos(x)+i\sin(x) \end{align}

This is not very rigorous, so here is how to make it rigorous. Instead of using $\Delta x$ infinitesimal, use the identity $e^{i\Delta x}=\cos(\Delta x)+i\sin(\Delta x)+\Delta x^2 \delta(\Delta x)$ where $\delta(\Delta x)\rightarrow 0$ as $\Delta x\rightarrow 0$. Then the limit comes out as

\begin{align} e^{ix}&=\lim_{\Delta x\rightarrow 0} (\cos(i\Delta x)+i\sin(i\Delta x)+\Delta x^2 \delta(\Delta x))^{x/\Delta x}\\&=\lim_{\Delta x\rightarrow 0} \left[(\cos(x)+i\sin(x))+\frac{x}{\Delta x}(\cos(x-\Delta x)+i\sin(x-\Delta x))\Delta x^2\delta(\Delta x)+\cdots\right] \end{align}

by the binomial theorem. All terms after the first will go to zero.

Joshua Meyers
  • 1,193
  • 6
  • 16
  • I don't think this is what we want. – Simply Beautiful Art Dec 18 '15 at 01:33
  • @SimpleArt Can you elaborate? – Joshua Meyers Dec 18 '15 at 02:13
  • 1
    the question concerns whether or not the induction technique is correct or not. Not whether the given equation is true or false – Simply Beautiful Art Dec 19 '15 at 01:16
  • My answer is a method of proving Euler's formula that I think maps onto the attempted method, because it starts with the infinitesimal case and uses it for the finite case. It is not just an unrelated argument that Euler's formula is true. I also directly answer the follow-up question about how we know Euler's formula for the infinitesimal case. – Joshua Meyers Dec 19 '15 at 04:19