103
$\begingroup$

I am teaching Calc I, for the first time, and I haven't seriously revisited the subject in quite some time. An interesting pedagogy question came up: How misleading is it to regard $\frac{dy}{dx}$ as a fraction?

There is one strong argument against this: We tell students that $dy$ and $dx$ mean "a really small change in $y$" and "a really small change in $x$", respectively, but these notions aren't at all rigorous, and until you start talking about nonstandard analysis or cotangent bundles, the symbols $dy$ and $dx$ don't actually mean anything.

But it gives the right intuition! For example, the Chain Rule says $\frac{dy}{du} \cdot \frac{du}{dx}$ (under appropriate conditions), and it looks like you just "cancel the $du$". You can't literally do this, but it is this intuition that one turns into a proof, and indeed if one assumes that $\frac{du}{dx} \neq 0$ this intuition gets you pretty close.

The debate about how rigorous to be when teaching calculus is old, and I want to steer clear of it. But this leaves an honest mathematical question: Is treating $\frac{dy}{dx}$ as a fraction the road to perdition, for reasons beyond the above, and which have not occurred to me?For example, what (if any) false statements and wrong formulas will it lead to?

(Note: Please don't worry, I have no intention of telling students that $\frac{dy}{dx}$ is a fraction; only, perhaps, that it can usually be treated as one.)

$\endgroup$
12
  • 6
    $\begingroup$ Couple of related posts of possible interest @ math.se: math.stackexchange.com/questions/21199/is-dy-dx-not-a-ratio math.stackexchange.com/questions/21869/… $\endgroup$
    – user11000
    Aug 23, 2011 at 13:41
  • 7
    $\begingroup$ Whether or not it can be viewed as a fraction, I think $dy/dx$ is a poor choice of notation for Calc I. The problem is illustrated by the need for the parenthetical remark "under appropriate conditions" in your third paragraph. Using $dy/dx$ means that students have to struggle with both the notion of derivative and the intricacies of the notation. Unfortunately, the notation is so prevalent that it is unreasonable to postpone the notation until Calc III or Diff Eq, where it actually comes in handy. Oh well... $\endgroup$ Aug 23, 2011 at 14:21
  • 7
    $\begingroup$ I entirely disagree with François G. Dorais about this. See my answer below. $\endgroup$ Aug 23, 2011 at 17:54
  • 9
    $\begingroup$ I would argue that the notation actually leads students to make fewer mistakes. In my experience, students who do not use this notation have trouble computing the derivative of the reciprocal of a function (the formula $(f^{-1})'(y) = \frac{1}{f'(f^{-1}(y))}$ is cumbersome compared to $\frac{dx}{dy} = \frac{1}{\frac{dy}{dx}}$), or performing a change of variable in an integral. $\endgroup$ Feb 29, 2012 at 21:57
  • 18
    $\begingroup$ Teaching Calc I for the first time: Do not deviate from the textbook. Not even one tiny bit. Not even by an infinitesimal dx. $\endgroup$ Oct 7, 2013 at 20:42

13 Answers 13

62
$\begingroup$

You can think of $x$ and $y$ as smooth functions on a one-dimensional manifold of states of some system that you are thinking about, then $dx$ and $dy$ are differential forms. In any open region where $dx$ does not vanish we can say that $dy/dx$ is the unique smooth function such that $(dy/dx)dx=dy$; in other words, $dy/dx$ is $dy$ divided by $dx$. Of course you don't want to tell the students that, but it does clear up the logical question as asked.

[Added later:] this approach also gives a clear picture of what goes wrong with partial derivatives: if your state space has dimension $n>1$, then $dy$ and $dx$ lie in a vector space of dimension $n$, and you cannot divide them to get a number. I think it's a bit fussy to worry too much about notation for derivatives in one variable, but traditional notation for partial derivatives is horrendous, especially in any context where you might want to hold different variables constant in different places, such as Maxwell's relations in thermodynamics ( http://en.wikipedia.org/wiki/Maxwell_relations )

$\endgroup$
5
  • 2
    $\begingroup$ There's a somewhat sneaky mistake to be made with the operator $d/dy$ (as applied to $x$). Multidimensionally, what you want to do is take directional derivatives $D_{\vec v}$. These have the property (on smooth functions) that $D_{\vec v+\vec w} = D_{\vec v} + D_{\vec w}$. If one confuses the one-form $dy$ with the vector field in the direction $y$ (by misusing the standard metric), one can get confused about whether $d/d(2y) = 1/2 d/dy$ vs. $2 d/dy$. Really, $dy$ and $d/dy$ should be sections of dual bundles, and the reciprocal suggests that correctly, as in Neil's answer. $\endgroup$ Aug 23, 2011 at 14:34
  • 30
    $\begingroup$ Why don't you want to tell the students that? This is something that has been boggling me lately; students are already taught to think this way -- e.g. to work with dependent variables, many learn to think of $dy/dx$ as a fraction (although that is rarely the intent of the teacher), and are even taught to work with differential forms (e.g. as a device for doing integration by parts, or even change of variable). Is it really such a bad idea to actually teach this way of thinking explicitly, rather than have the students try to absorb it through osmosis? $\endgroup$
    – user13113
    Aug 24, 2011 at 6:13
  • 2
    $\begingroup$ @Hurkyl: Well, you don't want to throw manifold theory at them at this point, which requires far more math including non-Euclidean geometry, which is serious overkill at this point. But if we are only considering functions defined on and to good ol' $\mathbb{R}$ or $\mathbb{R}^n$, is there a way to simplify the development and avoid the manifold theory? $\endgroup$ Mar 5, 2016 at 9:34
  • 7
    $\begingroup$ If the state space has dimension n > 1 you can still divide n-forms. For example in n=2 you can do (dx /\ df) / (dx /\ dy), which turns out to be the partial derivative df/dy holding x constant. $\endgroup$
    – Jules
    Dec 31, 2018 at 21:43
  • $\begingroup$ @Jules, I think the claim isn't so much that you can never divide 1-forms as that you can't always divide (non-0) 1-forms, as you can do in a 1-dimensional space. $\endgroup$
    – LSpice
    Feb 21, 2020 at 15:17
54
$\begingroup$

I am fine with using the notion of cancellation of fractions to help students remember the chain rule, but it is dangerous to be too cavalier with this idea. For example, suppose $F(x,y)=0$ defines $y$ implicitly as a function of $x$. Then $$\frac{dy}{dx}=-\frac{\frac{\partial F}{\partial x}}{\frac{\partial F}{\partial y}}.$$ Naive cancellation gives the wrong sign!

$\endgroup$
10
  • 3
    $\begingroup$ This is almost an orthogonal issue; the problem is that $\frac{dy}{dx}$ along some path in XY space needn't match $\frac{\partial y}{\partial x} = 0$. $\endgroup$ Aug 23, 2011 at 21:27
  • 55
    $\begingroup$ In more detail, we have this awful notational convention for partial derivatives of not explicitly indicating which variables are being held fixed (leaving that implicit to the context), resulting in endless conflative confusion (suppose one is working with the function $f(x,y,t)=2x+3y+4t$, but looking at its values as one moves along the line $y=5x$. Should $\frac{\partial f}{\partial x}=2$ or should $\frac{\partial f}{\partial x}=2+3*5$? This ambiguity ALWAYS trips up students on multivariable chain rule problems, needlessly, due to the confusing way they're standardly written.) $\endgroup$ Aug 23, 2011 at 21:29
  • 6
    $\begingroup$ Sridhar, I suggest you write your nice comments more comprehensively in an answer! $\endgroup$ Aug 24, 2011 at 1:57
  • 4
    $\begingroup$ @SridharRamesh, I enjoyed your comment immensely. But can you suggest a better alternative? $\endgroup$ May 14, 2016 at 16:24
  • 2
    $\begingroup$ @JimConant - does this formula have a name? I have seen it mentioned on a few questions, but I've never been able to find a name or citation for the formula. $\endgroup$
    – johnnyb
    Jan 4, 2018 at 21:12
30
$\begingroup$

Treating dy/dx as a fraction is the gateway drug to treating ${\partial y}/{\partial x}$ as a fraction. This plus a little more notational confusion leads students to conclude that if $U(x,y)$ is a function of two variables, then along a level curve of $U$ we have

$$dy/dx = {\partial U/\partial x\over\partial U/\partial y}$$ by "cancelling the ${\partial U}$'s''.

$\endgroup$
2
  • 30
    $\begingroup$ The problem here is the unfortunate notation for partial derivatives: the top $\partial U$ represents $dU=U_x dx+U_y dy$ with $dy$ set to zero, while the bottom $\partial U$ represents $dU$ with $dx$ set to zero; the $\partial$ represents a different differential operator in the two cases, even though it is written with the same symbol. This is somewhat orthogonal, though, to the issue of whether it is appropriate to view all these ratio-looking things as ratios; the problem isn't in treating ratio-type things as ratios, but in not notationally distinguishing different operators. $\endgroup$ Aug 23, 2011 at 22:02
  • 1
    $\begingroup$ If this argument worked, it would work equally well if you weren't moving on a level curve, but were instead following some other path. $\endgroup$
    – Will Sawin
    Dec 24, 2011 at 21:09
28
$\begingroup$

What would Euler say?

I tell first-year calculus students that Leibniz and Euler considered $dy$ and $dx$ to be infinitely small increments of $y$ and $x$, but that was found to be problematic in the 19th century, in more complicated problems than those considered in 1st-year calculus.

Then later I say that if $x = \tan\theta$ then $$\frac{dx}{d\theta} = \sec^2\theta = 1 + \tan^2\theta = 1 + x^2.$$ If $$ \frac{dx}{d\theta} = 1 + x^2, $$ then $$ \frac{d\theta}{dx} = \frac{1}{1+x^2}, $$ so we have the derivative of the arctangent function.

Then I ask if anyone can say what step in the argument might be questionable. With the right very mild hints, someone will recall that $dx$ and $d\theta$ are not actual numbers, so taking reciprocals that way might be questionable. And then I point out that this is another use of the chain rule.

$\endgroup$
2
  • 2
    $\begingroup$ Wait is this a counterexample in saying why treating differentials as fractions is wrong, or is it an example showing how manipulating it could give valid results regardless? $\endgroup$
    – Max0815
    Oct 13, 2023 at 4:54
  • 1
    $\begingroup$ @Max0815 : It's more of a suggestion that it's right but we don't entirely understand why. Certainly there are ways of making infinitesimals rigorous, but I suspect all of them capture less than the whole truth. $\endgroup$ Oct 13, 2023 at 21:23
16
$\begingroup$

A first answer to "how misleading": more than one will simplify and get $\frac{dy}{dx}=\frac{y}{x}$. A more serious objection is, thinking the derivative as the ratio of two infinitesimal increments $dy$ and $dx$ without the convenient foundation may lead a freshman student to the conclusion that every function is differentiable (if I can think to quantities $dy$ and $dx$, what's wrong in a harmless algebraic operation on them).

This does not mean one has to avoid $\frac{dy}{dx}$, but instead of using it to introduce the derivative "because it gives the right intuition", I would prefer a more rigorous definition, introducing the Leibnitz' notation only later, justifying it because it is formally consistent with the theorems about the derivatives of compositions and inverses of functions.

Personally, I prefer the definition via first order expansion: $f$ has derivative $m$ at $x$ if $f(x+h)=f(x)+mh+o(h)$ as $h\to 0$; as to the above mentioned composition rule, it is even more intuitive: the affine approximation of a composition is the composition of the affine approximations. (I happen to talk here on this point of view).

$\endgroup$
15
$\begingroup$

I always explain in terms of linear approximation. The "derivative" of $f(x)$ is the function $f'(x)$ for which linear approximation holds, i.e. if we change $x$ to $\Delta x$ then how does $f(x)$ change? $$ f(x+\Delta x) = f(x)+ f'(x)\,\Delta x + o(\Delta x)^2 $$ The example I give my section students is $100.17^2 \approx 10034$ Do we care about the extra $0.0289$? probably not.

Also real world data is not continuous time, so we are always estimating the rate of things.

The infinitesimal point of view is useful in math an physics. One exercise is to check Green's theorem $\oint P\,dx + Q\,dy = \iint \left( \frac{\partial Q}{\partial x} - \frac{\partial P}{\partial y}\right)\, dx\, dy$ by integrating on/in an infinitesimal rectangle of width $\Delta x$ and height $\Delta y$.

I also recommend Infinitesimal Calculus by James M. Henle and Eugene M. Kleinberg as a point of view on how to teach Calc I & II,

$\endgroup$
3
  • 17
    $\begingroup$ +1 for the appearance of 10034, my former Zip Code. $\endgroup$ Aug 24, 2011 at 5:41
  • 5
    $\begingroup$ The right remainder is $o(h)$, not $O(h^2)$ $\endgroup$ Sep 20, 2015 at 7:42
  • $\begingroup$ With this approach it's easy to give an analytic derivation of the product rule, and motivated high school students can come up with the correct geometric illustration of the derivation, doing better than Leibniz who with his purely algebraic approach got it wrong at first. Newton, since he interwove geometry and analysis, got it right from the beginning. In addition, with the finite difference approach, students can use numerical approximations as checks on and guidance for their symbolic analysis. The differential forms approach can wait until differential geometry if ever elected. $\endgroup$ Jul 18, 2020 at 9:08
13
$\begingroup$

If there is a well-defined tangent line to a function, then $dy/dx$ is the slope of that line, and this slope is manifestly a fraction. You can introduce (e.g.) the chain rule using this sort of thinking by noting that the slope of $y = n(mx+b)+c = mnx + (nb+c)$ is $mn$. Or the product rule by noting that the slope of $y = (mx+b)(nx+c) = mnx^2+(cm+bn)x + bc$ at $x=0$ is $mc+nb$, and by translation (but be careful here) this gives the product rule in general (as well as implying the quotient rule by replacing $nx+c$ with $-x/n+d$. No mucking around with the limit definition to get these results, just elementary analytic geometry.

$\endgroup$
1
  • 4
    $\begingroup$ Of course you'll also have to be careful about higher-order terms to make these into proofs, but it can surely be done. $\endgroup$ Aug 23, 2011 at 14:41
9
$\begingroup$

What's most misleading about Leibnizian notation is its implicit context dependence. After you get over that hurdle, it will be easy to safely think of $dy/dx$ as a fraction.

In the context of $y=f(x)$, you think of $dx$ either as an arbitrary nonzero infinitesimal also called $\Delta x$---I did this, using Keisler's book last fall---or as a nonzero real $\Delta x$ small enough for whatever your accuracy you currently need. Either way, $dy$ is defined as $f'(x)dx$, where $f'(x)$ is defined as the usual limit of difference quotients $\Delta y/\Delta x$. Of course, in the $x=g(y)$ context, the meanings of $dx$ and $dy$ switch, as do the meanings of $\Delta x$ and $\Delta y$. In the $z=h(x,y)$ context, the meanings of $dx$, $dy$, $\Delta x$, and $\Delta y$ change yet again.

The "small enough, but not infinitely small" approach is what you'll find in standard calculus textbooks, with a section devoted to the distinction between $\Delta y$ and $dy$ (in the $y=f(x)$ context).

That said, this fall I'm planning to de-emphasize $dy/dx$ as much as I can get away with. Whether I use the little-o notation or not, I will push hard (with lots of numerical examples) on the $\Delta y=f'(x)\Delta x+o(\Delta x)$ definition of $f'(x)$, and how this makes the chain rule true but not trivial.

If $y=f(x)=x^2$ and $dx=\Delta x$ is small (but not infinitely small this time around), then $\Delta(x^2)$ equals $(x+\Delta x)^2-x^2$ equals $2x\Delta x+\Delta x^2$ equals $2x\Delta x+(\mathrm{small})\Delta x$, so $dy=2x\ dx$ and $f'(x)=2x$. In the context of $y=f(u)$ and $u=g(x)$, my presentation of the chain rule will just be that a first-order approximation of a first-order approximation is a first-order approximation:

\begin{align*} \Delta y&=f'(u)\Delta u+(\mathrm{small}_1)\Delta u\\\\ &=f'(u)(g'(x)\Delta x+(\mathrm{small}_2)\Delta x)+(\mathrm{small}_1)(g'(x)\Delta x+(\mathrm{small}_2)\Delta x)\\\\ &=f'(u)g'(x)\Delta x+(\mathrm{small})\Delta x \end{align*} No fractions here!

$\endgroup$
8
$\begingroup$

I find $dy/dx$ misleading because it treats $x$ and $y$ as similar objects.

When you use this notation, you lose the important point that $y$ is a function of $x$; instead you end up looking at $x$ and $y$ as related quantities.

I think it is important for calculus students to get the idea that differentiation is an operation that takes one function and produces a new function. In that way, it is fundamentally different from addition (or unary negation) of numbers (which is not the same thing as addition of functions).

Note that I am a lot more interested in (theoretical) computer science than (any form of) physics - this may bias my point of view.

$\endgroup$
8
  • 5
    $\begingroup$ I take it implicit differentiation is not highly regarded in Computer Science? Gerhard "I Would Have Thought Otherwise?" Paseman, 2011.08.23 $\endgroup$ Aug 24, 2011 at 4:29
  • 6
    $\begingroup$ But people often really do think of $x$ and $y$ as "related quantities". And in the appropriate mode of thought, $d$ really is fundamentally the same sort of thing as addition. (but it acts to turn "differentiable numbers" into "differential forms") e.g. one way to make this precise is in the language of the category of sheaves on $\mathbb{R}$, in which "real number" really means means "continuous function $\mathbb{R} \to \mathbb{R}$. $\endgroup$
    – user13113
    Aug 24, 2011 at 6:27
  • 1
    $\begingroup$ Gerhard - there is a right way to do implicit differentiation from this viewpoint, but it takes more space than a comment allows. $\endgroup$ Aug 24, 2011 at 7:05
  • 1
    $\begingroup$ The "operator" notation $D_x[f]$ makes the "function of a function" idea abundantly clear. $\endgroup$ May 21, 2014 at 9:42
  • 2
    $\begingroup$ I do not think you should remind that y is a function of x if it is not necessary; so for me this is a feature!bug. $\endgroup$ Jun 9, 2015 at 22:34
7
$\begingroup$

A note from a publication (my own) that occurred several years after this question was asked. $\frac{dy}{dx}$ can be considered a fraction of differentials.

You can think of differentials as infinitesimal values that are related to each other. Non-standard analysis showed that although 19th century mathematics viewed infinitesimals as problematic, they can be easily treated as ordinary mathematical objects, capable of division, multiplication, etc.

There is no problem treating $\frac{dy}{dx}$ as a fraction, but there is a problem in higher-order derivatives and differentials, but that is because we are using a notation that doesn't support it. If you take the idea of $\frac{dy}{dx}$ being a fraction seriously, then, to find the second derivative, you are taking the derivative of a fraction. Therefore, you have to apply the quotient rule. If you apply the quotient rule to $\frac{dy}{dx}$ you do not get the typical result of $\frac{d^2y}{dx^2}$. Instead, you get: $$\frac{d^2y}{dx^2} - \frac{dy}{dx}\frac{d^2x}{dx^2}$$ Or, written less ambiguously: $$\frac{d(d(y))}{(d(x))^2} - \frac{d(y)}{d(x)}\frac{d(d(x))}{(d(x))^2}$$ When written this way, the second derivative can be considered actual fractions just like the first derivative. Third and higher derivatives are even uglier, because you are taking the derivative of that.

You can see more details of this in "Extending the Algebraic Manipulability of Differentials", Dynamics of Continuous, Discrete and Impulsive Systems, Series A: Mathematical Analysis 26(3):217-230, 2019. And, if anyone is concerned for its validity, it had a further review in Mathematics Magazine 92(5), pp. 396–397 in their "Reviews" section.

$\endgroup$
6
$\begingroup$

I just wish to share, that when I was an undergrad student I felt pretty much satisfied reading the introductory chapters of the ODE book by Arnol'd.

He tackled the obscurity of $\frac{dy}{dx}$ just by defining 1-forms (in a quite understandable fashion) as linear functions on the tangent space and, furthermore, making the connection to the definition of a derivative. For me, personally, it was pretty enjoyable, as I was also a bit dazed by the notion of fractions of infinitesimally small quantities.

$\endgroup$
1
$\begingroup$

I like the notation dy/dx because many formulas and computations become more clear and easy. However, I never thought of this as a fraction. Instead

(1) It seems to me that it would better to think of (d/dx)(y) instead of dy/dx in order to avoid misinterpretations. So, d/dx is another notation for the derivative, and df/dx is preferable to f'(x) because it points out what variable we are using. Hence, instead of the cumsy way of differentiating y=sin (x+1) by steps one can think of y=sin z, with z=x+1 and apply dy/dx= dy/dz . dz/dx. Here dy/dx only means the derivative of the function y=y(x).

(2) It then alllows us to write dy=f'(x)dx, which is coherent with using dy/dx as if it was a fraction. The formula dy=f'(x)dx is coherent with the theory of differential forms, that is, we have two 1-forms related by a function at each point. Moreover, it shows very well the idea that we have a linear function approximating the original f, with slope f'(x), but we are changing coordinates in order to put the origin at the point (x, y=y(x)).

(3) Finally, this has nothing to do with infinitesimals, excepting that the derivative dy/dx is the limit of Δy/Δx when Δx->0.

(4) Analogously for the notation udv when integrating by parts. In my opinion, the only reason for why we mathematicians prefer the notation f'(x) is because we are used to the notation $g\circ f$ for the composition of two maps, which would be another important matter for discussion.

$\endgroup$
-2
$\begingroup$

It's VERY misleading to regard dy/dx as a mere fraction, and I believe this is one of the major pitfalls of Leibniz notation.

What it is: dy/dx is a fraction with a condition built in!

The condition is that dy is the change in y ( which we call dy ) CAUSED by a change in x ( dx ). The dy is dependent on the dx. A better way to think of dy/dx is to think of it as a function, instead, where you would plug in a dx, get an intermediate dy, and then return the ratio of dy/dx.

An example where the fraction analogy breaks down:

a) Picture two basis vectors, e1 and e2, that have the same magnitude ( but not necessarily unit-length ), and that are NOT orthogonal ( they are oblique to one another ). For the sake of visualization, let's imagine that the angle between them, theta, is equal to 60 degrees.

b) Because the vectors have the same magnitude, we can say that cos( theta ) = cos( e1, e2 ) = dy/dx = dx/dy = 0.5 ( cos(60) )

...That's right! In this situation ( |e1| = |e2| ): dy/dx = dx/dy, without necessarily having to be equal to 1. If dy/dx was a mere fraction, that would be shocking.

Why it's true: the dy in the numerator of dy/dx is not the same dy as the one in the denominator of dx/dy. The former was dependent on dx, but not the latter.

c) As an aside, for the general case, when |e1| != |e2|,

cos( theta ) = dSy / dSx = ( |e2| / |e1| ) * ( dy / dx ) = ( |e1| / |e2| ) * ( dx / dy ) = dSx / dSy

where dSx is the element of arclength along x, and equals dx * |e1| = dSx. In other words, dSx is how long in metric space, one dx in what I call "component space" amounts to ( when multiplied by |e1| ).

while dSy in dSy / dSx is how much of a change in Sy ( the metric or "length" scalar field subtanted by the y axis ) we get when we make a small change in Sx ( dSx ) along the x axis.

Once again, we run the risk of getting confused with Leibniz notation.

While we can calulate dSx = dx * |e1| and dSy = dy * |e2|, seperately, their ratio ( dSy ) / ( dSx ) is not the same as dSy/dSx = d/dSx(Sy). In dSy/dSx, the dSy is dependent on dSx, and therefore on dx.

$\endgroup$
3
  • $\begingroup$ I don't quite understand either your explanation or your example. In fact, just as in the fraction world, $dx/dy=1/(dy/dx)$, which is usually referred to as the inverse function theorem. Overall, this looks like Marx' advice to Cauchy how to improve calculus. $\endgroup$ Jan 13, 2015 at 10:25
  • $\begingroup$ That was Marx' advice :) $\endgroup$ Jan 13, 2015 at 10:44
  • $\begingroup$ I never read either, though I recognize Cauchy as a famous Mathematician, so I'll take that as a compliment of sorts; even though my explanation is not coming across. What is dx? A change in the x component. Such a change in the x component ( dx ) produces a length in metric space of ( dx * |e1| = dSx ). dSy/dSx is the directional derivative along the x axis, in metric space. This ratio is equal to the cosine of the angle between the two oblique non-unit basis vectors in my example, and is same as dSx/dSy, which if the angle was 60deg = 0.5. So dSy/dSx = dSx/dSy. I can send a sketch... $\endgroup$
    – Namo
    Jan 13, 2015 at 11:00

Not the answer you're looking for? Browse other questions tagged or ask your own question.