Series

The derivative — the series

9 essays on one idea, from the one that introduces it to the one that assumes the rest.
  1. Secants closing on the tangent to x². Secant lines through x = 1 and a second point 1.2, 0.8, 0.5, 0.28, 0.12 away, with the slope of each. They approach 2, the derivative there.

    The slope of a single point

    A slope needs two points. A derivative is the slope at one. The construction that bridges the gap is a sequence of secants, and the whole difficulty of calculus is in what "the limit of that sequence" is allowed to mean.

    part 1 · analysis
  2. 6 cosines, and a curve with no tangent anywhere. Partial sums of a sum of cosines whose amplitudes shrink geometrically and whose frequencies grow faster. Each term adds finer detail; the curve converges and its slopes do not.

    A curve with a corner at every point

    Continuity means a curve can be drawn without lifting the pen. Differentiability means it has a tangent. The first was assumed to nearly imply the second until 1872, when Weierstrass exhibited a curve that is continuous everywhere and has a tangent nowhere — and it is a sum of cosines.

    part 2 · analysis
  3. A rectangle grown on two sides. A rectangle x by √x, with both sides grown by the change a step of h makes. The new area is the old one, two strips, and a small corner rectangle that has both increments in it.

    A rectangle grown on two sides

    A product of two changing quantities is the area of a rectangle whose sides both move. The extra area is two strips and a corner, and the whole of the product rule is the observation that the corner is negligible and the strips are not.

    part 3 · analysis
  4. A curved map of the plane, and the flat one that fits it at a point. The map (x² − y², 2xy) carrying a small square patch of grid. Beside it, the image of the same patch under the linear map given by the matrix of partial derivatives, drawn dashed on top of the curved image.

    The flat map that fits closest

    A derivative is usually met as a number, which works because a line through a point is described by one. In more than one dimension the object that plays the same role is a linear map, and the number was always a one-by-one instance of it.

    part 4 · analysis
  5. eˣ and its inverse, reflected in the diagonal. A curve, the line y = x, and the curve reflected in it — which is the graph of the inverse function. Tangents are drawn at matched pairs of points, and the two slopes at each pair multiply to one.

    The slope of the mirror image

    Undoing a function is reflecting its graph in the diagonal, and a reflection turns a slope into its reciprocal. That single observation supplies the derivative of every inverse — the logarithm, the roots, the inverse trigonometric functions — without differentiating any of them.

    part 5 · analysis
  6. Secants closing on the slope of the folium x³ + y³ = 3xy. The curve x³ + y³ = 3xy with secants from (1.001, 0.348) of slopes 1.274, 0.953, 0.830, 0.777, approaching the tangent slope 0.744 given by the equation's partial derivatives.

    A slope for a curve that is no function

    The folium x³ + y³ = 3xy loops back over itself, so no formula y = f(x) describes it, and yet at almost every point it has a perfectly good tangent. Differentiating the equation as it stands gives the slope, −(∂F/∂x) ÷ (∂F/∂y), and the only points where that fails are the ones where the curve turns vertical or crosses itself — which are exactly the points where it stops being a graph.

    part 6 · analysis
  7. The slope of x² sin(1/x): discontinuous at nought, and never jumping. The derivative of x^2 sin(1/x) on [−0.12, 0.12], with the value at nought marked and the level 0.5 crossed 77 times.

    A slope can swing but never jump

    A function can have a slope at every point without that slope changing continuously: x² sin(1/x) has slope nought at the origin and a slope that swings between −1 and 1 however close to the origin it is taken. What a slope cannot do is jump. Darboux proved in 1875 that a derivative takes every value between any two of its values, so a step is never a derivative — and the only way a slope can be discontinuous is by oscillating.

    part 7 · analysis
  8. Four ways to a slope, and the one that subtracts. f'(1.5)=4.053427893898620; forward best 2.60e-9 at h=1.0e-8; central best 6.41e-12 at h=3.2e-6; complex step at 1e-200: 1.00e-17; dual: 1.00e-17.

    A slope without a subtraction

    The definition of a slope subtracts two nearly equal numbers and divides by a tiny one, and a computer that follows it loses half its digits however carefully the step is chosen. Two pieces of arithmetic avoid the subtraction altogether. Step sideways into the complex numbers by 10⁻²⁰⁰ and read off the imaginary part; or calculate with numbers a + bε in which ε² = 0, and the slope rides along exactly. Run backwards, the second gives all thousand slopes of a function of a thousand variables for under five times the cost of the function.

    part 8 · analysis
  9. The cost of running a computation backwards with a few stored states. Minimal recomputation to reverse chains of length up to 200 with 1, 2, 3, 4, 6 slots; at length 200: 1: 19900, 2: 2470, 3: 1305, 4: 948, 6: 680.

    Running a computation backwards with little memory

    The reverse sweep that gives a whole gradient for the price of a few evaluations needs the computation's intermediate states in reverse order, and storing them all can cost more memory than the machine has. Keep only a few and recompute the rest, and the cheapest schedule has an exact answer found by Andreas Griewank in 1992: with c stored states and no step run more than r times, the longest chain that can be reversed is the binomial coefficient C(c + r, c). Pascal's triangle is the table of what memory buys.

    part 9 · analysis

All series