Jordan Decomposition

Every bounded-variation function splits into a difference of two increasing functions. Drag the probe and watch the identity hold pointwise.

Record the total rise and the total fall separately. Subtract falls from rises to get the current height. The variation function is the running record of rises.

For example, money now in a 🐷 piggy bank is the total πŸ’° money put in minus the total taken out. Rises and falls are counted on separate tallies.

f=gβˆ’h,g(x)=Vax(f),β€…β€Šh=gβˆ’ff = g - h, \quad g(x) = V_a^x(f), \; h = g - f
f(x)=x2sin⁑1xΒ (x>0),β€…β€Šf(0)=0f(x) = x^2\sin\frac{1}{x}\ (x > 0), \; f(0) = 0

Part of these learning paths

Examples & Exercises

Assume ff is of bounded variation on [a,b][a, b].

Prove that with g(x)=Vax(f)g(x) = V_a^x(f) and h(x)=g(x)βˆ’f(x)h(x) = g(x) - f(x), both gg and hh are increasing, so f=gβˆ’hf = g - h (one direction of the Jordan decomposition).

Vax2(f)=Vax1(f)+Vx1x2(f)(a≀x1<x2≀b)V_a^{x_2}(f) = V_a^{x_1}(f) + V_{x_1}^{x_2}(f) \qquad (a \le x_1 < x_2 \le b)
h(x2)βˆ’h(x1)=Vx1x2(f)βˆ’(f(x2)βˆ’f(x1))β‰₯0h(x_2) - h(x_1) = V_{x_1}^{x_2}(f) - \bigl(f(x_2) - f(x_1)\bigr) \ge 0
At x0=0.7x_0 = 0.7: the plot shows f=0.4851f = 0.4851, g=0.6776g = 0.6776, h=0.1926h = 0.1926 β€” both gg and hh climb as you move right, and their difference is exactly ff.
Show proof
  1. First prove additivity of variation: for a≀x1<x2≀ba \le x_1 < x_2 \le b, Vax2(f)=Vax1(f)+Vx1x2(f)V_a^{x_2}(f) = V_a^{x_1}(f) + V_{x_1}^{x_2}(f).

    The inequality ≀\le goes as follows: given any partition PP of [a,x2][a, x_2], if x1x_1 is absent we may insert it, since the sum can only grow (∣f(v)βˆ’f(u)βˆ£β‰€βˆ£f(x1)βˆ’f(u)∣+∣f(v)βˆ’f(x1)∣\lvert f(v) - f(u)\rvert \le \lvert f(x_1) - f(u)\rvert + \lvert f(v) - f(x_1)\rvert).

    A partition containing x1x_1 has its sum split into a sum over [a,x1][a, x_1] plus a sum over [x1,x2][x_1, x_2], each bounded by the variation on that interval; taking the supremum over PP gives ≀\le.

  2. The reverse β‰₯\ge: concatenating a partition P1P_1 of [a,x1][a, x_1] with a partition P2P_2 of [x1,x2][x_1, x_2] gives a partition P1βˆͺP2P_1 \cup P_2 of [a,x2][a, x_2] whose sum is the sum of the two sums.

    Taking suprema on each interval gives β‰₯\ge.

    All suprema are finite since ff is BV.

  3. Now g(x)=Vax(f)g(x) = V_a^x(f) is increasing: for x1<x2x_1 < x_2, additivity gives g(x2)=g(x1)+Vx1x2(f)β‰₯g(x1)g(x_2) = g(x_1) + V_{x_1}^{x_2}(f) \ge g(x_1) (variation is nonnegative).

    And h(x2)βˆ’h(x1)=(g(x2)βˆ’g(x1))βˆ’(f(x2)βˆ’f(x1))=Vx1x2(f)βˆ’(f(x2)βˆ’f(x1))h(x_2) - h(x_1) = (g(x_2) - g(x_1)) - (f(x_2) - f(x_1)) = V_{x_1}^{x_2}(f) - (f(x_2) - f(x_1)); but Vx1x2(f)β‰₯∣f(x2)βˆ’f(x1)∣β‰₯f(x2)βˆ’f(x1)V_{x_1}^{x_2}(f) \ge \lvert f(x_2) - f(x_1)\rvert \ge f(x_2) - f(x_1) from the single partition {x1,x2}\{x_1, x_2\}, so h(x2)βˆ’h(x1)β‰₯0h(x_2) - h(x_1) \ge 0 β€” hh is increasing too.

    The identity f=gβˆ’hf = g - h is the definition of hh.

Prove the converse: if g,hg, h are increasing on [a,b][a, b] and f=gβˆ’hf = g - h, then ff is of bounded variation with Vab(f)≀(g(b)βˆ’g(a))+(h(b)βˆ’h(a))V_a^b(f) \le (g(b) - g(a)) + (h(b) - h(a)).

∣f(xi)βˆ’f(xiβˆ’1)βˆ£β‰€(g(xi)βˆ’g(xiβˆ’1))+(h(xi)βˆ’h(xiβˆ’1))\lvert f(x_i) - f(x_{i-1})\rvert \le (g(x_i) - g(x_{i-1})) + (h(x_i) - h(x_{i-1}))
Vab(f)≀(g(b)βˆ’g(a))+(h(b)βˆ’h(a))V_a^b(f) \le (g(b) - g(a)) + (h(b) - h(a))
At x0=0.7x_0 = 0.7: gg climbs from 00 to 0.67760.6776 and hh from 00 to 0.19260.1926 β€” their sum 0.87020.8702 bounds the variation of ff on [0,0.7][0, 0.7], 0.67760.6776, exactly as the inequality below says.
Show proof
  1. Take any partition a=x0<x1<β‹―<xn=ba = x_0 < x_1 < \dots < x_n = b of [a,b][a, b].

    On each piece, ∣f(xi)βˆ’f(xiβˆ’1)∣=∣(g(xi)βˆ’g(xiβˆ’1))βˆ’(h(xi)βˆ’h(xiβˆ’1))βˆ£β‰€βˆ£g(xi)βˆ’g(xiβˆ’1)∣+∣h(xi)βˆ’h(xiβˆ’1)∣\lvert f(x_i) - f(x_{i-1})\rvert = \lvert (g(x_i) - g(x_{i-1})) - (h(x_i) - h(x_{i-1}))\rvert \le \lvert g(x_i) - g(x_{i-1})\rvert + \lvert h(x_i) - h(x_{i-1})\rvert (triangle inequality). Since gg and hh are increasing, each difference is nonnegative, so the absolute values drop away.

  2. Summing over i=1,…,ni = 1, \dots, n, βˆ‘i=1n∣f(xi)βˆ’f(xiβˆ’1)βˆ£β‰€βˆ‘i=1n(g(xi)βˆ’g(xiβˆ’1))+βˆ‘i=1n(h(xi)βˆ’h(xiβˆ’1))\sum_{i=1}^{n} \lvert f(x_i) - f(x_{i-1})\rvert \le \sum_{i=1}^{n}(g(x_i) - g(x_{i-1})) + \sum_{i=1}^{n}(h(x_i) - h(x_{i-1})), and the two sums telescope to (g(b)βˆ’g(a))+(h(b)βˆ’h(a))(g(b) - g(a)) + (h(b) - h(a)). This bound is partition-independent, so taking the supremum over all partitions gives Vab(f)≀(g(b)βˆ’g(a))+(h(b)βˆ’h(a))<∞V_a^b(f) \le (g(b) - g(a)) + (h(b) - h(a)) < \infty β€” ff is of bounded variation.

Prove the Jordan decomposition is never unique: if f=gβˆ’hf = g - h with g,hg, h increasing, then for any increasing uu, f=(g+u)βˆ’(h+u)f = (g + u) - (h + u) is another one.

f=gβˆ’h=(g+u)βˆ’(h+u),uΒ increasingf = g - h = (g + u) - (h + u), \qquad u\ \text{increasing}
At x0=0.7x_0 = 0.7: adding the same increasing uu (say u(x)=xu(x) = x) to both the gg and hh curves shifts them up in lockstep, leaving the difference ff untouched β€” there are infinitely many decompositions.
Show proof
  1. A sum of increasing functions is increasing: if x1<x2x_1 < x_2 then g(x1)≀g(x2)g(x_1) \le g(x_2) and u(x1)≀u(x2)u(x_1) \le u(x_2), so (g+u)(x1)≀(g+u)(x2)(g+u)(x_1) \le (g+u)(x_2) β€” and the same for h+uh + u.

  2. Then (g+u)βˆ’(h+u)=gβˆ’h=f(g + u) - (h + u) = g - h = f β€” the uu cancels.

    Since there are infinitely many increasing functions uu (e.g. u(x)=xu(x) = x, u(x)=x2u(x) = x^2, u(x)=2u(x) = 2 constant), there are infinitely many Jordan decompositions.

    Uniqueness would need an extra normalization (e.g. g(a)=0g(a) = 0).

11. Integration