Calculate integrals through level-set measures and distinguish strong from weak Lp control.
§4.3Estimated reading time: 12 min
An integral records both the size of a function and the amount of space on which that size occurs. A distribution function separates these two pieces of information: it asks how large the set is on which a function exceeds a given threshold. Summing the resulting layers recovers integrals and also reveals why a tail estimate may be too weak to imply integrability.
Let (X,M,μ) be a σ-finite measure space, not necessarily finite.
Definition 4.3.1
(Distribution function)
For a measurable real or complex function, or a nonnegative extended-valued measurable function, f, its distribution function is
df(t)=μ{x:∣f(x)∣>t},t>0.
It is nonincreasing, possibly infinite. We use Tonelli's theorem for nonnegative functions on the product of (X,μ) with Lebesgue measure on (0,∞); both factors are σ-finite. No finiteness of μ(X) is required in the formula below, and both sides may be infinite.
Recovering an integral from its layers
Theorem 4.3.2
(Layer-cake formula)
If p>0 and f is as above, then
∫X∣f∣pdμ=p∫0∞tp−1df(t)dt.(1)
Proof
(Write the power as an integral and exchange the order)
For every x, including infinite values with the extended-integral convention,
∣f(x)∣p=∫0∣f(x)∣ptp−1dt=∫0∞ptp−11{t<∣f(x)∣}dt.
The integrand is nonnegative and measurable on the product: the set {(x,t):t<∣f(x)∣} is measurable. Integrating in x first turns 1{t<∣f(x)∣} into df(t), and Tonelli's theorem exchanges the integrals, giving (1). □
Common mistake
(Treating >t and ≥t as interchangeable at a fixed threshold)
For a fixed threshold the sets {∣f∣>t} and {∣f∣≥t} may have different measures: for f=1[0,1] on R, df(1)=0 while μ{∣f∣≥1}=1. The two conventions agree only after integration in t: for each fixed x the indicators 1{t<∣f(x)∣} and 1{t≤∣f(x)∣} differ only at the single value t=∣f(x)∣, so either convention gives the same right side of (1) by Tonelli. Pointwise statements such as Markov's inequality below must use the convention they state.
The explorer below slices a six-point simple function into horizontal layers and compares the layer sum, Markov's inequality at a chosen threshold, and the weak tail quantity with the direct integral, under either convention for df.
Worked example 4.3.3
(Layers of the identity function)
On (0,1) with Lebesgue measure take f(x)=x. For 0<t<1, {x>t}=(t,1), so df(t)=1−t; for t≥1, df(t)=0. Formula (1) gives, for p>0,
∫01xpdx=p∫01tp−1(1−t)dt=p(p1−p+11)=p+11,
which agrees with the direct computation.
Worked example 4.3.4
(A power singularity and its critical exponent)
On (0,1) with Lebesgue measure, take f(x)=x−1/2. Since x−1/2>t exactly when x<t−2, df(t)=1 for 0<t≤1 and df(t)=t−2 for t>1. For p>0,
∫01x−p/2dx=p∫01tp−1dt+p∫1∞tp−3dt.
The first integral equals 1; the second is finite exactly when p<2, and then equals p/(2−p), so the total is 2/(2−p), matching ∫01x−p/2dx. At the borderline p=2 the second integral is 2∫1∞t−1dt, which diverges logarithmically. In general, a tail exactly equal to Ct−p for large t, with C>0, makes the integrand in (1) proportional to 1/t and hence gives a divergent integral.
Strong and weak tail bounds
The layer-cake formula compares the whole integral with the tail; a single layer gives a cruder but pointwise bound.
Theorem 4.3.5
(Markov inequality)
For p>0, measurable f, and t>0,
tpdf(t)≤∫X∣f∣pdμ.(2)
If f∈Lp, the stronger conclusion
tpdf(t)≤∫{∣f∣>t}∣f∣pdμ⟶0(t→∞)(3)
holds.
Proof
(Integrate a pointwise comparison)
Pointwise tp1{∣f∣>t}≤∣f∣p1{∣f∣>t}≤∣f∣p; integrate. If ∫∣f∣p<∞, then ∣f∣<∞ almost everywhere, so ∣f∣p1{∣f∣>t}→0 almost everywhere as t→∞, with integrable majorant ∣f∣p; dominated convergence gives (3). □
For 0<p<1, Lp here means the class of functions with finite integral of ∣f∣p; we do not use a norm.
Definition 4.3.6
(Weak-Lp tail quantity)
For p>0,
Kp(f)=t>0suptdf(t)1/p.
f is weak-Lp if Kp(f)<∞.
If Kp(f)≤K<∞, then df(t)≤(K/t)p for every t>0. Inequality (2) says that Kp(f)≤∥f∥p (with ∥f∥p=(∫∣f∣p)1/p), so a finite strong Lp integral implies a finite weak quantity.
Worked example 4.3.7
(Weak but not strong)
On (0,1) let f(x)=x−1/p. Since x−1/p>t exactly when x<t−p, df(t)=1 for 0<t≤1 and df(t)=t−p for t>1. Hence tdf(t)1/p=t for t≤1 and =1 for t>1, so Kp(f)=1. But ∫01∣f∣p=∫01x−1dx=∞. The converse of the strong-to-weak implication fails even on a finite measure space.
In that example tpdf(t)=1 for large t, so the vanishing conclusion (3) already fails. One might hope that (3) characterizes Lp; it does not.
Counterexample mode
A vanishing tail is still not enough
Claim: if tpdf(t)→0 as t→∞ on a finite measure space, then f∈Lp. On (0,e−1), let f(x)=x−1/p(log(e/x))−1/p, so fp=1/(xlog(e/x)). Since xlog(e/x) has derivative −logx>0, f decreases. When t is large, its strict superlevel set is (0,at) with f(at)=t, where at↓0 and tpat=1/log(e/at)→0. Nevertheless, substituting u=log(e/x),
∫0e−1∣f∣pdx=∫0e−1xlog(e/x)dx=∫2∞udu=∞.
So (3) is necessary but not sufficient for strong integrability; the correct characterization is the layer-cake formula (1) itself.
For p>1, weak control still gives useful local integrability.
Proposition 4.3.8
(Weak control integrates on finite-measure sets when p>1)
Let p>1, Kp(f)≤K and μ(E)=a<∞. Then
∫E∣f∣dμ≤p−1pKa1−1/p.(4)
Proof
(Layer cake with the minimum of two tail bounds)
Apply (1) with exponent 1 to ∣f∣1E, whose distribution function is μ(E∩{∣f∣>t})≤min{a,(K/t)p}:
∫E∣f∣dμ=∫0∞μ(E∩{∣f∣>t})dt≤∫0∞min{a,(K/t)p}dt.
If a=0 or K=0, the left side is zero. Otherwise split at t0=Ka−1/p, where the two bounds cross. The pieces are at0=Ka1−1/p and ∫t0∞Kpt−pdt=Kpt01−p/(p−1)=Ka1−1/p/(p−1), which sum to p−1pKa1−1/p. □
Common mistake
(Expecting the same local bound for weak-L1)
For p=1 the tail integral ∫t0∞Kt−1dt in this argument diverges, and the failure is genuine. On (0,1), f(x)=1/x has df(t)=min{1,1/t}, so K1(f)=1, yet ∫01dx/x=∞ on this set of measure 1. Weak-L1 does not imply local integrability.
Practice
Quick check
(For f(x)=x−1/3 on (0,1), find the values of r>0 for which f∈Lr and compute K3(f).)
Use the weak-but-not-strong example with p=3.
Solution · Answer
∫01x−r/3dx<∞ exactly when r/3<1, so f∈Lr for 0<r<3. With p=3, f=x−1/p, so K3(f)=1 although f∈/L3.
Quick check
(If μ(E)=4, p=2 and K2(f)≤3, what bound does (4) give for ∫E∣f∣?)
Substitute into (4).
Solution · Answer
p−1pKa1−1/p=2⋅3⋅41/2=12, so ∫E∣f∣≤12.
Quick check
(Let f∈Lp, p>0. Prove μ({∣f∣>t})→0 as t→∞ and say whether a finite-measure hypothesis is needed.)
Apply (2).
Solution · Answer
Markov gives df(t)≤t−p∫∣f∣p→0. Finiteness of μ(X) is unnecessary; only ∫∣f∣p<∞ is used.
Quick check
(On R with Lebesgue measure, is f(x)=1 weak-Lp for some p>0? Compute df.)
Look at thresholds below 1.
Solution · Answer
df(t)=∞ for 0<t<1 and df(t)=0 for t≥1. Hence Kp(f)=∞ for every p: a bounded function need not be weak-Lp on an infinite measure space, because its low layers have infinite measure. Formula (1) correctly gives ∫∣f∣p=∞.
ExploreMATH4050 · d_f
Slicing a simple function into layers
Six points x₁,…,x₆ carry weights μᵢ and values |f(xᵢ)|. The distribution function is d_f(t) = μ{|f| > t}. Choose the data, the exponent p and a threshold t, and compare the direct integral Σ μᵢ|f(xᵢ)|^p with the layer sum Σ_j M_j (v_j^p − v_{j−1}^p), which equals p∫t^{p−1}d_f(t)dt.
Data
Exponent p
Convention for d_f
Points of the measure space
Point
Weight μ
Value |f|
x1
x2
x3
x4
x5
x6
Step graph of d_f
Each shaded box is one layer: it spans (v_{j−1}, v_j] on the t axis and has height M_j. The dashed vertical line marks the threshold t; the table gives the same numbers.
Layers of the distribution function
Layer j
t in (v_{j−1}, v_j]
M_j = μ{|f| ≥ v_j}
v_j^p − v_{j−1}^p
M_j (v_j^p − v_{j−1}^p)
1
(0, 1]
5/2
1
5/2
2
(1, 3]
3/2
2
3
3
(3, 5]
1/2
2
1
μ{|f| > t} = 1/2; μ{|f| ≥ t} = 3/2. t is a level, so the two conventions differ here.
Direct integral Σ μᵢ|f|^p
13/2
Layer sum
13/2
Both totals agree
yes
d_f(t) with the chosen convention
1/2
Markov: t^p d_f(t)
3/2
t^p d_f(t) ≤ ∫|f|^p dμ
yes
Weak quantity K_p = sup_t t d_f(t)^(1/p)
9/2 (approached as t increases to t = 3)
K_p^p = max_j v_j^p M_j
9/2
K_p^p ≤ ∫|f|^p dμ, so K_p ≤ (∫|f|^p dμ)^(1/p)
yes
p = 1: direct total 13/2, layer sum 13/2 (equal). At t = 3, t^p d_f(t) = 3/2 ≤ 13/2. K_p^p = 9/2 ≤ 13/2.
Why the readings agree
On (v_{j−1}, v_j) the function d_f is constant, equal to M_j = μ{|f| ≥ v_j}, because no point has value in that open interval. Hence p∫t^{p−1}d_f(t)dt = Σ_j M_j ∫_{v_{j−1}}^{v_j} p t^{p−1}dt = Σ_j M_j (v_j^p − v_{j−1}^p), and regrouping the same sum by points gives Σ μᵢ|f(xᵢ)|^p.
Markov's inequality t^p d_f(t) ≤ ∫|f|^p holds because |f|^p ≥ t^p on {|f| > t} (and on {|f| ≥ t}). Taking the supremum over t, K_p^p = sup_t t^p d_f(t) ≤ ∫|f|^p. On a finite step function the supremum is reached at a level v_j: it is attained under the ≥ convention and only approached from below under the > convention.
Six points, weights 0, 1/4, 1/2, 1 or 2, integer values 0 to 9, and thresholds 0 to 10 in steps of 0.1. Integer p gives exact fractions; p = 1/2 gives rounded decimals with tolerance 1e-9. Only the layers with positive weight are listed; points of weight zero change nothing.