02.20.05 · analysis / littlewood-paley-interpolation

Oscillatory Integrals and the Method of Stationary Phase

shipped3 tiersLean: none

Anchor (Master): Stein, Harmonic Analysis (Princeton 1993) Ch. VIII-IX; Hörmander, The Analysis of Linear Partial Differential Operators I (Springer 1990) §7.7; Sogge, Fourier Integrals in Classical Analysis 2e (Cambridge 2017) §1-2

Intuition Beginner

An oscillatory integral adds up a wave that wiggles faster and faster. Picture the integral of $cos (λ x^{2})$ over a stretch of the real line, where $λ$ is a large number. When $λ$ is big the cosine flips between plus one and minus one extremely quickly, so neighbouring slivers of the curve nearly cancel: a positive bump is almost immediately undone by the negative bump right next to it. The bigger $λ$ is, the more thorough the cancellation, and the smaller the total integral. The whole subject is the art of measuring exactly how much survives the cancellation.

The one place cancellation fails is where the wave momentarily stops oscillating. The phase of $cos (λ x^{2})$ is $λ x^{2}$ , and its speed of change is the derivative $2 λ x$ , which is zero at $x = 0$ . Near that point the wave hovers at a single value instead of racing through its cycle, so the contributions add up instead of cancelling. Such a point is called a stationary point of the phase, and it is where almost all of the surviving integral comes from. This is the same principle that makes light follow the shortest path in optics: the paths near the true ray have nearly equal travel time, so they reinforce, while all the other paths cancel.

Why bother measuring the leftover so carefully? Because these integrals are everywhere: they are the Fourier transforms that decode signals, the wave packets that solve the wave and Schrödinger equations, and the sums that count lattice points in number theory. Knowing that an oscillatory integral shrinks like $1$ over the square root of $λ$ , rather than merely shrinking, is the difference between a useless bound and a sharp theorem. The van der Corput lemmas turn the vague phrase "rapid oscillation kills the integral" into exact rates that depend only on how fast the phase bends.

Visual Beginner

Picture the graph of $cos (λ x^{2})$ for a large $λ$ . Far from the origin the curve is a blur of tightly packed up-and-down wiggles, and the signed area under it almost vanishes because each crest is matched by a neighbouring trough. Right at the origin the wiggles stretch out into a single broad hump: the phase $λ x^{2}$ is nearly flat there, so the curve lingers near the value one instead of flipping sign. That broad central hump is the only part contributing real area.

   value
     1 |        .-''-.        far from 0: fast wiggles, area cancels
       |       /      \      vvvvvvvvvvv
     0 |~~~~~~/        \~~~~~/\/\/\/\/\/\/\
       |     stationary point (x = 0)
    -1 |________________________________ x
              the broad central hump
              is where the integral lives

The width of the central hump scales like $1/ λ$ , and its height is fixed near one, so its area is of order $1/ λ$ . That single picture explains the headline rate: an integral with a single non-degenerate stationary point decays like $λ^{- 1/2}$ in one dimension. Flatter phases (where higher derivatives also vanish) make a wider hump and slower decay; sharper bends make a narrower hump and faster decay.

Worked example Beginner

We estimate the size of the model integral $A (λ)$ , the signed area under $cos (λ x^{2})$ as $x$ runs from $- 1$ to $1$ , for large $λ$ , and watch the $1/ λ$ rate appear.

Step 1. Change variables to expose the natural scale. The phase is $λ x^{2}$ , and it varies by one full unit when $x$ moves a distance of about $1/ λ$ near the origin. Set $x = u / λ$ , so $d x = d u / λ$ . The area becomes $1/ λ$ times the signed area under $cos (u^{2})$ as $u$ runs from $- λ$ to $λ$ .

Step 2. Look at this rescaled area, the signed area under $cos (u^{2})$ over the symmetric window of half-width $λ$ . As $λ$ grows, the window widens to the whole line, and the total signed area under $cos (u^{2})$ across the whole line is a known finite number — the Fresnel value, equal to the square root of $π /2$ . So for large $λ$ the bracketed factor settles down to a fixed constant near $1.25$ .

Step 3. Multiply back the prefactor. The whole area is therefore approximately $\frac{1}{λ} \times π /2$ , which shrinks like $1/ λ$ .

Step 4. Sanity-check the size against the picture. The central hump has width about $1/ λ$ and height about one, so its area is about $1/ λ$ — exactly the rate the change of variables produced. The tails beyond the hump contribute a smaller amount because they oscillate and cancel.

Step 5. Compare with a phase that has no stationary point. The signed area under $cos (λ x)$ from $x = 1$ to $x = 2$ equals $\frac{s i n ( 2 λ ) - s i n ( λ )}{λ}$ , which shrinks like $1/ λ$ — faster than $1/ λ$ . Here the phase $λ x$ has derivative $λ$ , never zero, so there is no stationary point and the cancellation is more complete.

What this tells us: the presence of a stationary point slows the decay from $1/ λ$ to $1/ λ$ . The flat spot in the phase is what keeps the area from vanishing faster. The rate is set entirely by how the phase behaves near where its derivative is zero.

Check your understanding Beginner

Exercise (easy, multiple choice).

For large $λ$ , the signed area under $cos (λ x^{2})$ from $x = - 1$ to $x = 1$ shrinks at what rate?

A. $1/ λ$ B. $1/ λ$ C. $1/ λ^{2}$ D. It stays roughly constant.

Hint

The phase $λ x^{2}$ has a stationary point at $x = 0$ (its derivative $2 λ x$ vanishes there). The width of the reinforcing central hump scales like $1/ λ$ and its height is fixed, so its area scales the same way.

Answer

B. $1/ λ$ .

Feedback-correct: yes. A single non-degenerate stationary point in one dimension gives $λ^{- 1/2}$ decay, set by the width of the central hump where the phase is flat. Feedback-wrong: $1/ λ$ is the faster rate you get with no stationary point at all (a phase like $λ x$ whose derivative never vanishes); $1/ λ^{2}$ is faster still and does not occur for a smooth phase with a single ordinary critical point; the integral does not stay constant because away from the origin the rapid oscillation does cancel.

Exercise (easy, true-false).

True or false: if the phase of an oscillatory integral never has a flat spot (its derivative is never zero) on the region of integration, then the integral decays faster than any single power of $λ$ .

Hint

When the derivative of the phase is bounded away from zero, you can integrate by parts repeatedly. Each integration by parts pulls out one more factor of $1/ λ$ , and a smooth, compactly supported amplitude lets you repeat the trick as many times as you like.

Answer

True. With a smooth amplitude that vanishes near the endpoints and a phase whose derivative stays away from zero, repeated integration by parts produces a factor of $1/ λ$ each time with no leftover boundary terms. Since the process never stops, the integral is smaller than any fixed power of $1/ λ$ — this is the principle of non-stationary phase. A stationary point (a flat spot of the phase) is exactly what blocks this argument and forces the slower power-law decay.

Exercise (easy, multiple choice).

In optics, why does light travel along the path of stationary travel-time (Fermat's principle), as opposed to some random nearby path?

A. Nearby paths have wildly different travel times, so they reinforce. B. Nearby paths have nearly equal travel times, so their waves reinforce, while all other paths cancel. C. Light has no wave behaviour, so only one path is allowed. D. The shortest path is always a straight line regardless of the medium.

Hint

The travel-time plays the role of the phase. Where the travel-time is stationary, neighbouring paths share almost the same phase and add up; everywhere else the phases are spread out and cancel.

Answer

B. At the stationary path the travel-time barely changes between neighbours, so their light waves arrive in step and reinforce; along every other path the travel-times differ enough that the waves arrive out of step and cancel. This is the optical face of the stationary-phase principle: the surviving contribution comes from where the phase is flat. The straight-line answer is wrong in a medium with varying speed (the ray bends), and light very much does have wave behaviour — that is the whole point.

Formal definition Intermediate+

Throughout, $λ > 0$ is a large real parameter, $ψ \in C_{c}^{\infty} (R^{n})$ is a smooth amplitude (also called a cutoff) with compact support, and $φ : R^{n} \to R$ is a smooth real phase defined on a neighbourhood of $supp ψ$ . The Fourier-transform convention is that of 02.10.04.

Definition (oscillatory integral of the first kind). The oscillatory integral with phase $φ$ and amplitude $ψ$ is $$ I(\lambda) = \int_{\mathbb{R}^n} e^{i \lambda \varphi(x)} \psi(x) , dx . $$ The integral converges absolutely for each fixed $λ$ because $ψ$ is integrable; the content of the theory is the asymptotic behaviour as $λ \to \infty$ , governed entirely by the critical set ${x : \nabla φ (x) = 0}$ of the phase inside $supp ψ$ .

Definition (stationary and non-degenerate critical points). A point $x_{0} \in supp ψ$ is a stationary point (critical point) of $φ$ if $\nabla φ (x_{0}) = 0$ . The stationary point is non-degenerate if the Hessian $φ^{''} (x_{0}) = (\partial_{j} \partial_{k} φ (x_{0}))_{j, k}$ is an invertible symmetric matrix, $det φ^{''} (x_{0}) \neq = 0$ . The signature $sgn φ^{''} (x_{0})$ is the number of positive eigenvalues minus the number of negative eigenvalues of $φ^{''} (x_{0})$ .

Definition (oscillatory integral operator). Given a real phase $φ (x, y)$ smooth on a neighbourhood of $supp ψ$ with $ψ \in C_{c}^{\infty} (R^{n} \times R^{n})$ , the associated oscillatory integral operator is $$ (T_\lambda f)(x) = \int_{\mathbb{R}^n} e^{i \lambda \varphi(x, y)} \psi(x, y) f(y) , dy , \qquad f \in C_c^\infty(\mathbb{R}^n). $$ The phase is non-degenerate on $supp ψ$ if the mixed Hessian $\partial_{x y}^{2} φ = (\partial_{x_{j}} \partial_{y_{k}} φ)_{j, k}$ satisfies $det \partial_{x y}^{2} φ (x, y) \neq = 0$ for all $(x, y) \in supp ψ$ . The object of interest is the operator norm $∥ T_{λ} ∥_{L^{2} \to L^{2}}$ as $λ \to \infty$ .

Notation and conventions

$I (λ) = \int e^{iλ φ} ψ$ : an oscillatory integral of the first kind; $λ \to \infty$ is the regime studied.
$φ$ , $ψ$ : the real phase and the compactly supported smooth amplitude; $φ$ is normalised to be real-valued, so $∣ e^{iλ φ} ∣ = 1$ .
$φ^{''} (x_{0})$ , $det φ^{''}$ , $sgn φ^{''}$ : the Hessian matrix of the phase at a critical point, its determinant, and its signature.
$\partial_{x y}^{2} φ$ : the mixed Hessian of a two-variable phase $φ (x, y)$ ; its non-vanishing determinant is the non-degeneracy hypothesis for $T_{λ}$ .
$d σ$ : the surface measure on a smooth hypersurface $S \subset R^{n}$ ; $d σ (ξ) = \int_{S} e^{- i x \cdot ξ} d σ (x)$ its Fourier transform.
$a ≲ b$ means $a \leq C b$ for a constant $C$ depending only on the indicated data (dimension, finitely many derivatives of $φ$ , seminorms of $ψ$ ), uniformly in $λ$ .

Counterexamples to common slips

The decay rate is not $λ^{- 1/2}$ in every dimension. In $R^{n}$ , a single non-degenerate stationary point gives $λ^{- n /2}$ , because each of the $n$ orthogonal directions contributes one factor of $λ^{- 1/2}$ .
The van der Corput rate $λ^{- 1/ k}$ requires a lower bound on the $k$ -th derivative, not on the $k$ -th derivative at a single point. The hypothesis $∣ φ^{(k)} ∣ \geq 1$ must hold throughout the interval; otherwise the constant blows up.
Non-stationary phase does not require $φ$ to be a polynomial or the amplitude to be analytic. Smoothness and compact support of $ψ$ together with $\nabla φ \neq = 0$ on $supp ψ$ already give faster-than-polynomial decay.
The first van der Corput lemma ( $k = 1$ ) needs the monotonicity of $φ^{'}$ , not merely $∣ φ^{'} ∣ \geq 1$ . Without monotonicity the bound fails; the standard statement assumes $φ^{'}$ is monotone on the interval.

Key theorem with proof Intermediate+

Theorem (van der Corput lemma; van der Corput 1921, Stein §VIII.1 Proposition 2 ^{[Stein Ch. VIII]}). Let $φ : [a, b] \to R$ be smooth and real. Suppose either

(i) $k = 1$ and $∣ φ^{'} (x) ∣ \geq 1$ for all $x \in [a, b]$ , with $φ^{'}$ monotone; or

(ii) $k \geq 2$ and $∣ φ^{(k)} (x) ∣ \geq 1$ for all $x \in [a, b]$ .

Then there is a constant $c_{k}$ depending only on $k$ such that $$ \left| \int_a^b e^{i \lambda \varphi(x)} , dx \right| \le c_k , \lambda^{-1/k} $$ for all $λ > 0$ , with $c_{k}$ independent of $a$ , $b$ , and $φ$ .

Proof. The proof is by induction on $k$ , with the case $k = 1$ supplying the integration-by-parts engine and the inductive step subdividing the interval around the points where the derivative is small.

Case $k = 1$ . Write $g (x) = 1/ (iλ φ^{'} (x))$ , so that $\frac{d}{d x} e^{iλ φ} = iλ φ^{'} e^{iλ φ}$ gives $e^{iλ φ} = g (x) \frac{d}{d x} e^{iλ φ}$ . Integrate by parts: $$ \int_a^b e^{i\lambda\varphi}, dx = \int_a^b g(x), \frac{d}{dx}\big(e^{i\lambda\varphi}\big), dx = \Big[ g(x) e^{i\lambda\varphi} \Big]_a^b - \int_a^b g'(x) e^{i\lambda\varphi}, dx . $$ The boundary term is bounded by $∣ g (a) ∣ + ∣ g (b) ∣ \leq 2/ λ$ since $∣ φ^{'} ∣ \geq 1$ . For the remaining integral, $g^{'} (x) = - φ^{''} (x) / (iλ φ^{'} (x)^{2})$ , so $∣ g^{'} (x) ∣ = ∣ φ^{''} (x) ∣/ (λ φ^{'} (x)^{2})$ . Because $φ^{'}$ is monotone, $φ^{''}$ does not change sign, and $$ \int_a^b |g'(x)|, dx = \frac{1}{\lambda} \int_a^b \frac{|\varphi''(x)|}{\varphi'(x)^2}, dx = \frac{1}{\lambda} \left| \int_a^b \frac{d}{dx}\Big( \frac{-1}{\varphi'(x)} \Big) dx \right| = \frac{1}{\lambda}\left| \frac{1}{\varphi'(a)} - \frac{1}{\varphi'(b)} \right| \le \frac{1}{\lambda}, $$ using $∣ φ^{'} ∣ \geq 1$ at both endpoints. Together the two contributions give $∣ \int_{a}^{b} e^{iλ φ} ∣ \leq 3/ λ = 3 λ^{- 1}$ , so $c_{1} = 3$ .

Inductive step $k - 1 \to k$ (for $k \geq 2$ ). Assume the bound for $k - 1$ and suppose $∣ φ^{(k)} ∣ \geq 1$ on $[a, b]$ . Since $φ^{(k - 1)}$ has derivative of absolute value at least one, it is strictly monotone, so it has at most one zero $c \in [a, b]$ . Fix a small $δ > 0$ to be chosen. Off the interval $(c - δ, c + δ)$ we have $∣ φ^{(k - 1)} (x) ∣ \geq δ$ , because $φ^{(k - 1)}$ moves away from its zero at unit rate at least. On each such complementary interval, apply the inductive hypothesis to the phase $φ / δ$ — whose $(k - 1)$ -st derivative is bounded below by one — obtaining $$ \left| \int e^{i\lambda\varphi}, dx \right| = \left| \int e^{i(\lambda\delta)(\varphi/\delta)}, dx \right| \le c_{k-1} (\lambda\delta)^{-1/(k-1)}, $$ on each of the (at most two) complementary intervals. On the central interval $(c - δ, c + δ)$ of length $2 δ$ we use the length bound $2 δ$ . Adding, $$ \left| \int_a^b e^{i\lambda\varphi}, dx \right| \le 2\delta + 2 c_{k-1}(\lambda\delta)^{-1/(k-1)}. $$ If $φ^{(k - 1)}$ has no zero in $[a, b]$ the central term is absent and the bound only improves. Optimise in $δ$ : the two terms balance when $δ \sim λ^{- 1/ k}$ , and substituting $δ = λ^{- 1/ k}$ yields a bound of the form $c_{k} λ^{- 1/ k}$ with $c_{k} = 2 + 2 c_{k - 1}$ a constant depending only on $k$ . $□$

Bridge. The van der Corput lemma builds toward the entire quantitative theory of oscillation, and it appears again in the stationary-phase expansion below, where the $k = 2$ rate $λ^{- 1/2}$ is the leading term at a non-degenerate critical point. The foundational reason the proof works is the interplay of two ideas: integration by parts converts oscillation into decay wherever the phase moves quickly, and subdivision quarantines the bad set where a lower derivative is small. This is exactly the mechanism that the stationary-phase asymptotic makes precise — it isolates the critical point in a shrinking neighbourhood of width $λ^{- 1/2}$ and applies non-stationary phase outside it. The central insight is that the rate $λ^{- 1/ k}$ is governed by the order of vanishing of the phase: a phase whose first $k - 1$ derivatives can vanish but whose $k$ -th cannot decays no slower than $λ^{- 1/ k}$ , and this generalises directly to sublevel-set estimates and to the Newton-polyhedron analysis of degenerate phases in several variables. Putting these together, the lemma is the one-dimensional atom from which the higher-dimensional theory — stationary phase, oscillatory integral operators, and Fourier decay of surface measures — is assembled by slicing, scaling, and summing.

Exercises Intermediate+

Exercise 3 (medium, symbolic).

Compute the leading stationary-phase term of $I (λ) = \int_{R} e^{iλ x^{2}} ψ (x) d x$ where $ψ \in C_{c}^{\infty} (R)$ with $ψ (0) = 1$ . Identify the constant, including the phase factor.

Hint

The Fresnel integral $\int_{R} e^{iλ x^{2}} d x = π / λ e^{iπ /4}$ . The amplitude only contributes its value at the critical point to leading order.

Answer

The phase $φ (x) = x^{2}$ has a non-degenerate critical point at $x = 0$ with $φ^{''} (0) = 2 > 0$ , signature $+ 1$ . The general stationary-phase formula gives, to leading order, $$ I(\lambda) \sim \left( \frac{2\pi}{\lambda} \right)^{1/2} |\varphi''(0)|^{-1/2} e^{i \frac{\pi}{4} \operatorname{sgn}\varphi''(0)} \psi(0) = \left( \frac{2\pi}{\lambda} \right)^{1/2} \frac{1}{\sqrt{2}} e^{i\pi/4} = \sqrt{\frac{\pi}{\lambda}}, e^{i\pi/4}. $$ This matches the exact Fresnel value when $ψ \equiv 1$ on a neighbourhood of $0$ , confirming the constant $π / λ e^{iπ /4}$ . Rubric: full credit for the magnitude $π / λ$ and the phase $e^{iπ /4}$ from the signature.

Exercise 4 (medium, short-answer).

Prove the principle of non-stationary phase: if $\nabla φ \neq = 0$ on $supp ψ \subset R^{n}$ with $ψ \in C_{c}^{\infty}$ , then $I (λ) = \int e^{iλ φ} ψ = O (λ^{- N})$ for every $N$ .

Hint

Define the first-order operator $L = \frac{1}{iλ ∣\nabla φ ∣ ^{2}} \sum_{j} (\partial_{j} φ) \partial_{j}$ , which satisfies $L e^{iλ φ} = e^{iλ φ}$ . Integrate by parts using the transpose $L^{t}$ , which lowers $λ$ by one power each time.

Answer

On $supp ψ$ the vector field coefficients $a_{j} = (\partial_{j} φ) /∣\nabla φ ∣^{2}$ are smooth because $∣\nabla φ ∣$ is bounded below. The operator $L = (iλ)^{- 1} \sum_{j} a_{j} \partial_{j}$ satisfies $L e^{iλ φ} = (iλ)^{- 1} \sum_{j} a_{j} (iλ \partial_{j} φ) e^{iλ φ} = e^{iλ φ} \sum_{j} a_{j} \partial_{j} φ = e^{iλ φ}$ , using $\sum_{j} (\partial_{j} φ)^{2} /∣\nabla φ ∣^{2} = 1$ . Hence $e^{iλ φ} = L^{N} e^{iλ φ}$ for every $N$ . Its transpose is $L^{t} ψ = - (iλ)^{- 1} \sum_{j} \partial_{j} (a_{j} ψ)$ , a first-order operator carrying one factor $λ^{- 1}$ and preserving compact support. Integrating by parts $N$ times with no boundary terms (compact support), $$ I(\lambda) = \int (L^N e^{i\lambda\varphi})\psi = \int e^{i\lambda\varphi} (L^t)^N \psi. $$ Each application of $L^{t}$ produces a factor $λ^{- 1}$ times a smooth compactly supported function, so $(L^{t})^{N} ψ = λ^{- N} g_{N}$ with $g_{N} \in C_{c}^{\infty}$ independent of $λ$ . Therefore $∣ I (λ) ∣ \leq λ^{- N} ∥ g_{N} ∥_{L^{1}}$ , which is $O (λ^{- N})$ for every $N$ . Rubric: full credit for constructing $L$ , the transpose integration by parts, and the per-step gain of $λ^{- 1}$ .

Exercise 5 (medium, short-answer).

Use the first van der Corput lemma ( $k = 1$ ) to bound $\int_{1}^{2} e^{iλ x^{2}} d x$ and compare with the crude length bound.

Hint

On $[1, 2]$ the derivative $φ^{'} (x) = 2 x$ satisfies $∣ φ^{'} ∣ \geq 2 \geq 1$ and is monotone. Apply the $k = 1$ lemma directly.

Answer

On $[1, 2]$ , $φ (x) = x^{2}$ has $φ^{'} (x) = 2 x \in [2, 4]$ , so $∣ φ^{'} ∣ \geq 1$ and $φ^{'}$ is increasing (monotone). The first van der Corput lemma gives $∣ \int_{1}^{2} e^{iλ x^{2}} d x ∣ \leq c_{1} λ^{- 1} = 3 λ^{- 1}$ . (One can sharpen: integration by parts with $g = 1/ (iλ \cdot 2 x)$ gives a bound of order $λ^{- 1}$ with smaller constant.) The crude bound is the length of the interval, namely $1$ , with no decay in $λ$ . The lemma improves this to $O (λ^{- 1})$ because there is no stationary point on $[1, 2]$ — the critical point $x = 0$ lies outside the interval, so this is a non-stationary-phase situation and the decay is the fast rate $λ^{- 1}$ , not the stationary rate $λ^{- 1/2}$ . Rubric: full credit for the $λ^{- 1}$ bound and the comparison.

Exercise 6 (medium, symbolic).

Let $S = {∣ x ∣ = 1} \subset R^{n}$ be the unit sphere with surface measure $d σ$ . State the sharp decay rate of $d σ (ξ)$ as $∣ ξ ∣ \to \infty$ and explain which geometric feature controls it.

Hint

Write $d σ (ξ) = \int_{S} e^{- i x \cdot ξ} d σ (x)$ as an oscillatory integral with phase $- x \cdot ξ /∣ ξ ∣$ restricted to the surface. The stationary points are where the normal to $S$ is parallel to $ξ$ ; there are exactly two such antipodal points, and the relevant Hessian is the second fundamental form (the curvature).

Answer

The sharp rate is $∣ d σ (ξ) ∣ ≲ ∣ ξ ∣^{- (n - 1) /2}$ , and this is attained (not merely an upper bound) for the sphere. Writing $ξ = ∣ ξ ∣ ω$ , the phase $x \mapsto - x \cdot ω$ on the $(n - 1)$ -dimensional surface $S$ has stationary points exactly where the outward normal is $\pm ω$ — the two antipodal points $\pm ω \in S$ . At each, the Hessian of the restricted phase is the second fundamental form (Gaussian curvature) of $S$ , which is non-degenerate precisely when $S$ has non-vanishing Gaussian curvature. The sphere has curvature one everywhere, so both critical points are non-degenerate, each contributing $∣ ξ ∣^{- (n - 1) /2}$ by stationary phase in $n - 1$ dimensions. The decay rate is therefore set by the curvature of the surface: a flat piece (zero curvature) would give a degenerate critical point and slower decay, while non-vanishing curvature is exactly the condition for the full $(n - 1) /2$ rate. Rubric: full credit for the rate $(n - 1) /2$ and the identification of curvature (the surface Hessian) as the controlling feature.

Exercise 7 (hard, short-answer).

Prove the Hörmander $L^{2}$ bound: if $φ (x, y)$ is a non-degenerate phase ( $det \partial_{x y}^{2} φ \neq = 0$ on $supp ψ$ ) and $ψ \in C_{c}^{\infty} (R^{n} \times R^{n})$ , then $∥ T_{λ} ∥_{L^{2} \to L^{2}} ≲ λ^{- n /2}$ .

Hint

Form $T_{λ}^{*} T_{λ}$ , whose kernel is $K (y, y^{'}) = \int e^{iλ (φ (x, y) - φ (x, y^{'}))} ψ (x, y) \overline{ψ (x, y^{'})} d x$ . Apply non-stationary phase in $x$ : the gradient $\nabla_{x} (φ (x, y) - φ (x, y^{'}))$ is non-zero unless $y$ is close to $y^{'}$ , because the mixed Hessian is invertible. Then bound the kernel by Schur's test.

Answer

Consider $T_{λ}^{*} T_{λ}$ , an integral operator with kernel $$ K(y, y') = \int_{\mathbb{R}^n} e^{i\lambda(\varphi(x,y) - \varphi(x,y'))}, \psi(x,y)\overline{\psi(x,y')}, dx. $$ The phase in $x$ is $Φ (x) = φ (x, y) - φ (x, y^{'})$ with $\nabla_{x} Φ = \nabla_{x} φ (x, y) - \nabla_{x} φ (x, y^{'})$ . By the mean value theorem and the non-degeneracy $det \partial_{x y}^{2} φ \neq = 0$ , the map $y \mapsto \nabla_{x} φ (x, y)$ is a local diffeomorphism with Jacobian bounded below, so $∣ \nabla_{x} Φ (x) ∣ ≳ ∣ y - y^{'} ∣$ uniformly on $supp ψ$ . Apply the non-stationary-phase estimate (Exercise 4) in the $x$ -integral: for every $N$ , $$ |K(y, y')| \lesssim_N \big(1 + \lambda|y - y'|\big)^{-N}, $$ since each integration by parts gains $λ^{- 1} ∣ \nabla_{x} Φ ∣^{- 1} ≲ (λ ∣ y - y^{'} ∣)^{- 1}$ and the crude bound $∣ K ∣ ≲ 1$ covers $λ ∣ y - y^{'} ∣ ≲ 1$ . Now apply Schur's test: $sup_{y} \int ∣ K (y, y^{'}) ∣ d y^{'} ≲ \int (1 + λ ∣ y - y^{'} ∣)^{- N} d y^{'} ≲ λ^{- n}$ for $N > n$ , by scaling $z = λ (y - y^{'})$ . By symmetry the same bound holds for $sup_{y^{'}} \int ∣ K ∣ d y$ . Schur's test then gives $∥ T_{λ}^{*} T_{λ} ∥_{L^{2} \to L^{2}} ≲ λ^{- n}$ , hence $∥ T_{λ} ∥_{L^{2} \to L^{2}} = ∥ T_{λ}^{*} T_{λ} ∥^{1/2} ≲ λ^{- n /2}$ . Rubric: full credit for the $T^{*} T$ kernel, the gradient lower bound from the mixed Hessian, the non-stationary-phase kernel decay, and the Schur-test conclusion.

Exercise 8 (hard, short-answer).

Show that the one-dimensional stationary-phase rate $λ^{- 1/2}$ at a non-degenerate critical point is sharp: it cannot be improved to $λ^{- 1/2 - ϵ}$ for any $ϵ > 0$ .

Hint

Take the model phase $φ (x) = x^{2}$ and an amplitude $ψ$ equal to one near $0$ . Compute the leading term explicitly via the Fresnel integral and observe its magnitude is exactly of order $λ^{- 1/2}$ , not smaller.

Answer

Take $φ (x) = x^{2}$ and $ψ \in C_{c}^{\infty} (R)$ with $ψ \equiv 1$ on $[- 1, 1]$ . Split $I (λ) = \int_{∣ x ∣ \leq 1} e^{iλ x^{2}} d x + \int_{∣ x ∣ > 1} e^{iλ x^{2}} ψ d x$ . On $∣ x ∣ > 1$ the phase has $φ^{'} (x) = 2 x$ with $∣ φ^{'} ∣ \geq 2$ and is monotone on each side, so by the first van der Corput lemma (and a non-stationary-phase argument with the smooth cutoff) that part is $O (λ^{- 1})$ . The main part is $\int_{- 1}^{1} e^{iλ x^{2}} d x = λ^{- 1/2} \int_{- λ}^{λ} e^{i u^{2}} d u \to λ^{- 1/2} \int_{R} e^{i u^{2}} d u = λ^{- 1/2} π e^{iπ /4}$ , using the Fresnel integral. Therefore $I (λ) = π e^{iπ /4} λ^{- 1/2} + O (λ^{- 1})$ , so $∣ I (λ) ∣$ is bounded below by a positive constant times $λ^{- 1/2}$ for all large $λ$ . No bound of the form $λ^{- 1/2 - ϵ}$ can hold: the explicit leading term has magnitude exactly of order $λ^{- 1/2}$ . Rubric: full credit for the explicit lower bound via the Fresnel computation and the separation of the non-stationary tail.

Advanced results Master

Theorem (stationary-phase asymptotic expansion; Hörmander §7.7 Theorem 7.7.5 ^{[Hörmander §7.7]}). Let $φ \in C^{\infty}$ be real with a single non-degenerate critical point $x_{0} \in supp ψ$ , $\nabla φ (x_{0}) = 0$ , $det φ^{''} (x_{0}) \neq = 0$ , and no other critical point on $supp ψ$ . Then for every $N$ , $$ I(\lambda) = \int e^{i\lambda\varphi}\psi = e^{i\lambda\varphi(x_0)} \left( \frac{2\pi}{\lambda} \right)^{n/2} \frac{e^{i\frac{\pi}{4}\operatorname{sgn}\varphi''(x_0)}}{|\det\varphi''(x_0)|^{1/2}} \sum_{j=0}^{N-1} \lambda^{-j} L_j \psi + O(\lambda^{-n/2 - N}), $$ where each $L_{j}$ is an explicit differential operator of order $2 j$ evaluated at $x_{0}$ , with $L_{0} ψ = ψ (x_{0})$ and $L_{1}$ built from $φ^{''}$ and the third and fourth derivatives of $φ$ .

The leading coefficient packages the three pieces of data the integral can see at $x_{0}$ : the scale $λ^{- n /2}$ from the width of the reinforcing neighbourhood in each of $n$ directions; the Gaussian normalisation $∣ det φ^{''} ∣^{- 1/2}$ from the curvature of the phase; and the Maslov phase $e^{iπ sgn φ^{''} /4}$ recording the number of positive versus negative eigenvalues. The expansion is the precise analytic content of the optical principle that light concentrates where the travel-time is stationary.

Theorem (Hörmander $L^{2}$ bound for oscillatory integral operators; Hörmander 1973, Stein §IX.1 ^{[Stein Ch. VIII]}). Let $T_{λ} f (x) = \int e^{iλ φ (x, y)} ψ (x, y) f (y) d y$ with $ψ \in C_{c}^{\infty} (R^{n} \times R^{n})$ and $det \partial_{x y}^{2} φ \neq = 0$ on $supp ψ$ . Then $∥ T_{λ} ∥_{L^{2} (R^{n}) \to L^{2} (R^{n})} \leq C λ^{- n /2}$ , with $C$ depending on finitely many derivatives of $φ$ and seminorms of $ψ$ .

The $λ^{- n /2}$ rate is exactly half the $L^{1} \to L^{\infty}$ elementary gain doubled through $T^{*} T$ , and the non-degenerate mixed Hessian is the operator-theoretic analogue of a non-degenerate critical point. This bound is the engine behind local smoothing for the wave equation, $L^{2}$ restriction estimates, and the Carleson-Sjölin theorem for Bochner-Riesz multipliers; the same phase-non-degeneracy hypothesis recurs in each.

Theorem (Fourier decay of curved surface measures; Stein §VIII.3, Sogge §1.2 ^{[Sogge §1-2]}). Let $S \subset R^{n}$ be a smooth compact hypersurface with non-vanishing Gaussian curvature, and $d μ = ψ d σ$ a smooth surface-carried measure. Then $∣ d μ (ξ) ∣ \leq C (1 + ∣ ξ ∣)^{- (n - 1) /2}$ , and the exponent $(n - 1) /2$ is sharp.

The curvature hypothesis enters as the non-degeneracy of the Hessian of the restricted linear phase at the two antipodal points where the surface normal aligns with $ξ$ . Vanishing curvature degrades the rate: a hypersurface with a flat point obeys only the slower bound dictated by the order of contact, computed via the van der Corput / Newton-polyhedron machinery. This curvature-decay dictionary is the quantitative core of the Fourier restriction and Kakeya circles of problems.

Synthesis. These three theorems are one phenomenon read at three altitudes, and putting these together is the substance of oscillatory-integral harmonic analysis. The stationary-phase expansion is the local model; the operator bound is its uniform-in-parameter lift; the surface-decay estimate is its geometric incarnation. The foundational reason they cohere is that each reduces, after a change of variables, to controlling a Gaussian-type integral against the Hessian of a phase, and the rate is dual to the order of vanishing of that phase — this is exactly the principle the van der Corput lemma isolates in one variable. The central insight is that curvature is the geometric name for phase non-degeneracy: the determinant of the surface's second fundamental form, the determinant of the phase Hessian, and the determinant of the mixed Hessian all play the identical role of a Gaussian normalisation, and where any of them vanishes the decay slows in a way governed by the Newton polyhedron. This generalises in two directions that the curriculum develops downstream — to Fourier integral operators, where the phase carries a Lagrangian and the Maslov index globalises the signature factor, and to restriction-extension theory, where the $(n - 1) /2$ surface-decay exponent is the input to the Stein-Tomas theorem. The bridge from this unit to the wider subject is the recognition that the entire apparatus of microlocal analysis — wave-front sets, propagation of singularities, the parametrices of 02.14.02 — is built by feeding non-degenerate phases into these same estimates and tracking the geometry of their critical sets.

Full proof set Master

Proposition (one-dimensional stationary phase, leading term). Let $φ \in C^{\infty} (R)$ be real with $φ^{'} (x_{0}) = 0$ , $φ^{''} (x_{0}) \neq = 0$ , and no other critical point on $supp ψ$ , $ψ \in C_{c}^{\infty}$ . Then $$ \int e^{i\lambda\varphi}\psi, dx = e^{i\lambda\varphi(x_0)}\left(\frac{2\pi}{\lambda}\right)^{1/2}\frac{e^{i\frac{\pi}{4}\operatorname{sgn}\varphi''(x_0)}}{|\varphi''(x_0)|^{1/2}},\psi(x_0) + O(\lambda^{-3/2}). $$

Proof. Translate so that $x_{0} = 0$ and subtract the constant $φ (0)$ from the phase, absorbing $e^{iλ φ (0)}$ as an overall factor; assume henceforth $φ (0) = 0$ , $φ^{'} (0) = 0$ , $φ^{''} (0) = a \neq = 0$ .

Step 1 (Morse normal form). By the Morse lemma, there is a smooth change of variables $u = u (x)$ near $0$ with $u (0) = 0$ , $u^{'} (0) = 1$ , such that $φ (x) = \frac{1}{2} a u^{2}$ — that is, the phase becomes exactly quadratic in the new coordinate. The map $x \mapsto u$ is a diffeomorphism on a neighbourhood of $0$ ; let $J (u)$ be the Jacobian $d x / d u$ , a smooth function with $J (0) = 1$ . Choosing $ψ$ supported in that neighbourhood (the rest contributes $O (λ^{- \infty})$ by non-stationary phase since $φ^{'} \neq = 0$ off $0$ ), the integral becomes $\int e^{iλa u^{2} /2} \tilde{ψ} (u) d u$ with $\tilde{ψ} (u) = ψ (x (u)) J (u) \in C_{c}^{\infty}$ , $\tilde{ψ} (0) = ψ (0)$ .

Step 2 (Gaussian evaluation). Reduce to the Fresnel integral. For $b \neq = 0$ , $$ \int_{\mathbb{R}} e^{i b u^2}, du = \left(\frac{\pi}{|b|}\right)^{1/2} e^{i\frac{\pi}{4}\operatorname{sgn} b}, $$ obtained by rotating the contour in the complex plane and reducing to the real Gaussian $\int e^{- t^{2}} d t = π$ ; the phase $e^{iπ sgn b /4}$ is the boundary value of the contour rotation. With $b = λa /2$ this gives $\int e^{iλa u^{2} /2} d u = (2 π / (λ ∣ a ∣))^{1/2} e^{iπ sgn (a) /4}$ .

Step 3 (amplitude expansion). Write $\tilde{ψ} (u) = \tilde{ψ} (0) + u ρ (u)$ with $ρ \in C_{c}^{\infty}$ . The term $\tilde{ψ} (0) \int e^{iλa u^{2} /2} d u$ produces the asserted leading coefficient times $ψ (0)$ . The remainder $\int e^{iλa u^{2} /2} u ρ (u) d u$ is integrated by parts once: $u e^{iλa u^{2} /2} = (iλa)^{- 1} \frac{d}{d u} e^{iλa u^{2} /2}$ , so the term equals $- (iλa)^{- 1} \int e^{iλa u^{2} /2} ρ^{'} (u) d u$ , which by Step 2 applied to the amplitude $ρ^{'}$ is $O (λ^{- 1} \cdot λ^{- 1/2}) = O (λ^{- 3/2})$ . Collecting, the integral equals the stated leading term plus $O (λ^{- 3/2})$ . $□$

Proposition (sublevel-set estimate; the integrated form of van der Corput). Let $φ \in C^{\infty} ([a, b])$ with $∣ φ^{(k)} ∣ \geq 1$ on $[a, b]$ for some $k \geq 1$ . Then for all $α > 0$ , $∣ {x \in [a, b] : ∣ φ (x) ∣ \leq α} ∣ \leq C_{k} α^{1/ k}$ , with $C_{k}$ independent of $φ$ and $[a, b]$ .

Proof. Argue by induction on $k$ . For $k = 1$ : $∣ φ^{'} ∣ \geq 1$ makes $φ$ strictly monotone, so ${∣ φ ∣ \leq α}$ is an interval on which $φ$ ranges over a set of length at most $2 α$ ; since $∣ φ^{'} ∣ \geq 1$ , the preimage has length at most $2 α = 2 α^{1/1}$ , giving $C_{1} = 2$ . For the step, suppose the bound for $k - 1$ and $∣ φ^{(k)} ∣ \geq 1$ . Then $φ^{(k - 1)}$ is strictly monotone with at most one zero $c$ ; off $(c - β, c + β)$ one has $∣ φ^{(k - 1)} ∣ \geq β$ , and rescaling $φ$ by $β$ and applying the inductive hypothesis bounds the sublevel set there by $C_{k - 1} (α / β)^{1/ (k - 1)}$ on each complementary piece. Adding the central interval of length $2 β$ gives $∣ {∣ φ ∣ \leq α} ∣ \leq 2 β + 2 C_{k - 1} (α / β)^{1/ (k - 1)}$ ; optimising $β \sim α^{1/ k}$ yields $C_{k} α^{1/ k}$ . The bound is the measure-theoretic twin of the van der Corput lemma: a lower bound on a derivative forces the phase to spend little time near any value, which is what makes its exponential average small. $□$

Connections Master

The composition law for pseudodifferential operators in 02.14.02 is proved by performing stationary phase in the intermediate Fourier variable; the asymptotic expansion in $i^{- ∣ α ∣} \partial_{ξ}^{α} a \partial_{x}^{α} b$ is exactly the stationary-phase series of this unit applied to the quadratic phase $- y \cdot ζ$ . That unit uses the method as a microlocal tool; this unit supplies the standalone theory.
The Fourier transform of 02.10.04 is the oscillatory integral $f (ξ) = \int e^{- 2 π i x \cdot ξ} f (x) d x$ with linear phase $- x \cdot ξ$ ; its Riemann-Lebesgue decay is the crudest non-stationary-phase statement, and the surface-measure decay theorem here is the curvature-quantified refinement that the linear-phase case cannot see.
The $L^{p}$ -interpolation machinery of this chapter feeds the oscillatory operator bounds: the $L^{2} \to L^{2}$ rate $λ^{- n /2}$ combines with the elementary $L^{1} \to L^{\infty}$ bound through interpolation built on the $L^{p}$ spaces of 02.07.06 to yield the full range of mixed-norm oscillatory estimates used in restriction theory.

Historical & philosophical context Master

The method of stationary phase descends from Stokes and Kelvin's nineteenth-century analyses of dispersive wave trains, where the dominant contribution to a wave packet was traced to the frequencies of stationary group velocity. Lord Kelvin gave the principle its first systematic statement in 1887 in his study of ship-wave patterns, isolating the stationary-phase points as the carriers of the asymptotics ^{[Sogge §1-2]}. The arithmetic side matured independently: J. G. van der Corput's 1921 paper Zahlentheoretische Abschätzungen introduced the derivative-test lemmas to bound exponential sums in the study of the lattice-point and divisor problems, supplying the uniform constants that make the estimates a genuine tool rather than a heuristic ^{[van der Corput 1921]}.

The modern synthesis is due to Hörmander, whose 1971 theory of Fourier integral operators placed stationary phase at the centre of microlocal analysis and whose 1973 paper isolated the non-degenerate-mixed-Hessian condition for the $λ^{- n /2}$ operator bound ^{[Hörmander §7.7]}. Stein's 1993 monograph organised the real-variable theory — oscillatory integrals of the first and second kind, curvature and Fourier decay, the connection to restriction and Bochner-Riesz — into the form taught today ^{[Stein Ch. VIII]}. The throughline from Kelvin's ship waves to the Fourier restriction conjecture is the recurring discovery that the geometry of a phase's critical set, measured by Hessians and Newton polyhedra, controls the quantitative decay of oscillatory averages.

Bibliography Master

@book{stein1993harmonic,
  author    = {Stein, Elias M.},
  title     = {Harmonic Analysis: Real-Variable Methods, Orthogonality, and Oscillatory Integrals},
  publisher = {Princeton University Press},
  series    = {Princeton Mathematical Series},
  number    = {43},
  year      = {1993}
}

@article{vandercorput1921,
  author  = {van der Corput, J. G.},
  title   = {Zahlentheoretische Abschätzungen},
  journal = {Mathematische Annalen},
  volume  = {84},
  pages   = {53--79},
  year    = {1921}
}

@book{hormander1990analysis1,
  author    = {Hörmander, Lars},
  title     = {The Analysis of Linear Partial Differential Operators I},
  publisher = {Springer},
  series    = {Grundlehren der mathematischen Wissenschaften},
  number    = {256},
  edition   = {2},
  year      = {1990}
}

@article{hormander1973oscillatory,
  author  = {Hörmander, Lars},
  title   = {Oscillatory integrals and multipliers on $FL^p$},
  journal = {Arkiv för Matematik},
  volume  = {11},
  pages   = {1--11},
  year    = {1973}
}

@book{sogge2017fourier,
  author    = {Sogge, Christopher D.},
  title     = {Fourier Integrals in Classical Analysis},
  publisher = {Cambridge University Press},
  series    = {Cambridge Tracts in Mathematics},
  number    = {210},
  edition   = {2},
  year      = {2017}
}

@book{steinshakarchi2011functional,
  author    = {Stein, Elias M. and Shakarchi, Rami},
  title     = {Functional Analysis: Introduction to Further Topics in Analysis},
  publisher = {Princeton University Press},
  series    = {Princeton Lectures in Analysis},
  number    = {IV},
  year      = {2011}
}

Prerequisites

02.14.02
02.10.04
02.07.06

Tier anchors

beginner: Stein-Shakarchi, Functional Analysis (Princeton Lectures in Analysis IV) §VIII.1; physical intuition from the stationary-phase principle in optics (Fermat's principle) and the ringing of a struck bell
intermediate: Stein, Harmonic Analysis: Real-Variable Methods, Orthogonality, and Oscillatory Integrals (Princeton 1993) §VIII.1-2; Stein-Shakarchi, Functional Analysis §VIII; Wolff, Lectures on Harmonic Analysis §1-2
master: Stein, Harmonic Analysis (Princeton 1993) Ch. VIII-IX; Hörmander, The Analysis of Linear Partial Differential Operators I (Springer 1990) §7.7; Sogge, Fourier Integrals in Classical Analysis 2e (Cambridge 2017) §1-2

References

Stein — Harmonic Analysis: Real-Variable Methods, Orthogonality, and Oscillatory Integrals · Princeton University Press 1993; Ch. VIII (Oscillatory integrals of the first kind) and Ch. IX (Oscillatory integral operators)
van der Corput — Zahlentheoretische Abschätzungen · Math. Ann. 84 (1921) 53-79; the derivative-test estimates for exponential sums and integrals
Hörmander — The Analysis of Linear Partial Differential Operators I · Springer, Grundlehren 256, 2e 1990; §7.7 (The method of stationary phase) and §7.6 (oscillatory integrals)
Stein-Shakarchi — Functional Analysis (Princeton Lectures in Analysis IV) · Princeton University Press 2011; Ch. VIII (Oscillatory integrals)
Sogge — Fourier Integrals in Classical Analysis · Cambridge Tracts in Mathematics 210, 2e 2017; Ch. 1-2 (stationary phase, the Carleson-Sjölin theorem, and restriction)
Hörmander — Oscillatory integrals and multipliers on FL^p · Ark. Mat. 11 (1973) 1-11; oscillatory integral operators and L^2 decay

Estimated time

beginner: 18m
intermediate: 45m
master: 85m