The Quantum Propagator

Every tool this chapter has built answers the same question — given the state now, what is it later? — and answers it operator-first: $\hat{U}(t)$ acting on kets. This article asks the question in the laboratory's own coordinates. If a particle's amplitude is smeared over positions, how much amplitude arrives at point $x$ at time $t$ from each launching point $x'$ at time $t'$ ? The answer is one matrix element of the evolution operator,

K(x,t;x',t')=\langle x|\hat{U}(t,t')|x'\rangle,

the propagator — the amplitude for the elementary journey $(x',t')\to(x,t)$ , from which every journey is assembled by superposition. The object earns its keep through its properties, and this article proves the full pack. The propagator is the impulse response of the Schrödinger equation: a solution in its own right, seeded by a δ-spike. It composes: the amplitude to go far is the sum over stopovers of amplitudes to go near — Chapman–Kolmogorov for waves, and precisely Huygens' three-century-old picture of wavelets begetting wavelets, machine-verified here to $10^{-14}$ in a box. It diagonalizes over energy: the spectral form $K=\sum_n\psi_n(x)\psi_n^*(x')e^{-iE_n(t-t')/\hbar}$ turns any solved spectrum into a dynamics machine. And iterating the composition — slicing the journey into infinitely many stopovers — opens the door Feynman walked through in 1942; we stop respectfully at the doorway, since that story is told at the Schrödinger-equation article. Two exactly solvable kernels — the free particle's spreading Gaussian and the oscillator's refocusing Mehler kernel — get the next two articles to themselves; the resolvent, $K$ 's energy-domain twin, closes the run. This article owns the object, its laws, and the craft of reading them.

Key takeaways

The propagator $K(x,t;x',t')=\langle x|\hat{U}(t,t')|x'\rangle$ is the amplitude for the elementary journey $(x',t')\to(x,t)$ ; the state assembles as $\psi(x,t)=\int K(x,t;x',t')\,\psi(x',t')\,dx'$ .
Spectral form for static $\hat{H}$ : $K=\sum_n\psi_n(x)\psi_n^*(x')\,e^{-iE_n(t-t')/\hbar}$ — diagonalize once, propagate forever.
Initial condition: $K\to\delta(x-x')$ as $t\to t'$ — the propagator is the Schrödinger equation's impulse response (a δ-seeded solution of the TDSE in $(x,t)$ ).
Composition law: $K(3;1)=\int dx_2\,K(3;2)K(2;1)$ — every point of the intermediate wave re-emits: Huygens' principle as an operator identity (machine-checked $10^{-14}$ ).
Backward reading: $K(x,t;x',t')^*=K(x',t';x,t)$ — conjugation swaps departure and arrival.
Iterating composition over infinitely many time slices previews the path integral (Dirac 1933 → Feynman 1942; story owned elsewhere).
Key idea: dynamics is a weighted rebroadcast — every point of the present wavefunction emits the kernel as its wavelet, and the future is their superposition.

The core insight

The propagator $K(x,t;x',t')=\langle x|\hat{U}(t,t')|x'\rangle$ is the transition amplitude between spacetime points: a δ-seeded solution of the Schrödinger equation that propagates any state by superposition, composes over intermediate times, and expands over the energy spectrum as $\sum_n\psi_n(x)\psi_n^*(x')e^{-iE_n(t-t')/\hbar}$ .

Where this lives today

Kernel language is the working dialect wherever dynamics must be computed many times from one solve. Semiconductor and quantum-well simulation propagates carriers with spectral kernels exactly like equation (3); attosecond spectroscopy reconstructs electron motion as kernel-propagated wave packets; and quantum chemistry's split-operator codes are literally composition (Step 2) run on alternating kinetic and potential half-kernels, thousands of slices per molecular vibration. At the research frontier the kernel is the object of interest in itself: out-of-time-order correlators probe how $K$ spreads information (quantum chaos), tensor-network methods compress many-body kernels, and "learning the propagator" — fitting $K$ from data with its properties pack as constraints — is a live program in quantum device characterization. The audit of Example 2 is, in miniature, how those learned kernels are validated.

The physical picture

Drop a pebble in still water and the ring that spreads is the pond's answer to a point disturbance — its impulse response. Know that one ring and you know everything: any initial splash is a crowd of pebbles, and the later surface is the sum of their rings. The propagator is quantum mechanics' ring. Seed the Schrödinger equation with a position spike $\delta(x-x')$ and $K(x,t;x',t')$ is what evolves; seed it with any $\psi$ and linearity assembles the future from kernel copies weighted by the present.

Figure

The kernel as impulse response — every point of the present re-emits, and the future is the superposition

The composition law upgrades the picture from metaphor to mechanism. To propagate from $t_1$ to $t_3$ , stop anywhere at $t_2$ : the amplitude is the product of legs, summed over every stopover. This is Huygens' wavelets exactly — each point reached by the wave becomes a fresh source — and it is worth pausing on what quantum mechanics adds to the 1678 optics: the amplitudes are complex, so stopovers interfere. Most stopover chains cancel; the survivors trace, in the classical limit, the paths of stationary phase. Iterate the slicing $t_1\to t_2\to\cdots\to t_N$ and the sum-over-stopovers becomes a sum over histories — the path-integral doorway, glimpsed and deliberately not entered (the Dirac-remark-to-Feynman-thesis story lives in the Schrödinger equation article's portrait).

Figure

Composition — amplitude to go far is the interfering sum over stopovers; iterated, it becomes a sum over histories

The object and its laws

Definition first, in the chapter's own grammar. With $\hat{U}(t,t')$ the evolution operator (The Time-Evolution Operator) and $|x\rangle$ the position kets of Position Representation:

K(x,t;x',t')=\langle x|\hat{U}(t,t')|x'\rangle \qquad (1)

where: $K$ is a complex amplitude density (dimension 1/length in 1D — one factor of the position-ket normalization), read as arrive at $x$ , having launched from $x'$ . It propagates states by pure linear algebra — insert the continuous resolution $\int dx'\,|x'\rangle\langle x'|=\hat{1}$ (Continuous Spectra and Dirac Delta Normalization's license) into $|\psi(t)\rangle=\hat{U}|\psi(t')\rangle$ and project onto $\langle x|$ :

\psi(x,t)=\int_{-\infty}^{\infty}K(x,t;x',t')\,\psi(x',t')\,dx' \qquad (2)

where: the integral is the weighted rebroadcast of the hero figure — the kernel against the past, summed over launch points. Equation (2) is the whole practical point: one function of two positions and one time interval replaces the operator exponential for every initial condition. For a time-independent Hamiltonian the kernel inherits the chapter's spectral machinery. Insert the energy resolution $\sum_n|n\rangle\langle n|$ (assembled by Spectral Decomposition of Time Evolution, consumed here) into (1):

K(x,t;x',t')=\sum_n \psi_n(x)\,\psi_n^*(x')\,e^{-iE_n(t-t')/\hbar},\qquad \psi_n(x)=\langle x|n\rangle \qquad (3)

where: each term is a standing wave carrying its own phase clock — the position-space shadow of $\hat{U}=\sum_n e^{-iE_nt/\hbar}|n\rangle\langle n|$ . Equation (3) converts any solved spectrum into a dynamics machine, and it is the form every exhibit in this article computes with.

Derivation AThe spectral form of the kernel

We derive equation (3) and read off what its structure commits us to. The strategy is two insertions of already-banked machinery: the energy-basis assembly of the evolution operator (owned by Spectral Decomposition of Time Evolution, consumed here — its derivation is never re-run) and the position-representation dictionary of Position Representation. The instructive part is not the two lines of algebra but the audit afterward: the sum is an amplitude sum, its diagonal terms are static, and every piece of motion lives in the cross terms — facts that the box exhibit then displays numerically.

Step 1. Start from the assembled evolution operator for static $\hat{H}$ (cited): $\hat{U}(t-t')=\sum_n e^{-iE_n(t-t')/\hbar}|n\rangle\langle n|$ . Take the position matrix element demanded by definition (1):

K(x,t;x',t')=\sum_n e^{-iE_n(t-t')/\hbar}\,\langle x|n\rangle\langle n|x'\rangle

Step 2. Translate with the dictionary $\langle x|n\rangle=\psi_n(x)$ , $\langle n|x'\rangle=\psi_n^*(x')$ :

K(x,t;x',t')=\sum_n\psi_n(x)\,\psi_n^*(x')\,e^{-iE_n(t-t')/\hbar}

Step 3. Audit the structure. Set $t=t'$ : the sum becomes $\sum_n\psi_n(x)\psi_n^*(x')=\delta(x-x')$ — the completeness relation in position language, so the δ initial condition is built in. Set $x=x'$ and integrate: $\int K(x,t;x,0)\,dx=\sum_n e^{-iE_nt/\hbar}$ — the kernel's trace is a pure spectral fingerprint, a fact the Green-functions article will spend. Why the cross terms carry all the motion: term $n$ alone contributes $|\psi_n|^2$ -type structure with a phase that cancels in any single-term modulus — one clock is silent. Interference between different $n$ 's produces every beat, carpet and revival; the spectral kernel is the chapter's clocks-on-a-wall picture, projected into position space. *Verified: the two-mode ISW propagation through this form matches the analytic phase evolution to $10^{-10}$ ; the $t\to t'$ limit reproduces a band-limited test function ( $2\times10^{-4}$ at 60 modes); composition and reversal hold to $10^{-14}$ / $10^{-12}$ (article verifier).*

(For continuous spectra the sum becomes an integral by the box-bridge of Continuous Spectra; the free particle's closed form is the next article's derivation.)

Go deeper · Derivation B · 1 minThe properties pack — initial condition, composition, TDSE, reversal

Derivation BThe properties pack — initial condition, composition, TDSE, reversal

We derive the four laws that make $K$ usable, each inherited from one property of $\hat{U}$ by taking position matrix elements. The strategy is deliberate bookkeeping: nothing new is proven about dynamics — the point is to watch operator facts (identity at $t=t'$ , group composition, the operator Schrödinger equation, unitarity) turn into integral identities for a two-point function. Sameness of content, change of register: this is the position-representation dictionary doing its job.

Step 1. Initial condition. At $t=t'$ , $\hat{U}=\hat{1}$ , so

K(x,t';x',t')=\langle x|x'\rangle=\delta(x-x')

Why this names $K$ the impulse response: combined with Step 3, $K$ is the solution of the TDSE whose initial data is a position spike — the pond's ring. The δ is the continuum orthonormality of Continuous Spectra, cited.

Step 2. Composition. The group law $\hat{U}(t_3,t_1)=\hat{U}(t_3,t_2)\hat{U}(t_2,t_1)$ (The Time-Evolution Operator's requirement) with the continuous resolution inserted between the factors:

K(x_3,t_3;x_1,t_1)=\int dx_2\;K(x_3,t_3;x_2,t_2)\,K(x_2,t_2;x_1,t_1)

Why the stopover integral is Huygens with phases: each intermediate point re-emits with the amplitude it received; the complex sum lets stopovers cancel or reinforce — the interference that optics' real-amplitude wavelets never had.

Step 3. $K$ solves the TDSE in its first slot. Apply $i\hbar\partial_t$ to (1) and use $i\hbar\,\partial_t\hat{U}=\hat{H}\hat{U}$ :

i\hbar\,\frac{\partial K}{\partial t}=\langle x|\hat{H}\hat{U}|x'\rangle=\hat{H}_x\,K(x,t;x',t')

with $\hat{H}_x$ the differential operator of the position dictionary acting on the $x$ -dependence.

Step 4. Reversal. Unitarity $\hat{U}^\dagger(t,t')=\hat{U}(t',t)$ gives

K(x,t;x',t')^*=\langle x'|\hat{U}(t',t)|x\rangle=K(x',t';x,t)

— conjugating the kernel swaps departure and arrival: the film played backwards. Verified: all four machine-checked on the ISW spectral kernel (60 modes) — composition to relative $2\times10^{-14}$ ; δ-limit reproducing a band-limited test function ( $2\times10^{-4}$ at $t=10^{-6}$ ); the pointwise TDSE residual $i\partial_tK+\frac{1}{2}\partial_x^2K$ at $4\times10^{-5}$ relative (15-mode truncation, fine grid); $K(t)^\dagger=K(-t)$ to $10^{-12}$ (article verifier).

Reading a kernel: the box as exhibit

The infinite square well makes every abstraction concrete because its spectral sum (3) is explicit: $\psi_n=\sqrt{2/L}\sin(n\pi x/L)$ , $E_n=n^2E_1$ . Three readings, all machine-run. The δ-seed reading: truncate the sum at 60 modes and apply $K(t\to0)$ to a test function — it reproduces the function's 60-mode content to $2\times10^{-4}$ ; the kernel is a δ-spike smeared only by the truncation, the band-limit honesty that any numerical kernel carries. The propagation reading: feeding the two-mode state $(\psi_1+\psi_2)/\sqrt2$ through equation (2) reproduces the analytic phase evolution to $10^{-10}$ — and at half the beat period the centroid lands mirror-opposite its start ( $0.320L\to0.680L$ , exactly $L-x_0$ ), the sloshing that Spectral Decomposition anatomized, now produced by a kernel integral rather than a mode sum. The structure reading: because the box spectrum is commensurate, $K$ itself revives — at $T_{\rm rev}$ the kernel returns to a δ (up to the mirror/parity structure of the half-revival), the carpet physics of 0892 seen from the kernel's side; its worked numbers stay there. The kernel's practical register is worth stating plainly: equation (3) is how quantum dynamics is actually computed when spectra are known — quantum-well optoelectronics, molecular vibronic propagation, every "diagonalize once, evolve many" code path. The two kernels nature hands you in closed form — free space and the harmonic well — are the next two articles; everything else is (3) plus numerics or the semiclassical craft the APRX chapter builds.

Limiting cases

case	result	meaning
$t\to t'$	$K\to\delta(x-x')$	impulse response; nothing has propagated yet
$\hat{H}$ static	spectral form (3)	one diagonalization = all dynamics
$V=0$	$K_0=\sqrt{\frac{m}{2\pi i\hbar t}}\,e^{im(x-x')^2/2\hbar t}$	evaluated next article (its derivation, its checks)
$V=\frac12m\omega^2x^2$	Mehler kernel	exact, refocusing — two articles on
composition iterated $N\to\infty$	sum over histories	the path-integral doorway (story owned by 0825's band)
conjugation	$K^*(x,t;x',t')=K(x',t';x,t)$	microreversibility of amplitudes

Common trapThe propagator is an amplitude, not a probability

$|K(x,t;x',0)|^2$ is not "the probability the particle went from $x'$ to $x$ " — and treating it so breaks everything the kernel is for. Probabilities attach to complete journeys of a state: you must assemble $\psi(x,t)=\int K\psi_0$ first and square after the integral, so different launch points interfere before any modulus is taken. (The classical-stochastic object that composes with probabilities — a Markov transition density — obeys Chapman–Kolmogorov without phases; the quantum kernel's law looks identical but runs on amplitudes, and that single difference carries all of interference.) The related trap: the spectral sum (3) is not a sum of "probabilities of taking level $n$ " — its terms are complex and their cross-terms are exactly the beats, carpets and revivals of this chapter.

Worked examples

Example 1 — kernel dynamics in a 0.5 nm quantum box. An electron sits in an ISW of width $L=0.5$ nm in the state $(\psi_1+\psi_2)/\sqrt2$ . Use kernel language end to end.

Step 1. The clock scale: $E_1=\pi^2\hbar^2/2m_eL^2=$ 1.50 eV — four times the 1 nm well's 0.376 eV, by the $1/L^2$ law.

Step 2. The kernel: two terms of (3) suffice for this state, with Bohr gap $E_2-E_1=3E_1$ . The relevant clock is the beat period $T_{21}=h/3E_1=$ 0.917 fs.

Step 3. Propagate by equation (2) at $t=T_{21}/2$ (machine-run, 60-mode kernel): the centroid moves from $0.320L$ to $0.680L$ — mirror-imaged through the center, $\langle x\rangle(T/2)=L-\langle x\rangle(0)$ exactly (checked to $10^{-6}$ ).

Step 4. Read it physically: the packet sloshes across a half-nanometer box roughly once per femtosecond — the attosecond-to-femtosecond traffic that ultrafast spectroscopy now films directly. Every number came from one kernel; a different initial state would reuse it untouched. (The full three-mode carpet and revival numbers of the 1 nm well belong to Spectral Decomposition of Time Evolution.)

Example 2 — auditing a kernel before trusting it. You are handed a numerically computed $K$ for some 1D device potential. Certify it with the properties pack, quantitatively.

Step 1. δ-test: apply $K(\varepsilon)$ for tiny $\varepsilon$ to a smooth test function; demand return to the function within the mode-truncation error. The ISW reference run returns $2.4\times10^{-4}$ at $\varepsilon=10^{-6}$ (natural units) with 60 modes — your tolerance scales with your basis.

Step 2. Composition test: pick two intervals and check $\int K(3;2)K(2;1)dx_2$ against $K(3;1)$ . Reference: relative error $2\times10^{-14}$ — this one should sit at machine precision, since it tests only the group law, not discretization.

Step 3. TDSE residual: finite-difference $i\hbar\partial_tK+\frac{\hbar^2}{2m}\partial_x^2K-VK$ pointwise; reference residual $4\times10^{-5}$ relative on a fine grid. This is the test that catches a wrong potential or a dropped factor of $\hbar$ .

Step 4. Reversal test: $K(t)^\dagger$ vs $K(-t)$ ( $10^{-12}$ reference) — catches non-unitary time-stepping (a lossy integrator fails here first). Four cheap checks, each tied to one law of Derivation A: a kernel that passes all four is trustworthy for any initial state, which is precisely the economy the propagator was invented for.

Origins · 1 minChristiaan Huygens and how this was discovered

Christiaan Huygens

1629–1695 · The Hague, Paris

The composition law of this article — every point reached by a wave becomes a source — was proposed for light in 1678 by a Dutch diplomat's son who was, by then, Europe's most complete natural philosopher. Huygens' Traité de la Lumière (presented 1678, printed 1690) modeled light as a disturbance in an ether and asserted the construction now bearing his name: each point of a wavefront emits a secondary wavelet, and the later front is their common envelope. With it he explained reflection, refraction, and — his showpiece — the double refraction of Iceland spar, deriving the ellipsoidal wavelets that anisotropic crystals demand. What his century lacked was the concept his construction begged for: interference between the wavelets. Huygens took envelopes where quantum mechanics takes phased sums; Fresnel supplied the missing phases for optics in 1818 (a story told at the next article), and quantum mechanics inherited the completed pattern wholesale — equation (2) is Huygens' construction with complex amplitudes, and the composition law is his wavelets-beget-wavelets run as an exact operator identity. That a bookkeeping picture for seventeenth-century optics became the load-bearing structure of quantum dynamics — and, iterated, of the path integral — is among the longest-running returns on a single physical idea. (His pendulum-clock mechanics wears a different portrait in the mechanics track; here he is the wavelet man only.)

History

1678

— Huygens presents the wavelet construction to the Académie; Traité de la Lumière prints in 1690

1818

— Fresnel adds phases and interference to the wavelets; diffraction becomes computable

1900

— Bachelier's thesis composes transition densities — Chapman–Kolmogorov in probability

1926

— Schrödinger's equation; its Green-function solutions carry the quantum kernel implicitly

1933

— Dirac's remark: the short-time amplitude "corresponds to" the classical action's exponential

1942

— Feynman's thesis iterates the composition law into the sum over histories

Going further

References. Sakurai & Napolitano, Modern Quantum Mechanics, 2nd ed., §2.6 · Shankar, Principles of Quantum Mechanics, 2nd ed., Ch. 5 & 8 · Cohen-Tannoudji, Diu & Laloë, Quantum Mechanics, Vol. I · Feynman & Hibbs, Quantum Mechanics and Path Integrals, Ch. 2–3 · Huygens, Traité de la Lumière (1690).

Provenance. Definitions and equations cross-checked against Sakurai & Napolitano Sec. 2.6, Shankar Ch. 5 & 8, and Cohen-Tannoudji Vol. I. A verifier should: (1) redo the Gaussian momentum integral to confirm the free-particle prefactor $(m/2\pi i\hbar t)^{1/2}$ and exponent $im(x-x')^2/2\hbar t$ ; (2) verify the harmonic-oscillator (Mehler) kernel reduces to the free kernel as $\omega\to0$ ; (3) confirm the half-line time transform yields the resolvent with $E-E_n+i0^+$ denominators; (4) check the coincidence limit gives $\delta(x-x')$ . All numbers are standard textbook values; none fabricated. For this edition, item (4) plus the full properties pack were machine-executed on the ISW spectral kernel (numpy, 60 modes): composition $2\times10^{-14}$ ; δ-limit $2\times10^{-4}$ band-limited; pointwise TDSE residual $4\times10^{-5}$ ; reversal $10^{-12}$ ; two-mode propagation vs analytic $10^{-10}$ with the mirrored centroid $0.320L\to0.680L$ ; and the physical anchors ( $E_1=1.50$ eV, $T_{21}=0.917$ fs at $L=0.5$ nm) from CODATA constants. Items (1)–(3) are deliberately NOT executed here: the free-kernel integral is owned by Propagator for the Free Particle (next serial), the Mehler kernel and its $\omega\to0$ limit by Propagator for the Harmonic Oscillator, and the resolvent transform by Green Functions in Quantum Mechanics — each cited by name. Per the chapter ownership map: the spectral assembly is Spectral Decomposition of Time Evolution's (consumed); the continuous resolution and δ-orthonormality are Continuous Spectra and Dirac Delta Normalization's; the path-integral story stays with the Time-Dependent Schrödinger Equation's Feynman band; retarded/advanced causal structure → Retarded and Advanced Green Functions BY NAME.

Builds on: The Time-Evolution Operator · Spectral Decomposition of Time Evolution · Position Representation · Continuous Spectra and Dirac Delta Normalization

Leads to: Propagator for the Free Particle and Gaussian Spreading · Propagator for the Harmonic Oscillator · Green Functions in Quantum Mechanics · Retarded and Advanced Green Functions

Is the propagator the same thing as a Green's function?

Same family, different boundary conditions — and the distinction is the next-but-one article's whole subject. $K$ solves the TDSE with a δ initial condition and exists for all $t$ ; the (retarded) Green function solves the inhomogeneous equation $(i\hbar\partial_t-\hat{H})G=\delta(t)\delta(x-x')$ and vanishes before the source fires — it equals $-\frac{i}{\hbar}\theta(t)K$ up to convention. In casual usage "propagator" and "Green function" blur; in calculations the $\theta(t)$ and the resulting $i\varepsilon$ analytic structure matter enormously. This chapter keeps them in separate articles for exactly that reason.

Why does the composition law hold exactly even though the stopover points are not "where the particle was"?

Because it is bookkeeping for operators, not a story about trajectories. $\hat{U}(t_3,t_1)=\hat{U}(t_3,t_2)\hat{U}(t_2,t_1)$ is exact; inserting $\int|x_2\rangle\langle x_2|dx_2$ is the resolution of identity, not a measurement. No collapse occurs at $t_2$ — all stopovers remain coherent, which is why they interfere. Had you measured position at $t_2$ , the phases would decohere and you would get the classical Chapman–Kolmogorov law for probabilities instead: the difference between the two laws is literally the difference between inserting identity and inserting a measurement.

What does the kernel look like for a time-dependent Hamiltonian?

The definition (1) survives untouched — $\hat{U}(t,t')$ just becomes the time-ordered exponential of Time-Ordered Exponential and the Dyson Series — and so do the composition law, the δ-limit and the TDSE property (with $\hat{H}(t)$ ). What dies is the spectral shortcut (3), since there is no single eigenbasis to expand in; $K$ then depends on $t$ and $t'$ separately, not just their difference. Practical routes: Dyson/Magnus expansions for the kernel, Floquet quasi-energy sums for periodic driving, or direct split-operator numerics — which are themselves the composition law iterated.

How does a kernel encode bound states versus scattering states at the same time?

Additively. For a Hamiltonian with both (a finite well, say), the spectral form splits into $\sum_{\rm bound}\psi_b(x)\psi_b^*(x')e^{-iE_bt/\hbar}$ plus a continuum integral over scattering states with δ-normalization (the mixed-spectra doctrine of Continuous Spectra). At long times the continuum piece disperses away from any finite region while the bound sum keeps beating forever — watch a kernel-propagated packet and the transient that leaves is continuum, the persistent slosh that remains is the bound-state subspace. The energy-domain version of this split — poles versus branch cut — is the Green-functions article's opening exhibit.

Is K measurable, or just a calculational device?

Interferometrically measurable, in the same sense any amplitude is: compare a system evolved along two arms and the relative phase of kernel values shows up in fringes. Modern examples are direct: attosecond streaking reconstructs the phase of electron wave packets (kernel-propagated), and "process tomography" in quantum computing reconstructs the full evolution operator of a qubit register — literally measuring the finite-dimensional analogue of $K$ , matrix element by matrix element, from interference data. The kernel is no more (and no less) abstract than the wavefunction it propagates.

Why bother with K when I can just diagonalize H and use the spectral sum for my state?

For one state, don't bother — the mode sum is cheaper. The kernel pays when the same system faces many initial conditions (device simulation, pump–probe scans over launch parameters), when you need the response structure itself (DOS from the kernel's trace, coming in the Green-functions article), when no clean eigenbasis exists (driven systems — the kernel survives, the mode sum doesn't), or when you are headed for semiclassics (the stationary-phase analysis of kernels is the WKB/Van Vleck road, previewed next article). The kernel is the system's dynamics factored out of any particular state — that factorization is the product being sold.