Type-3 and Type-4 Generating Functions

Two members of the generating-function family remain, and they answer a question the first two quietly dodged: what if the natural handle on the old state is its momentum, not its position? Type 1 and type 2 both steer with the old coordinate $q$ . Types 3 and 4 steer with the old momentum $p$ — paired with the new coordinate ( $F_3(p,Q,t)$ ) or the new momentum ( $F_4(p,P,t)$ ). Nothing deep changes: each is one more Legendre transform applied to the master identity, the same $\pm qp$ and $\pm QP$ bookkeeping that built $H$ from $L$ . But the sign pattern shifts in a way worth mastering once and never again by memorization: whenever the old momentum becomes an independent argument, the old coordinate is recovered with a minus sign, $q=-\partial F_3/\partial p=-\partial F_4/\partial p$ . Why keep two more types on the shelf? Coverage. Every type has a blind spot where its argument pair degenerates: type 1 cannot make the identity, type 2 cannot make the full coordinate–momentum exchange. The momentum-driven types own exactly the maps the others fumble — the sign flip $Q=-q,\ P=-p$ falls out of $F_3=pQ$ in one line, and the phase-space quarter-turn $Q=p,\ P=-q$ out of $F_4=pP$ . Together the four form an atlas: every canonical transformation is representable by at least one chart, possibly different types for different degrees of freedom. The taxonomy and master table live in Generating Functions and the Four Types; this page owns the momentum-side machinery and its own worked transformations — including the harmonic oscillator solved from the momentum side, landing on the same action–angle chart as the type-1 route.

Key takeaways

$F_3(p,Q,t)$ and $F_4(p,P,t)$ complete the family: the old momentum becomes an independent argument in place of the old coordinate.
Type-3 rules: $q=-\partial F_3/\partial p$ , $P=-\partial F_3/\partial Q$ ; type-4 rules: $q=-\partial F_4/\partial p$ , $Q=+\partial F_4/\partial P$ . The universal Hamiltonian rule $K=H+\partial F/\partial t$ holds for all four types.
They are Legendre transforms of the first two: $F_3=F_1-qp$ and $F_4=F_2-qp=F_3+QP$ ; the recurring minus on $q$ is the standard Legendre sign.
$F_3=pQ$ generates the sign flip $Q=-q$ , $P=-p$ (a 180° phase-space rotation); $F_4=pP$ generates the exchange $Q=p$ , $P=-q$ (a 90° rotation) — the two staple maps types 1 and 2 handle awkwardly or not at all.
Existence conditions: $\det(\partial^2F_3/\partial p\,\partial Q)\neq0$ , $\det(\partial^2F_4/\partial p\,\partial P)\neq0$ ; when one type degenerates, another covers the map.
Key idea: types 3 and 4 are Legendre transforms of types 1 and 2 that replace the old coordinate by the old momentum as an independent variable — and the minus signs in $q=-\partial F_3/\partial p$ and $q=-\partial F_4/\partial p$ are the direct signature of that transform.

The core insight

Type-3 and Type-4 are Legendre transforms of Type-1 and Type-2 that replace the old coordinate $q$ by the old momentum $p$ as an independent variable. The minus signs in $q=-\partial F_3/\partial p$ and $q=-\partial F_4/\partial p$ are the direct signature of that Legendre transform.

Where this lives today

The canonical quarter-turn $(q,p)\mapsto(Q,P)=(p,-q)$ is implemented, or closely approximated, in many physical systems. In quantum optics, a light mode's quadratures $\hat X$ and $\hat Y$ are a conjugate pair — literally the $q$ and $p$ of an oscillator — and letting the mode evolve a quarter cycle, or shifting the local oscillator's phase by $90^\circ$ in a homodyne detector, enacts precisely the $F_4=pP$ exchange: what was position noise is read out as momentum noise. Squeezed-light interferometry — since 2019 a permanent upgrade in LIGO and Virgo — lives on choosing which rotated quadrature carries the squeezing, i.e. on dialing a one-parameter family of type-4-like rotations to put quantum noise where the measurement is not looking. The same map runs in classical hardware: a converging lens performs an optical Fourier transform between its focal planes, exchanging transverse position for transverse spatial frequency — a literal quarter-turn in ray phase space, generalized in modern beam optics to fractional Fourier transforms of arbitrary angle. When an optical engineer cascades lenses to reshape a beam's phase-space footprint, she is composing momentum-driven canonical transformations, Goldstein's Table 9.1 running at the speed of light.

The physical picture

For each conjugate pair — old and new — a generating function must pick one steering variable. Two pairs, two choices each: a four-entry menu. That is the entire taxonomy, and it is worth seeing as a grid:

Figure

The four-types grid — each type picks one steering variable per pair; types 3 and 4 are the momentum-driven row

The picture to keep for the two flagship maps is rotation of the phase plane. The exchange $Q=p,\ P=-q$ is a quarter-turn: the position axis rotates onto the momentum axis. Apply it twice and you get the sign flip $Q=-q,\ P=-p$ — a half-turn. Apply it four times and you are home: the exchange is a fourth root of the identity, exactly like multiplying by $i$ in the complex plane. This is not a metaphor. In the $(q,p)$ plane with suitable units, $F_4=pP$ implements the rotation matrix by $90^\circ$ , and the harmonic oscillator's own time evolution — which circulates phase points around ellipses — interpolates the quarter-turn continuously.

Figure

Momentum-driven staple maps as rotations of the phase plane — F4 gives the quarter-turn, F3 the half-turn

Why does driving with the old momentum bring a minus sign? Because the Legendre swap $p\,\mathrm{d}q=\mathrm{d}(pq)-q\,\mathrm{d}p$ trades the term $+p\,\mathrm{d}q$ for $-q\,\mathrm{d}p$ : the coefficient of the new independent differential is minus the old conjugate. There is nothing to memorize once you can write that one identity — every sign in the four-type table regenerates from it in seconds.

The momentum-side rules, built from the master identity

All four types descend from one statement — the old and new phase-space one-forms differ by an exact differential:

p\,\mathrm{d}q-P\,\mathrm{d}Q+(K-H)\,\mathrm{d}t=\mathrm{d}F \qquad (1)

where: $q,p$ — old pair; $Q,P$ — new pair; $H,K$ — old and new Hamiltonians; $F$ — the generating function, whose natural arguments are decided by which differentials survive on the left. As written, the surviving differentials are $\mathrm{d}q$ and $\mathrm{d}Q$ — the type-1 configuration. Legendre transforms rearrange them. To make the old momentum independent, swap the first term:

F_3(p,Q,t)=F_1(q,Q,t)-qp \qquad (2)

where: $-qp$ — the Legendre term exchanging $q$ for $p$ , the mirror image of $H=p\dot q-L$ exchanging $\dot q$ for $p$ (see Legendre Transformation). Carrying the differentials through (Derivation A) yields the type-3 rules:

q=-\frac{\partial F_3}{\partial p},\qquad P=-\frac{\partial F_3}{\partial Q},\qquad K=H+\frac{\partial F_3}{\partial t} \qquad (3)

where: each derivative holds the other member of $(p,Q)$ fixed. Read the signs structurally: the old slot is now a momentum, so its conjugate $q$ returns negated; the new slot is still a coordinate, so — exactly as in type 1 — its conjugate $P$ returns negated too. Both minus signs trace to the orientation of (1). One more Legendre transform, on the new pair this time, produces the all-momentum type:

F_4(p,P,t)=F_2(q,P,t)-qp=F_3(p,Q,t)+QP \qquad (4)

where: the two routes — old-side swap of $F_2$ , or new-side swap of $F_3$ — must agree, and Derivation C checks that they do. The working rules:

q=-\frac{\partial F_4}{\partial p},\qquad Q=+\frac{\partial F_4}{\partial P},\qquad K=H+\frac{\partial F_4}{\partial t} \qquad (5)

where: old-momentum slot negates its conjugate (as in type 3); new-momentum slot returns its conjugate positively (as in type 2). The Hamiltonian rule never changes: Legendre transforms shuffle phase-space slots and leave the time slot untouched. Existence is a non-degeneracy condition, one per type:

\det\!\left(\frac{\partial^{2}F_3}{\partial p\,\partial Q}\right)\neq0,\qquad \det\!\left(\frac{\partial^{2}F_4}{\partial p\,\partial P}\right)\neq0 \qquad (6)

where: the mixed Hessians must be invertible for the implicit relations to be solvable for the map. When your target transformation kills one Hessian, it generically leaves another alive — that is the atlas at work.

Derivation AType-3 rules straight from the one-form identity

We are deriving the type-3 relations without any detour through $F_1$ : starting from the master identity (1), we perform the Legendre swap on the old pair and read off coefficients. This is the cleanest template for generating any of the four types — learn this once and the whole table becomes reproducible on demand. The step to watch is the absorption of the exact piece $\mathrm{d}(pq)$ into a redefinition of the generating function; that redefinition is what "changing type" means.

Step 1. Start from the master identity, with $F$ the type-1-configured generator:

p\,\mathrm{d}q-P\,\mathrm{d}Q+(K-H)\,\mathrm{d}t=\mathrm{d}F

Step 2. Swap the old-pair term using the product rule, $p\,\mathrm{d}q=\mathrm{d}(pq)-q\,\mathrm{d}p$ :

-q\,\mathrm{d}p-P\,\mathrm{d}Q+(K-H)\,\mathrm{d}t=\mathrm{d}(F-pq)

Why this is the whole trick: the exact differential $\mathrm{d}(pq)$ can be moved across the equation and bundled into the generator, because "generator plus any function of phase space" is still a legitimate boundary term. The bundle $F-pq$ is, by inspection of the left side, naturally a function of $(p,Q,t)$ — its surviving differentials.

Step 3. Define $F_3\equiv F-pq$ and match coefficients of the independent differentials $\mathrm{d}p$ , $\mathrm{d}Q$ , $\mathrm{d}t$ :

\frac{\partial F_3}{\partial p}=-q,\qquad \frac{\partial F_3}{\partial Q}=-P,\qquad \frac{\partial F_3}{\partial t}=K-H

Step 4. Sanity-check on the linear generator $F_3=pQ$ : the rules give $q=-Q$ and $P=-p$ — the sign flip, a half-turn of phase space. Bracket check in old variables: $\{Q,P\}=\partial_qQ\,\partial_pP-\partial_pQ\,\partial_qP=(-1)(-1)-0=1$ ✓. Note: the same two-step template with $+QP$ instead of $-pq$ regenerates the type-2 rules, confirming that the same Legendre-transform procedure generates the rules for every generating-function type.

Go deeper · Derivation B · 1 minThe exchange map from F4 and its canonicity

Derivation BThe exchange map from F4 and its canonicity

We are constructing the canonical transformation generated by the simplest all-momentum generator, $F_4=pP$ , and verifying canonicity by the fundamental Poisson bracket. This map — the phase-space quarter-turn — is the signature transformation of the momentum-driven types: type 2 cannot represent it at all (its $(q,P)$ pair degenerates on it), while for $F_4$ it is the baseline, playing the role $F_2=qP$ plays for the identity. Watch how few moves it takes.

Step 1. Apply the type-4 rules to $F_4=pP$ :

q=-\frac{\partial F_4}{\partial p}=-P,\qquad Q=+\frac{\partial F_4}{\partial P}=p

Step 2. Solve for the new variables: $Q=p$ and $P=-q$ . Coordinate becomes momentum, momentum becomes minus coordinate — a rotation of the phase plane by $90^\circ$ . Why the minus sits where it does: a rotation must preserve orientation. The naive swap $Q=p$ , $P=q$ is a reflection — it reverses the sense of circulation in phase space, and would flip the sign of every Poisson bracket. The minus is not decorative; it is the difference between a canonical map and an anticanonical one.

Step 3. Verify with the fundamental bracket, computed in the old chart:

\{Q,P\}_{q,p}=\frac{\partial Q}{\partial q}\frac{\partial P}{\partial p}-\frac{\partial Q}{\partial p}\frac{\partial P}{\partial q}=(0)(0)-(1)(-1)=1

Step 4. Iterate the map: applying it twice sends $(q,p)\to(p,-q)\to(-q,-p)$ — the type-3 sign flip; four applications return the identity. Successive applications generate the cyclic group $C_4$ : four quarter-turns return the identity. The bracket value $\{Q,P\}=1$ confirms canonicity, and the composition check — two applications of the quarter-turn send $(q,p)\to(-q,-p)$ — agrees with the $F_3=pQ$ sign-flip map.

Go deeper · Derivation C · 1 minTwo routes to F4 must agree — an internal consistency check

Derivation CTwo routes to F4 must agree — an internal consistency check

We are confirming that the two definitions in equation (4) — $F_4=F_2-qp$ (swap the old pair of type 2) and $F_4=F_3+QP$ (swap the new pair of type 3) — generate identical transformation rules. This is more than pedantry: it verifies that the four types form a consistent commuting square of Legendre transforms, so the atlas has no seams. The strategy is to compute $\mathrm{d}F_4$ along both routes and compare coefficients.

Step 1. Route one. From the type-2 differential $\mathrm{d}F_2=p\,\mathrm{d}q+Q\,\mathrm{d}P+(K-H)\,\mathrm{d}t$ , subtract $\mathrm{d}(qp)=p\,\mathrm{d}q+q\,\mathrm{d}p$ :

\mathrm{d}(F_2-qp)=-q\,\mathrm{d}p+Q\,\mathrm{d}P+(K-H)\,\mathrm{d}t

The $p\,\mathrm{d}q$ terms cancel; surviving differentials $(\mathrm{d}p,\mathrm{d}P,\mathrm{d}t)$ — correct type-4 arguments.

Step 2. Route two. From the type-3 differential $\mathrm{d}F_3=-q\,\mathrm{d}p-P\,\mathrm{d}Q+(K-H)\,\mathrm{d}t$ , add $\mathrm{d}(QP)=P\,\mathrm{d}Q+Q\,\mathrm{d}P$ :

\mathrm{d}(F_3+QP)=-q\,\mathrm{d}p+Q\,\mathrm{d}P+(K-H)\,\mathrm{d}t

The $P\,\mathrm{d}Q$ terms cancel. The right-hand sides of Steps 1 and 2 are identical.

Step 3. Match coefficients in either expression:

q=-\frac{\partial F_4}{\partial p},\qquad Q=+\frac{\partial F_4}{\partial P},\qquad K=H+\frac{\partial F_4}{\partial t}

Why the agreement was guaranteed: Legendre transforms on independent pairs commute — the old-pair swap touches only $p\,\mathrm{d}q$ terms, the new-pair swap only $P\,\mathrm{d}Q$ terms. The four types are the four corners of a commuting square, and $F_4$ is reachable along either edge. Both routes reproduce the same type-4 rules, and traversing the square in the reverse direction — $F_4\to F_3\to F_1$ — returns the type-1 rules, confirming that the four types form a closed commuting square of Legendre transforms.

Limiting cases

Case	Result	Meaning
Sign flip	$F_3=pQ\ \Rightarrow\ Q=-q,\ P=-p$	half-turn of phase space; type 1 cannot generate it
Exchange	$F_4=pP\ \Rightarrow\ Q=p,\ P=-q$	quarter-turn; the baseline map of the momentum-driven types
Momentum-side point transformation	$F_3=-p\,f(Q,t)\ \Rightarrow\ q=f(Q,t),\ P=p\,\partial f/\partial Q$	any invertible relabeling written from the inverse side — useful when $f$ is simple but $f^{-1}$ is ugly
Time-independent generator	$\partial F/\partial t=0\ \Rightarrow\ K=H$	universal across all four types
Identity	no $F_4$ exists ( $p$ , $P$ dependent there)	the exchange map's blind spot mirror-images type 1's; use $F_2=qP$
Degenerate Hessian (6)	type unavailable	switch types, or mix types across degrees of freedom

Common trapThe minus signs are not interchangeable — each lives in a specific slot

The four-type sign table is the most-botched piece of bookkeeping in this chapter, usually because students memorize a slogan ("types 3 and 4 have minus signs") instead of slots. The law: an independent old momentum returns its conjugate negated, $q=-\partial F/\partial p$ (types 3 and 4 alike); an independent new coordinate returns its conjugate negated, $P=-\partial F/\partial Q$ (types 1 and 3 alike); the other two slots are positive. Mixing this up produces maps that are anticanonical — orientation-reversing — and every downstream Poisson bracket comes out $-1$ . When in doubt, do not trust memory: write the one-form identity (1), perform the Legendre swaps your argument choice requires, and read the signs off the surviving differentials. Twenty seconds, zero risk.

Worked examples

1 · The quarter-turn on a numbered oscillator (confidence builder)

Apply $F_4=pP$ to the oscillator $H=p^2/2m+\tfrac12 m\omega^2q^2$ with $m=1.0\ \mathrm{kg}$ , $\omega=2.0\ \mathrm{rad/s}$ , at the phase point $q=0.30\ \mathrm{m}$ , $p=0.80\ \mathrm{kg\,m/s}$ . Find the new variables and $K$ , and verify both canonical equations numerically.

Step 1. Run the type-4 rules: $q=-\partial F_4/\partial p=-P$ gives $P=-q=-0.30$ ; $Q=\partial F_4/\partial P=p=0.80$ . The quarter-turn in numbers: $(0.30,\,0.80)\to(0.80,\,-0.30)$ .

Step 2. Build the new Hamiltonian by substituting $p=Q$ , $q=-P$ (no explicit time, so $K=H$ re-expressed):

K(Q,P)=\frac{Q^{2}}{2m}+\tfrac12 m\omega^{2}P^{2}

Value check: $K=\frac{0.64}{2.0}+\tfrac12(1.0)(4.0)(0.09)=0.32+0.18=0.50\ \mathrm{J}$ , equal to $H=0.32+0.18=0.50\ \mathrm{J}$ ✓. Note the roles have traded: the kinetic-looking term now carries $Q$ , the spring-looking term carries $P$ — the oscillator is self-similar under the quarter-turn, which is precisely its phase-space rotational symmetry.

Step 3. First canonical equation: $\dot Q=\partial K/\partial P=m\omega^{2}P=(1.0)(4.0)(-0.30)=-1.2$ . Directly: $\dot Q=\dot p=-m\omega^{2}q=-(1.0)(4.0)(0.30)=-1.2$ ✓.

Step 4. Second canonical equation: $\dot P=-\partial K/\partial Q=-Q/m=-0.80$ . Directly: $\dot P=-\dot q=-p/m=-0.80$ ✓. The rotated chart runs on canonical equations with the substituted $K$ , no extra work — and the numbers confirm each sign in the type-4 rules.

2 · The oscillator solved from the momentum side (exam level)

Solve the oscillator of the type-1 article — $m=0.50\ \mathrm{kg}$ , $\omega=4.0\ \mathrm{rad/s}$ , $q_0=0.30\ \mathrm{m}$ , $p_0=0.80\ \mathrm{kg\,m/s}$ at $t=0$ — using the type-3 generator $F_3(p,Q)=\dfrac{p^{2}}{2m\omega}\cot Q$ . Find the action, the initial angle, and $q$ at $t=0.50\ \mathrm{s}$ ; the answer must match the type-1 route.

Step 1. Run the type-3 rules. $q=-\partial F_3/\partial p=-\dfrac{p}{m\omega}\cot Q$ and, using $\mathrm{d}(\cot Q)/\mathrm{d}Q=-1/\sin^{2}Q$ , $P=-\partial F_3/\partial Q=\dfrac{p^{2}}{2m\omega\sin^{2}Q}$ — positive, as an action must be.

Step 2. Invert: $p=\sqrt{2m\omega P}\,\sin Q$ from the $P$ -relation, then $q=-\sqrt{2P/m\omega}\,\cos Q$ from Step 1. Substitute into $H$ :

H=\frac{p^{2}}{2m}+\tfrac12 m\omega^{2}q^{2}=\omega P\sin^{2}Q+\omega P\cos^{2}Q=\omega P

Same rectified Hamiltonian as the type-1 route — the chart differs (this angle is offset by a quarter period), the physics does not.

Step 3. Numbers. Energy: $E=\frac{0.64}{1.0}+\tfrac12(0.50)(16)(0.09)=0.64+0.36=1.00\ \mathrm{J}$ , so the action is $P=E/\omega=0.250\ \mathrm{J\,s}$ — identical to the type-1 computation, as it must be (the action is chart-independent). Amplitudes: $\sqrt{2P/m\omega}=\sqrt{0.25}=0.50\ \mathrm{m}$ , $\sqrt{2m\omega P}=\sqrt{1.0}=1.0\ \mathrm{kg\,m/s}$ .

Step 4. Initial angle from this chart's map: $\cos Q_0=-q_0/0.50=-0.60$ and $\sin Q_0=p_0/1.0=0.80$ , so $Q_0$ sits in the second quadrant: $Q_0=\pi-0.9273=2.2143\ \mathrm{rad}$ . (Compare $\phi_0=0.6435$ in the type-1 chart: $Q_0-\phi_0=1.5708=\pi/2$ exactly — the two angle charts differ by a quarter turn, the signature of the $F_1\to F_3$ Legendre swap.)

Step 5. March and map back: $Q(0.50)=\omega t+Q_0=2.0+2.2143=4.2143\ \mathrm{rad}$ , and

q(0.50)=-0.50\cos(4.2143)=-0.50\times(-0.4777)=\mathbf{0.239\ m}

matching the type-1 route's $0.239\ \mathrm{m}$ to the digit ✓. Cross-check against direct integration, $q(t)=q_0\cos\omega t+(p_0/m\omega)\sin\omega t=0.30(-0.4161)+0.40(0.9093)=0.239\ \mathrm{m}$ ✓. Two different generating functions, two different intermediate charts, one physical answer.

Origins · 1 minPaul Dirac and how this was discovered

Paul Dirac

1902–1984 · Cambridge and Tallahassee

No one took the equal standing of $q$ and $p$ — the symmetry that types 3 and 4 formalize — more seriously than Dirac. His 1926 thesis and the 1927 "transformation theory" rebuilt the new quantum mechanics as a theory of transformations between representations: the position representation and the momentum representation are two charts on one state space, connected by a unitary map whose classical shadow is exactly the exchange-type canonical transformation of this page (the kernel $e^{ipq/\hbar}$ is a generating function in disguise — Dirac says so explicitly in his 1933 paper on the Lagrangian in quantum mechanics, the seed Feynman grew into the path integral). Famously taciturn — his Cambridge colleagues coined the "dirac", one word per hour — he assembled The Principles of Quantum Mechanics (1930) around the canonical formalism: Poisson brackets promoted to commutators, canonical transformations to unitaries, the classical types surviving as the freedom to quantize in any chart. Engineering-trained at Bristol before Cambridge, he kept an engineer's economy: when asked how he found the relativistic electron equation, he answered that he found it beautiful. The momentum-side generating functions are a small room in the palace he built, but the palace's floor plan — physics as geometry of transformations — is his.

History

1787

— Legendre introduces the transform that swaps a variable for its conjugate slope, the engine behind all four types

1834

— Hamilton casts dynamics in canonical form; coordinates and momenta become interchangeable bookkeeping

1843

— Jacobi's lectures systematize generating functions; the full table of types becomes standard equipment

1925

— Matrix mechanics puts q and p on strictly equal algebraic footing, vindicating the momentum-driven charts

1927

— Dirac's transformation theory: quantum representations connected by unitaries descend from canonical transformations

1933

— Dirac's Lagrangian paper identifies the quantum kernel with a classical generating function, seeding the path integral

Going further

References. Goldstein, Poole & Safko, Classical Mechanics (3rd ed.), §9.1 Table 9.1 · Landau & Lifshitz, Mechanics (Vol. 1, 3rd ed.), §45 · Morin, Introduction to Classical Mechanics, Ch. 15 (canonical transformations) · Taylor, Classical Mechanics, Ch. 13 The Legendre-transform definitions $F_3=F_1-qp$ and $F_4=F_2-qp=F_3+QP$ reproduce the partial-derivative relations listed above; the Poisson bracket $\{Q,P\}=1$ for the $F_4=pP$ example confirms canonicity; and the universal rule $K=H+\partial F/\partial t$ holds across all four types. The only numerical constant entering the general theory is the bracket value $1$ . The numeric examples — quarter-turn at $(q,p)=(0.30,0.80)$ with $m=1.0\ \mathrm{kg}$ , $\omega=2.0\ \mathrm{rad/s}$ ; momentum-side oscillator with $m=0.50\ \mathrm{kg}$ , $\omega=4.0\ \mathrm{rad/s}$ — are worked out in the sections above, and the second recovers $q(0.50)=0.239\ \mathrm{m}$ , matching the type-1 route.

Builds on: Definition of Canonical Transformations · Generating Functions and the Four Types · Type-1 Generating Function F1(q,Q,t) · Type-2 Generating Function F2(q,P,t) Leads to: Symplectic Condition · Poisson Brackets as a Test for Canonicity · Symplectic Group and Phase-Space Geometry

Where does the minus sign in q = −∂F₃/∂p really come from? I keep losing it.

From one line of calculus: $p\,\mathrm{d}q=\mathrm{d}(pq)-q\,\mathrm{d}p$ . When you absorb $\mathrm{d}(pq)$ into the generator to make $p$ independent, the surviving term is $-q\,\mathrm{d}p$ — coefficient minus $q$ . Every sign in the four-type table is the coefficient of a surviving differential after such swaps, so the reliable procedure is: write the one-form identity, do the swaps your argument choice demands, read the coefficients. It reproduces the whole of Goldstein's Table 9.1 in under a minute and is immune to memory failure.

Honestly, will I ever use F3 or F4, or are they textbook completeness?

Less often than $F_2$ , genuinely — but "rarely" is not "never," and the situations are characteristic. You reach for them when the momentum is the natural handle: writing a point transformation from its inverse side ( $F_3=-p\,f(Q,t)$ when $q=f(Q)$ is simple but $f^{-1}$ is not), generating phase-space rotations and exchanges ( $F_4$ ), and in perturbation-theory constructions where a chart must keep some momenta as labels. They also carry conceptual weight out of proportion to their usage: the exchange map they own is the classical shadow of the position–momentum Fourier duality in quantum mechanics.

Why is there no generating function with arguments (q, p) — both old variables?

Because such a function could not encode a map. A generator works by mixing one label from each side, so that its cross-derivatives tie old to new; the pair $(q,p)$ already coordinatizes the old phase space completely, and any $F(q,p)$ would carry zero information about where points go. The four legitimate types are exactly the four old/new mixed pairs. (Same reason $(Q,P)$ is excluded.) The deeper statement: a canonical transformation is a Lagrangian submanifold of the doubled phase space, and the types are the four coordinate charts on it that project nicely.

Can I use different types for different degrees of freedom in one transformation?

Yes, and for many-body problems you frequently must. The generator is a sum $F=\sum_i g_i$ in which each $g_i$ mixes its own pair in whatever type is non-degenerate there — say type 2 for the coordinates you keep and type 4 for a pair you rotate. The differentials add, so the rules apply slot by slot. Goldstein makes this point with a two-freedom example ( $F=q_1P_1+p_2Q_2$ : identity on the first pair, exchange-like on the second). The atlas metaphor is literal: charts are chosen locally, per pair.

Is the F4 exchange map related to the Fourier transform in quantum mechanics?

Directly. Quantize the exchange $Q=p,\ P=-q$ : the unitary that implements it maps the position representation to the momentum representation — it is the Fourier transform, up to phase conventions. Dirac's transformation theory makes the correspondence exact: unitary changes of representation descend from canonical transformations, and mixed-variable kernels like $e^{ipq/\hbar}$ play the role of generating functions. The oscillator makes it vivid: its quarter-period evolution operator is the fractional Fourier transform, interpolating what $F_4$ does in one jump.

The sign flip Q = −q, P = −p is canonical, but is plain q → −q with p untouched also canonical?

No — and checking is instructive. For $Q=-q$ , $P=p$ : $\{Q,P\}=(-1)(1)-0=-1\neq1$ . That map is a reflection of the phase plane (orientation-reversing), so it is anticanonical: it preserves areas in magnitude but flips their sign, and it maps Hamilton's equations onto themselves only with $t\to-t$ . Parity in mechanics acts on coordinates and momenta together, $(q,p)\to(-q,-p)$ — the half-turn $F_3=pQ$ generates — precisely so that it stays canonical.

Why does F4 fail on the identity when F2 fails on the exchange? Is there a pattern?

A clean duality. A type exists when its argument pair can independently label a phase point through the map. For the identity, $P=p$ : the pair $(p,P)$ is one variable twice — $F_4$ 's Hessian degenerates. For the exchange, $P=-q$ makes $(q,P)$ degenerate — $F_2$ fails. In general, applying the exchange map to any transformation swaps which types can represent it ( $F_1\leftrightarrow F_3$ , $F_2\leftrightarrow F_4$ up to signs), which is why the family needs exactly its four members: they are one construction seen through the quarter-turn's eyes.

Do types 3 and 4 change the rule for time-dependent Hamiltonians?

No. $K=H+\partial F/\partial t$ holds verbatim for all four types, with the partial derivative taken at fixed values of the type's own two arguments. The reason is structural: the Legendre transforms that convert between types add terms like $qp$ or $QP$ that carry no explicit time dependence, so the $\mathrm{d}t$ coefficient in the one-form identity is never touched. What does change between types is only which variables you hold fixed while computing $\partial F/\partial t$ — a bookkeeping point that matters when your generator's arguments are themselves time-dependent expressions.