University of Ljubljana

Faculty of Mathematics and Physics

Feynman’s formulation of Quantum mechanics

Author: Vasja Susič

Advisor: prof. dr. Anton Ramšak

December 2009

Abstract
This paper is about Feynman’s path integral formulation of quantum mechanics. It mostly deals
with the basic concept of the kernel, both for one–particle and for multiparticle systems. The
paper also includes a brief overview of the classical action and the wave function formulation
of quantum mechanics, which will be needed throughout our discussion. The treatment of
path integrals is separated from Feynman’s formulation, so both topics can be presented in a
more precise manner. Also, some brief arguments for the reduction of Feynman’s formulation
of quantum mechanics to classical mechanics in the macroscopic limit are made. Finally,
the Schrödinger equation for a single particle in one dimension is derived from Feynman’s
formulation.
V S̌ Feynman’s formulation of Quantum mechanics 2(19)

Contents
Introduction 2

1 Preliminaries 4
1.1 On Variational Calculus and the Classical Action . . . . . . . . . . . . . . . . . . 4
1.2 On Schrödinger’s Formulation of Quantum Mechanics . . . . . . . . . . . . . . . 5
1.3 A Heuristic Approach to Path Integrals . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Path Integral Formulation of Quantum mechanics 9

2.1 A Justification Through Thought Experiments . . . . . . . . . . . . . . . . . . . . 9
2.2 The Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.2 Chain Rule for Events . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 The Kernel of Multiparticle Systems . . . . . . . . . . . . . . . . . . . . . . . . . 13

3 Connections to other physical theories 14

3.1 The classical limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Equivalence to Schrödinger’s formulation . . . . . . . . . . . . . . . . . . . . . . 16

Conclusion 18

References 19

Introduction
By the end of the 19th century, it had become apparent that a considerable amount of experimental
evidence directly contradicted the prevalent physical theories of that time. What ensued was a new
kind of physics, which was so much different, we label the post–1900 physics as modern physics,
as opposed to classical physics. Even Newtonian mechanics, the flagship of classical physics, was
supplanted by Quantum Mechanics and Einstein’s theory of relativity.
In 1900 Max Planck hypothesized that atoms emit energy in “discrete energy elements”. This
enabled him to work around the problem in the calculation of black body radiation, which was
falsely predicted to be infinite by the classical theory of that time. In 1905 Albert Einstein found
an explanation for the photoelectric effect by postulating “discrete energy packets” of light called
photons. Louis De Broglie was the first to propose that matter behaves as waves in 1923 [1]. This
all brought about the idea of particle–wave duality: in some experiments, particles behaved like
discrete packets, while in others they behaved like waves.
The first major theoretical framework that appropriately incorporated these ideas came in
1925, when physicists Werner Heisenberg, Max Born and Pascual Jordan developed the matrix
formulation of quantum mechanics, while Erwin Schrödinger developed the wave function
formulation of quantum mechanics in 1926 [1]. In matrix mechanics, the physical properties of a
particle were interpreted in terms of a time dependent matrix. On the other hand, wave mechanics
interpreted particle properties through a function — the wave function — which is the solution of a
partial differential equation called the Schrödinger equation. At first, that was a puzzling situation,
because the two approches to quantum mechanics were seemingly distinct, yet both made accurate
V S̌ Feynman’s formulation of Quantum mechanics 3(19)

predictions. Soon, Schrödinger managed to show the equivalence of the two approaches, with
the final reconciliation coming from Paul Dirac in 1930 [2]. By 1932, the first mathematically
coherent theory had been fully developed by mathematician John von Neumann, building on the
earlier work done by David Hilbert on function spaces [3]. Von Neumann realized that the wave
function could be thought of as a point in an infinitely dimensional Hilbert space of functions.
And due to Hermann Weyl’s work on Lie groups and representation theory, the matrix form was
identified as a representation of the wave function in a certain basis of the Hilbert space. Thus
the two formulations of quantum mechanics, the matrix formulation and wave formulation, were
proved to be equivalent in a rigorous manner.
In 1948, a thrid approach to quantum mechanics was introduced by the American physicist
Richard Feynman, building on the work previously done by Paul Dirac [4]. The new development
is now called the path integral formulation of quantum mechanics, or alternatively Feynman’s
formulation of quantum mechanics. This approach is an extension of the notion of classical action.
In classical mechanics, we can define the action as an integral of the Lagrangian, a function
which “summarizes the dynamics of a system” [5]. The classical trajectory of a particle (or
any physical system in general) can be obtained by Hamilton’s principle, which states that the
trajectory is such that the action (as a function of trajectory) is stationary [6]. This formulation of
classical mechanics is called Lagrangian mechanics, and it’s main tool is the calculus of variations
(a part of mathematical analysis).
Feynman’s approach to quantum mechanics is trajectory based. Unfortunately, the trajectory
of a physical system is not unique. Instead, every trajectory is taken into account, even the ones
that would seem impossible in the classical sense. Then a sum of all these paths is calculated
via path integrals [4]. A sum will represent the probabilitiy of finding the particle in a certain
location. Although it is far easier to use the Schrödinger equation for calculations in most cases,
the importance of Feynman’s formulation should not be underestimated, since his approach is of
great importance in theoretical physics. Because it uses the notion of action, it has benefits similar
to those found in Lagrangian mechanics, when one can have a quick overview of the underlying
physics by examining the structure of the Lagrangian and its symmetry properties. Feynman’s
formulation also gives a better intuitive understanding of quantum mechanics by linking it with
classical mechanics in a way far more apparent than the wave function and matrix formulations do.
An established formal mathematical theory of Feynman’s formulation of quantum mechanics
does exist, but is quite complex, mostly due to the notion of the path integral. The Lebesgue
approach to integration cannot be performed for path integrals in a straight–forward manner, due
to the trouble of finding a suitable measure on the infinite–dimensional space of functions. A
reader mostly interested in path integrals should refer to [7, 8].
V S̌ Feynman’s formulation of Quantum mechanics 4(19)

Figure 1: The men behind the three formulations of quantum mechanics (from left to right): Werner
Heisenberg (1901-1976) [9], Erwin Schrödinger (1887-1961) [10] and Richard Feynman (1918-
1988) [11].

1 Preliminaries
1.1 On Variational Calculus and the Classical Action
Variational calculus deals, among other things, with the extremal values of a functional. Consider
a function L ∈ C 2 ([ta , tb ] × R × R), by which we mean a twice continuously differentiable function
[ta , tb ] × R × R −→ R, where ta , tb ∈ R. For a real function q ∈ C 1 ([ta , tb ]) we define

Z tb
S (q) B L (t, q(t), q̇(t)) dt. (1)
ta

The functional S is a mapping C 1 ([ta , tb ]) −→ R. At this point we fix the end points of the functions
q under consideration: we are interested only in such functions q that q(ta ) = a and q(tb ) = b for a
given pair of real numbers a and b. The problem is posed as follows: for which funtion q does the
functional S take a minimal value? There is a theorem, which states the necessary condition for q
(see [12], p. 47): if such a minimizer funtion exists (we shall denote this function q̄), then it must
satisfy the differential equation

∂L   d " ∂L  #
0 = t, q̄(t), q̄˙ (t) − t, q̄(t), q̄˙ (t) . (2)
∂q dt ∂q̇
We call equation (2) the Euler–Lagrange equation, and it is a second order ordinary differential
equation for q̄. In a more general case, we could define a real function fq : R −→ R, with fq (α) B
S (q̄ + αq). We say that the functional S has an extremal value in q̄, when for all real functions fq
the expression fq0 (0) = 0 holds. A more general theorem (than the one stated previously) identifies
the Euler-Lagrange equation (2) as the necessary condition for the extremal of S (see [13], p. 37).
Sometimes, the search for the stationary path q̄ is summed up by the mathematically somewhat
dubious equation δS = 0, and the expression of the first variation of the functional being zero.
V S̌ Feynman’s formulation of Quantum mechanics 5(19)

The formulation of classical mechanics in terms of variational methods is called Lagrangian

mechanics. In this framework, S is called the classical action and L is called the Lagrangian.
The Lagrangian of a system is a function of the generalized coordinates qi , which we choose
by applying all the holonomic constraints (those of the form f (xi , . . . , t) = 0) on the regular
coordinates xi (for a N–particle system in three dimensions, xi are the 3N space coordinates
of the particles). When the generalized coordinates qi are independent (with no nonholonomic
constraints), and the forces acting on the system are conservative, so that their potential V is
explicitly dependent only on qi (and not q̇i or t), the Lagrangian can be written as T − V, where T
is the expression for the total kinetic energy of a system. The main postulate here is Hamilton’s
principle, which states that the actual trajectory q̄ of the system, descibed by a Lagrangian, is the
one in which the action S (q) is extremal ([13], p. 36). This means the trajectory in physical reality
is the solution of the Euler–Lagrange equation.
A more general situation represents a system described by many variables qi , all with specified
end–points. Let i ∈ {1, 2, . . . , n}, where n ∈ N. Then the Lagrangian is expressed in terms of
generalized coordinates qi and their first derivatives. In this case, the action is expressed as

Z tb
S (q1 , . . . , qn ) B L [t, q1 (t), q̇1 (t), . . . , qn (t), q̇n (t)] dt. (3)
ta

Hamilton’s principle for such a system again states, that the trajectories q̄i are extremal values of
the action S . Every trajectory thus obeys its own Euler–Lagrange equation ([13], p. 31):

∂L  d ∂L 
 " #
∀i ∈ {1, . . . , n} : t, q̄i (t), q̄˙ i (t) = t, q̄i (t), q̄˙ i (t) . (4)
∂qi dt ∂q̇i

1.2 On Schrödinger’s Formulation of Quantum Mechanics

In Schrödinger’s formulation, the states of a quantum system are described by a wave function Ψ.
The wave function Ψ(r, t) for a single particle is a function R3 × R −→ C, where r will denote a
point in space and t a point in time. If c ∈ C, then cΨ represents the same state as Ψ ([13], p. 247).
The squared absolute value |Ψ(r, t)|2 expresses the probability density for the particle at time t to
be near r. Because the probability of finding a particle in all space is 1, spatial normalization of
the wave functions is demanded:

Z
|Ψ(r, t)|2 dV = 1 ∀t ∈ R. (5)
R3

We should remember from section 1.1, that in classical mechanics, a differential equation
determines the actual trajectory for a particle. In Schrödinger’s formulation of quantum mechanics,
the notion of a trajectory is meaningless, since the wave function can provide only probabilities. In
fact, due to Heisenberg’s uncertainty principle δxδp ≥ ~/2, we cannot know at any given time both
a particle’s position x and its momentum p to an arbitrary degree of accuracy. Because the notion
of a trajectory implies we know both the exact position and velocity, it is inconsistent with the
quantum mechanics of Schrödinger. However, the wave function must satisfy its own differential
equation. We call this equation the Schrödinger equation and it takes the following form (see [14],
p. 27):
V S̌ Feynman’s formulation of Quantum mechanics 6(19)

∂Ψ
i~ = ĤΨ. (6)
∂t

The symbol i denotes the imaginary unit, ~ = h/2π, where h is Planck’s constant, and Ĥ is the
Hamiltonian operator, acting on the space of functions Ψ. If we write Hamilton’s operator for a
single particle with mass m in the potential V(r) explicitly, we get the following well known form
of equation (6):

∂Ψ h ~2 i
i~ (r, t) = − ∇2 + V(r) Ψ(r, t). (7)
∂t 2m
The symbol ∇2 = ∂2 /∂x2 + ∂2 /∂y2 + ∂2 /∂z2 denotes the Laplace differential operator.
In Schrödinger’s quantum mechanics, the functions Ψ form a Hilbert space; a Hilbert space
is a vector space equipped with an inner product (and it is complete as a metric space, where the
metric is naturally defined by the norm, which in turn is naturally defined by the inner product).
The inner product between two functions Ψ1 and Ψ2 is defined by an integral:

Z
hΨ2 | Ψ1 i (t) B Ψ∗2 (r, t)Ψ1 (r, t) dV. (8)
R3

The symbol ∗ denotes complex conjugation.

Physical quantities, which are functions in classical mechanics, become hermitian operators
in quantum mechanics. We define the average value of an operator  as hÂi B hΨ | ÂΨi ([14],
p. 27).
Up until now, we have been describing the quantum mechanics of a single particle. Suppose
we have n particles (n ∈ N), each with its own coordinate ri , i = 1, . . . , n. The states of a
multiparticle system are described by a wave function Ψ : R3n × R −→ C. The squared absolute
value |Ψ(r1 , . . . , rn , t)|2 expresses the probability density for finding at a given time t each of the
particles i near ri . The wave function is again spatially normalized:

Z
|Ψ(r1 , . . . , rn , t)|2 dV1 . . . dVn = 1 ∀t ∈ R. (9)
R3n

The Schroödinger equation in the form (6) still holds, if one takes the Hamiltonian operator for
multiple particles. In its explicit form, the Schrödinger equation for n–particles is the following
([14], p. 28):

n 
∂Ψ ~2 2 
"X #
i~ (r1 . . . , rn , t) = − ∇ + V(r1 , . . . , rn ) Ψ(r1 , . . . , rn , t). (10)
∂t i=1
2mi i

Here, mi signifies the mass of the i–th particle, while ∇2i = ∂2 /∂xi2 + ∂2 /∂y2i + ∂2 /∂z2i .
V S̌ Feynman’s formulation of Quantum mechanics 7(19)

1.3 A Heuristic Approach to Path Integrals

Path integrals are a type of integral, where we integrate in a space of continuous functions (paths).
They are usually defined alongside Feynman’s formulation of quantum mechanics; we will treat
them here separately as a mathematical subject. The integration “over all paths” is in stark contrast
with the well known Riemann integral, where we integrate in the space of real numbers R or a
multidimensional space Rn , for n ∈ N. An alternative to the Riemann integral is the Lebesgue
integral, which allows us to integrate real functions with their domains in a measurable space. In
case of path integrals, we are hard pressed to define a measure in the space of paths R −→ R,
mainly because the paths do not form a finite dimensional vector space over R. Consequently, we
shall attempt to describe path integrals with an analogy to Riemann integrals. Our approach will
be heuristic in nature (based on [15], p. 31–35), since the intent is to give the reader an intuitive
feel for path integrals. The reader should not be discouraged by the lack of mathematical rigor in
the following paragraphs.
We shall first consider the Riemann integral of functions R −→ R. The main idea of this
integral is to calculate the area under the curve in a plot of y = f (x) of the integrated function.
Suppose we are interested only in the area, where x ∈ [xa , xb ]. We shall follow a series of steps,
where we calculate a regular sum in each of the steps. In the n–th step, we divide the interval
[xa , xb ] into a partition of n equally long intervals. For a given n we define

xb − xa
 B , (11)
n
xi B xa + i i ∈ {0, 1, 2, . . . , n}. (12)

We can interpret  as the length of the intervals in the n–th step, while xi are partition points.
The given definitions also imply x0 = xa and xn = xb . The Riemann integral of f is in this case
(somewhat informally) defined as the following:

Z xb n−1
X
f (x) dx B lim  f (xi ). (13)
xa n→∞
i=0

Of course the limit on the right–hand side of the above equation does not necessarily exist. The
Riemann integral is defined only for such functions f where the limit exists. It turns out that all
functions [a, b] −→ R, which have finitely many points of discontinuity on [a, b], can be integrated
(see [16], p. 126).
Now we turn our attention to path integrals. Our first problem will be, how to characterize a
continuous function q(t) : R −→ R in a finite manner. The objects we will be integrating will be
functionals F(q) : RR −→ R. As in the Riemann integral, we shall specify the endpoints of the
region over which we integrate. We specify the path endpoints by fixing q(ta ) = a and q(tb ) = b.
We again divide the limiting process into steps. In the n–th step, we divide the time interval into n
smaller intervals in a similar manner as we did with the Riemann integral (see Figure 2):

tb − ta
 B , (14)
n
ti B ta + i i ∈ {0, 1, 2, . . . , n}, (15)
xi B q(ti ) i ∈ {0, 1, 2, . . . , n}. (16)
V S̌ Feynman’s formulation of Quantum mechanics 8(19)

The time values ti are the partition points of the interval [ta , tb ]. We have t0 = ta and tn = tb .
We characterize a path q by specifying the values xi , or in other words, by specifying its value
at specific equidistant points in time. Of course, we already know that x0 = a and xn = b. This
means we must specify n − 1 values: xi for i ∈ {1, . . . , n − 1}. These values can be any real number.
However, the functions q are specified by the values in all time points between ta and tb . We now
define values q(t) for times other than ti in such a way, that the motion of a particle moving along
such a path between xi and xi+1 is uniform (its velocity between ti and ti+1 is constant, the path
“consists of straight lines”). We can write that explicitly as

xi+1 − xi
∀i ∈ {0, . . . , n − 1} : q(t) = xi + t · for t ∈ [ti , ti+1 ]. (17)


Figure 2: We characterize a path by splitting the time interval into a partition of equally long
intervals, and specifying the values xi at times ti , while demanding uniform motion between them.

A sum over all possible paths (actually all paths with line segments) in step n will thus translate
into a sum for all possible values xi for each i ∈ {1, . . . , n − 1}. Because xi can change by an
infinitesimal amount, a sum for each xi will be calculated by integration over xi . Analogous to
the normalizing constant  (it changes from step to step) in the Riemann integral, we will here use
the symbol A for the normalizing constant in path integrals. This constant ensures, that the limit,
where the number of steps approaches ininity, exists. This means that A depends on , which in
turn depends on n.
We now define the path integral (the reader should note the notation used on the left–hand side)
as

Z b0 Z
1 dx1 dxn−1
F[b0 ,a0 ] (q) Dq B lim F[b0 ,a0 ] (q) ... . (18)
a0 n→∞ A Rn−1 A A
In the above definition, a0 = (ta , a) and b0 = (tb , b), and by F[b0 ,a0 ] we mean the functional F
restricted to such functions q, for which q(ta ) = a, q(tb ) = b and q consists of line segments with
specified values in partition points xi .
V S̌ Feynman’s formulation of Quantum mechanics 9(19)

In general, we do not know how A relates to , but we will calculate it for a special case in
section 3.2. We will otherwise not concern ourselves too much with A, because in practice we
usually do not calculate the path integral by the form given in (18), but rather use tricks on a case
by case basis ([15], p. 41–74).
Taking uniform motion for the path between two points xi could present difficulties, if the
integrated functional F assumes a twice differentiable function (path) q in its domain, or in other
words, if F depends on q̈. Namely, our path with segments of uniform motion is not in general
twice differentiable, because the “acceleration” at times ti is infinite, therefore the second derivative
at those times does not exist. Since all functionals used will depend on the classical action S , which
in turn demands only the existence of the first derivative in the Lagrangian, the problem of taking
uniform paths between points xi will not be an issue.
Also, it is mathematically especially convenient, that F(q) is an oscillatory functional (its real
values oscillate around zero), so that many different paths cancel each other out. This makes it
easier to achieve the finitness of the path integral.

2 Path Integral Formulation of Quantum mechanics

2.1 A Justification Through Thought Experiments
The two slit experiment is a famous thought experiment in quantum physics (see Figure 3).
Suppose we have a source of electrons at plane A. The electrons come out of this source with
roughly the same energies, but in different directions. We have a screen at B, with two narrow
holes (slits). Further away, we have a plane C, where we detect the coming electrons. We move
the detector in plane C, so that we measure the intensity of the current of electrons at different
positions in the plane. In classical physics, we would expect most of the electrons to be near two
distinct locations at C, depending on which slit in the screen B they passed through. If we actually
perform such an experiment, it turns out the result in plane C is an interference pattern. That is
an unexpected result from the perspective of classical physics: interference patterns are formed by
interfering waves, while we expect the electron to behave like a point particle. Therefore, we have
to find a way to give the electron wave–like properties.

Figure 3: The configuration of the two slit experiment, with the source of electrons at A, the screen
with the two slits at B, and the detector at C. The experimental result for the probability density
P(y) of detecting electrons at y is on the far–right — it is an interference pattern.

Perhaps our first attempt at explaining the interference pattern at C would be by proposing
interference between different electrons, while they pass through the two slits. If the intensity of
V S̌ Feynman’s formulation of Quantum mechanics 10(19)

the electron source at A is sufficiently reduced, no more than a single electron would be passing
through the screen at B toward the detector C at any given time. The detector would in this
case record single pulses, corresponding to single particles. One could perform this variation
of the experiment by counting how many particles come at different positions in the plane C in
a given amount of time. By plotting the results, one would again obtain an interference pattern,
even though single electrons were passing through the slits. This means each electron somehow
”interferes with itself”: it does not simply go through one of the two slits as a point particle in
classical physics would. The mathematics describing such behaviour is known from the study of
waves: each wave–source is attributed an amplitude function, say φ1 for the first source and φ2 for
the second. In our experiment, the two slits in the screen B are our sources (for the “waves of the
electron”). These amplitude functions take the known form of the exponential function C · ei2πs/λ ,
where C is the amplitude, i the imaginary unit, λ the wavelenght of the wave, and s the distance
between the wave–source and the location, where the amplitude is being calculated. Note that the
value of the amplitude function is a complex number. The principle of superposition states that the
amplitude of the newly constructed wave is a sum of the amplitudes from the two separate sources:
φ = φ1 + φ2 .
In the field of optics, a light detector does not measure the amplitude of the electric field, but
rather the intensity, which is proportional to the square of the absolute value |E|2 of the amplitude
E. Analogously to optics, our detector at plance C measures the intensity of the coming electrons
in the form of electrical current, this means we are not actually measuring φ, but |φ|2 . The number
of electrons detected at a specific location is thus directly proportional to |φ|2 . We say that the
number of electrons detected near a location divided by the number of all electrons detected in
plane C is the probability of detecting the electron at that location. Therefore, the probability P (or
rather probability density) of finding the electron at a specific location is proportional to |φ|2 . By
renormalizing the function φ we can get P = |φ|2 . That is why we call φ the probability amplitude.
The main idea in the “two slit experiment”, was to calculate the probability amplitude of the
electron by adding the two separate probability amplitudes of the two possible paths the electron
could take: the path through the first and second slit. Now we shall try to generalize the result.
If we drill a third hole into the screen at B, the electron can take three possible paths: through
the first, the second, or the third slit. That means we have 3 probability amplitudes, each for a
corresponding slit, and the total probability amplitude at the detector as the sum φ = φ1 + φ2 + φ3 .
If we have n holes in the screen, the total probability amplitude will be the sum of n separate
amplitudes (corresponding to each of the holes).
An even further generalization would be to have multiple screens (see Figure 4), say B,D and E.
A particle detected at C could have passed through a given screen at any of its holes. This means we
have to take into account the multiple choices at each screen. Since the choices between different
screens are independent, this means the number of possible paths through a series of screens is
worked out by multiplying the number of choices for each screen. Again, the probability density
at the detector C would be the sum of the separate amplitudes of all different paths the particle
could take through the screens. By taking more and more screens, one would have to calculate a
sum over more and more paths, where one would specify the path by attributing a hole for each of
the screens. The number of possible paths is also increased by repeatedly drilling more and more
holes into any of the screens.
If there was nothing but empty space between the source of the electrons and the detector,
one could imagine that as a limiting case with infinitely many screens (arbitrarily close in every
location), but each screen has infinitely many holes, so that the particle can pass through anywhere
V S̌ Feynman’s formulation of Quantum mechanics 11(19)

Figure 4: We can modify the two slit experiment by adding new screens and drilling new holes,
which results in many possible paths (three such paths are drawn on the left–hand side). If the
number of screens and holes increases, a possible path can become very complex (right–hand
side).

in the screen. A sum over all possible paths through the screens would in this case be a sum over
all possible paths in euclidean space; the sum would actually be a path integral. For each path q
there would be a probability amplitude φ(q). The form of φ in the general case is not immediately
apparent, even more so if the particle moves in fields of conservative forces. It turns out that the
following case for φ is consistent with the regular quantum mechanics of Schrödinger:

i
φ(q) = e ~ S (q) . (19)

The contribution of each path thus has the same amplitude (the absolute value of each probability
density is 1), but different paths contribute different phase shifts, which depend on the classical
action S . The total probability amplitude in a point in space is calculated as q φ(q), where
P
we take the sum over all possible paths q from the starting point to the point in question. The
form in equation (19) is not inconsistent with our previous thought experiments on the probability
amplitude in the case of screens: the Lagrangian is classically a constant of motion for a free
particle, and in that case, we took the sum only over “the relevant paths” of uniform motion. The
exponential of the action of such paths thus reduces to a simple phase factor.

2.2 The Kernel

2.2.1 Definition
We denote the previously discussed probability amplitude of a particle as K[b0 ,a0 ] , where a0 = (ta , a)
and b0 = (tb , b). Very loosely speaking, we could regard the square of the absolute value of the
kernel as the probability density of finding the particle at the location b at time tb , if the particle
was located at a at time ta . The function K[b0 ,a0 ] is also called the kernel.
According to Feynman’s formulation, the kernel of a particle is determined in the following
way ([15], p. 34):
Z b0 Z tb
i
K[b0 ,a0 ] = e ~ S [b0 ,a0 ] (q) Dq, where S [b0 ,a0 ] (q) = L (t, q(t), q̇(t)) dt. (20)
a0 ta

The domain of S [b0 ,a0 ] are continuous functions q with fixed endpoints q(ta ) = a and q(tb ) = b. Due
to the heuristic manner of our definition of the path integral, we will not enter into the issue of the
V S̌ Feynman’s formulation of Quantum mechanics 12(19)

existence of the above integrals. Suffice it to say that it is the oscillatory nature of the integrated
functionals that makes the evaluation of the path integrals possible.
It should be noted, the kernel in equation (20) relates to nonrelativistic quantum mechanics,
also with no account for the spin of particles. In a relativistic quantum theory, one could still use
the notion of a kernel as a sum over all paths. In this paper, we shall be content with Feynman’s
formulation of nonrelativistic quantum mechanics with no spin.
The notion of a probability of finding a particle in a cetrain location is strange in the context
of classical physics, because classically the exact position can be determined. However, because
the exact position of the particle is known at time ta , the uncertainty of the position at that time is
0. That means, according to Heisenberg’s uncertainty principle δqδpq ≥ ~/2, the uncertainty in
velocity (which is proportional to momentum) is infinite. Thus the particle’s unknown velocity can
be of any value (in nonrelativistic physics the speed of light c is not the limit), and the particle can
be found anywhere in space at a later time tb .

2.2.2 Chain Rule for Events

Suppose we choose a time tc , so that ta < tc < tb , and c0 = (tc , c). We can then write:

Z tb
S [b0 ,a0 ] (q) = L (t, q(t), q0 (t)) dt, (21)
ta
Z tc Z tb
S [b0 ,a0 ] (q) = L (t, qa (t), q0a (t)) dt + L (t, qb (t), q0b (t)) dt, (22)
ta tc
S [b0 ,a0 ] (q) = S [b0 ,c0 ] (qb ) + S [c0 ,a0 ] (qa ). (23)

In the first term on the right–hand side of equation (23), we take the restriction qa = q|[ta ,tc ] with
endpoints q(ta ) = a and q(tc ) = c, while the second term is the restriction qb = q|[tc ,tb ] with endpoints
q(tc ) = c and q(tb ) = b. The differential can now be written as Dq = dc Dqa Dqb : we integrate
over all paths from a0 to c0 , integrate over all possible paths from c0 to b0 , then integrate over all
possible midpoints c at time tc (see Figure 5). We can then write:

Z Z b0 Z c0
K[b0 ,a0 ] = dc Dqb Dqa e(i/~)S [c0 ,a0 ] (qa )+(i/~)S [b0 ,c0 ] (qb ) , (24)
R c0 a0
Z Z c0 ! Z b0 !
i i
K[b0 ,a0 ] = dc e ~ S [c0 ,a0 ] (qa )
Dqa e ~ S [b0 ,c0 ] (qb )
Dqb . (25)
R a0 c0

We recognize the two factors in the integral as kernels. We can write the so called chain rule for
two events as

Z
K[b0 ,a0 ] = K[b0 ,c0 ] K[c0 ,a0 ] dc. (26)
R

The name kernel for K is now justified, for the above equation can be viewed as an integral
transform with the kernel function (also called nucleus) K[b0 ,c0 ] . Generalizing the chain rule result,
by picking n specific times ta < tc1 < · · · < tcn < tb , with c0i = (tci , ci ), we get
V S̌ Feynman’s formulation of Quantum mechanics 13(19)

Figure 5: In this demonstration of the chain rule, a sum over all paths between two points is
substituted by two sums over paths from an endpoint to the midpoint (tc , c); then we also have to
take the sum over all possible locations c for a specified tc .

Z  
K[b0 ,a0 ] = K[b0 ,c0n ] K[c0n ,c0n−1 ] · · · K[c02 ,c01 ] K[c01 ,a0 ] dc1 . . . dcn . (27)
Rn

One could choose for times tci the times ti from the definition of path integrals in step n: ti =
ta + i(tb − ta )/n. In step n, we would be integrating (n − 1)–times over variables c1 , . . . , cn−1 . By
sending n to infinity, the (n−1)-time integral would actually become a path integral, and the infinite
product of kernels would become the probability amplitude. We define c00 = a0 and c0n = b0 . Then
one could write the probability amplitude alternatively as (see [15], p. 38)

n−1
Y
φ(q) = lim K[ci+1 ,ci ] . (28)
n→∞
i=0

The function q in the amplitude has specified values q(ti ) = ci .

2.3 The Kernel of Multiparticle Systems

Until now, we have viewed the kernel K[b0 ,a0 ] as a function relating to a single particle. Moreover,
the path of the particle was a continuous function into one–dimensional space R. We are now
interested in a general case, when we have multiple particles in multi–dimensional space, or
perhaps even a general quantum system, described by generalized coordinates.
Suppose we have a system, which is classically described by n (independent) generalized
coodrinates q1 , . . . , qn , and with the classical action S (q1 , . . . , qn ) (which is obtained by integrating
the Lagrangian in the form of equation (3)). Again, we shall fix the endpoints of the functions in
question: for i ∈ {1, . . . , n} we define a0i B (ta , ai ), b0i B (tb , bi ) and demand qi (ta ) = ai , qi (tb ) = bi .
Note, that we specify endpoints for the same timepoints ta and tb for all functions qi . We shall write
the action of a system with specified endpoints as S [b01 ,a01 ;...;b0n ,a0n ] (q1 , . . . , qn ). The kernel (probability
amplitude) can be expressed as a multiple path integral (a generalization of eq. in [15], p. 66):

Z b01 Z b0n i
S [b0 ,a0 ;...;b0n ,a0n ] (q1 ,...,qn )
K[b01 ,a01 ;...;b0n ,a0n ] = ··· e~ 1 1 Dq1 . . . Dqn . (29)
a01 a0n
V S̌ Feynman’s formulation of Quantum mechanics 14(19)

Again, we can view this kernel as representing the wavefunction of the system at time tb in the
state qi (tb ) = bi , when the system was measured at time ta of being in the state qi (ta ) = ai .
A notable special case of a system with multiple degrees of freedom is one with a separable
action: the action S as a functional of multiple paths can be expressed as the sum of functionals S i
of one variable qi :

n
X
S [b01 ,a01 ;...;b0n ,a0n ] (q1 , . . . , qn ) = i
S [b 0 ,a0 ] (qi ). (30)
i i
i=1

If such a separation is possible, then the kernel can be written in the following way
(generalization of eq. in [15], p. 67):

Z b01 Z b0n i Pn i
S [b 0 ,a0 ] (qi )
K[b01 ,a01 ;...;b0n ,a0n ] = ··· e ~ i=1
i i Dq1 . . . Dqn , (31)
a01 a0n
n 0
Y bi i S i 0 0 (qi )
Z
K[b01 ,a01 ;...;b0n ,a0n ] = ~ [b ,a ]
e i i Dqi , (32)
i=1 a0i
Yn
K[b01 ,a01 ;...;b0n ,a0n ] = i
K[b 0 ,a0 ] . (33)
i i
i=1

If the action can be separated into a sum of actions of a single variable, then the kernel K of a
multi–variable system is the product of the kernels K i , which correspond to actions of a single
variable.
In Schrödinger’s quantum mechanics, the wavefunction exhibits certain symmetries under
particle exchanges. Namely, the wavefunction is symmetric under the exchange of two bosons
(particles with integer spin) and antisymmetric under the exchange of two fermions (particles
with half–integer spin). The product of kernels does not exhibit these types of symmetry under
particle exchanges, but rather conserves the type of symmetry of a pre–existing wavefunction in an
integral transformation, when we propagate the wavefunction to a later time with the kernel (see
equation (36) and section 3.2 for an explanation).

3 Connections to other physical theories

3.1 The classical limit
Quantum mechanics should be structured in such a way, that it reduces to classical mechanics in the
macroscopic limit. We shall make some brief arguments, why this is true in the case of Feynman’s
formulation of quantum mechanics.
In the path integral formulation of quantum mechanics, a particle takes no one path. Rather,
the probability amplitude is calculated by a sum over all paths q, with each path contributing just
a different phase factor (equation (20)). In classical physics, the particle takes a specific path q̄,
called the trajectory. An important question is, how is q̄ any different from other paths in the
macroscopic limit of Feynman’s formulation.
In the macroscopic limit, masses, spatial dimensions and times are large, which also means
the classical action S is large, at least compared to ~. One could also obtain the classical limit by
V S̌ Feynman’s formulation of Quantum mechanics 15(19)

demanding ~ → 0. Either way, S  ~, which means S /~  2π. Considering only paths with
the same endpoints, if one slightly changes the path from q to q + δq, then this can be observed in
the change of the action from S to S + δS . The phase factor between the probability amplitudes
δS
of these two paths is ei ~ . If q is not an extremal of S , then even relatively slight changes δq in
the path q bring about a large enough δS , so that δS /~ is comparative to 2π. Due to the periodic
nature (the period is 2π) of the exponential function of imaginary numbers, the phase shift between
the two paths is equally likely of being any value on the unit circle in complex numbers C. The
real and imaginary parts of the exponential function are the functions cosine and sine, respectively.
Considering paths similar to a given non–stationary q, the sine and cosine oscillate quickly around
the value 0, which means that neighboring paths cancel each other out: their net contribution to the
kernel is close to zero (see Figure 6).

Figure 6: In the macroscopic limit, neighboring paths cancel each other out, unless they are in the
neighborhood of the trajectory q̄, where δS = 0.

One path is exempt from this analysis: the classical trajectory. For this path, δS = 0 in the
first order, because this path is stationary. This means its neighboring paths don’t have a change
in the phase factor: all paths near the classical trajectory contribute a very similar probability
amplitude. This fact has consequences on the kernel. Suppose we are interested in K[b0 ,a0 ] , or
ultimately in the probability density at time tb at location b. If b ≈ q̄(tb ), then the classical trajectory
and its neighboring paths bring a substantial probability amplitude into the kerel, while all other
paths cancel each other out. The net contribution off all paths is equal to the contribution of the
classical trajectory, which means the particle is verly likely to be in the place specified by classical
mechanics. On the other hand, if b is far from q̄(tb ), then no possible paths with the specified
endpoints are stationary, which means they are cancelled out by their neighborhoods: the net
contribution to the kernel is 0. That also means the probability density is near zero: the particle is
unlikely to deviate from its trajectory in the classical limit.
It has already been mentioned that paths similar to the classical trajectory also contribute to the
kernel. They contribute, as long as their action is similar to the action S (q̄) in the units of ~; in
other words, paths contribute, while roughly δS < ~. This means the kernel considerably differs
from zero in a neighborhood of q̄(tb ): the path of a particle in the classical limit is still a little fuzzy.
It is very likely that the particle will be observed near its classical trajectory to within around ~
of S (q̄): this is consistent with Heisenberg’s uncertainty principle of quantum mechanics, for we
cannot know exactly a particle’s position and velovity at the same time, and the measure for this
uncertainty is ~.
V S̌ Feynman’s formulation of Quantum mechanics 16(19)

When not in the classical limit, when we are on an atomic scale, the action S is comparable to
~, which means all paths have to be considered.

3.2 Equivalence to Schrödinger’s formulation

It can be shown that Feynman’s and Schrödinger’s formulation of quantum mechanics are
equivalent. It will not be discussed here, how to get from Schrödinger to Feynman, but it can
be done (see [7], p. 154). We shall rather show, how to get to the wave functions and to the
Schrödinger equation (we follow the work in [15], p. 76–78).
The kernel K[b0 ,a0 ] contains both the starting point and the endpoint of the particle. Suppose we
are interested only in the endpoint, while we are not interested in the place of origin of the particle.
This new concept will be the wave function (we define x0 = (t, x)):

Ψ(x, t) B K[x0 ,a0 ] . (34)

Note, that the wave function written above is a wave function for a single particle in
onedimensional space. The wave function for a kernel with multiple degrees of freedom xi would
be defined as:

Ψ(x1 , . . . , xn , t) B K[x10 ,a01 ;...;xn0 ,a0n ] . (35)

Again, xi0 = (t, xi ) in the above definition. Ψ in the above definitions (34) and (35) are the
mathematical objects familiar to us from Shrödinger’s formulation. Now, we shall try to derive the
Schrödinger equation (7) for a single particle in one dimension (the most simple case).
First, we can rewrite equation (26), by defining x20 = b0 , x00 = a0 , x10 = c0 , x20 = (t2 , x2 ),
x0 = (t0 , x0 ), x10 = (t1 , x1 ) and recognizing two kernels as wave functions: K[x20 ,x00 ] = Ψ(x2 , t2 ) and
0

K[x10 ,x00 ] = Ψ(x1 , t1 ). We get:

Z
Ψ(x2 , t2 ) = K[x20 ,x10 ] Ψ(x1 , t1 ) dx1 . (36)
R

Again, we can interpret the kernel K[x20 ,x10 ] as the kernel of an integral transform; we could also
view K as the propagator, because it tells us how to propagate the wavefuntion from an earlier time
t1 to a later time t2 . Alternatively, we can also view the kernel K[x0 ,x10 ] as a wave funtion Ψ(x, t) with
the initial value Ψ(x, t1 ) = δ(x − x1 ), where δ(x) is the Dirac delta function. It should be noted that
not all wavefunctions (as known from Schrödinger’s quantum mechanics) are kernels, but only
those wavefunctions with the peculiar initial condition of a delta function (these wavefunctions
can not be normalized). However, equation (36) holds for all wavefunctions, even those we did not
define in the Feynman formalism by equation (34).
The obtained equation (36) is an integral equation, while we would like to get a differential
equation. That is why we take the two timepoints very close together: t1 = t and t2 = t + . We will
use an approximation for the kernel with a short time–gap between the specified endpoints, as if
the action on the short time interval is constant (evaluated in the spatial midpoint (x1 + x2 )/2, with
the velocity being equal to the average velocity (x2 − x1 )/, with only one path x(t) in the sum—the
straight path):
V S̌ Feynman’s formulation of Quantum mechanics 17(19)

Z (t+,x2 )
i 1 ~i R t+ L (t,x(t), ẋ(t)) dt 1 ~i L (t+ 2 , x2 +x1 x2 −x1
2 ,  ).
K[x20 ,x10 ] = e ~ S (q) Dq ≈ e t ≈ e (37)
(t,x1 ) A A
If we insert (37) into (36), while also inserting the Lagrangian for a single particle in one
dimension L (t, x, ẋ) = m ẋ2 /2 − V(x, t) (m is the mass of the particle, while V is the potential), we
get:

Z
1 ~i L (t+ 2 , x2 +x1 x2 −x1
2 ,  ) Ψ(x , t) dx .
Ψ(x2 , t + ) = e 1 1 (38)
R A
Z ! !
1 ~i  m(x2 −x2 1 )2 − ~i V( x1 +x2 ,t+  )
Ψ(x2 , t + ) = e 2 e 2 2 Ψ(x1 , t) dx1 . (39)
R A

The exponential of the first factor contains (x2 − x1 )2 /, which is very large, if x2 and x1 are
not very close. This means that for x1 not very near x2 , the integrated function oscillates (all other
functions in the integral are smooth) and the total contribution of the part of the integral far from
x2 is zero. We define x2 = x and make the substitution x1 = x2 + η, suspecting contributions only
for η near 0.

Z ! !
1 ~i mη22 − ~i V(x+ 2η ,t+ 2 )
Ψ(x, t + ) = e e Ψ(x + η, t) dη. (40)
R A
√
We can see that the phase change of a few radians in the phase factor will require η ∼ 2~/m.
We will be interested only in approximations in the first order of , or consequently of order η2 in
η. The approximation will be done by taking only the first few terms in the Taylor series around
the point (x, t) for functions Ψ and V:

i mη2
∂Ψ ∂Ψ η2 ∂2 Ψ
Z  2 ∂V η ∂V
 !
e ~ 2 − ~i V(x,t)+ 2 ∂t + 2 ∂x +O( )
2
Ψ(x, t) +  + O( 2 ) = e Ψ(x, t) + η + + O(η ) dη,
3
∂t R A ∂x 2 ∂x2

∂Ψ ∂Ψ η2 ∂2 Ψ
Z ! !
1 ~i mη22 i
Ψ(x, t) +  = e 1 − V(x, t) Ψ(x, t) + η + dη. (41)
∂t R A ~ ∂x 2 ∂x2
If we send  → 0, the equation (41) must still hold, and we can determine the constant A by
comparing the two terms with Ψ:

!2
Ψ(x, t) ∞
Ψ(x, t) 2πi~
Z
imη2
Ψ(x, t) = e 2~ dη = (42)
A −∞ A m
!1/2
2πi~
=⇒ A = . (43)
m
The integral in (42) looks similar to an integral of a Gaussian function, but with an imaginary
variance. The integral is over an infinite region and does not converge in the mathematical sense. A
meaningful value for this integral can still be obtained by taking the value around which the integral
V S̌ Feynman’s formulation of Quantum mechanics 18(19)

RR imη2 R∞ 2
2~ dη oscillates, as R → ∞. This can be evaluated by a trick, if we integrate −λη imη
−R
e −∞
e e 2~ dη,

which converges for all λ > 0, and sending λ → 0 in the result. By the same principles, the values
of two other integrals are obtained (the first is zero, because the integrated function is of odd parity
in η; the result from the second one is from [15], p.78):

1 ∞ imη
Z 2
e 2~ η dη = 0, (44)
A −∞
1 ∞ imη
Z 2 i~
e 2~ η2 dη = . (45)
A −∞ m
Now, knowing the value of the above integrals, we return to equation (41):

∂Ψ 1 ∞ imη ∂Ψ ∂ Ψ
! Z Z ∞ 2 2 Z ∞ 2 !
i 2 1 imη 1 1 imη
Ψ(x, t) +  = 1 − V(x, t) Ψ(x, t) e 2~ dη + e 2~ η dη + e 2~ η dη ,
2
∂t ~ A −∞ ∂x A −∞ 2 ∂x2 A −∞

∂Ψ ∂Ψ 1 ∂2 Ψ i~
! !
i
Ψ(x, t) +  = 1 − V(x, t) Ψ(x, t) · 1 + ·0+ · , (46)
∂t ~ ∂x 2 ∂x2 m
∂Ψ ~2 ∂2
!
i~ = − + V(x, t) Ψ. (47)
∂t 2m ∂x2
We have obtained the desired end–result: equation (47) is the Schrödinger equation for a particle
in 1 dimension in the presence of potentials, and we have shown it follows from Feynman’s
formulation of quantum mechanics.

Conclusion
Since its inception, Feynman’s formulation has been of great importance in theoretical physics.
One advantage over Shcrödinger’s formulation is in the use of the Lagrangian instead of the
Hamiltonian. In the Schrödinger equation, we relate the time derivative of the wavefunction with
the Hamiltonian. However, in relativity, time is not independent of the reference frame. Thus the
Hamiltonian is not very appropriate for a relativistic theory. On the other hand, the Lagrangian
can be written in a manifestly relativistic form, making it the preferred choice in describing a
relativistic quantum theory.
An extension of Feynman’s approach has also been carried out in quantum field theory. In
this case, the sum over all possible paths is replaced by a sum over all possible scalar fields φ :
R3 × R −→ R. The action S now has a domain of all fields, with integration of the Lagrangian
density over a region in space–time. A formal mathematical framework dealing with integration
over all fields still remains to be developed [4].
Finally, path integrals have also established a remarkable link between quantum mechanics
and stochastic processes. The first hint of similarity comes from the Schrödinger equation, since it
is essentially a diffusion equation with an imaginary diffusion constant. Path integrals themselves
take the similarity further, since they are also used in a number of areas in statistical physics, such as
the study of Brownian motion (for the description of random walks of particles), in polymer science
(for random chains), and even outside statistical physics in the study of financial markets. In the
V S̌ Feynman’s formulation of Quantum mechanics 19(19)

1970s, there was a grand synthesis of quantum field theory and statistical field theory. Namely,
the partition function in statistical mechanics can be obtained by performing the “Wick rotation”
t → it on Feynman’s path integrals [4].
In all likelihood, the path–integral will remain an important tool in many areas of theoretical
physics in the future.

References
[1] http://en.wikipedia.org/wiki/History of quantum mechanics (11/29/2009)

[2] http://en.wikipedia.org/wiki/Mathematical formulation of quantum mechanics

(11/29/2009)

[3] http://en.wikipedia.org/wiki/John von Neumann (11/29/2009)

[4] http://en.wikipedia.org/wiki/Path integral formulation (11/29/2009)

[5] http://en.wikipedia.org/wiki/Lagrangian (11/29/2009)

[6] http://en.wikipedia.org/wiki/Hamilton’s principle (11/29/2009)

[7] M. Chaichian and A. Demichev, Path Integrals in Physics, Volume 1: Stochastic Processes
and Quantum Mechanics (Institute of Physics Publishing, Bristol and Philadelphia, 2001).

[8] S. Mazzucchi, Mathematical Feynman Path Integrals And Their Application (World Scientific
Publishing, Singapore, 2009).

[9] [image] http://www.fourmilab.ch/documents/figures/heisenberg.jpg

(11/15/2009)

[10] [image] http://www.flavinscorner.com/portrait-schrodinger.jpg (11/15/2009)

[11] [image] http://boringrise.files.wordpress.com/2009/04/richard-feynman.jpg

(11/15/2009)

[12] B. Dacorogna, Introduction To The Calculus Of Variations (Imperial College Press, London,
2004).

[13] A. Wachter and H. Hoeber, Compendium of Theoretical Physics (Springer, New York, 2006).

[14] F. Schwabl, Quantum Mechanics (Springer, Berlin, 2007).

[15] R. P. Feynman and A. R. Hibbs, Quantum Mechanics and Path Integrals (McGraw–Hill,
1965).

[16] W. Rudin, Principles of Mathematical Analysis, Third Edition (McGraw–Hill Kogakusha,

Tokyo, 1976).

[17] http://en.wikipedia.org/wiki/Functional integration (11/15/2009)