INDICES OF QUADRATIC PROGRAMS OVER

REPRODUCING KERNEL HILBERT SPACES FOR

FUN AND PROFIT
arXiv:2412.18201v1 [math.OC] 24 Dec 2024

GEOFFREY HUTINET AND J. E. PASCOE†

Abstract. We give an abstract perspective on quadratic pro-

gramming with an eye toward long portfolio theory geared toward
explaining sparsity via maximum principles. Specifically, in opti-
mal allocation problems, we see that support of an optimal distri-
bution lies in a variety intersect a kind of distinguished boundary
of a compact subspace to be allocated over. We demonstrate some
of its intelligence by using it to solve mazes and interpret such
behavior as the underlying space trying to understand some hy-
pothetical platonic index for which the capital asset pricing model
holds.

1. Introduction
The manucript contains animations which may appear static in feature-
incomplete pdf viewers.
Stock market indices, such as the S&P 500, Dow Jones Industrial
Average, NASDAQ Composite index and the like, are supposed to
in some way capture the overall health and behavior of the market.
The so-called “Bogleheads” (followers of the philosophy of Jack Bogle
[13, 14]) often make stronger claims such as one cannot beat such and
such an index (such as the S&P 500) in the long run. The point being
that such an allocation reflects the broad allocation of the resources of
Date: December 25, 2024.
2020 Mathematics Subject Classification. 30C15, 30C80, 47B32, 46E22, 91G10,
90C20.
Key words and phrases. modern portfolio theory, topiary, kernel embedding of
measures, positive allocation problem, long investment, reproducing kernel Hilbert
spaces, geometric function theory, maze solving.
† Partially supported by National Science Foundation DMS Analysis Grant
2319010. Thanks to Mathematisches Forschungsinstitut Oberwolfach workshops
“Real Algebraic Geometry with a View toward Koopman Operator Methods” and
“Non-commutative Function Theory and Free Probability.” Part of this research
was performed while the author was visiting the Institute for Pure and Applied
Mathematics (IPAM), which is supported by the National Science Foundation
(Grant No. DMS-1925919).
1
2 G. HUTINET AND J. E. PASCOE

society, and perhaps that, in principle, the growth in our real resources
cannot be too outpaced within the market, if only because the outpac-
ing portion becomes synonymous with the market as the residual parts
of the market become impossible to resolve by comparison. (Such can
be viewed as a sort of efficient market hypothesis, see [32].)
On the other hand, many proponents of “value” advocate for a nar-
row range of extremely well-researched long investments, with a general
preference for determinism over speculative large payoffs. (Large spec-
ulative payoffs are often sought in the arena of “growth” investing,
e.g. [8].) Advocates of such a philosophy often cite [19] as founda-
tional. Here an investment being long means that you actually own
something, as opposed to investment based on more exotic financial
products, such as shorts, options, pet insurance and other derivative
products.
We give some formal justification to the apparently contradictory
reasoning– our invisible index theorem says that the optimal portfolio
will approximate some broad index if possible, but because of geomet-
ric features of the space of securities, the optimal portfolio may remain
sparse. Moreover, any optimal portfolio behaves somewhat like an in-
dex, and all assets that are not used are worse than some capital asset
pricing model of that index would require. Sparsity, in turn, suggests
that investing fundamentally is massively multiplayer– even if we
remove the assets underlying ones optimal portfolio, good approxima-
tions to the invisible index may be available using wildly different secu-
rities which fit together to make a coherent portfolio is a superficially
different way.

1.1. Diversification. First, we will review the formal justification for

diversification (essentially the arithmetic-geometric mean inequality,
that parallel beats sequential, etc.) and the consequent modern port-
folio theory.

1.1.1. One coin. Consider the following game: one is allowed to bet
any amount of money on a single coin flip, if one wins they get back
double their bet, otherwise, one gets two fifths their bet back, rounding
down any fractional pennies. For example, without loss of generality,
assume we always pick heads and always bet all of our money. Starting
with one dollar we flip the coin ten times and flip the sequence
HT T T T T HHHT
where each H represents a head and T represents a tail. We can track
how much money we would have in the following table.
 TOPIARISM 3

0 1 2 3 4 5 6 7 8 9 10
Flip - H T T T T T H H H T
Value $1.00 $2.00 $0.80 $0.32 $0.12 $0.04 $0.01 $0.02 $0.04 $0.08 $0.03
Clearly, on such a play we have gotten somewhat unlucky.
To analyze the game, let us first summarize the possible outcomes
of a single flip in a table.
H T
Outcome 2 2/5
Each outcome has a likelyhood or probability of happening.
H T
Chance 50% 50%

On glance, the game seems good. On average, we make 2+2/5 2

= 65
times our money for every bet made, that is the expected value of an
individual round where we bet one unit of value is to get an extra fifth
of a unit back. Obviously a great deal.
On the other hand, suppose again that we always go all in and the
difference between the number of heads and tails is less than or equal
to 10%. Let 100 be the number of flips so far. At most we have 55
heads, and at minimum 45 tails. Under such rosy assumptions, we will
have  45 50
450

55 2 4
2 = 45 = ≈ 0.9450 ≈ 0.045
5 5 59/10
units of value for every one put in intitially. (Here the 2n+n/10 term

n−n/10
comes from compounding the wins, and 52 from compounding
losses.) Thus, in the long run with such a strategy, we lose money.
From expectation we might have expected (6/5)100 ≈ 82, 817, 974 units
of value. Incongruous.
One sees parodies of “growth” in the regime of games as those above–
some may even have speculated that every individual company, no
matter how successful, is destined to die. The dedicated reader who
simulates such a game may observe that it looks like it might have
actually been “growing” during some periods, but also goes to 0. One
should not follow ghosts whole-heartedly into the abyss– those that
embody the bull in the china shop.
1.1.2. Martingale gambling. Suppose there is a game where you can bet
any amount and if you win, you double your bet, and on a loss you lose
everything. No prior assumptions on the odds except that the chance of
winning is positive and constant. The martingale strategy supposedly
employed by gamblers was to double ones bet on a loss. Thus, as the
potential outcome of the next game had double the potential upside,
they would be accounting of all their lost prior bets with a potential
to gain what was to be gained on the first game itself.
4 G. HUTINET AND J. E. PASCOE

1.1.3. Going all-in lets the house play a martingale. If one continually
goes all in, we see the exponential increase in bets that occurs in a
martingale strategy. Note that after some point successive bets will
need to be made with loans.
That is, companies and banks can play more aggressive, high-risk
strategies due to their extreme capacity for leverage. Presumably, when
one invests in something that itself has debt, although they may have
no obligation to pay, they still accept indirect risk due to the prior claim
of the bondholders. One gauges how successful their use of leverage has
been within the context of their overall strategy. As large amounts of
debt and leverage are already baked into assets, analysis becomes very
delicate when one wants to use leverage on top of that.

1.1.4. Two coins. Consider again the following game: one is allowed
to bet any amount of money on a single coin flip, if one wins they get
back double their bet, otherwise, one gets two fifths their bet back,
rounding down any fractional pennies. Now assume one can play two
games at once, splitting your money between the coins however you
like. For example, without loss of generality, assume we always pick
heads and always bet all of our money. We can track how much money
we would have in the following table.
0 1 2 3 4 5 6 7 8 9 10
Flip - H T T T T T H H H T
Flip - H T T H T T H T H H
Value $1.00 $2.00 $0.80 $0.32 $0.38 $0.14 $0.04 $0.08 $0.04 $0.08 $0.04

We have added a second coin to our original data in Section 1.1.1

and thus inherited some of its bad luck. We have made a penny more
though.
To analyze the game, let us first summarize the possible outcomes
of a single flip in a table.
Outcome H T
H 2 6/5
T 6/5 2/5

Each outcome has a likelyhood or probability of happening.

Probability H T
H 25% 25%
T 25% 25%

On glance, the game seems good but similar in ruin to the first game
as we only got a penny more on the small simulation above. On average,
we make 2+6/5+6/5+2/5
4
= 65 times our money for every bet made, that
is the expected value of an individual round where we bet one unit of
value is to get an extra fifth of a unit back. Still obviously a great deal.
On the other hand, suppose again that we always go all in. Let 100
be the number of flips so far, assume we have flipped 27 double tails,
50 mixed heads and tails and 23 double heads, a somewhat average
 TOPIARISM 5

outcome. We will have

 27  50
23 2 6
2 ≈ 1.38
5 5
units of value at the end.

1.1.5. Magic coins. Consider again the following game: one is allowed
to bet any amount of money on a single coin flip, if one wins they get
back double their bet, otherwise, one gets two fifths their bet back,
rounding down any fractional pennies. Now assume one can play two
games at once, splitting your money between the coins however you like.
Moreover, these coins are enchanted so that they always flip opposite
outcomes. For example, without loss of generality, assume we always
pick heads and always bet all of our money.
We can track how much money we would have in the following table.
0 1 2 3 4 5 6 7 8 9 10
Flip - H T T T T T H H H T
Flip - T H H H H H T T T H
Value $1.00 $1.20 $1.44 $1.72 $2.06 $2.47 $2.96 $3.55 $4.26 $5.11 $6.13

We took our original unlucky sequence from Section 1.1.1 and turned
it into quite a good one.
To analyze the game, let us first summarize the possible outcomes
of a single flip in a table.
Outcome H T
H 2 6/5
T 6/5 2/5

Each outcome has a likelyhood or probability of happening.

Probability H T
H 0% 50%
T 50% 0%

The outcome in terms of our value is deterministic. We earn a fifth of

our money each time we play. Thus, by combining two anti-correlated
coins, we have created a riskless game.
For general configurations of coins, see the theory of Kelly strategies
[22].

1.2. Modern portolio theory. Given a vector of average returns ψi

with covariances ρij modern portfolio theory suggests that we maximize
X 1X
ωi ψi − ωi ωj ρij
2
where ωi sum to 1. If one assumes that the underlying securities grow as
geometric Brownian motions, the maximum corresponds to the portfo-
lio with maximum asymptotic median growth rate. Namely, geometric
6 G. HUTINET AND J. E. PASCOE

Brownian motion satisfies the stochastic differential equation:

dS
= mdt + σdBt .
S
At time t,
2 /2)t+σB
S = e(m−σ t

√
where Bt is a Brownian motion which grows like t. Thus, to find the
optimal portfolio for long term median growth, one sees the need to
maximize m − σ 2 /2 which is given explicitly by the formula above.
The following principle explains the reward for diversification as
a reward for risk management (or generation determinism [for lower
bounds]): In terms of long term median growth, given a basket of se-
curities with approximately equal returns, one is stochastically paid for
destroying variance at a rate on one variance destroyed gives one half
unit of median return.
The capital asset pricing model states that
ψi − r = βi (ψM − r)
where ψM is the return of the market portfolio (often modelled by an
index such as the S&P 500) and r represents the return on a risk-
free asset (usually modelled by something like a treasury bill,) and βi
is the covariance of the market portfolio with asset i divided by the
variance of the market portfolio. The supposed point being that to
beat the market one needs to choose stocks with more market risk, in
the sense that their covariance with the market portfolio is higher than
the variance of the market portfolio. Fama-French type models add
extra terms to better fit data– in principle, these are factors affecting
the return that cannot be captured by mean-variance analysis. See [18]
for futher overview.
The approach to capital asset pricing modeling does not take into ac-
count an important phenomenon– that a portfolio must cohere. When
one selects an outfit, some things do not work together– if you need to
add some particular accessory, one may need to remove other articles
to have the ensemble make sense. (Where harmony may be thought
of as approximating the magic coins above.) We rectify such by deal-
ing with coherent fragments of the market directly, which will be our
topiaric index theory.

1.3. The maximum principle. The maximum principle states that

in a steady state of a diffusion type process, the maximum concentra-
tion must occur on the boundary. (That is, if there were a concentra-
tion maximum at some interior point, then it would have diffused to
 TOPIARISM 7

other points in the interior.) Such occur in general in complex vari-

ables and functional analysis, where it is often called the distinguished
or Shilov boundary. We shall define an appropriate frontier for which
the maximum principle holds in quadratic programming.

1.4. Main results. We establish that optimal portfolios must be cho-

sen from some potentially small set we call the green frontier in our
green topiary theorem. We also show that if in some expanded
universe there is a totally diversified index fund as the optimal port-
folio, then finding the optimal portfolio in a restricted class of assets
is merely trying to approximate the fluctuations of the index, which
we call the invisible index theorem. In general, we treat portfolios
which are internally consistent and assets which are inconsistent with
such must perform worse than the capital asset pricing model would
predict, which we call the capital asset pricing inequality.

2. Preliminaries on Hilbert spaces

We now recall some notions from the theory of Hilbert spaces.
Let H be a real vector space. Call ⟨·, ·⟩ : H × H → R an inner
product if
(1)⟨x, x⟩ ≥ 0,
(2)If ⟨x, x⟩ = 0, then x = 0,
(3)⟨x, y⟩ = ⟨y, x⟩,
(4)⟨αx + βz, y⟩ = α⟨x, y⟩ + β⟨z, y⟩.
p
We call ∥x∥ = ⟨x, x⟩ the norm.
We a vector
P∞ space with an inner product P∞ a real Hilbert space if
whenever n=1 ∥xn ∥ converges then n=1 xn converges.
Examples include:
(1) Rn , where ⟨(x1 , . . . , xn ), (y1 , . . . , yn )⟩ =
P
xi yi , (Euclidean space)
(2) If X, Y are random variables with finite variance, ⟨X, Y ⟩ =
Cov(X, Y ),
(3) Rn , A a positive semidefinite matrix, where ⟨x, y⟩ = y T Ax,
The set of square summable sequences where ⟨(x1 , x2 , . . .), (y1 , y2 , . . .)⟩ =
(4) P
xi yi , called ℓ2 . R1
(5) The set of functions on the unit interval such that 0 |f |2 dx <
R1
∞. where ⟨x, y⟩ = 0 f gdx. (Called L2 .)
Say x ⊥ y or x is perpendicular to y if ⟨x, y⟩ = 0. Hilbert spaces
satisfy the Pythagorean theorem, where it is obtained via algebra in
the following classic proof.
8 G. HUTINET AND J. E. PASCOE

Theorem 2.1 (The Pythagorean theorem). x ⊥ y if and only if

∥x∥2 + ∥y∥2 = ∥x + y∥2 .
Proof.
∥x∥2 + ∥y∥2 = ⟨x, x⟩ + ⟨y, y⟩
= ⟨x, x⟩ + 2⟨x, y⟩ + ⟨y, y⟩
= ⟨x + y, x + y⟩ = ∥x + y∥2
□
Because of the relationship to coavariance, perpendicularity can help
capture the notion of statistical independence. Another important fact
is the following inequality.
Theorem 2.2 (Cauchy-Schwarz inequality). ⟨x, y⟩ ≤ ∥x∥∥y∥.
The following will be very powerful for us.
Theorem 2.3 (The Riesz representation theorem). Let H be a Hilbert
space. If λ : H → R is bounded and linear, that is |λ(x)| ≤ C∥x∥, then
there exists a unique λ̃ ∈ H such that
λ(x) = ⟨x, λ̃⟩.
Definition 2.4. A real reproducing kernel Hilbert space H is
a Hilbert space of real-valued functions on some domain Ω such that
point evaluation at each α ∈ Ω is a bounded linear functional.
We call the element realizing this via the Riesz representation theo-
rem kα . That is,
f (α) = ⟨f, kα ⟩.
See [29] for reference.
Examples include:
(1) Euclidean space, where we view vectors as functions on {1, . . . , n},
(2) ℓ2 ,
(3) NOT L2 of [0, 1] (such functions are merely defined almost ev-
erywhere, that is point evaluations are not defined)
(4) P
The space of functions on the complex unit disk of the form
∞ n
P 2
c
−∞ n z where |c n < ∞ and cn = c−n . Here,
|
1 + zw
kw (z) = Re .
1 − zw
Called the real Hardy space of harmonic functions on the
disk.
 TOPIARISM 9

Definition 2.5. Let H be real reproducing kernel Hilbert space on

some domain Ω. Let µ be a compactly supported distribution on Ω.
We associate the embedded distribution µ ∈ H via the Riesz rep-
resentation theorem of the functional given by integration against µ.
That is,
Z
λµ (f ) = f dµ = ⟨f, µ⟩.

(We abuse notation to treat these two objects the same.)

We call the map taking a measure to its corresponding Hilbert space
element the kernel embedding of distributions.
We define the space of embedded measures to be the weak clo-
sure of the image of the kernel embedding of distributions.
P P
For a finite sum of point masses ν = αi δxi we view ν = αi kxi
and
X
⟨f, ν⟩ = αi f (xi ).
See [33] for information on kernel embeddings and learning applica-
tions. Our results explain sparsity (thus fast to evaluate) in the positive
analog of support vector machines. (Instead of doing the job of sep-
arating two sets, instead one can think of this as wanting to find the
outline of a set.)

3. Topiarism
Our basic problem of study is as follows.
Definition 3.1. Let H be real reproducing kernel Hilbert space on
some domain Ω. Let ψ be a continuous function on Ω. Let K ⊆ Ω be
compact. We define the aesthetic objective to be
Z
O(µ) = ψdµ − ∥µ∥2 /2.

We call the maximizer of O over all distributions the topiary of K

with respect to ψ over H, denoted topiary(K).
The minimizer may not be unique as a measure, but the embedded
measure is unique. When ψ + r ∈ H for some r ∈ R, we have that
maximizing the aesthetic objective is equivalent to minimizing ∥ψ +
r − µ∥.
That is, minimizing
∥ψ − µ∥2 = ∥ψ∥2 − 2⟨ψ, µ⟩ + ∥µ∥2 ,
10 G. HUTINET AND J. E. PASCOE

is equivalent to maximizing
Z
O(µ) = ψdµ − ∥µ∥2 /2

as ψ is fixed. That is, from the closest measure perspective, the latter
aesthetic formulation we have adopted need not require ψ ∈ H.
An important example is the problem of optimization of long port-
folios (and other operations research problems) following Markowitz
modern portfolio theory [18]. (The asymptotic growth rate of a geo-
metric Brownian motion is equal to its mean minus its variance over
two.) The name topiary is chosen in relation to the financial concept
of hedging.
Lemma 3.2. Let H be real reproducing kernel Hilbert space on some
domain Ω. Let ψ be a continuous function on Ω. Let K ⊆ Ω be compact.
The aesthetic objective satisfies
Z
DO(µ)[δx − µ] = ψ(x) − µ(x) − ψ(t) − µ(t)dµ(t).

Thus, we call
Z
ιµ = ψ − µ − ψ(t) − µ(t)dµ(t)

the aesthetic margin. We define the score S(µ) of µ to be the supre-

mum of the aesthetic margin.
Theorem 3.3. Let H be real reproducing kernel Hilbert space on some
domain Ω. Let ψ be a continuous function on Ω. Let K ⊆ Ω be compact.
The aesthetic margin
R of topiary(K) is 0 on its support and nonpositive
on K. We call ψ(t)−µ(t)dµ(t) the available topiaric rate, denoted
rK .
We define the topiaric index of K, denoted Topiary(K), to be the
preimage of zero for the aesthetic margin of the topiary. That is,
Topiary(K) = ι−1
topiary(K) (0).

We call any set arising as the topiaric index of some K a topiaric

index.
Proof. The aesthetic margin must be be nonpositive at an optimum.
If it were negative at a point its support, one would be able to scale
up outside of a neighborhood of the point and scale down within a
neighborhood of the point to to increase the aesthetic objective. □
Call a set of the form {x|ψ(x) − f (x) = C} for some f ∈ H, C ∈ R a
marginal hypersurface. We see the following principle: A topiaric
index for a set K lies on a marginal hypersurface.
 TOPIARISM 11

4. The capital asset pricing inequality

The connection to nonpositive functions allows one to interface with
Herglotz type theory as was done in [26] and replicate approaches to
the Julia-Caratheodory theory [21, 15, 35] as in [28, 27] following the
work of Agler, McCarthy and Young [6], Agler, Tully-Doyle and Young
[7] and Agler and McCarthy [2] which give examples where regularity
of a function forces the regularity of some underlying Agler kernels. In
complex variables, we may also thus have connections to the theory of
stable polynomials, their geometry, and the relation to complex func-
tion theory as in [11, 12, 5, 4]. For a general introduction to advanced
kernel methods and geometry beyond that of basic repoducing kernel
Hilbert spaces, see [3]. The point being that with an appropriate view,
one should see a theory relating the behavior on the topiaric index of
K to behavior on the rest of K. A concrete problem would be to un-
derstand the Fama-French theory (or general arbitrage pricing theory,
although the von Neumann elephant trunk-wiggling phenomenon there
probably obfuscates the matter) in terms of the capital asset pricing
inequality below.
In the case of portfolio optimization, if there is a risk free asset in
the optimal portfolio, the available topiaric rate is equal to the risk free
rate.
Theorem 4.1 (Capital asset pricing inequality). Let H be real repro-
ducing kernel Hilbert space on some domain Ω. Let ψ be a continuous
function on Ω. Let K ⊆ Ω be compact. Let µ ̸= 0 be the topiary of K.
Define the topiaric beta of x ∈ Ω to be
topiary(K)(x)
β(x) = .
∥ topiary(K)∥2
Note β ∈ H. For every x ∈ K we have that
Z 
ψ(x) − rK ≤ β(x) ψdµ − rK

where equality holds exactly on the topiaric index of K.

R
Proof. Let µ be the topiary of K. Note ψ − µ − R − µ(t)dµ(t) ≤ 0.
ψ(t)
µ
2
R
Also, ∥µ∥ = ψdµ − rK . So, ψ − rK = µ = ∥µ∥2 ( ψdµ − rK ). □

The classical capital asset pricing model [18] is essentially the as-
sumption that whatever assets being analyzed form a topiaric index
and contain a risk free asset. We note that “alpha” based analysis
with respect to a topiaric index returns alpha uniformly nonpositive,
12 G. HUTINET AND J. E. PASCOE

that is Z 
α(x) = ψ(x) − rK − β(x) ψdµ − rK

is nonpositive.
The exact elaboration of Julia-Caratheodory-Wolff theory in such a
setting is unclear, especially with respect to more sophisticated notions
such as horocycles. As with the approach of Agler, McCarthy and
Young of Julia-Caratheodory type theorems [6], we derive estimates
using Cauchy-Schwarz. A somewhat simplistic interpretation is that
they describe regularity of functions near boundary optima.
Theorem 4.2 (Topiaric Julia-Caratheodory inequality). Let H be real
reproducing kernel Hilbert space on some domain Ω. Let ψ be a contin-
uous function on Ω. Write d(x, y) = ∥kx − ky ∥. Let K ⊆ Ω be compact.
Let µ be the topiary of K. Let x be in the topiaric index of K. For every
y ∈ K such that d(x, y) ̸= 0 we have that
ψ(y) − ψ(x) µ(y) − µ(x)
−∥ψ∥ ≤ ≤ ≤ ∥µ∥
d(y, x) d(y, x)
where ∥ψ∥ if formally infinite if ψ ∈ / H. Here, the middle inequality is
an equality if y is also in the topiaric index.
Proof. Note,
Z 
ψ(x) − rK = β(x) ψdµ − rK = µ(x),

and Z 
ψ(y) − rK ≤ β(y) ψdµ − rK = µ(y).

Subtracting,
ψ(y) − ψ(x) ≤ µ(y) − µ(x).
Thus, the middle inequality is satisfied. Now,
µ(y) − µ(x) = ⟨µ, ky − kx ⟩ ≤ ∥µ∥∥ky − kx ∥
by the Cauchy-Schwartz inequality. Substituting d(x, y) = ∥kx − ky ∥
and rearranging gives the right hand inequality. Similarly, if so defined
ψ(y) − ψ(x) = ⟨ψ, ky − kx ⟩ ≥ −∥ψ∥∥ky − kx ∥
by Cauchy-Schwartz. □
Informally, we may interpret ψ as more contractive that µ.
 TOPIARISM 13

µ(y)

(µ(x), ψ(x))
ψ(y)
rK

Figure 1. Our topiaric Julia-Caratheodory inequality

says that the elements in the topiaric frontier lie on a line
with slope 1, depicted in red, elements of K, depicted in
brown must be below the line, whereas elements in Ω not
in K may be above or below, depicted in violet. The line
must intersect the ψ(y) axis at the available topiaric rate
rK . Thus, the topiaric Julia-Caratheodory inequality is
somewhat analogous to the classical security market line,
[18].
14 G. HUTINET AND J. E. PASCOE

µ(y)

(µ(x), ψ(x))
ψ(y)
rK

Figure 2. If instead we have that the remainder of Ω

lies on the axis, we see that the topiary has a small in-
ner product with the market, which in interpretations as
covariance gives that it is independent of the rest of the
space. (One can imagine a variety of reasons for such
in the securities context, innovation, new ideas, fraud,
memes and so on.)

5. The invisible index theorem

We can become more comfortable with both the empirical failure of
the capital asset pricing model and the obviously contradictory held-
fast belief in such if we employ the belief that there is some, perhaps
hypothetical, bigger “platonic” invisible index which is supported
everywhere out there where the model is saved. (Called such in anal-
ogy with the classical invisible hand.) Note, for example, no investor
has access to all assets– matters such as salience, sanctions and other
barriers. See Roll’s critique [31] on lack of testability of the classical
capital asset pricing model. Note nothing obstructs the topiary of such
a topiaric index minus a few key elements from being quite sparse.
Theorem 5.1 (Invisible index theorem). Let H be real reproducing
kernel Hilbert space on some domain Ω. Let ψ be a continuous function
on Ω. Let K1 ⊆ K2 ⊆ Ω be compact. Suppose K2 is a topiaric index.
The topiary of K1 minimizes ∥µ − topiary(K2 )∥.
R
Proof. We see to minimize ∥µ−topiary(K2 )∥2R= ∥µ∥2 −2 topiary(K2 )dµ+
∥ topiary(K2 )∥2 . Thus, we want to maximize topiary(K2 )dµ−∥µ∥2 /2
As K2 is an index topiary(K2 ) = ψ + r for some r ∈ R, and thus we
have our original maximization problem. □
 TOPIARISM 15

To complement the above we note that, if ψ ∈ H, then maximizing

our aesthetic objective is equivalent to minimizing ∥ψ − µ∥. Namely, if
ψ is an embedded distribution, then ψ is the topiary.

6. On the green frontier

We now define the appropriate distinguished boundary for our present
consideration.
Definition 6.1. Let H be real reproducing kernel Hilbert space on
some domain Ω. Let ψ be a continuous function on Ω. Let K ⊆ Ω be
compact. Let U be the collection of all open dense subsets of the space
of embedded measures on K. Let
MU = {x ∈ K|ιµ (x) = sup ιµ for some µ ∈ U }.
K
We define the green frontier by the formula
\
Green(K) = MU .
U ∈U

Note that given a reproducing kernel Hilbert space of harmonic func-

tions with harmonic ψ that Green(K) ⊆ ∂K by the maximum prin-
ciple. Essentially, the green frontier is similar to the Shilov boundary
in functional analysis, or the distinguished boundary in complex anal-
ysis. Moreover, via Krein-Milman, the green frontier is a subset of
the extreme points of K if we view K as a subset of H via the kernel
embedding.
Theorem 6.2 (Green topiary theorem). Let H be real reproducing
kernel Hilbert space on some domain Ω. Let ψ be a continuous function
on Ω. Let K ⊆ Ω be compact. There exists a measure supported on the
green frontier of K which embeds to topiary(K).
The green topiary theorem is an immediate consequence of the ef-
fectiveness of gradient ascent given by the update inequality, Theorem
7.3. As the support must lie on a marginal hypersurface, we have that
the support of the topiary is contained in the intersection of the green
frontier with a marginal hypersurface– such is deeply related the the
theme of semi-stability. In classical semi-stability regimes, we often
have extremely constrained geometry and in particular low dimension-
ality [11, 12, 5, 23].

7. Approximation
We now give some perspectives on approximating the topiary, espe-
cially in light of empirical sparsity.
16 G. HUTINET AND J. E. PASCOE

7.1. Ascent methods. First we expand on our discussion of gradient

ascent type methods.
Lemma 7.1. Let H be real reproducing kernel Hilbert space on some
domain Ω. Let ψ be a continuous function on Ω. Let K ⊆ Ω be compact.
We have that
S(µ) ≥ O(topiary(K)) − O(µ) + ∥ topiary(K) − µ∥2 /2.
Proof. Integrating the aesthetic margin with respect to the topiary
gives exactly the right hand side, thus there is at least one x with
aesthetic margin that large. □
The following is a calculation.
Lemma 7.2. Let H be real reproducing kernel Hilbert space on some
domain Ω. Let ψ be a continuous function on Ω. Let K ⊆ Ω be compact.
Let µ be a distribution on K. The optimal measure µt = (1 − t)µ + tδx
occurs at
ιµ (x)
t= .
∥kx − µ∥2
Moreover,
ιµ (x)2
O(µt ) − O(µ) = .
2∥kx − µ∥2
Combining the two previous we see the following.
Theorem 7.3 (Update inequality). Let H be real reproducing kernel
Hilbert space on some domain Ω. Let ψ be a continuous function on
Ω. Let K ⊆ Ω be compact. Let µ be a distribution on K. Choose x an
optimum of the aesthetic margin. For the optimal measure
µt = (1 − t)µ + tδx
we have that
[O(topiary(K)) − O(µ) + ∥ topiary(K) − µ∥2 /2]2
O(µt ) − O(µ) ≥ .
2∥kx − µ∥2
The following follows immediately from the previous.
Theorem 7.4 (Topiaric law of large numbers). Let H be real repro-
ducing kernel Hilbert space on some domain Ω. Let ψ be a continuous
function on Ω. Let K ⊆ Ω be compact. Optimally generate a sequence
of measures µn as in the update inequality. The aesthetic objective of
µn converges to that of topiary(K) with rate at worst O(1/n). The se-
quence µn converges at rate at worst O(1/n1/4 ). If ψ ∈ H, then µn
converges at rate at worst O(1/n1/2 ).
Compare to [16] which gives results on the dumptruck-sandpile-
moving Wasserstein distance for measure approximation .
 TOPIARISM 17

7.1.1. Greedy constructions. The greedy construction of the topiary

greedn is constructed by taking the convex combination of greedn−1
and a point mass at a maximizer contained in Green(A) of ι(greedn−1 )
such the aesthetic objective of that convex combination in maximized.
Choosing greed0 in Green(A) is advised. The greedy construction con-
verges to a measure with support in Green(A).
Define the growth set of A to be
Grow(A) = {x ∈
/ A|ι(topiary(A)) > 0}.
Observation 7.5. Let B be compact. Suppose Grow(A) ∩ B ̸= ∅.
Then, Grow(A) ∩ Topiary(B) ̸= ∅
Let hedge(A) be a finite signed measure with total variation 1 such
that ι(µ)(x) ≡ 0. The hedge need not be defined, we call sets A where
it is defined prunable. Let Prune(A) be the support of the negative
part of hedge(A).
Observation 7.6. Suppose A is prunable. Suppose Prune(A) ̸= ∅.
Then,
Prune(A) ∩ Topiary(A) ̸= Prune(A)
The hedge and the topiary coincide exactly on topiaric frontiers.
That is, if we know the topiaric frontier in advance, the problem is
purely algebraic.
Observation 7.7. hedge(Topiary(A)) = topiary(A).
The second greedy construction of the topiary sgreedn is con-
structed by taking the convex combination of sgreedn−1 and a point
mass at the maximizer of ι(sgreedn−1 ) such the aesthetic objective of
that convex combination in maximized, and then eliminating any point
masses in the distribution one by one that are now disadvangtageous
and scaling the rest of the measure. Doing such minimizes the follow-
ing chronic zig-zag-drag problem, which occurs when one chooses to
greedily include elements which are not in the topiaric frontier.
7.2. Algebraic approaches and combinatorics. One can consider
the problem algebraically for finite sets as well. The key being when
one wants to augment some proposed support K with some new x while
preserving the nullity of the aesthetic margin on its support, one may
need to remove some elements from K as they may be redundant or
incompatiable.
Theorem 7.8 (Discussion theorem). Let H be real reproducing kernel
Hilbert space on some domain Ω. Let ψ be a continuous function on
18 G. HUTINET AND J. E. PASCOE

bad0

bad1
bad2
bad3

topiary

Figure 3. Take our reproducing kernel to correspond

to the grammian of (−3, 1), (0, 2) and (2, 1) and assume
ψ = 0. Starting from a bad seed and applying the greedy
algorithm, we see a chronic zig-zag-drag pattern. If we
had chosen a proper intial state, the algorithm in fact
halts in one step. Here, the aesthetic objective is merely
the negative of the norm squared over two.

Ω. Let K ⊆ Ω be a finite topiaric index. Suppose topiary(K) has

positive aesthetic margin at some x. There exists B ⊆ K such that
O(topiary(B ∪ {x})) > O(topiary(K)).
Proof. Take B to be the topiary of K ∪ {x}. □
The proof is hardly enlightening or effective. However, the following
approach is:
(1) Initialize µ to be topiary(K).
(a) Find a signed distribution ν such that the mass of K is
1 on the support of µ such that the embedded measures
δx (y) = ν(y) for y in the support of µ.
(b) Find the maximum t such that µ + t(δx − ν) is a positive
measure and ι(µ + t(δx − ν))(x) ≥ 0
(c) If ιµ+t(δx −ν) (x) = 0, we are done. Otherwise, replace µ with
µ + t(δx − ν) and repeat.
However, we caution that such an approach may not produce the opti-
mal B, and we do not get the runtime guarantees given in the topiaric
law of large numbers. If we remove 0 or 1 elements from K to obtain
B indeed it must be optimal. Thus, in terms of adding mass at the
optimal x in a way that preserves the constancy of aesthetic margin on
its support except perhaps at x, we see that there is a sort of ko rule as
in the game of Go– if we add x to remove y, we can not immediately
desire to add back y. Because of the lack of meaningful runtime esti-
mates for such an algorithm, at least those apparent to the author, one
 TOPIARISM 19

fears there may be some metaphorical ko fights, where it adds x and

removes y then adds some other point, then adds back y and removes
x, then adds some other point, then adds back x to remove y and so
on.
7.3. Post hoc no-backtracking presentations. We now discuss the
theory of dissection and construction of finite topiaric indices.
Theorem 7.9 (Finite accessibility). Let H be real reproducing kernel
Hilbert space on some domain Ω. Let ψ be a continuous function on Ω.
Let K ⊆ Ω be a finite topiaric index. There exists an element x such
that K \ {x} is a finite topiaric index.
Proof. Consider the topiaric index K0 which is a proper subset of K
and optimizes O(topiary(K0 )). As K is a topiaric index, the update
inequality gives that there exists a measure with higher aesthetic ob-
jective supported on K0 ∪ {x} for some x ∈ K. Thus, the topiaric index
of K0 ∪ {x} has higher aesthetic objective and thus can not be a proper
subset of K and thus must be equal to K. □
As an immediate corollary, we see that a topiaric index is a union of
nested topiaric indices gained by inductively adding elements. Impor-
tantly, this also means that one may also deconstruct a topiaric index
in a particular order piece by piece.
Corollary 7.10 (Finitary no removal constructability). Let H be real
reproducing kernel Hilbert space on some domain Ω. Let ψ be a contin-
uous function on Ω. Let K ⊆ Ω be finite and a topiaric index. There is
a well-ordering on K such that each initial segment is a topiaric index.
The point being that if one knew the topiaric frontier in advance, one
could come up with an explanation of it such that we start with one
thing, add another which is compatiable repeatedly, namely needing no
removals. Such a presentation makes the portfolio look predetermined,
it is more expedient to be understood, and witnessed to be correct.
It is unclear if there is an infinite analog of no removal constructabil-
ity, which we give a formal problem for below.
Problem 1 (Existence of a topiaric branch). Let H be real reproduc-
ing kernel Hilbert space on some domain Ω. Let ψ be a continuous
function on Ω. Let K ⊆ Ω be compact. Does there exist a sequence
points (xn )∞
n=1 such that each Kn = {x1 , . . . , xn } is a topiaric index
and topiary(Kn ) → topiary(K) weakly.
Moreover, one expects in “real data,” whatever that means, for there
to outliers or elements with otherwise unreliable behavior, which some-
what explains why models might fail, but suggests that one wants to
20 G. HUTINET AND J. E. PASCOE

Figure 4. The concrete version of updating via the dis-

cussion theorem, where x is chosen to maximize the aes-
thetic margin. Here the space H is the real Hardy space
on the disk. The goldenrod curve represents the set K,
and the yellow part is where the aesthetic margin is neg-
ative, the purple part where it is postive. The red points
represent the current support. The pink point repre-
sents the most recent point we added. We have taken
ψ = 0. Note that each component of the purple parts
(and yellow parts for that matter) touches the boundary,
and that 0 is always in the purple part as it represents
somewhat of an invisible index. We will discuss how the
combination solves mazes later on. One may look at this
as trying to “positively classify” K. In the animation, we
see it start with a stereotype of K which persists and is
rarely covering much of K except the extremes. Slowly,
it resolves more and works to cover gaps it has missed,
sometimes making a large shift in its understanding. We
have neglected to depict the Julia-Caratheodory theorem
as ψ = 0, so the Julia-Caratheodory theorem projects the
goldenrod set to a line x = −y, under which the image of
the support is the origin. The goldenrod set is projected
to the ray the left the origin, and 0 is projected to some
point to the right of the origin.
 TOPIARISM 21

thus find large topiaric indices within your space, without regard to
true optimality.

Problem 2 (Broad index problem). Given a set K find the largest top-
iaric index possible.

8. Examples
8.1. Minotaur in the hedge maze. Consider the real Fock space
with kernel A(z, w) = ℜez,w . Let ψ = 0. Let M ⊆ C be a compact
set with path-connected complement, and have 0 ∈ / M . Starting at
0 following the gradient of the aesthetic objective of topiary(M ) gives
a path connecting 0 to ∞ not intersecting M. We note that if the
boundary of M is a smooth analytic curve, the topiaric index is finite,
as any accumulation on a compact smooth closed real analytic curve
would force the function to be globally constant as the aesthetic margin
of the topiary is constant exactly on the topiaric index. We also note
that the contour of 0 of the aesthetic objective touches M exactly at
the topiaric index of M. If M is the closure of its interior, indeed such
a contour must be tangent. (Indeed, the topiary behaves somewhat
like a support vector machine from the theory of machine learning
and cybernetic intelligence. See [30] for insight into support vector
machines.)
Of course, on any set winding around 0, we must have the topiary be
constant. Thus, by the invisible index theorem, we are approximating
the kernel of 0. Recursively defined mazes can encode universal com-
puters. (In the sense that if there is path from some particular point,
then some particular clause is true. That is, one can use a tree to write
down the letters of a possible proof which deposits into a pool when
it finds a legitimate proof.) There is at least some weak analogy of a
market absent of shocks behaving like a diffusion process (and likely
such has the stocks as waterwheels gathering the flux.)
One can do similarly over other spaces of harmonic functions in two
variables or to reach other locales by minimizing ∥µ − δα ∥ where α is
some point of interest and one wants to find a path from the origin to
the boundary.
22 G. HUTINET AND J. E. PASCOE

Figure 5. The harmonic conjugates of the approximat-

ing sequence coming from the concrete implementation
of the discussion theorem. The goldenrod part is our
“maze” and plot is colored by distance from the value
at 0. Due to conformality, (or equivalently satisfaction
of the Cauchy-Riemann equations) the resulting bright
curve through the origin represents a gradient descent
path for the topiary.

It is helpful to constrast with the typical way we see mazes resolved

physically in terms of static flows [1]. Typically, these involve putting a
large amount of stuff (such as water or charge, or even a growing slime
mold, although its true processes are more complicated) through some
maze which is fundamentally inoculated from the stuff. Eventually,
because of some fundamental nature of this stuff, it finds an efficient
path to flow through the maze. Fundamentally, these implement a
breadth first search called Lee’s algorithm as noted in [1]. In some
cases, one can observe optimal path heat or light up[9], which is pretty
fast essentially because Avogadro’s number is so large. We lack any real
insulation on our maze, and yet we obtain a similar gradient solving
the maze– the invisible hand is caught in the cookie jar anyway. With
the weak analogy that markets exist to diffuse value, one conjectures
there may exist some analogous interpretation in terms of value flows.
Analogies may also be made, and perhaps extended to deal with some
shocks and stimuli within the control systems philosophy of biological
and social systems, with the work of Doyle et al. [17].
 TOPIARISM 23

8.2. Trichotomy. Lets look at the image of the point mass at 0 again
in the real Fock space. Fix M ⊆ C compact with analytic boundary.
Again let ψ be 0. There are three possibilities:
(1) that M winds around the origin, in which case one could take
a measure supported on a continua of points,
(2) that we choose some finite collection of points from the bound-
ary of M,
(3) that 0 is in M, thus we take a point mass at 0.
One can find caricatures of various “value investing” theories in each,
with the first essentially corresponding to constructing a broad index
fund, the second corresponding to deep research and careful selection,
the last being buying an index fund. Difference in strategy may be
explained by difference in salience, which can be manifested as differ-
ences of M. In light of the invisible index theorem, they all approximate
correspond to buying the invisible true index fund.
8.3. Topiaric portfoliation. Such an optimization problem arises nat-
urally in portfolio theory, as the median growth rate of a geometric
Brownian motion is given by µ − σ 2 /2 where µ is the mean growth
and σ 2 is the variance. (That is, the kernel matrix arises from some
covariance matrix and the ψ gives the performances.) Assuming the
presence of risk-free assets such as cash or bonds of firm credit, one
sees that an asset with maximum mean must be in the green frontier,
which helps avoid the chronic zig-zag-drag problem. In portfolio theory,
the hedge corresponds to the Markowitz optimal portfolio [25, 34] for
choice of risk tolerance parameter two, and other risk tolerances may be
obtained by scaling the kernel. (Those correspond to the optimal port-
folio keeping at least some fixed fraction in risk-free reserve, or playing
with at some fixed amount of margin.) Optimal portfolios with non-
negative weights are often supported on small sets, whereas those with
arbitrary weights are diffuse [20, 10], hence in practice, greedy compu-
tation of the topiary is a low dimensional problem which is significantly
easier and more robust. Note also that large covariance matrices are
often ill-conditioned, and thus the Markowitz optimal portfolio may be
unknowable with much certainty.
Given that large amounts of empirical data will contain “lucky” and
“unlucky” outliers, one may want to mildly correct the data according
to your risk belief. For example, if the return of a security exceeds
its variance by more that the risk free rate, one would be incentivized
to go all in on such a security, buying none of the risk free asset.
One can correct such by either decreasing the mean, increasing the
variance, or both and retain the positive semidefiniteness of the kernel.
24 G. HUTINET AND J. E. PASCOE

Such is fundamentally different from the Markowitzian notion of risk

tolerance which is hard to reconcile with the growth of geometric
Brownian motions.

Problem 3 (Conservative data treament problems). Develop a theory

of how solutions to quadratic programs behave when one takes ψ is
replaced by ψ̃ such that ψ̃ ≤ ψ and the norm ∥ · ∥H is replaced by ∥ · ∥H̃
such that ∥ · ∥H ≤ ∥ · ∥H .

Moreover, since the optimal portfolio takes its maximum value on

its support, assuming identical means, it must be almost paradoxically
least correlated with the securities contained in the portfolio. In terms
of portfolio theory, our median growth optimization is a bit different
from more traditional Markowitz efficient frontier [25] or Sharpe ratio
[34] based methods. We view it as beneficial, although perhaps only
heuristically, that we take positive convex combinations of securities,
as that avoids some extrapolatory error in terms of globally minimiz-
ing the objective over portfolios with arbitrary weights summing to 1.
From a finance point of view, it also avoids some of the fundamental
problems with shorts and margin, although certainly the perspective
on optimization here might shed some insight there, especially when
hidden fees are taken into account. Taking into account the theory
laid out above gives reasonably fast performance. (Informally, we will
comment if one does not, one will observe extremely slow convergence
due to chronic zig-zag-drag.)
Note too, that if we change the method via which return and covari-
ance are obtained, the topiary will change. Thus too, if the currency
changes, the topiaric portfolio will change. Moreover, what it takes
for one to raise their score may be what another needs to depart for a
rise. Thus, certain mutualism exists due to uneven opportunity, reso-
lution, or information. Additionally, as the topiaric frontier is sparse,
one expects there may be many locations to compete in– the game is
massively multiplayer.
 TOPIARISM 25

Figure 6. An animation of our Julia-Cartheodory the-

orem for real securities data. Data is provided with-
out warranty of correctness, and the labels have been
anonymized, and is meant as a demonstration of the
theorem. Specifically, it neither represents financial ad-
vice nor is it intended for trading purposes. However,
we would encourage further empirical investigation. The
green dot represents the optimal portfolio, and the (often
obscured) red point is fixed, representing an instrument
with fixed yield. Here the excess marginal return is the
aesthetic margin and the covariant risk is the covariance
of each security with the portfolio.

8.4. Around organizational structure. Operations research prob-

lems that organizations must (at least implicitly) solve are analogous
to portfolio optimization. These generally deal with how organizations,
such as businesses, governments, and so on, allocate resources.
Organizations must commit resources to unreliable sources in order
to function. For example, in a foraging type task, one may want to
visit multiple areas to look for a desired resource, but not look at areas
for which there is little or no utility.
26 G. HUTINET AND J. E. PASCOE

The “going well together” analogy works here, and informs what
suborganizations can survive on their own. Suborganizations coming
from one division may not be able to thrive.

Figure 7. Lateral suborganizations can be better than

subdivisions such as manufacturing, advertising and so
on as they piggyback off the risk reduction of the ambient
organization. That is, groups of parallel vertically inte-
grated business pathways likely have more of a chance of
surviving on their own.

Figure 8. The no removal finite constructability theo-

rem suggests one may posess as pathway to divest from
individual parts as long as one does it in a particular
order. Note, however, it may be necessary to do some
reallocation in the remainder, and fitness still decreases
as we remove parts.

We see some heuristic justification for why businesses do not diver-

sify more. For example, they end up in an organization that can not
grow except by needing to cut something first, even if the ultimate
goal would be broad diversification. Note that a business entity faces
significantly more reallocation costs than an investor in liquid equities
such as stocks.
8.5. The adaptive market hypothesis. In terms of adaptive market
hypotheses [24], we see the problem as maxmimizing
Z
ψdµ − ∥µ − ν∥2 /2

where ν is a distribution on a set K1 representing the present and µ is

a distribution on K2 representing the past. That is, one can optimize
 TOPIARISM 27

determinacy and return similarly, although you must optimize over two
measures instead of one. The point being that we expect allocations
which are similar to superpositions of allocations of different securities
from the past under our kernel embedding to have similar returns.

References
[1] Andrew Adamatzky. Physical maze solvers. all twelve prototypes implement
1961 lee algorithm. Emergent computation: A festschrift for Selim G. Akl,
pages 489–504, 2017.
[2] J. Agler and J.E. McCarthy. Hankel vector moment sequences and the non-
tangential regularity at infinity of two variable Pick functions. Trans. Amer.
Math. Soc, 366(3):1379–1411, 2014.
[3] J. Agler and J.E. McCarthy. Pick Interpolation and Hilbert Function Spaces.
American Mathematical Society, Providence, 2002.
[4] J. Agler and J.E. McCarthy. Distinguished varieties. Acta Math., 194:133–153,
2005.
[5] J. Agler, J.E. McCarthy, and M. Stankus. Toral algebraic sets and function
theory on polydisks. J. Geom. Anal., 16(4):551–562, 2006.
[6] J. Agler, J.E. McCarthy, and N.J. Young. A Carathéodory theorem for the
bidisk using Hilbert space methods. Math. Ann., 352:581–624, 2012.
[7] J. Agler, R. Tully-Doyle, and N.J. Young. Boundary behavior of analytic func-
tions of two variables via generalized models. Indag. Math. (N.S.), 23:995–1027,
2012.
[8] David Alger. Raging bull: how to invest in the growth stocks of the 90s. Irwin
Professional Pub, 1992.
[9] Simon Ayrinhac. The electron in the maze. Shortest Path Solvers. From Soft-
ware to Wetware, pages 409–420, 2018.
[10] Michael J Best and Robert R Grauer. Positively weighted minimum-variance
portfolios and the structure of asset expected returns. Journal of Financial
and Quantitative Analysis, 27(4):513–537, 1992.
[11] K. Bickel, J. E. Pascoe, and A. Sola. Derivatives of rational inner functions: ge-
ometry of singularities and integrability at the boundary. Proc. London Math.
Soc., 116:281–329, 2018.
[12] Kelly Bickel, James Eldred Pascoe, and Alan Sola. Level curve portraits of
rational inner functions. Annali della Scuola Normale Superiore di Pisa (Classe
Scienze), Serie V, 21:449–494, 2020.
[13] John C Bogle. Common sense on mutual funds: New imperatives for the in-
telligent investor. John Wiley & Sons, 1999.
[14] John C Bogle. Bogle on mutual funds: New perspectives for the intelligent
investor. John Wiley & Sons, 2015.
[15] C. Carathéodory. Über die Winkelderivierten von beschraänkten analytischen
Funktionen. Sitzunber. Preuss. Akad. Wiss., pages 39–52, 1929.
[16] Eustasio Del Barrio, Evarist Giné, and Carlos Matrán. Central limit theorems
for the Wasserstein distance between the empirical and the true distributions.
Annals of Probability, pages 1009–1071, 1999.
28 G. HUTINET AND J. E. PASCOE

[17] John C Doyle and Marie Csete. Architecture, constraints, and behavior. Pro-
ceedings of the National Academy of Sciences, 108(supplement 3):15624–15630,
2011.
[18] Jack Clark Francis and Dongcheol Kim. Modern portfolio theory: Foundations,
analysis, and new developments. John Wiley & Sons, 2013.
[19] Benjamin Graham. Security analysis, 1934.
[20] Richard C Green. Positively weighted portfolios on the minimum-variance fron-
tier. The Journal of Finance, 41(5):1051–1068, 1986.
[21] G. Julia. Extension nouvelle d’un lemme de Schwarz. Acta Math., 42:349–355,
1920.
[22] John L Kelly. A new interpretation of information rate. The Bell System Tech-
nical Journal, 35(4):917–926, 1956.
[23] Greg Knese. Integrability and regularity of rational functions. Proceedings of
the London Mathematical Society, 111(6):1261–1306, 2015.
[24] Andrew Lo. Adaptive markets: Financial evolution at the speed of thought.
Princeton University Press, 2017.
[25] Harry M Markowitz et al. The optimization of a quadratic function subject to
linear constraints. Naval research logistics Quarterly, 3(1-2):111–133, 1956.
[26] J. E. Pascoe. The inverse problem for kernel means. preprint.
[27] J. E. Pascoe. An inductive Julia-Carathéodory Theorem for Pick functions in
two variables. Proc. Edin. Math. Soc., 61(3):647–660, 2018.
[28] JE Pascoe, Meredith Sargent, and Ryan Tully-Doyle. A controlled tangen-
tial julia–carathéodory theory via averaged julia quotients. Analysis & PDE,
14(6):1773–1795, 2021.
[29] Vern I Paulsen and Mrinal Raghupathi. An introduction to the theory of repro-
ducing kernel Hilbert spaces, volume 152. Cambridge university press, 2016.
[30] Tomaso Poggio, Steve Smale, et al. The mathematics of learning: Dealing with
data. Notices of the AMS, 50(5):537–544, 2003.
[31] Richard Roll. A critique of the asset pricing theory’s tests part i: On past and
potential testability of the theory. Journal of financial economics, 4(2):129–
176, 1977.
[32] Paul A Samuelson. Challenge to judgment. The Journal of Portfolio Manage-
ment, 1(1):17–19, 1974.
[33] Bernhard Scholkopf and Alexander J Smola. Learning with kernels: support
vector machines, regularization, optimization, and beyond. MIT press, 2018.
[34] William F Sharpe. Capital asset prices: A theory of market equilibrium under
conditions of risk. The journal of finance, 19(3):425–442, 1964.
[35] J. Wolff. Sur une généralisation d’un théorème de Schwarz. C.R. Acad. Sci.
Paris, 183:500–502, 1926.
Email address, Geoffrey Hutinet: [email protected]

Department of Mathematics, Drexel University, 3141 Chestnut St,

Philadelphia, PA 19104
Email address, J. E. Pascoe: [email protected]