Help with math terminology #2

[hotvette]

I have another dilemma with terminology that is puzzling and would appreciate some advice.

Consider the following truncated Taylor Series:
$$\begin{equation*}
f(\vec{z}_{k+1}) \approx f(\vec{z}_k)
+ \frac{\partial f(\vec{z}_k)}{\partial x} \Delta x
+ \frac{\partial f(\vec{z}_k)}{\partial \beta_1} \Delta \beta_1
+ \frac{\partial f(\vec{z}_k)}{\partial \beta_2} \Delta \beta_2 + \dots
+ \frac{\partial f(\vec{z}_k)}{\partial \beta_n} \Delta \beta_n
\end{equation*}$$
where:
$$\begin{align*} f &= f(\vec{z}\,) &
\vec{z} &= (x;\vec{\beta}\,) \\
\vec{z}_k &= (x;\vec{\beta}\,)_k &
\vec{z}_{k+1} &= (x;\vec{\beta}\,)_{k+1} \\
\Delta x &= x_{k+1} - x_k &
\Delta \beta_j &= (\Delta \beta_j)_{k+1} - (\Delta \beta_j)_k \\
\vec{\beta} &= (\beta_1, \beta_2, \dots, \beta_n)
\end{align*}$$
Then form the following function with the truncated Taylor Series:
$$\begin{equation*}
L = \sum_{i=0}^m \left[ f_i
+ \frac{\partial f_i}{\partial x_i} \Delta x_i
+ \frac{\partial f_i}{\partial \beta_1} \Delta \beta_1
+ \frac{\partial f_i}{\partial \beta_2} \Delta \beta_2 + \dots
+ \frac{\partial f_i}{\partial \beta_n} \Delta \beta_n - y_i
\right]^2 \end{equation*}$$
where:
$$\begin{align*}
f_i &= f(z_k) &
z_k &= (x_i; \beta)_k \\
\Delta x_i &= (x_i)_{k+1} - (x_i)_k &
\Delta \beta_j &= (\Delta \beta_j)_{k+1} - (\Delta \beta_j)_k \\
\vec{x} &= (x_1, x_2, \dots, x_m) &
\vec{y} &= (y_1, y_2, \dots, y_m)
\end{align*}$$
The $\partial x_i$ in the second term of the sum matches the $\Delta x_i$ but doesn't seem right because it implies $f=f(\vec{x};\vec{\beta})$ which isn't true, but using $\partial x$ instead doesn't seem right either because it doesn't match $\Delta x_i$. How to handle?

 

[anuttarasammyak]

I guess that L would be :

Does it make sense ?

 

[hotvette]

Thanks for you comment (and catching the error on the sum index). If I wanted to make the expression for L more compact, it would be:
$$\begin{equation*}
L = \sum_{i=1}^m \left[ f_i - y_i + \frac{\partial f_i}{\partial x_i} \Delta x_i
+ \sum_{k=1}^n \frac{\partial f_i}{\partial \beta_k} \Delta \beta_k \right]^2
\end{equation*}$$
but the dilemma remains with $\partial x_i$. It's a strange situation; $f$ has only one $x$ as an independent variable but the sum involves $m$ values of $x$. I'm tempted to use:
$$\frac{\partial f_i}{\partial x} \Delta x_i$$ but it just doesn't look right.

 

[anuttarasammyak]

In my hand writing I introduced j as another dummy index for sum as you did with k. Does it help you ?

 

[hotvette]

I don't think so. The dilemma really shows up when L is expanded, where we let $r_i = (f_i - y_i)$:
$$\begin{align*}
L &= \left[ r_1 + \frac{\partial f_1}{\partial x} \Delta x_1
+ \sum_{k=1}^n \frac{\partial f_1}{\partial \beta_k} \Delta \beta_k \right]^2
+ \left[ r_2 + \frac{\partial f_2}{\partial x} \Delta x_2
+ \sum_{k=1}^n \frac{\partial f_2}{\partial \beta_k} \Delta \beta_k \right]^2
\\
&+ \dots
+ \left[ r_m + \frac{\partial f_m}{\partial x} \Delta x_m
+ \sum_{k=1}^n \frac{\partial f_m}{\partial \beta_k} \Delta \beta_k \right]^2
\end{align*}$$
or (?)
$$\begin{align*}
L &= \left[ r_1 + \frac{\partial f_1}{\partial x_1} \Delta x_1
+ \sum_{k=1}^n \frac{\partial f_1}{\partial \beta_k} \Delta \beta_k \right]^2
+ \left[ r_2 + \frac{\partial f_2}{\partial x_2} \Delta x_2
+ \sum_{k=1}^n \frac{\partial f_2}{\partial \beta_k} \Delta \beta_k \right]^2
\\
&+ \dots
+ \left[ r_m + \frac{\partial f_m}{\partial x_m} \Delta x_m
+ \sum_{k=1}^n \frac{\partial f_m}{\partial \beta_k} \Delta \beta_k \right]^2
\end{align*}$$
I think the first one is mathematically correct because $f_i=f(x_i;\vec{\beta})=f(x;\vec{\beta})|_{x_i}$, but it seems to me in violation of nomenclature for Taylor Series. Maybe that's OK?

 

[anuttarasammyak]

Let me say what I think I understand.
$f_i$ is n+1 variable function $f_i(x_i;\beta_1,\beta_2,...,\beta_n)$ and partial derivatives are taken keeping other n variables are constant. i.e.
$$\frac{\partial }{\partial x_i}|_{\beta_1,\beta_2,...\beta_n}$$
where usual way of
$$\frac{\partial }{\partial x_1}|_{x_2,x_3,...x_m}$$
does not apply, and
$$\frac{\partial }{\partial \beta_1}|_{x_i,\beta_2,...\beta_n} := (\frac{\partial }{\partial \beta_1})_i$$
,and so on. I introduced RHS to make it clear that these partial derivative operators are different for different i.

In full details
$$L=\sum_{i=1}^m [\ r_i+[\ \triangle x_i \frac{\partial }{\partial x_i}|_{\beta_1,\beta_2,..,\beta_n}+\sum_{j=1}^n \triangle \beta_j \frac{\partial }{\partial \beta_j}|_{x_i,\beta_1,...,\beta_{j-1},\beta_{j+1} ..,\beta_n} \ ]\ f_i\ \ ]^2$$
Using the above said convention
$$L=\sum_{i=1}^m [\ r_i+[\ \triangle x_i \frac{\partial}{\partial x_i}+\sum_{j=1}^n \triangle \beta_j (\frac{\partial}{\partial \beta_j})_i\ ]\ f_i\ \ ]^2$$

[EDIT]
Let us introduce m+n dimension vector
$$\mathbf{q}=(q_1,q_2,...q_{m+n})=(x_1,x_2,..,x_n,\ \beta_1,\beta_2,...,\beta_m)$$
$$L=\sum_{i=1}^m [\ r_i+ \sum_{k=1}^{m+n} \triangle q_k \frac{\partial f_i}{\partial q_k}\ ]^2$$
$$=\sum_{i=1}^m [\ r_i+ \triangle \mathbf{q} \cdot \frac{\partial }{\partial \mathbf{q}}f_i\ ]^2$$
We get a simple formula at the expense of many zeros in the sum because $f_i$ is function of $x_i$ and not of $x_k$ so
$$\frac {\partial f_i}{\partial x_k}=0$$
with no regard to partial differential conditions.

This formula is useful for multiple-x's-function of $f_i$ case, i.e.
$$f_i(x_1,x_2,..,x_n,\ \beta_1,\beta_2,...,\beta_m)$$
which is extended from the original
$$f_i(x_i,\ \beta_1,\beta_2,...,\beta_m)$$

 

[hotvette]

I decided to simplify and go back to fundamentals as a way to try to sort this out. Below is the result. Start with the following definition of truncated Taylor Series:
$$\begin{equation*} f(x) \approx f(x_0) + \frac{d f(x_0)}{dx} (x-x_0) \end{equation*}$$
then the following should be valid (where each $x_i$ is a distinct point and $x'_i$ is nearby $x_i$):
$$\begin{align*}
&f(x'_0) \approx f(x_0) + \frac{d f(x_0)}{dx} (x'_0-x_0) \\
&f(x'_1) \approx f(x_1) + \frac{d f(x_1)}{dx} (x'_1-x_1) \\
&f(x'_2) \approx f(x_2) + \frac{d f(x_2)}{dx} (x'_2-x_2) \\
&f(x'_m) \approx f(x_m) + \frac{d f(x_m)}{dx} (x'_m-x_m)
\end{align*}$$
Form the sum:
$$\begin{align*}
&L = \left[ f(x_1)-y_1 + \frac{df(x_1)}{dx} (x'_1-x_1) \right]^2
+ \left[ f(x_2)-y_2 + \frac{df(x_2)}{dx} (x'_2-x_2) \right]^2 \\
&+ \dots + \left[ f(x_m) - y_m + \frac{d f(x_m)}{dx} (x'_m-x_m) \right]^2
\\\\
&L = \sum_{i=1}^m \left[ f(x_i) - y_i + \frac{d f(x_i)}{dx} (x'_i-x_i) \right]^2 \\
&L = \sum_{i=1}^m \left[ f_i - y_i + \frac{df_i}{dx} \Delta x_i \right]^2
\end{align*}$$
where we let $f(x_i) = f_i$ and $(x'_i-x_i) = \Delta x_i$. Is there any flaw in the above? I can't see any.

 

[anuttarasammyak]

WRT the last formula I would like to propose
$$L=\sum_{i=1}^m[f_i-y_i-f'_i\triangle x_i]^2$$
where
$$f'_i=f'(x_i)=\frac{df}{dx}(x_i)$$
to avoid possible confusion of $x$ or $x_i$ in the derivative denominator.

With m sets values of $\{x_i,\triangle x_i\}$ given and functions f(x) f'(x)and y(x) in
$$y_i=y(x_i)$$
known, I can calculate L.

Now I noticed my misunderstanding of you in the previous posts.
I took $x_i$ is one of variables $x_1,x_2,....$. Now I know $x_i$ is a value of single variable x.
I would like to amend the formula of L as
$$L=\sum_{i=1}^m [\ r_i+\ \triangle x_i \frac{\partial f}{\partial x}(x_i;\beta_1,\beta_2,..,\beta_n)|_{\beta_1,\beta_2,..,\beta_n}+\sum_{j=1}^n \triangle \beta_j \frac{\partial f}{\partial \beta_j}(x_i;\beta_1,\beta_2,..,\beta_n)|_{x,\beta_1,...,\beta_{j-1},\beta_{j+1} ..,\beta_n} \ \ ]^2$$
We may write it in brevity
$$=\sum_{i=1}^m [\ r_i+\ \triangle x_i f_{x\ i}+\sum_{j=1}^n \triangle \beta_j f_{\beta_j\ i}\ ]^2$$
with convention for function
$$\frac{\partial f}{\partial \alpha}:=f_\alpha$$
and for its output value wrt input value $x_i$
$$\frac{\partial f}{\partial \alpha}(x_i):=f_{\alpha\ i}$$
omitting mention of $\beta$s

Notations
$$\frac{df}{dx_i},\frac{df_i}{dx},\frac{df_i}{dx_i}$$
are misleading because suffixed ones are not variables but the values which are inappropriate as
$$\frac{df}{d\ 2},\frac{d\ 10}{dx},\frac{d 10}{d\ 2}$$are.

 

[hotvette]

Does the following work?
\begin{equation*}
L = \sum_{i=1}^m \left[ f_i - y_i + \Delta x_i \cdot \frac{\partial f}{\partial x} (x_i; \vec{\beta})
+ \sum_{k=1}^n \Delta \beta_k \cdot \frac{\partial f}{\partial \beta_k} (x_i; \vec{\beta}) \right]^2
\end{equation*}
with the clarification that:
\begin{equation*}
\frac{\partial f}{\partial x} (x_i; \vec{\beta}) = \frac{\partial f}{\partial x} |_{x_i; \vec{\beta}}
\end{equation*}

I had a completely different line of thought. The following is directly from my Advanced Calculus book (Kaplan, fourth edition 1957), p370:
\begin{equation*}
f(x,y) = f(x_1, y_1) + \left[ \frac{\partial f}{\partial x}(x-x_1) + \frac{\partial f}{\partial y}(y-y_1) \right] + \dots
\end{equation*}
all derivatives evaluated at $(x_1, \,y_1)$.

But, what if I wanted to know $f(x, y)$ for specific values of $x$ and $y$, say $x=3$ and $y=4$. Isn't the following valid (as long as $3$ is close to $x_1$ and $4$ is close to $y_1$)?

\begin{equation*}
f(3,4) = f(x_1, y_1) + \left[ \frac{\partial f}{\partial x}(3-x_1) + \frac{\partial f}{\partial y}(4-y_1) \right] + \dots
\end{equation*}
all derivatives evaluated at $(x_1, \,y_1)$.

 

[anuttarasammyak]

Yes, it seems to work. Possible concerns are:
$\beta_k$ is used as both variable in partial derivative formula and their specific values in $\vec{\beta}$. (Below I used \boldsymbol font for the variable but it does not make a big difference in appearance.)
$(\triangle x)_i$ might be better than $\triangle x_i$ to avoid possible misinterpretation of $\triangle$ multiplied by $x_i$
$$\frac{\partial f}{\partial x}(x_i;\vec{\beta}):=\frac{\partial f}{\partial x}|_{\vec{\beta}}(x_i)$$
$$\frac{\partial f}{\partial \boldsymbol{\beta}_k}(x_i;\vec{\beta}):=\frac{\partial f}{\partial\boldsymbol{\beta}_k}|_{x_i,\beta_1,...\beta_{k-1},\beta_k,...,\beta_n}(\beta_k)$$

 

[anuttarasammyak]

But, what if I wanted to know f(x,y) for specific values of x and y, say x=3 and y=4. Isn't the following valid (as long as 3 is close to x1 and 4 is close to y1)?

f(3,4)=f(x1,y1)+[∂f∂x(3−x1)+∂f∂y(4−y1)]+…
all derivatives evaluated at (x1,y1).

In order to avoid possible misinterpretation I would like to write it as
$$f(3,4)=f(x_1,y_1)+(3-x_1)\frac{\partial f}{\partial x}(x_1,y_1)+(4-y_1)\frac{\partial f}{\partial y}(x_1,y_1)+...$$
where (A)g(B) means (A) multiplied by g(B) where B is variable(s) of function g.

 

[hotvette]

Thanks for the discussion. It was very helpful.

 

[hotvette]

I think I found a solution to the dilemma. A caution was posted earlier not to confuse variables and values, but the point didn't really hit me until now. I was using $\beta_k$ as a variable but $x_i$ as a specific value of $x$. I think what resolves the inconsistent use of terminology is to define a set of functions $f_i = f_i(x_i; \vec{\beta})$, where $x_i$ and $\beta_k$ are variables, $\vec{\beta} = (\beta_1, \beta_2, \dots, \beta_{k-1}, \beta_k, \beta_{k+1}, \dots, \beta_n)$, and $\vec{x} = (x_1, x_2, \dots, x_{i-1}, x_i, x_{i+1}, \dots, x_m)$ . Letting subscripts $p$ and $p+1$ represent iterates of specific values of the variables, the Truncated Taylor Series expansions of each $f_i$ might be:
\begin{align*}
f_1 &\approx f_1(x_{1p};\vec{\beta}_p) + \Delta x_1 \cdot \frac{\partial f_1}{\partial x_1} (x_{1p};\vec{\beta}_p)
+ \sum_{k=1}^n \Delta \beta_k \cdot \frac{\partial f_1}{\partial \beta_k} (x_{1p};\vec{\beta}_p)
\\
f_2 &\approx f_2 (x_{2p};\vec{\beta}_p) + \Delta x_2 \cdot \frac{\partial f_2}{\partial x_2} (x_{2p};\vec{\beta}_p)
+ \sum_{k=1}^n \Delta \beta_k \cdot \frac{\partial f_2}{\partial \beta_k} (x_{2p};\vec{\beta}_p)
\\ \vdots \\
f_m &\approx f_m (x_{mp};\vec{\beta}_p) + \Delta x_m \cdot \frac{\partial f_m}{\partial x_m} (x_{mp};\vec{\beta}_p)
+ \sum_{k=1}^n \Delta \beta_k \cdot \frac{\partial f_m}{\partial \beta_k} (x_{mp};\vec{\beta}_p)
\end{align*}
Also:
\begin{align*}
\vec{x}_p &= (x_1, x_2, \dots, x_m)_p = (x_{1p}, x_{2p}, \dots, x_{mp} ) \\
\vec{\beta}_p &= (\beta_1, \beta_2, \dots, \beta_n)_p = (\beta_{1p}, \beta_{2p}, \dots, \beta_{np} ) \\
\Delta x_i &= (x_i)_{p+1} - (x_i)_p \\
\Delta \beta_k &= (\beta_k)_{p+1} - (\beta_k)_p \\
\Delta \vec{x} &= (\Delta x_1, \Delta x_2, \dots, \Delta x_m) \\
\Delta \vec{\beta} &= (\Delta \beta_1, \Delta \beta_2, \dots, \Delta \beta_n) \\
\vec{x}_{p+1} &= \vec{x}_{p} + \Delta \vec{x} \\
\vec{\beta}_{p+1} &= \vec{\beta}_{p} + \Delta \vec{\beta}
\end{align*}
The function $L$ then becomes:
\begin{equation*}
L = \sum_{i=1}^m \left[ f_i (x_{ip};\vec{\beta}_p) + \Delta x_i \cdot \frac{\partial f_i}{\partial x_i} (x_{ip};\vec{\beta}_p)
+ \sum_{k=1}^n \Delta \beta_k \cdot \frac{\partial f_i}{\partial \beta_k} (x_{ip};\vec{\beta}_p) - \widetilde{y}_i \right]^2
\end{equation*}
where $\widetilde{y}_i$ are constants. I plan to work out a way to get rid of the double subscripts. Per the last reply on my other post, I will also consider $x_i^{(p)}$ and $\beta_k^{(p)}$ to denote the values.